Real-time torque filtering and noise reduction method and system for electric actuator under edge computing framework

Abstract

The application relates to the technical field of edge computing torque filtering, and particularly discloses a real-time torque filtering and noise reduction method and system for an electric actuator under an edge computing framework, which comprises the following steps: in response to a starting or configuration request of an electric actuator, an industry-level data benchmark issued by a cloud is received, weighted calibration is performed in combination with local test data, and initial filtering parameters are generated; multi-source sensor data of the electric actuator are collected and time sequence alignment is performed, initial linear filtering is performed on a torque signal in the multi-source sensor data by using the initial filtering parameters, and a linear estimation signal is obtained. By introducing a two-way heterogeneous verification mechanism based on a linear shadow predictor, the system can sensitively capture a slight mutation under a pseudo-stationary state during operation of an extremely simple model, and the problem that a lightweight model is prone to over-smoothing and losing transient response details is solved.

Description

Technical Field

[0001] This invention relates to the field of edge computing torque filtering technology, specifically to a real-time torque filtering and noise reduction method and system for electric actuators under an edge computing framework. Background Technology

[0002] With the development of the Industrial Internet of Things (IIoT), electric actuators, as core components of precision industrial control, directly determine the stability of production lines through the accuracy and response speed of their torque control. Existing electric actuators typically employ traditional linear filters, such as low-pass filters, Kalman filters, or static parameter neural network models, to denoise the torque signals acquired by sensors.

[0003] However, in scenarios where edge computing resources are limited, traditional linear filters struggle to cope with the aliasing of nonlinear electromagnetic interference and mechanical coupling noise, often resulting in control signal lag or excessive residual noise. While static large-scale deep learning models have strong noise reduction capabilities, their high inference latency makes them unable to meet millisecond-level real-time control requirements. If lightweight small models are forced to be used for real-time performance, the filtering effect is prone to significant degradation under complex operating conditions due to insufficient feature extraction capabilities.

[0004] More seriously, in the high-precision control and health management of electric actuators, there exists a fundamental conflict between smooth control and feature preservation. To ensure the smoothness of control actions, algorithms tend to forcefully smooth out high-frequency fluctuations in the signal; however, early mechanical faults, such as bearing micro-cracks and gear wear, often exist in the form of weak high-frequency features. Existing filtering schemes usually cannot distinguish between useless high-frequency noise and critical fault precursors, inevitably leading to the accidental rejection of fault feature signals in the pursuit of control smoothness, or the introduction of excessive control jitter in order to preserve features. This makes it difficult to simultaneously achieve high-precision real-time noise reduction and full life-cycle fault early warning at the edge. Summary of the Invention

[0005] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the purpose of this invention is to propose a real-time torque filtering and noise reduction method and system for electric actuators under an edge computing framework, effectively resolving the contradiction between limited computing resources on the edge side and the need for complex nonlinear noise suppression.

[0006] To achieve the above objectives, a first aspect of the present invention proposes a real-time torque filtering and noise reduction method for electric actuators under an edge computing framework, comprising the following steps:

[0007] In response to the start-up or configuration request of the electric actuator, it receives industry-level data benchmarks from the cloud, performs weighted calibration in combination with local test data, and generates initial filtering parameters;

[0008] The multi-source sensor data of the electric actuator is collected and time-aligned. The torque signal in the multi-source sensor data is subjected to initial linear filtering using the initial filtering parameters to obtain a linear estimation signal.

[0009] Obtain the current load level index and signal-to-noise ratio index. Based on the comparison results between the load level index, the signal-to-noise ratio index and the preset threshold, select and call the corresponding sub-network from the preset deep learning network architecture to perform nonlinear noise suppression on the linearly estimated signal and output the intermediate signal.

[0010] The proportion of nonlinear noise in the intermediate signal is calculated, and the post-filtering strategy is dynamically switched according to the preset interval to which the proportion of nonlinear noise belongs. The intermediate signal is smoothed to generate the target control torque, and a control command is constructed based on the target control torque.

[0011] Before issuing the control command, the system checks whether the preset fault conditions are met based on the fault severity classification model. If so, a cross-device collaborative fault tolerance process is triggered to correct the control command; otherwise, the control command is issued directly.

[0012] To achieve the above objectives, a second aspect of the present invention proposes a real-time torque filtering and noise reduction system for electric actuators under an edge computing framework, the system comprising:

[0013] The cloud-edge collaborative calibration module is used to receive cloud-based benchmark values ​​and perform weighted calculations in conjunction with local test data to generate initial filter parameters;

[0014] The edge real-time processing module is used to collect multi-source sensor data, perform linear filtering using the initial filtering parameters, and dynamically call the deep learning sub-network for nonlinear compensation and post-filtering according to the load level and signal-to-noise ratio.

[0015] The cross-device collaboration module is used to perform fault tolerance compensation based on the fault severity classification model and to perform cross-device load balancing scheduling based on the load difference factor.

[0016] The cloud optimization module is used to construct an industry condition map and receive status data uploaded from the edge to iteratively optimize the parameters of the deep learning network architecture.

[0017] To achieve the above objectives, a third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described real-time torque filtering and noise reduction method for electric actuators under an edge computing framework.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] The real-time torque filtering and noise reduction method and system for electric actuators under the edge computing framework of this invention introduces a dual-path heterogeneous verification mechanism based on a linear shadow predictor. The system can keenly capture tiny abrupt changes in the pseudo-stationary state during the operation of the simplified model, which solves the problem that the lightweight model is prone to over-smoothing and losing transient response details.

[0020] Meanwhile, this scheme utilizes the conflict resolution mechanism of gradient orthogonal projection to forcibly separate the gradient directions of the smoothing task and the fault extraction task during the model training phase, fundamentally resolving the conflict between waveform smoothing and fault feature preservation. This enables the scheme to ensure highly smooth and stable actuator output torque while losslessly preserving key high-frequency features for fault diagnosis, achieving a balance between low-latency real-time control and high-sensitivity health management. Attached Figure Description

[0021] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0022] Figure 1 This is a flowchart illustrating the real-time torque filtering and noise reduction method for electric actuators under the edge computing framework provided by the present invention.

[0023] Figure 2 This is a schematic diagram comparing the frequency response characteristics of the electric actuator before and after cloud-edge collaborative weighted calibration in the real-time torque filtering and noise reduction method for electric actuators under the edge computing framework provided by this invention.

[0024] Figure 3 This is a schematic diagram comparing the smoothness of the deep learning network output signal before and after introducing the torque fluctuation constraint loss function in the real-time torque filtering and noise reduction method for electric actuators under the edge computing framework provided by this invention.

[0025] Figure 4 This is a schematic diagram of the VMD mode decomposition and signal reconstruction effect in the real-time torque filtering and noise reduction method for electric actuators under the edge computing framework provided by the present invention when the proportion of nonlinear noise is greater than the first threshold.

[0026] Figure 5 This is a schematic diagram illustrating the signal reconstruction effect before and after cross-device data sharing and completion triggered by a severe sensor fault in the real-time torque filtering and noise reduction method for electric actuators under the edge computing framework provided by this invention.

[0027] Figure 6This is a schematic diagram of the model consistency deviation and instantaneous spectral entropy change rate response before and after the pseudo-stationary escape mechanism is triggered in the real-time torque filtering and noise reduction method for electric actuators under the edge computing framework provided by this invention.

[0028] Figure 7 This is a schematic diagram comparing the convergence of fault feature retention rate between the gradient orthogonal projection conflict resolution mechanism and the traditional gradient descent method in the real-time torque filtering and noise reduction method for electric actuators under the edge computing framework provided by this invention.

[0029] Figure 8 This is a schematic diagram illustrating the implementation of the real-time torque filtering and noise reduction system for electric actuators under the edge computing framework provided by this invention.

