A needle knife robot operation fault prediction method and system
By monitoring the command and feedback velocity sequences of the needle knife robot, and using deep learning and reinforcement learning models for fault prediction, the problem of identifying endogenous mechanical decay characteristics and exogenous load disturbances under high impedance conditions was solved, thus achieving accuracy in fault prediction and reliability in the control system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- EMERGENCY GENERAL HOSPITAL
- Filing Date
- 2026-06-17
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies cannot accurately identify endogenous mechanical decay characteristics and exogenous load disturbances in the high impedance interference environment of needle knife robots, resulting in false alarms or missed alarms in fault prediction, and cannot achieve predictive perception in complex surgical conditions.
By monitoring the polarity switching state of the command speed sequence, the zero-speed observation window is opened using the commutation zero-crossing trigger point, the feedback speed sequence is collected, a health feature vector is generated, and deep learning and reinforcement learning models are used for fault prediction, and control commands are dynamically adjusted to compensate for mechanical decay.
This achieves effective decoupling of intrinsic characteristics under high impedance conditions, improves the accuracy and robustness of fault prediction, and ensures the reliability of the control system and the accuracy of surgical trajectory.
Smart Images
Figure CN122480979A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of control system health management technology, and in particular relates to a method and system for predicting operational faults of a needle knife robot. Background Technology
[0002] As a precision medical device, the operational stability and command execution accuracy of the underlying control circuit of the current needle knife robot are directly related to the safety of the operation. Existing technology determines the spatial deviation between the two by comparing the displacement command sequence issued by the controller with the set of position coordinates returned by the actuator, and judges whether there is an abnormality in the system based on whether the deviation exceeds a preset threshold.
[0003] However, in typical operating conditions such as cutting high-density tissues or piercing deep fascia with needle knife robots, the aforementioned threshold monitoring criteria generate logical constraints at the physical level. The needle knife operation process is accompanied by severe nonlinear load pulses. The feedback signal distortion caused by such external loads is usually greater in amplitude than the physical displacement deviation in the early stage of wear of the transmission mechanism. At this time, the intrinsic physical decay characteristics and the extrinsic load disturbances overlap in the time and frequency domains. The residual calculation method based on the full amount of operating data is limited by the signal-to-noise ratio distortion, making it difficult to achieve the physical separation of intrinsic characteristics under complex surgical conditions.
[0004] Analysis shows that existing technologies have the following limitations in application: 1. The hysteresis characteristics caused by mechanical wear in the monitoring data are highly mixed with the feedback spikes caused by tissue impedance, making it impossible for the model to accurately identify latent decline trends; 2. The static threshold judgment criteria cannot be compatible with the background feature drift caused by individual differences in the actuator, which easily leads to false alarms or missed alarms; 3. The all-time monitoring path is limited by the masking effect of external loads on the transmission gap characteristics within and outside the sampling window, resulting in the prediction system lacking predictive perception in the early stage of the equipment entering a sub-healthy state. In addition to the hardware limitations of transmission mechanism wear, the software control method also has shortcomings. Chinese invention patent application CN119326514A discloses a method for compensating for transmission gaps in surgical instruments. The compensation method, device, and slave controller determine the motion direction and speed based on the change in the command angle, and calculate the current cycle gap compensation angle according to a fixed proportional coefficient, which is then superimposed on the target command. The control mechanism relies on the premise that the transmission gap exhibits a pure kinematic linear mapping within the continuous motion phase. In the scenario of high impedance dynamic evolution of the needle knife penetrating deep fascia, external nonlinear loads induce structural elastic deformation and deep coupling with the reverse physical gap. Existing solutions extract compensation parameters throughout the continuous load-bearing period, failing to physically separate the feedback signal distortion caused by tissue impedance. The core premise is fundamentally mismatched with the actual boundary conditions, resulting in distorted compensation calculations when dealing with high-intensity alternating disturbances, and an inability to reconstruct purely intrinsic mechanical decay. Therefore, how to effectively decouple the intrinsic characteristics of the control system under high impedance interference environments and construct a decay trend evolution prediction mechanism with physical self-consistency has become the technical problem to be solved by this invention. Summary of the Invention
[0005] The present invention aims to solve the problem that the latent faults cannot be accurately predicted under complex operating conditions due to signal aliasing caused by endogenous mechanical degradation characteristics and exogenous load disturbances.