[0030] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0031] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0032] The following description, with reference to the accompanying drawings, outlines a real-time torque filtering and noise reduction method, system, and electronic device for electric actuators within an edge computing framework, according to embodiments of the present invention.

[0033] Example 1:

[0034] This embodiment details a real-time torque filtering and noise reduction method for electric actuators within an edge computing framework. This embodiment aims to address the technical challenges of existing industrial electric actuators under complex operating conditions, such as low torque control accuracy, difficulty in balancing noise suppression and feature preservation, and limited computing resources, through a comprehensive technical solution integrating cloud-edge collaboration, deep learning dynamic scheduling, and cross-device fault tolerance mechanisms.

[0035] like Figure 1 As shown, this embodiment specifically includes the following steps:

[0036] S1: Cloud-edge collaborative initialization and parameter calibration.

[0037] Specifically, the method in this embodiment first runs on an edge computing device, which can be an embedded high-performance computing unit deployed in an industrial field, such as an industrial gateway or intelligent controller equipped with a dedicated AI acceleration chip. When the electric actuator powers on or receives a configuration request from the host computer, the edge computing device does not immediately enter the working state, but first triggers an initialization process.

[0038] In response to the start-up or configuration request of the electric actuator, the edge computing device performs the first step, which is to receive the industry-level data benchmark sent from the cloud, perform weighted calibration in combination with local test data, and generate the initial parameters for filtering. This process is designed based on the idea of ​​cloud empowerment and edge correction.

[0039] For example, the specific execution process of receiving industry-level data benchmarks from the cloud, combining them with local test data to perform weighted calibration, and generating initial filtering parameters is as follows: The edge computing device establishes a secure connection with the cloud server via an industrial Ethernet or 5G network. The cloud server maintains a vast industry map, which records the common characteristics of similar electric actuators under different operating conditions. The edge computing device receives the inherent vibration frequency benchmark value output by the cloud based on the industry map. This inherent vibration frequency benchmark value is the theoretical resonant frequency point of this actuator model, derived from statistical analysis of massive historical data.

[0040] It is important to note that because the installation environment of each device varies, such as the rigidity of the base and the degree of mechanical wear, relying solely on cloud-based baseline values ​​is insufficient for accuracy. Therefore, edge computing devices need to obtain their local natural vibration frequency and local stopband frequency through no-load and load tests at the local edge. Specifically, the edge computing device controls the electric actuator to execute a preset no-load operation program and a load operation program. It uses a built-in high-frequency vibration sensor to collect time-domain waveforms and performs Fast Fourier Transform (FFT) analysis to accurately identify the actual resonance peak of the device, i.e., the local natural vibration frequency, and the noise frequency band boundary that needs to be suppressed, i.e., the local stopband frequency.

[0041] Specifically, to integrate the common experiences of the cloud and the unique characteristics of the edge, this embodiment employs a weighted calibration algorithm. Using preset weighting factors, the reference value of the natural vibration frequency and the local natural vibration frequency are weighted and summed to obtain the final natural vibration frequency. This calculation logic can be described by the following mathematical relationship: Let the final natural vibration frequency be... The natural vibration frequency reference value sent from the cloud is The locally measured local natural vibration frequency is The preset weighting factor is The calculation formula is:

[0042] ;

[0043] In the formula, It's a value between 0 and 1, used to adjust the level of trust in cloud or local data. For example, in the initial stages of using a new device, when the local data sample is small, a larger value can be set. The value relies more on cloud experience; however, after the device has been running for a period of time, the impact can be reduced. The value relies more on local test data.

[0044] like Figure 2 The comparison of the frequency response characteristics of the electric actuator before and after cloud-edge collaborative weighted calibration is shown. Figure 2 The horizontal axis represents frequency in Hertz, and the vertical axis represents vibration amplitude in decibels.

[0045] Figure 2 The blue dashed line represents the industry-level data benchmark sent from the cloud. The curve shows a smooth and standard single-peak resonance characteristic, reflecting the theoretical natural frequency position of this actuator under ideal conditions. However, due to the lack of consideration for the differences in specific installation environments, its peak position has an inherent deviation from the actual working conditions.

[0046] Figure 2 The thin solid gray line represents the local measured data collected by the edge end through no-load and load tests. It can be seen that although the curve accurately reflects the actual resonant frequency position of the equipment in the current physical environment, that is, it has shifted towards the high frequency direction, the waveform is superimposed with a large number of random spikes and high-frequency noise caused by the field environment. If it is directly used for filter design, it will easily lead to parameter jitter.

[0047] Figure 2 The thick red line represents the final calibration characteristic curve generated after processing by the cloud-edge collaborative weighted algorithm. This curve intelligently approximates the actual resonance point measured locally at the peak position, corrects the deviation of the cloud data, and inherits the excellent characteristics of the cloud benchmark in terms of waveform smoothness, effectively filtering out random interference in local measurements.

[0048] This change intuitively demonstrates the technical effect of the present invention in integrating cloud experience with local measurements by introducing a preset weighting factor. This ensures that the final generated initial filtering parameters can accurately avoid the actual mechanical resonance point and maintain the numerical stability of the parameters, thereby preventing instability of the control system caused by directly using noisy data.

[0049] After calculating the final natural vibration frequency, the edge computing device determines the passband frequency parameter of the filter based on the final natural vibration frequency and the local stopband frequency, and uses this passband frequency parameter as the initial filter parameter. This step ensures that the subsequent linear filter can accurately avoid the actuator's natural frequency, preventing the induction of mechanical resonance.

[0050] S2: Multi-source data acquisition and initial linear filtering.

[0051] After parameter calibration, the system enters the real-time operation phase. Specifically, the edge computing device collects multi-source sensor data from the electric actuator and performs time alignment. This multi-source sensor data includes not only the core torque sensor data, but also housing vibration data collected by the vibration sensor, motor drive current data collected by the current sensor, and winding temperature data collected by the temperature sensor. Since different sensors may have different data sampling rates and transmission delays, the edge computing device utilizes FPGA (Field-Programmable Gate Array) hardware timestamp technology to uniformly map all sensor data onto the same high-precision time axis, achieving microsecond-level time alignment.

[0052] For example, after data alignment, the edge computing device performs initial linear filtering on the torque signal in the multi-source sensor data using the initial filtering parameters to obtain a linear estimation signal. This initial linear filtering can employ an infinite impulse response (IIR) filter or an adaptive least squares (RLS) filter, and the filter's cutoff frequency and stopband characteristics are determined by the initial filtering parameters generated in the preceding steps. The main purpose of this step is to filter out obvious, high-frequency random white noise, providing a relatively clean raw signal—the linear estimation signal—for subsequent deep learning processing.

[0053] S3: Dynamic calling and nonlinear compensation of deep learning subnetworks.

[0054] This step aims to resolve the conflict between computing resources and filtering performance in existing technologies. Specifically, the edge computing device acquires the current load level index and signal-to-noise ratio index. Based on the comparison results of the load level index, the signal-to-noise ratio index, and a preset threshold, it selects and calls the corresponding sub-network from a preset deep learning network architecture to perform nonlinear noise suppression on the linearly estimated signal and outputs an intermediate signal.

[0055] The deep learning network architecture pre-configured in this embodiment is not a single, massive model, but rather a dynamic architecture consisting of a main network and multiple levels of sub-networks. To achieve precise scheduling, the system pre-sets a strict threshold system: a first load threshold, a second load threshold, a first signal-to-noise ratio (SNR) threshold, and a second SNR threshold. The first load threshold is greater than the second load threshold, and the first SNR threshold is less than the second SNR threshold. For example, the first load threshold can be set to 100% (full load), the second load threshold can be set to 50% (half load), the first SNR threshold can be set to 15dB, and the second SNR threshold can be set to 20dB.