[0006] In this technical solution, a method for predicting operational faults of a needle knife robot is implemented in a monitoring system including a signal processing unit and a remote computing unit, and includes the following steps: Step 101: Collect runtime sequence data, which includes instruction speed sequence, feedback speed sequence, and instruction position sequence; Step 102: Calculate the derivative of the command velocity sequence with respect to time to obtain the command acceleration value; Step 103: Monitor the polarity switching state of the command speed sequence, and determine the zero-crossing trigger point of the reversal by combining the command acceleration value. The zero-crossing trigger point of the reversal corresponds to the transient zero speed point when the motion direction of the actuator reverses. Step 104: In response to the zero-crossing trigger point of the reversal, open the zero-velocity observation window of a preset duration and collect the real-time velocity values of the feedback velocity sequence within the zero-velocity observation window; Step 105: Determine the duration during which the absolute value of the real-time speed is lower than the preset zero-speed dead zone threshold, in order to define the physical time delay parameter characterizing the backlash of the actuator transmission chain. Step 106: Align and map the physical delay parameters with the command position sequence in the spatial coordinate system to generate a health feature vector; Step 107: Input the health feature vector into the fault prediction model and obtain the health degradation index, which represents the evolution trajectory of the system's operating state, output by the fault prediction model. Step 108: When the health decline index exceeds the preset safety threshold, calculate the position compensation amount based on the offset determined by the health decline index, and superimpose the position compensation amount onto the instruction position sequence to generate a dynamic correction instruction.
[0007] Preferably, the process of determining the zero-crossing trigger point in step 103 includes the following sub-steps: Step 1031: Real-time comparison of adjacent sampling points of the command velocity sequence, and determination of the velocity polarity change interval where the product of adjacent sampling points is less than 0; Step 1032: Extracting the command acceleration value within the velocity polarity change interval, and determining the time corresponding to the velocity polarity change interval as the zero-crossing trigger point when the absolute value of the command acceleration value is greater than the preset acceleration threshold.
[0008] Preferably, the fault prediction model is composed of a deep learning layer and a reinforcement learning layer connected in series. The deep learning layer is used to extract the nonlinear trend parameters of the health feature vector, and the reinforcement learning layer is used to calibrate the nonlinear trend parameters based on historical operation logs in the cloud.
[0009] Preferably, step 106 further includes the following calibration process: step 1061: obtain the reference time delay sequence of the actuator in the fault-free initial operation phase; step 1062: use the reference time delay sequence to perform mean centering processing on the physical time delay parameters to eliminate the influence of the inherent assembly deviation of the transmission chain on the wear trend prediction.
[0010] Preferably, while acquiring real-time velocity values, the method also includes an adjustment step for the zero-velocity observation window: monitoring the numerical changes in the command velocity sequence; when the value of the command velocity sequence exceeds 50 mm / s, linearly increasing the sampling duration of the zero-velocity observation window by a preset step size.
[0011] Preferably, step 108 specifically includes: step 1081: comparing the health decline index with multiple preset risk ranges; step 1082: matching the corresponding step compensation gain according to the risk level determined by the comparison results.
[0012] Preferably, the stepped compensation gain includes: adjusting the feedforward gain coefficient in the dynamic correction command when the health decline index is in the first risk range, in order to offset the position lag caused by the reverse backlash.
[0013] Preferably, the stepped compensation gain includes: when the health decline index is in the second risk range, issuing a speed limiting command to the drive circuit to reduce the commutation speed of the actuator.
[0014] Preferably, the method further includes an alarm locking step: when the health degradation index reaches the failure threshold of 0.95, a warning signal for physical wear of the transmission chain is generated and the drive power is cut off, forcibly locking the motion degree of freedom of the actuator.
[0015] A fault prediction system for acupuncture robot is provided to implement the fault prediction method for acupuncture robot as described in claim 1. The system includes: a data acquisition module, a signal processing unit, a remote computing unit, and a drive compensation module. The data acquisition module is used to collect runtime sequence data, which includes the issued instruction speed sequence, feedback speed sequence, and instruction position sequence. The signal processing unit is used to calculate the derivative of the command velocity sequence with respect to time to obtain the command acceleration value, and to determine the reversal zero-crossing trigger point corresponding to the reversal of the motion direction of the controlled mechanism by combining the polarity switching state of the command velocity sequence. The signal processing unit is also used to open the zero-velocity observation window in response to the commutation zero-crossing trigger point, and to acquire the real-time velocity value of the feedback velocity sequence within the zero-velocity observation window in order to determine the physical time delay parameter; The remote computing unit, connected to the signal processing unit, is used to align and map physical delay parameters with instruction position sequences to generate health feature vectors and obtain the health degradation index output by the fault prediction model. The drive compensation module is used to calculate the position compensation amount based on the offset determined by the health decline index when the health decline index exceeds a preset safety threshold, and send the position compensation amount to the controlled mechanism for deviation correction.
[0016] Compared with existing technologies, the fault prediction method and system for needle knife robot of the present invention has the following advantages: 1. In the fault prediction of the needle knife robot, the polarity reversal transient of the command speed sequence is monitored in real time, and an independent time-domain observation window is established when the speed crosses zero. Combined with the measurement of the duration of the feedback speed within the zero-speed dead zone threshold, a deterministic physical correlation between the transmission chain backlash and the command response hysteresis is established. By utilizing the physical characteristics of the external tissue load being in the static friction transition period at the moment of the actuator reversal, the pure endogenous mechanical decay characteristics are separated from the high-intensity background noise without changing the hardware topology of the control system. This avoids false alarms caused by legitimate impedance fluctuations under high load conditions and ensures that the status monitoring process has a physical signal-to-noise ratio gain.
[0017] 2. The synchronous alignment of runtime sequence data by edge computing nodes, combined with the real-time correction of individual benchmarks by adaptive feature extraction algorithms, forms a background noise compensation mechanism for individual machine differences. This mechanism dynamically updates the global judgment boundary by extracting the residual features of the initial fault-free cycle of the equipment, enabling the prediction model to automatically adapt to the characteristic drift caused by different equipment due to manufacturing tolerances, assembly gaps, or long-term operating environment differences. This multi-level data calibration logic enhances the robustness of the fault prediction system to non-uniform operating conditions and solves the perception blind spot problem presented by traditional fixed threshold monitoring logic when facing dynamic alternating loads.
[0018] 3. By deeply coupling state change smoothing control with multi-level decay index, a closed-loop intervention chain from state perception to command correction is constructed. This chain automatically triggers feedforward gain adjustment or tiered clamping of drive speed and acceleration according to the predicted health index range, enabling the control system to actively avoid latent decay. This preventive dynamic compensation logic, before physical damage crosses the failure threshold, smooths the impact of accumulated mechanical errors on surgical trajectory accuracy by adjusting the control envelope. This not only reduces the risk of sudden jamming of the transmission mechanism, but also maintains the operational reliability of the control loop throughout the entire life cycle of the equipment. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating the steps of the method for predicting operational failures of the needle knife robot according to the present invention. Figure 2 This is a functional module interaction architecture diagram of the needle knife robot operation fault prediction system of the present invention. Detailed Implementation
[0020] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.
[0021] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, bottom, transverse, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.
[0022] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.