[0056] Specifically, the logical judgment process of the call is as follows: First, if the load level index is greater than the first load threshold and the signal-to-noise ratio index is less than or equal to the first signal-to-noise ratio threshold, this corresponds to an extremely harsh working condition of high load and strong noise. At this time, the edge computing device calls the main network containing the distillation and pruning architecture. Although the main network has undergone model compression, it still retains the most complete feature extraction capability and can forcibly recover the nonlinear characteristics of the torque signal from strong noise.

[0057] Secondly, if the load level indicator is between the first load threshold and the second load threshold, and the signal-to-noise ratio indicator is also between the first signal-to-noise ratio threshold and the second signal-to-noise ratio threshold, this corresponds to a normal operating condition with medium load and medium noise. In this case, the edge computing device invokes a pruned first-level sub-network of the attention layer. This first-level sub-network removes the computationally intensive attention mechanism module while retaining the backbone structure of the Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network, thus reducing inference latency while maintaining accuracy.

[0058] Furthermore, if the load level indicator is less than or equal to the second load threshold and the signal-to-noise ratio indicator is greater than or equal to the second signal-to-noise ratio threshold, this corresponds to a good operating condition of low load and low noise. At this time, the edge computing device calls a secondary sub-network pruned into a two-layer convolutional neural network and a long short-term memory network. This sub-network structure is extremely simplified and can complete inference quickly with extremely low power consumption.

[0059] Finally, if the system is determined to be operating under stable conditions with minimal noise (i.e., extremely low load and extremely high signal-to-noise ratio), the system calls a three-level subnetwork that retains only the core feature channels. This subnetwork may contain only a simple feature mapping layer, consuming almost no computational resources, and is used to prevent overfitting and unnecessary signal processing delays.

[0060] It is also important to note that the training process of the aforementioned deep learning network architecture is crucial. To balance noise reduction and feature preservation, the deep learning network architecture is trained based on a condition-adaptive composite loss function. This condition-adaptive composite loss function consists of a weighted sum of scale-invariant signal-to-noise ratio loss, mean square error loss, torque fluctuation constraint loss, and fault feature deviation loss. This composite design ensures that the network, during training, not only pursues waveform similarity (constrained by mean square error loss) but also signal purity (constrained by scale-invariant signal-to-noise ratio loss) and the integrity of key fault features (constrained by fault feature deviation loss).

[0061] Specifically, to prevent abrupt, non-physical jumps in the filtered signal, a torque fluctuation constraint loss is introduced. This torque fluctuation constraint loss is calculated by: calculating the absolute value of the difference between the filtered output signal at the current moment and the filtered output signal at the previous moment, and then summing and averaging these absolute values ​​over a preset time window. In mathematical terms, let N be the signal length or the number of sampling points within the time window. Let be the filtered output signal at time t. for The filtered output signal at time (i.e., the previous time) then represents the torque fluctuation constraint loss. The calculation formula can be expressed as:

[0062] ;

[0063] The physical meaning of this formula is to average the absolute value of the first-order difference (i.e., the rate of change) of the output signal, thereby penalizing any non-smooth abrupt changes and forcing the network to output a smooth torque curve.

[0064] like Figure 3 This diagram illustrates the comparison of the smoothness of the deep learning network's output signal before and after introducing the torque ripple constraint loss function. Figure 3 The horizontal axis represents the sampling time of the system operation in milliseconds, and the vertical axis represents the output torque amplitude in Newton-meters.

[0065] Figure 3 The medium gray dashed line represents the real torque reference under ideal conditions, showing the actual physical change trend of the electric actuator during dynamic operation.

[0066] Figure 3 The thin solid blue line represents the network output waveform when the model is trained using only the mean square error loss function without introducing torque fluctuation constraints. It can be clearly observed that although the curve follows the true benchmark in terms of overall value, there are a large number of severe sawtooth high-frequency jitters and spikes in local areas. This indicates that a single numerical approximation target cannot constrain the temporal continuity of the signal, and direct application will lead to motor current oscillation and mechanical wear.

[0067] Figure 3The thick red solid line represents the network output waveform obtained after training with the torque fluctuation constraint loss function proposed in this invention. Compared with the blue curve, this curve successfully eliminates non-physical high-frequency noise, exhibiting extremely high smoothness and continuity, while closely following the dynamic change trend of the real benchmark without significant phase lag. This comparative effect intuitively demonstrates that the torque fluctuation constraint loss function, through its penalty mechanism for the difference between adjacent time steps of the output signal, effectively forces the deep learning model to balance noise reduction while ensuring the compliance of control commands, thereby solving the technical problem that lightweight models at the edge computing end are prone to outputting discrete and abrupt signals.

[0068] S4: Post-filtering strategy based on the proportion of nonlinear noise.

[0069] Although the intermediate signal output by the deep learning network has removed most of the noise, nonlinear coupling noise or high-frequency artifacts may still remain under certain special conditions. Therefore, this embodiment designs a dynamic post-filtering stage. Specifically, the edge computing device calculates the proportion of nonlinear noise in the intermediate signal, and dynamically switches the post-filtering strategy according to the preset interval to which the proportion of nonlinear noise belongs, smoothing the intermediate signal to generate the target control torque.

[0070] For example, the method for calculating the proportion of nonlinear noise is as follows: the ratio of the short-time energy of the vibration reference signal to the short-time energy of the initial residual is calculated as the proportion of nonlinear noise. Here, the initial residual refers to the difference between the original input torque signal and the aforementioned linearly estimated signal. This ratio can reflect how much of the current noise is caused by nonlinear components due to mechanical vibration.

[0071] To implement the strategy switching, the system presets a first noise proportion threshold and a second noise proportion threshold, wherein the first noise proportion threshold is greater than the second noise proportion threshold. For example, the first threshold is set to 0.8, and the second threshold is set to 0.5. The specific strategy is as follows:

[0072] If the proportion of nonlinear noise is greater than the first noise proportion threshold, for example This indicates that nonlinear noise dominates the signal and the signal composition is extremely complex. Therefore, a joint filtering strategy combining Kalman filtering and improved variational mode decomposition (VMD) is employed. VMD can decompose the complex signal into several intrinsic mode functions (IMFs), thereby separating out the nonlinear noise, which is then combined with Kalman filtering for optimal estimation.

[0073] like Figure 4 This is a schematic diagram illustrating the variational mode decomposition and signal reconstruction results when the proportion of nonlinear noise exceeds the first threshold. Figure 4It contains three vertically arranged sub-charts, with the horizontal axis representing time in seconds and the vertical axis representing signal amplitude in Newton-meters.

[0074] The topmost sub-figure (a) shows the original input signal collected under extremely harsh working conditions. Due to the extremely high proportion of nonlinear noise, the real periodic torque waveform has been completely masked by a large low-frequency trend term and irregular random interference, which reflects the judgment logic for heavy noise scenarios in Example 1.

[0075] The middle sub-figure (b) shows the result of processing the original signal using the improved variational mode decomposition algorithm proposed in this invention. The orange dashed line represents the nonlinear trend noise mode that was successfully separated. This component concentrates most of the energy in the original signal but is an invalid interference. The green solid line represents the extracted torque feature main mode. It can be seen that the algorithm effectively strips away the target signal features that are covered by strong noise.

[0076] The bottom sub-figure (c) shows the final signal reconstruction result, where the black dotted line represents the true torque reference under ideal conditions, and the blue solid line represents the final reconstructed output signal after removing noise modes and combining Kalman filtering. The comparison clearly shows that even when the original signal is completely distorted, the joint filtering strategy employed in this invention can still highly restore the frequency and amplitude characteristics of the true torque, demonstrating that this technical solution possesses excellent noise suppression and feature recovery capabilities under complex operating conditions dominated by nonlinear noise.

[0077] If the proportion of nonlinear noise is between the first noise proportion threshold and the second noise proportion threshold, for example... This indicates the presence of a certain amount of nonlinear noise, but not enough to mask the main signal. In this case, a Kalman filtering strategy that adjusts the weights of the process noise covariance matrix is ​​adopted. By dynamically increasing the weights of the process noise covariance matrix (Q matrix), the filter's confidence in the system model decreases while its confidence in the observed data increases, thus better tracking signal changes.