[0023] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0024] A method for predicting operational failures of a needle knife robot, implemented in a monitoring system including a signal processing unit and a remote computing unit, includes the following steps: Step 101: Collect runtime sequence data, which includes instruction speed sequence, feedback speed sequence, and instruction position sequence; Step 102: Calculate the derivative of the command velocity sequence with respect to time to obtain the command acceleration value; Step 103: Monitor the polarity switching state of the command speed sequence, and determine the zero-crossing trigger point of the reversal by combining the command acceleration value. The zero-crossing trigger point of the reversal corresponds to the transient zero speed point when the motion direction of the actuator reverses. Step 104: In response to the zero-crossing trigger point of the reversal, open the zero-velocity observation window of a preset duration and collect the real-time velocity values of the feedback velocity sequence within the zero-velocity observation window; Step 105: Determine the duration during which the absolute value of the real-time speed is lower than the preset zero-speed dead zone threshold, in order to define the physical time delay parameter characterizing the backlash of the actuator transmission chain. Step 106: Align and map the physical delay parameters with the command position sequence in the spatial coordinate system to generate a health feature vector; Step 107: Input the health feature vector into the fault prediction model and obtain the health degradation index, which represents the evolution trajectory of the system's operating state, output by the fault prediction model. Step 108: When the health decline index exceeds the preset safety threshold, calculate the position compensation amount based on the offset determined by the health decline index, and superimpose the position compensation amount onto the instruction position sequence to generate a dynamic correction instruction.
[0025] Preferably, the process of determining the zero-crossing trigger point in step 103 includes the following sub-steps: Step 1031: Real-time comparison of adjacent sampling points of the command velocity sequence, and determination of the velocity polarity change interval where the product of adjacent sampling points is less than 0; Step 1032: Extracting the command acceleration value within the velocity polarity change interval, and determining the time corresponding to the velocity polarity change interval as the zero-crossing trigger point when the absolute value of the command acceleration value is greater than the preset acceleration threshold.
[0026] Preferably, the fault prediction model is composed of a deep learning layer and a reinforcement learning layer connected in series. The deep learning layer is used to extract the nonlinear trend parameters of the health feature vector, and the reinforcement learning layer is used to calibrate the nonlinear trend parameters based on historical operation logs in the cloud.
[0027] Preferably, step 106 further includes the following calibration process: step 1061: obtain the reference time delay sequence of the actuator in the fault-free initial operation phase; step 1062: use the reference time delay sequence to perform mean centering processing on the physical time delay parameters to eliminate the influence of the inherent assembly deviation of the transmission chain on the wear trend prediction.
[0028] Preferably, while acquiring real-time velocity values, the method also includes an adjustment step for the zero-velocity observation window: monitoring the numerical changes in the command velocity sequence; when the value of the command velocity sequence exceeds 50 mm / s, linearly increasing the sampling duration of the zero-velocity observation window by a preset step size.
[0029] Preferably, step 108 specifically includes: step 1081: comparing the health decline index with multiple preset risk ranges; step 1082: matching the corresponding step compensation gain according to the risk level determined by the comparison results.
[0030] Preferably, the stepped compensation gain includes: adjusting the feedforward gain coefficient in the dynamic correction command when the health decline index is in the first risk range, in order to offset the position lag caused by the reverse backlash.
[0031] Preferably, the stepped compensation gain includes: when the health decline index is in the second risk range, issuing a speed limiting command to the drive circuit to reduce the commutation speed of the actuator.
[0032] Preferably, the method further includes an alarm locking step: when the health degradation index reaches the failure threshold of 0.95, a warning signal for physical wear of the transmission chain is generated and the drive power is cut off, forcibly locking the motion degree of freedom of the actuator.
[0033] Preferably, a fault prediction system for a needle knife robot includes: a data acquisition module, a signal processing unit, a remote computing unit, and a drive compensation module; The data acquisition module is used to collect runtime sequence data, which includes the issued instruction speed sequence, feedback speed sequence, and instruction position sequence. The signal processing unit is used to calculate the derivative of the command velocity sequence with respect to time to obtain the command acceleration value, and to determine the reversal zero-crossing trigger point corresponding to the reversal of the motion direction of the controlled mechanism by combining the polarity switching state of the command velocity sequence. The signal processing unit is also used to open the zero-velocity observation window in response to the commutation zero-crossing trigger point, and to acquire the real-time velocity value of the feedback velocity sequence within the zero-velocity observation window in order to determine the physical time delay parameter; The remote computing unit, connected to the signal processing unit, is used to align and map physical delay parameters with instruction position sequences to generate health feature vectors and obtain the health degradation index output by the fault prediction model. The drive compensation module is used to calculate the position compensation amount based on the offset determined by the health decline index when the health decline index exceeds a preset safety threshold, and send the position compensation amount to the controlled mechanism for deviation correction.
[0034] Example 1: In the case of cutting high-density fascia tissue with a needle knife robot, the actuator faces severe nonlinear tissue impedance. Due to the reverse gap caused by physical wear inside the transmission chain, the gap produces microscopic hysteresis of command response during frequent reciprocating reversing movements. However, during the continuous load work phase, the external load pulse masks the feedback characteristics generated by the transmission gap, making it difficult for the fault prediction system to identify hidden decay.
[0035] To achieve physical-level decoupling between intrinsic wear characteristics and extrinsic load disturbances, the system utilizes the physical characteristic of the zero-crossing point of the operating speed to acquire the command speed sequence in real time through the signal processing unit. ,calculate The first derivative is used to monitor its polarity reversal state, thereby determining the zero-crossing trigger point for reversing the actuator's motion direction. Upon detecting the trigger point, a zero-speed observation window of a set duration is opened. During this directional transient, although the front-end needle knife continues to bear the nonlinear viscoelastic restoring force from the soft tissue, due to the inherent mechanical backlash within the actuator's transmission chain, the drive motor's minute rotational displacement at the initial stage of reversal is only used to traverse this physical space and has not yet established a rigid mechanical coupling with the front-end load. Therefore, within the extremely short time window before this reverse space is completed, the motor's feedback signal only reflects the intrinsic mechanical damping, thus effectively shielding the reverse transmission of external high-intensity soft tissue stress on the physical topology link. This allows the static friction conversion assumption to hold at the local microscale. Within the zero-speed observation window, the system collects the feedback velocity sequence. ,calculate The absolute value remains at the preset zero-speed dead zone threshold. The duration within which the physical time delay parameter characterizing the backlash of the actuator drive train is determined. This parameter is used to quantify the absolute time loss caused by the actuator crossing the physical wear idle distance.
[0036] Get The system maps the command position sequence to generate a health feature vector, which is then input into a fault prediction model composed of deep learning and reinforcement learning layers to obtain a health degradation index that characterizes the evolution trajectory of the system's operating state. Even under the interference of tissue impedance, this mechanism still extracts a gap increment feature of 0.02mm. When the health degradation index exceeds a preset safety threshold, the drive compensation module calculates the position compensation amount and superimposes it onto the command position sequence to generate a dynamic correction command to offset the physical gap, thereby achieving predictive maintenance of the underlying hardware physical aging.
[0037] Example 2: The effectiveness of the fault prediction method was verified on a needle knife robot test platform equipped with a high-precision servo motor, torque sensor, and 0.001mm resolution position encoder under a tissue cutting resistance condition of 150N. The data collected by the test platform came from the physical experimental system, where the sampling frequency of the servo motor was... The sampling frequency is set to 1000Hz. The design considerations are to balance the real-time performance of data acquisition with the processing load of the computing unit; to ensure the complete zero-speed crossing trajectory is captured and to avoid signal aliasing, when the speed reversal period of the actuator is between 5ms and 10ms, The median value of its criterion range of 500Hz to 2000Hz is selected, that is, the sampling period is set to 1ms.