[0078] If the proportion of nonlinear noise is less than the second noise proportion threshold, for example This indicates that the signal is relatively pure with few nonlinear components. Therefore, a moving average filtering strategy combining threshold suppression is employed. This is a computationally minimal approach, suppressing only outliers exceeding the threshold and performing simple smoothing to preserve the original dynamic characteristics of the signal to the greatest extent possible.

[0079] The signal obtained after this processing step is the final target control torque used for control. The edge computing device constructs control commands based on the target control torque, preparing to send them to the motor drive unit of the actuator.

[0080] S5: Fault detection and cross-device collaborative fault tolerance.

[0081] Before the control command is officially issued, a final check must be performed to ensure safety. Specifically, before issuing the control command, a fault severity grading model is used to check whether preset fault conditions are met. If so, a cross-device collaborative fault tolerance process is triggered to correct the control command; otherwise, the control command is issued directly.

[0082] The fault severity grading model in this embodiment not only focuses on individual data loss but also on the logical rationality of the data. When a severe fault is determined, such as a data deviation rate exceeding a set threshold or continuous missing data from sensors, it means that the current actuator's sensors have failed or the data is completely unreliable. In this case, forcibly executing commands may damage the equipment. Therefore, the system enables a cross-cluster healthy actuator data sharing and completion mechanism.

[0083] Specifically, in industrial settings, multiple actuators of the same type typically work collaboratively, such as multi-joint robotic arms or valve arrays. In this case, leveraging physical coupling or process correlations, the theoretical state of a faulty device can be inferred from the health data of neighboring equipment. An edge computing device acquires vibration and current data from associated healthy actuators. Then, based on preset mapping coefficients, the vibration data, current data, and constant terms are linearly weighted and combined. The calculation result is used as a completed signal to correct the control commands.

[0084] This completion logic can be described by the following mathematical formula: Let the completed signal be... The vibration data of the associated health actuator is The current data is The preset mapping coefficients are respectively , The constant term is The completed formula is:

[0085] ;

[0086] In the formula, This reflects the weight of the vibration signal's contribution to the torque. This reflects the weight of the current signal (which is strongly correlated with electromagnetic torque) in the output torque. This is the bias compensation term.

[0087] like Figure 5 This is a diagram illustrating the fault-tolerant completion effect based on cross-device multivariate mapping. Figure 5 The horizontal axis represents the operating time of the electric actuator in seconds, and the vertical axis represents the torque amplitude in Newton-meters.

[0088] Figure 5 The solid blue line on the left shows the continuous waveform collected by the system during the normal operation of the local sensor. At this time, the curve completely coincides with the gray dashed line representing the real physical torque reference, indicating that the collected data is accurate and reliable.

[0089] Figure 5 The black vertical dotted line marks the trigger point when the sensor experiences a severe malfunction. After this point, the blue solid line drops sharply to zero, indicating that the local sensor signal is completely lost.

[0090] Following closely behind, Figure 5 The solid red line on the right represents the completion signal generated after the system immediately starts the cross-device collaborative fault-tolerant process. This signal is obtained by multivariate weighted synthesis based on the vibration data and current data of the associated health actuator using preset linear mapping coefficients.

[0091] pass Figure 5 As can be clearly seen from the contrast between the solid red line and the dashed gray line, even though the local sensor has failed, the completed signal generated by cross-device multivariate mapping can still highly fit the fluctuation trend and amplitude characteristics of the real physical torque. This proves that the fault-tolerant mechanism can take over control by generating a high-confidence virtual signal through the correlation and deduction of cluster data under extreme conditions of single-point device failure, thereby ensuring the continuity and safety of the production line operation.

[0092] In this way, even if the sensors of this equipment completely fail, a safety fallback control command can be generated based on the status of the sibling equipment to maintain the minimum operation of the production line and avoid downtime accidents.

[0093] S6: Adaptive load balancing scheduling.

[0094] In addition to filtering and fault tolerance for individual devices, this embodiment also considers resource optimization at the edge computing cluster level.

[0095] Optionally, the method further includes a load adaptive balancing scheduling step. Because operating conditions in industrial settings are dynamic, some actuators may suddenly face high-load computing tasks, such as starting the main network, while other actuators remain idle.

[0096] Specifically, the master node of the edge computing device or cluster calculates the load difference factor of each actuator within the cluster, whereby the load difference factor is the difference between the maximum load rate and the minimum load rate within the cluster. Let the maximum load rate be... Minimum load rate Then the load difference factor To achieve fine-grained scheduling, the system presets a first difference threshold and a second difference threshold, wherein the second difference threshold is greater than the first difference threshold. For example, the first difference threshold is set to 30%, and the second difference threshold is set to 50%.

[0097] The scheduling strategy is as follows: if the load difference factor is between the first difference threshold and the second difference threshold, for example... This indicates a significant imbalance in cluster load. In this case, a layered iterative strategy is adopted, prioritizing the allocation of computing resources to high-load executors. Specifically, the allocation of edge computing resources is tilted towards high-load nodes, or some non-real-time computing tasks (such as log analysis) on high-load nodes are migrated to low-load nodes, while core filtering tasks remain local.

[0098] If the load difference factor is greater than the second difference threshold, for example This indicates extreme imbalance, with some nodes potentially on the verge of overload and crashing. In this situation, a master-slave collaborative mode is adopted, with the high-load executor at the core, calling upon low-load executors to provide auxiliary data. In this mode, the low-load executor not only contributes computing resources but may also directly provide pre-processed data to the high-load executor through the aforementioned cross-device collaborative mechanism, reducing the pre-processing overhead of the high-load executor.

[0099] In summary, this embodiment achieves a complete edge computing torque filtering scheme through the above steps and the close cooperation of system modules.

[0100] It's also important to note that in practical applications, the intermediate signal and the linear estimation signal mentioned above are names used to refer to the data stream at different processing stages. The linear estimation signal refers to the signal that has undergone RLS linear filtering, removing high-frequency white noise but still retaining nonlinear distortion; while the intermediate signal refers to the signal that has had its nonlinear characteristics recovered or compensated after processing by the deep learning network, but may have introduced algorithm artifacts. The final target control torque is the final usable signal after post-filtering and smoothing, which retains the true physical characteristics while removing algorithm artifacts.

[0101] In addition, mapping coefficients , , In real-world systems, these coefficients are not fixed constants, but can be identified and updated online based on historical normal operation data using least squares methods or regression analysis. However, at the moment a fault occurs, the system uses the most recently updated fixed value to ensure real-time performance.

[0102] The solution in this embodiment, particularly the introduction of a sub-network switching mechanism for different operating conditions in the deep learning network architecture, significantly reduces the average power consumption and computational latency at the edge. Simultaneously, by introducing... The loss function ensures from the training source that the model will not lose key fault features due to excessive smoothing. Meanwhile, the cross-device collaboration mechanism provides reliable redundancy protection for single-point failures, which is impossible to achieve in traditional standalone smart meters.

[0103] Finally, all thresholds in this embodiment, such as the first load threshold and the first signal-to-noise ratio threshold, can be modified through a configuration file during the system initialization phase to adapt to the needs of electric actuators with different power ratings and application scenarios (such as valve control, robotic arm joints, and AGV drive wheels). This parameterized design gives this solution strong versatility and scalability.

[0104] Example 2:

[0105] Building upon Example 1, this embodiment further addresses the pseudo-stationarity risk that three-level subnetworks in deep learning network architectures may encounter during long-term practical operation. It presents a deep learning technical solution incorporating linear shadow prediction and a dual-path heterogeneous verification mechanism. This embodiment aims to solve the oversmoothing distortion problem that may occur in minimally simplistic and lightweight models when facing sudden, small perturbations, ensuring that the electric actuator, while pursuing extreme energy saving and computational efficiency, does not lose its ability to perceive and respond to potential sudden changes in the system's state.

[0106] Specifically, the technical solution described in this embodiment is mainly applied to the stage where the system, according to the scheduling strategy described in Embodiment 1, determines that it is currently in a stable operating condition without significant noise and has already invoked the three-level sub-network, which retains only the core feature channels, to perform filtering operations. In this stage, to prevent the system from missing minor mechanical faults or load changes due to excessive pruning of the feature extraction capabilities of the three-level sub-network, this embodiment introduces a parallel monitoring mechanism named the pseudo-stationary escape and switching protection strategy.

[0107] For example, the core of this strategy lies in constructing a bypass monitor with extremely low computational cost but high physical interpretability. While the three-level subnetwork is invoked for filtering, the edge computing device utilizes idle logic gate resources or a very small amount of CPU time slices to launch a linear shadow predictor based on the Autoregressive Moving Average (ARMA) algorithm in parallel. This linear shadow predictor does not participate in the regular control signal output; its sole responsibility is to use the current input torque signal as an observation to generate a linear predicted torque based on linear statistical laws.

[0108] It is important to note that the autoregressive moving average algorithm was chosen as the core algorithm of the linear shadow predictor because it has extremely high prediction accuracy and extremely low computational cost when dealing with stationary time series. This algorithm assumes that the current torque value is a linear combination of the torque values ​​at several past times plus the current random disturbance.

[0109] Specifically, the linear shadow predictor operates according to the following mathematical model: Let The input torque signal at time t is ,but Linear predicted torque generated at each time step Equal to the past The weighted sum of the actual input values ​​at each moment and the past The sum of the weighted prediction errors at each time step. and These represent the autoregressive order and the moving average order, respectively, and are typically rounded to smaller integers at the edges to ensure real-time performance. In this way, the linear shadow predictor constructs a baseline curve reflecting the linear dynamic characteristics of the system.

[0110] Optionally, after obtaining the output of the three-level subnetwork, namely the intermediate signal in Example 1 and the output of the linear shadow predictor, the system enters the dual-path heterogeneous verification stage. The edge computing device calculates the model consistency deviation between the intermediate signal output by the three-level subnetwork and the linear predicted torque in real time. This model consistency deviation essentially reflects the difference in understanding of the same input signal between the deep learning nonlinear model and the traditional statistical linear model.

[0111] Specifically, the calculation of the model consistency deviation is not based on single-point differences, but on statistical characteristics within a sliding time window. Let the length of the current sliding window be... The intermediate signal output by the three-level subnetwork is The linearly predicted torque is Then the model consistency deviation Defined in window The root mean square value or absolute value integral of the difference between the two. When the system is truly in a stationary state, since the signal is mainly composed of low-frequency components, the outputs of the deep learning model and the linear model should highly overlap. It remains at an extremely low level. However, when the system encounters exotic features not learned by deep networks, such as nonlinear transient impacts caused by gear microcracks, deep networks may treat them as noise and forcibly smooth them out, while linear models, due to their sensitivity to abrupt changes, will produce large prediction residuals, leading to a significant separation between their output curves. It then rose rapidly.

[0112] In addition to time-domain model comparison, this embodiment also introduces frequency-domain complexity monitoring as a second layer of protection. The system synchronously calculates the instantaneous spectral entropy change rate of the input torque signal. Instantaneous spectral entropy is a physical quantity that measures the complexity of the signal's spectral distribution.

[0113] For example, the calculation process of instantaneous spectral entropy is as follows: First, a short-time Fourier transform or wavelet transform is performed on the input torque signal to obtain the current power spectral density distribution. Then, the power spectral density is normalized to satisfy the properties of a probability distribution. Next, the current instantaneous spectral entropy value is calculated using the definition formula of Shannon information entropy. If the signal is a single-frequency sine wave or white noise, its spectral entropy value is relatively stable; however, if a fault frequency component or modulation sideband with a specific structure is suddenly mixed into the signal, the energy distribution of the signal will be reorganized, causing the spectral entropy value to jump. The instantaneous spectral entropy change rate is the ratio of the absolute value of the difference between the current spectral entropy value and the previous spectral entropy value to the time interval. This indicator can keenly capture early fault signals with low energy (i.e., no significant change in signal-to-noise ratio) but complex structures.

[0114] After calculating the two key indicators mentioned above, the system executes its core decision logic. Specifically, the edge computing device compares the calculated indicators with preset safety thresholds. If the statistical value of the model consistency deviation exceeds the first safety threshold, or the instantaneous spectral entropy change rate exceeds the second safety threshold, the system will determine that the current operating condition is not truly stable, but rather in a pseudo-stable state with potential risks.

[0115] like Figure 6 This is a joint monitoring graph of model consistency deviation and instantaneous spectral entropy change rate under pseudo-stationary conditions. Figure 6 The horizontal axis represents the sampling time of the system operation in seconds, and the vertical axis represents the normalized amplitude of the monitoring index, which is dimensionless.

[0116] Figure 6 The black dashed line represents the system's preset safety trigger threshold, used to determine whether it is necessary to exit the current three-level sub-network.

[0117] Figure 6 The left-hand area shows the state of the electric actuator when it is in the pseudo-stationary operating range. At this time, the model consistency deviation represented by the blue solid line and the instantaneous spectral entropy change rate represented by the red solid line are both at extremely low levels and far below the safety threshold, indicating that the output of the deep learning sub-network is highly consistent with the output of the linear shadow predictor and the signal spectrum structure is stable.

[0118] Figure 6The peak pulse in the center shows the instantaneous response of the system when subjected to a slight mechanical disturbance. Although the disturbance is extremely difficult to detect in the time-domain waveform, due to the sensitivity of the linear model to sudden changes and the instantaneous increase in signal complexity, the two monitoring curves almost simultaneously jump sharply and break through the safety threshold.

[0119] The response characteristics of this joint monitoring intuitively demonstrate that the dual-path heterogeneous verification mechanism proposed in this invention can effectively identify abnormal states hidden beneath a stable appearance, thereby triggering model switching and escape strategies in a timely manner and avoiding the risk of control failure caused by over-smoothing of the lightweight model.

[0120] It is important to note that the logical judgment here uses OR logic, meaning that if either a model divergence in the time domain or a sudden change in complexity in the frequency domain occurs, the system considers the three-level subnetwork no longer applicable. The first and second safety thresholds are empirical values ​​calibrated based on a large amount of offline experimental data. For example, the first safety threshold is usually set to 2% to 5% of the rated torque, aiming to tolerate normal model errors but intercept abnormal deviations; the second safety threshold is set to be able to distinguish between normal operating condition fluctuations and typical mechanical shocks, such as the entropy increase limit caused by the initial stage of bearing spalling.

[0121] Once a pseudo-stationary state is determined, the system must take decisive measures to avoid control failure. Specifically, the edge computing device immediately terminates the invocation of the third-level sub-network and switches back to the main network. The main network, namely the network in Embodiment 1 that includes distillation and pruning architecture but retains complete feature extraction capabilities, has a large computational overhead but possesses extremely strong nonlinear fitting and feature preservation capabilities, and can correctly handle the current complex signal.

[0122] However, a direct model switch can cause a step jump in the control signal, which is unacceptable for precision electric actuators. Therefore, this embodiment designs a smooth transition mechanism. In the first control cycle after the switch, the system does not directly output the calculation results of the main network, but performs a weighted fusion operation. Specifically, the output value of the main network and the linear predicted torque are weighted and fused according to a preset confidence level, and used as the intermediate signal for the final output.

[0123] Optionally, the confidence weight is a dynamically adjusted parameter. At the moment of switching, since the main network has just re-entered, its internal state, such as the hidden layer state of the LSTM, may not have fully converged to the optimal operating point of the current signal, while the linear shadow predictor, because it has been running in parallel, has a continuous state. Therefore, in the first cycle of switching, the linear predictor can be given a higher weight, while the main network output can be given a lower weight.

[0124] For example, let the fused intermediate signal be... The main network output is The linearly predicted torque is The confidence weight is The fusion formula is: This can be expressed by the formula:

[0125] ;

[0126] As time passes after the switchover, for example, within the subsequent 3 to 5 control cycles, The value gradually increases linearly to one, thus achieving a seamless soft switch from linear statistical support to deep nonlinear dominance.

[0127] In summary, this second embodiment constructs a complete pseudo-stationary escape system by introducing a linear shadow predictor, dual-path heterogeneous verification, and a soft-switching fusion mechanism. This system effectively solves the oversmoothing risk that lightweight models may face during long-term operation in actual industrial settings, enabling electric actuator systems under the edge computing framework to enjoy the extreme low power consumption advantages brought by the three-level sub-network while possessing the high reliability and agile fault detection capabilities of the main network.

[0128] Example 3:

[0129] This embodiment, building upon Embodiments 1 and 2, further focuses on the core optimization of deep learning network architecture and the model iterative update process. This embodiment details a model parameter update method based on a gradient orthogonal projection conflict resolution mechanism, aiming to address the gradient competition and feature annihilation problems that may be encountered during model training within an edge computing framework when simultaneously pursuing two seemingly contradictory optimization goals: extremely smooth control torque and high-fidelity fault feature extraction.

[0130] Specifically, the technical solution described in this embodiment is mainly applied to the stage where the cloud optimization module performs offline training of the deep learning network, or the stage where the edge real-time processing module uses locally accumulated data to fine-tune the model online. As mentioned in Embodiment 1, the deep learning network architecture is trained based on a condition-adaptive composite loss function, which consists of a weighted sum of scale-invariant signal-to-noise ratio loss, mean square error loss, torque fluctuation constraint loss, and fault feature deviation loss. However, in actual mathematical optimization processes, simple weighted summation is often insufficient to handle extremely complex gradient conflict situations.

[0131] It is important to note that in the actual operating data of electric actuators, the torque ripple constraint loss aims to penalize all high-frequency fluctuations to achieve waveform smoothing, while the fault feature deviation loss aims to preserve specific fault features, which often manifest as weak high-frequency signals. When the system is in the early fault latency period, the objectives of these two loss functions may form completely opposite gradient directions mathematically and geometrically. If the two are directly added together and backpropagated, the huge gradient generated by the smoothing task may completely cancel out the small gradient generated by the fault extraction task, causing the model to sacrifice fault detection capability for smoothing.

[0132] To fundamentally address this issue, this embodiment introduces and executes a gradient orthogonal projection conflict resolution step when updating model parameters using the aforementioned adaptive composite loss function. This step is not performed at the level of loss function value calculation, but rather after the gradient is calculated via backpropagation and before the weight parameters are updated.

[0133] Specifically, in each training iteration or fine-tuning step, the system first calculates the gradient of each sub-loss function relative to the network shared layer parameters using the backpropagation algorithm. In this embodiment, the system calculates the first gradient vector corresponding to the torque fluctuation constraint loss and the second gradient vector corresponding to the fault feature deviation loss.

[0134] For example, suppose the current weight parameter space of the network is... The torque ripple constraint loss is The fault characteristic deviation loss is The system uses the chain rule to calculate... right The partial derivatives are used to obtain the first gradient vector. Similarly, calculate right The partial derivatives are used to obtain the second gradient vector. These two vectors, in the multidimensional parameter space, represent the descent direction that makes the waveform smoothest and the descent direction that preserves the fault characteristics most completely, respectively.

[0135] After obtaining these two key gradient vectors, the system does not rush to perform weighted merging, but instead first performs conflict detection in the geometric direction. Specifically, the system calculates the cosine similarity between the first gradient vector and the second gradient vector. Cosine similarity is an important indicator for measuring the consistency of the pointing directions of two vectors in multidimensional space. It is calculated by performing a dot product operation on the first gradient vector and the second gradient vector, and then dividing by the product of the magnitudes of the two vectors.

[0136] It is important to note that, according to the properties of vector algebra, a positive cosine similarity indicates that the angle between the two gradient directions is less than 90 degrees, meaning that their optimization objectives are compatible or at least do not conflict at the current parameter position. A zero cosine similarity indicates that the two are orthogonal and unrelated. However, if the cosine similarity is less than zero, it has a clear physical and geometric meaning, indicating a gradient conflict. This means that the angle between the first and second gradient vectors exceeds 90 degrees, exhibiting a negative correlation. In this case, updating parameters along the direction of the first gradient vector (the smoothing direction) will cause the fault features to deviate from the loss, thus compromising fault feature extraction; conversely, the opposite is also true.

[0137] To address this detected conflict, this embodiment employs a fault feature priority protection strategy. Specifically, the system performs a gradient orthogonal projection operation. The first gradient vector is projected onto the normal plane of the second gradient vector to remove components in the first gradient vector that are opposite in direction to the second gradient vector, resulting in a corrected first gradient vector. Here, the normal plane refers to the hyperplane formed by order variables perpendicular to the second gradient in parameter space. Through projection, we retain those components in the first gradient that promote smoothing without impairing fault features, while eliminating harmful components that attempt to erase fault features for the sake of smoothing.

[0138] For example, let the corrected first gradient vector be... The original first gradient vector is The second gradient vector is The calculation formula can then be expressed as:

[0139] ;

[0140] In the formula, It accurately represents the calculated original gradient vector corresponding to the torque ripple constraint loss; It accurately represents the calculated gradient vector corresponding to the loss caused by the deviation of the fault characteristics; Dot product operation representing vectors; This represents the square of the L2 norm of the second gradient vector, i.e., the square of the modulus. The fractional part of the formula... Calculated exist The projection coefficient in the direction, multiplied by Vector, that is, obtained exist The component in the direction. Since the cosine similarity was previously determined to be less than zero, this component is related to... In opposite directions. By starting from the original Subtracting this component from the middle, we get It no longer contains anything related to The antagonistic elements thus achieve geometrically and Orthogonal or non-negative correlation.

[0141] like Figure 7 The graph shows the convergence curve of fault feature retention rate under gradient orthogonal projection optimization. Figure 7 The horizontal axis represents the number of training iterations of the deep learning model, and the vertical axis represents the retention rate of the model for weak fault feature signals, in percentage.

[0142] Figure 7 The blue dashed line represents the feature retention rate trajectory when training the model using the traditional gradient descent method. It can be seen that as the number of iterations increases, the dominant gradient direction generated by the smoothing constraint term in the loss function conflicts with the gradient direction of weak fault features, causing fault features to be gradually regarded as noise and suppressed. The retention rate shows a significant downward trend and eventually stabilizes at a low level. This intuitively reflects the feature annihilation problem in the existing technology.

[0143] Figure 7 The solid red line represents the feature retention rate trajectory after introducing the gradient orthogonal projection conflict resolution mechanism described in this invention. This curve has remained at a high level of over 90% throughout the entire training period, indicating that by eliminating the adversarial component projection of the smooth gradient during backpropagation, the algorithm successfully locked key fault feature information while performing waveform smoothing optimization. This proves that the optimization strategy can effectively solve the gradient competition problem between multi-objective tasks, ensuring that the model has high smoothness while still possessing high sensitivity to fault perception.

[0144] Optionally, if the calculated cosine similarity is greater than or equal to zero in the above conflict detection step, it indicates that there is no gradient conflict, and the system does not need to perform a projection operation; it can simply set the corrected first gradient vector equal to the original first gradient vector. This conditional triggering mechanism ensures the computational efficiency of the algorithm, intervening only when necessary.

[0145] After gradient correction, the system proceeds to the final parameter update stage. Specifically, the weighted sum of the corrected first gradient vector and the second gradient vector is used as the backpropagation gradient to update the weight parameters of the deep learning network architecture. The system updates the weight parameters according to the weight coefficients set in Embodiment 1, for example, corresponding to... and Construct the final update gradient. Let the final gradient be... The weights are respectively and ,but This can be expressed as a formula:

[0146] ;

[0147] The system then inputs the final gradient into an optimizer such as Adam or SGD to perform a parameter update.

[0148] It should also be noted that the gradient orthogonal projection conflict resolution mechanism described in this embodiment can be applied not only to the updating of global parameters, but also to the more refined layered updating of the network.

[0149] For example, for shallow feature extraction layers in deep learning networks, such as convolutional layers, gradient conflicts may be relatively small because they are mainly responsible for extracting general textures and edges; however, for deeper decision or regression layers, gradient conflicts are often the most severe. Therefore, edge computing devices can be configured to enable the aforementioned projection computation only in the last few fully connected layers or LSTM layers of the network to achieve the best balance between computational resource consumption and optimization performance.

[0150] By implementing the method described in this embodiment, the filtering model of the electric actuator establishes a strict optimization discipline during training: under no circumstances should noise reduction and smoothing operations be performed at the expense of the saliency of fault features. This mechanism mathematically ensures that even weak early bearing fault signals masked by strong electromagnetic interference can be stubbornly locked and preserved during iterative model optimization, rather than being smoothed out as ordinary noise. This provides the most fundamental model algorithm support for achieving nonlinear noise suppression and cross-device collaborative fault tolerance, ensuring that the entire system maintains high-sensitivity fault detection capability and high-precision control smoothness throughout its entire lifecycle.

[0151] In summary, this embodiment, by detailing the complete process of gradient calculation, conflict detection, orthogonal projection, and formula correction, provides a solid algorithmic foundation for the realization of highly reliable intelligent filtering and health management in this invention.

[0152] Example 4:

[0153] like Figure 8 As shown, this embodiment will provide a detailed system-level implementation of the method flow described in embodiments one through three from the perspective of hardware architecture and functional module collaboration. This embodiment constructs a real-time torque filtering and noise reduction system for electric actuators under an edge computing framework. This system aims to solve the technical bottlenecks of insufficient single-point computing power, loose cloud-edge collaboration mechanism, and weak fault tolerance in existing industrial control architectures.

[0154] Specifically, this system adopts a layered distributed architecture in its physical deployment. Its core logical components are a cloud-edge collaborative calibration module, an edge real-time processing module, a cross-device collaborative module, and a cloud optimization module. These four modules are independent yet tightly coupled, working together to support high-precision and high-reliability torque control operations.

[0155] First, this system includes a cloud-edge collaborative calibration module. This module is typically deployed at the protocol conversion layer of the edge gateway or the front-end service layer in the cloud, acting as a bridge connecting macro-level industry experience with micro-level device characteristics. During system initialization, this module receives cloud-based benchmark values ​​and performs weighted calculations based on local test data to generate initial filter parameters. Specifically, this module has a built-in parameter synchronization engine that can retrieve industry benchmark maps generated from massive amounts of historical data in the cloud in real time, such as the inherent vibration frequency benchmark value mentioned in Embodiment 1. Simultaneously, this module controls the local actuator to perform idling and load tests, obtaining the local resonant frequency and stopband frequency through Fast Fourier Transform analysis. Subsequently, this module executes a weighted algorithm to fuse industry common data with local unique data, outputting high-precision passband frequency parameters. These initial parameters are immediately sent to the edge real-time processing module to ensure that subsequent linear filters can accurately avoid mechanical resonance points and prevent control divergence caused by parameter mismatch.

[0156] Secondly, as the core of data processing in this system, the edge real-time processing module is deployed in an embedded high-performance computing unit adjacent to the electric actuator, such as an industrial-grade edge controller equipped with a neural network acceleration unit. This module is mainly used to collect multi-source sensor data, perform linear filtering using the initial filtering parameters, and dynamically call deep learning sub-networks for nonlinear compensation and post-filtering based on load level and signal-to-noise ratio. In terms of hardware implementation, this module integrates a multi-channel high-speed analog-to-digital conversion interface, which can simultaneously acquire analog signals from torque sensors, vibration sensors, current sensors, etc., and achieve microsecond-level timestamp alignment through a built-in field-programmable gate array chip.

[0157] It is worth noting that the edge real-time processing module operates with complex scheduling logic. It not only performs the dynamic switching between the main network and multi-level sub-networks as described in Embodiment 1, but also embeds the linear shadow predictor and dual-path heterogeneous verification mechanism described in Embodiment 2. When this module calls the three-level sub-network that retains only the core feature channels to process stationary conditions, its internal watchdog process calculates the model consistency deviation and the instantaneous spectral entropy change rate in parallel. Once a pseudo-stationary state is detected, the module can forcibly switch to the main network in milliseconds and smooth the output signal through a weighted fusion algorithm, thus achieving a balance between low power consumption and high reliability at the hardware level.

[0158] Furthermore, to enhance the robustness of cluster operation, this system is equipped with a cross-device collaboration module. This module runs on an industrial fieldbus or time-sensitive network, breaking down traditional information silos between actuators. It performs fault tolerance and compensation based on a fault severity grading model and executes cross-device load balancing scheduling based on load difference factors. In actual operation, when a sensor on an actuator experiences a severe failure leading to data loss, this module immediately activates a data sharing mechanism to obtain real-time vibration and current data from physically adjacent or logically related healthy actuators. Utilizing the linear mapping relationship established in the aforementioned embodiments, this module can locally synthesize a high-confidence virtual torque signal to correct control commands and ensure that the production line does not stop due to a single point of failure. In addition, this module monitors the computational load rate of each node within the cluster in real time. When the load difference factor exceeds a preset threshold, it automatically migrates non-real-time computation tasks from high-load nodes to low-load nodes, achieving adaptive balancing of computing resources.

[0159] Finally, this system also includes a cloud optimization module deployed on a remote server or private cloud platform. This module is used to construct an industry condition map and receive status data uploaded from the edge nodes to iteratively optimize the parameters of the deep learning network architecture. This module is the brain of the entire system and has powerful offline training capabilities. It not only collects regular operating data uploaded by each edge node, but also focuses on those difficult-to-detect samples that cause large prediction deviations in the edge model. During model update training, this module strictly implements the gradient orthogonal projection conflict resolution mechanism described in Example 3. When updating weights through backpropagation, it forces orthogonalization of the gradient of the torque fluctuation constraint loss and the gradient of the fault feature deviation loss, ensuring that the newly generated model parameters improve smoothing performance without sacrificing sensitivity to weak fault features. The newly trained model will be pushed to each edge real-time processing module via over-the-air download technology to achieve system self-evolution.

[0160] In summary, the real-time torque filtering and noise reduction system for electric actuators under the edge computing framework described in Embodiment 4 achieves precise parameter initialization through the cloud-edge collaborative calibration module, millisecond-level adaptive filtering and anomaly protection through the edge real-time processing module, ensures high availability of the cluster through the cross-device collaboration module, and enables continuous algorithm iteration through the cloud optimization module. The organic combination of these four modules, together with the aforementioned three method embodiments, forms a tight logical loop, effectively resolving the conflict between control smoothness and fault feature preservation raised in the background art.

[0161] Example 5:

[0162] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0163] like Figure 9 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0164] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0165] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0166] The memory 103 is used to store a computer program corresponding to the real-time torque filtering and noise reduction method for electric actuators under the edge computing framework of the above embodiments of the present invention. This computer program is controlled and executed by the processor 101. The processor 101 is used to execute the computer program stored in the memory 103 to implement the content shown in the foregoing method embodiments.

[0167] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 9 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0168] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A real-time torque filtering and noise reduction method for electric actuators under an edge computing framework, characterized in that, Includes the following steps: In response to the start-up or configuration request of the electric actuator, it receives industry-level data benchmarks from the cloud, performs weighted calibration in combination with local test data, and generates initial filtering parameters; The multi-source sensor data of the electric actuator is collected and time-aligned. The torque signal in the multi-source sensor data is subjected to initial linear filtering using the initial filtering parameters to obtain a linear estimation signal. Obtain the current load level index and signal-to-noise ratio index. Based on the comparison results between the load level index, the signal-to-noise ratio index and the preset threshold, select and call the corresponding sub-network from the preset deep learning network architecture to perform nonlinear noise suppression on the linearly estimated signal and output the intermediate signal. The proportion of nonlinear noise in the intermediate signal is calculated, and the post-filtering strategy is dynamically switched according to the preset interval to which the proportion of nonlinear noise belongs. The intermediate signal is smoothed to generate the target control torque, and a control command is constructed based on the target control torque. Before issuing the control command, the system checks whether the preset fault conditions are met based on the fault severity classification model. If so, a cross-device collaborative fault tolerance process is triggered to correct the control command; otherwise, the control command is issued directly.

2. The method according to claim 1, characterized in that, The system receives industry-level data benchmarks from the cloud, combines them with local test data to perform weighted calibration, and generates initial filter parameters, including: Receives the inherent vibration frequency reference value output from the cloud based on industry maps; The local natural vibration frequency and local stopband frequency were obtained through no-load and load tests at the local edge. Using a preset weighting factor, the natural vibration frequency reference value and the local natural vibration frequency are weighted and summed to obtain the final natural vibration frequency; The passband frequency parameter of the filter is determined based on the final natural vibration frequency and the local stopband frequency, and this passband frequency parameter is used as the initial parameter of the filter.

3. The method according to claim 1, characterized in that, The step of selecting and calling the corresponding sub-network from the preset deep learning network architecture based on the comparison results of the load level index, the signal-to-noise ratio index, and the preset threshold includes: A first load threshold, a second load threshold, a first signal-to-noise ratio threshold, and a second signal-to-noise ratio threshold are preset, wherein the first load threshold is greater than the second load threshold, and the first signal-to-noise ratio threshold is less than the second signal-to-noise ratio threshold; If the load level index is greater than the first load threshold and the signal-to-noise ratio index is less than or equal to the first signal-to-noise ratio threshold, the main network containing the distillation and pruning architecture is invoked. If the load level indicator is between the first load threshold and the second load threshold, and the signal-to-noise ratio indicator is between the first signal-to-noise ratio threshold and the second signal-to-noise ratio threshold, then the first-level subnetwork with the attention layer pruned is invoked; If the load level index is less than or equal to the second load threshold and the signal-to-noise ratio index is greater than or equal to the second signal-to-noise ratio threshold, call the secondary sub-network that is pruned into a two-layer convolutional neural network and a long short-term memory network; If the condition is determined to be a stable operating condition with no obvious noise, the three-level sub-network that retains only the core feature channels is invoked.

4. The method according to claim 3, characterized in that, The method also includes a pseudo-stationary escape and handover protection strategy executed during the invocation of the third-level sub-network, specifically: While the three-level sub-network is called for filtering, a linear shadow predictor based on the autoregressive moving average algorithm is started in parallel to generate a linear predicted torque using the current input torque signal. The model consistency deviation between the intermediate signal output by the three-level sub-network and the linear predicted torque is calculated in real time, and the instantaneous spectral entropy change rate of the input torque signal is calculated simultaneously. If the statistical value of the model consistency deviation exceeds the first safety threshold, or the instantaneous spectral entropy change rate exceeds the second safety threshold, the current operating condition is determined to be a pseudo-stationary state. The call to the third-level sub-network is forcibly terminated and the main network is switched back. In the first control cycle after the switch, the output value of the main network and the linear predicted torque are weighted and fused according to the preset confidence weight to serve as the intermediate signal for the final output.

5. The method according to claim 1, characterized in that, The dynamic switching of the post-filtering strategy based on the preset interval to which the nonlinear noise proportion belongs includes: The ratio of the short-time energy of the vibration reference signal to the short-time energy of the initial residual is calculated and used as the proportion of nonlinear noise. A first noise percentage threshold and a second noise percentage threshold are preset, wherein the first noise percentage threshold is greater than the second noise percentage threshold; If the proportion of nonlinear noise is greater than the first noise proportion threshold, a joint filtering strategy of Kalman filtering and improved variational mode decomposition is adopted. If the proportion of nonlinear noise is between the first noise proportion threshold and the second noise proportion threshold, a Kalman filtering strategy that adjusts the weights of the process noise covariance matrix is ​​adopted. If the proportion of nonlinear noise is less than the second noise proportion threshold, a moving average filtering strategy that combines threshold suppression is adopted.

6. The method according to claim 1, characterized in that, The deep learning network architecture is trained based on the working condition adaptive composite loss function, which is composed of a weighted sum of scale-invariant signal-to-noise ratio loss, mean square error loss, torque fluctuation constraint loss and fault feature deviation loss. The torque fluctuation constraint loss is calculated by: calculating the absolute value of the difference between the filtered output signal at the current moment and the filtered output signal at the previous moment, and summing and averaging the absolute values ​​within a preset time window.

7. The method according to claim 6, characterized in that, When updating model parameters using the adaptive composite loss function, the method further includes a gradient orthogonal projection conflict resolution step, specifically: Calculate the first gradient vector corresponding to the torque fluctuation constraint loss and the second gradient vector corresponding to the fault feature deviation loss, respectively. Calculate the cosine similarity between the first gradient vector and the second gradient vector; If the cosine similarity is less than zero, it is determined that there is a gradient conflict. At this time, the first gradient vector is projected onto the normal plane of the second gradient vector to remove the component in the first gradient vector that is opposite in direction to the second gradient vector, and the corrected first gradient vector is obtained. The weight parameters of the deep learning network architecture are updated using the weighted sum of the corrected first gradient vector and the second gradient vector as the backpropagation gradient.

8. The method according to claim 1, characterized in that, The process of triggering cross-device collaborative fault tolerance to correct the control command includes: When a severe fault is determined to be a data deviation rate greater than the set threshold or a sensor continuously missing data, cross-cluster health actuator data sharing and completion is enabled. Acquire vibration and current data from the associated health actuator; Based on preset mapping coefficients, the vibration data, the current data, and the constant term are linearly weighted and combined, and the calculation result is used as the completed signal to correct the control command.

9. The method according to claim 1, characterized in that, The method further includes a load adaptive balancing scheduling step, specifically: Calculate the load difference factor for each actuator in the cluster, where the load difference factor is the difference between the maximum load rate and the minimum load rate in the cluster; A first difference threshold and a second difference threshold are preset, wherein the second difference threshold is greater than the first difference threshold; If the load difference factor is between the first difference threshold and the second difference threshold, a hierarchical iteration strategy is adopted to prioritize the allocation of computing resources to high-load executors. If the load difference factor is greater than the second difference threshold, a master-slave collaborative mode is adopted, with the high-load executor as the core, and the low-load executor is called to provide auxiliary data.

10. A real-time torque filtering and noise reduction system for electric actuators under an edge computing framework, characterized in that, For performing the method according to any one of claims 1 to 9, comprising: The cloud-edge collaborative calibration module is used to receive cloud-based benchmark values ​​and perform weighted calculations in conjunction with local test data to generate initial filter parameters; The edge real-time processing module is used to collect multi-source sensor data, perform linear filtering using the initial filtering parameters, and dynamically call the deep learning sub-network for nonlinear compensation and post-filtering according to the load level and signal-to-noise ratio. The cross-device collaboration module is used to perform fault tolerance compensation based on the fault severity classification model and to perform cross-device load balancing scheduling based on the load difference factor. The cloud optimization module is used to construct an industry condition map and receive status data uploaded from the edge to iteratively optimize the parameters of the deep learning network architecture.

Images