[0038] To simulate electromagnetic interference in a real surgical environment, Gaussian white noise with a signal-to-noise ratio of 20dB and power frequency interference harmonics at a frequency of 50Hz were superimposed on the runtime sequence data. The experiment was divided into a control group and an experimental group. The control group used a conventional path based on full-time residual monitoring, while the experimental group used an intrinsic state decoupling method based on zero-crossing dead zone isolation. The zero-velocity dead zone threshold was used in the experiment. The determination of this factor lies in balancing wear sensitivity with sensor background noise suppression; based on the static noise level of the measurement system, The value is set to 5 times the noise standard deviation, i.e., 0.5 mm / s, in order to extract micro-gap features while suppressing random fluctuations.
[0039] During the operation of the test group, the signal processing unit calculated the command speed sequence in real time. The polarity switching state, and the zero-crossing trigger point of the reversal is captured. A zero-velocity observation window with a duration of 20ms is opened; the feedback velocity sequence within the zero-velocity observation window is acquired. The calculated physical delay parameters The wear pattern exhibits a gradient evolution; when the manually set wear amount of the actuator transmission chain increases from 0.01mm to 0.05mm and then to 0.10mm, the test group measured... The times were 2.8ms, 10.2ms, and 19.5ms respectively; the data shows that... The test showed a linear correlation of 0.98 with the physical wear amount. In contrast, under the influence of a 150N tissue impedance pulse, the command-execution position residual of the control group fluctuated between 0.12mm and 0.45mm, and the signal distortion amplitude exceeded the intrinsic wear characteristic of 0.05mm. When the wear amount exceeded the performance inflection point of 0.18mm, due to the deterioration of the transmission chain stiffness, the duration of the feedback speed in the zero-speed dead zone tended to saturate. At this time, the health degradation index climbed from 0.85 to the failure threshold of 0.96. The test results confirmed that the fault prediction accuracy of the test group under the high tissue impedance interference environment was 97.2%, which was higher than the accuracy of 42.5% of the control group, and achieved stable locking of the physical degradation trajectory of the control system.
[0040] Example 3: This example combines Figures 1 to 2 This document describes a method and system for predicting operational faults in a needle knife robot. Figure 1 As shown, step 101 involves collecting runtime sequence data, including the command velocity sequence, feedback velocity sequence, and command position sequence. The process proceeds by calculating the derivative of the command velocity sequence with respect to time in step 102 (using a downward arrow) to obtain the command acceleration value. Then, step 103, using a downward arrow, monitors the command velocity polarity switching state. Combined with the command acceleration, the process determines the zero-crossing trigger point for the corresponding actuator's reversal. Finally, step 104, using a downward arrow, responds to the zero-crossing trigger point by opening a preset-length zero-velocity observation window and collecting the real-time velocity value within the window. Finally, step 105, using a downward arrow, confirms... The duration for which the absolute value of the real-time speed is below a preset dead zone threshold is defined to determine the physical delay parameter characterizing the backlash of the transmission chain. Then, via step 106 (pointing to the downward arrow), the physical delay parameter and the command position sequence are aligned and mapped in the spatial coordinate system to generate a health feature vector. Next, via step 107 (pointing to the downward arrow), the health feature vector is input into the fault prediction model to obtain a health degradation index characterizing the evolution trajectory of the system's operating state. Finally, via step 108 (pointing to the end of the downward arrow), when the health degradation index exceeds a preset safety threshold, the position compensation amount is calculated based on the offset and superimposed on the command position sequence to generate a dynamic correction command.
[0041] like Figure 2As shown, the module interaction logic and overall functional node layout used to support the operation of this scheme are illustrated. The main body of the diagram includes a centrally located main rectangle. A regular hexagonal frame containing the execution mechanism text stands on the left boundary of the main rectangle. Three lines extend from this hexagonal frame into the main rectangle, directly connecting to three elliptical frames within the main rectangle: one containing text for collecting runtime sequence data, one containing text for generating dynamic correction instructions, and one containing text for issuing speed limit instructions. Within the flow space of the main rectangle, the elliptical frame for collecting runtime sequence data extends two downward dashed arrows: one pointing to an elliptical frame containing text for determining physical delay parameters, and the other directly pointing to an elliptical frame containing text for obtaining the health decline index. The elliptical frame for determining physical delay parameters is then connected via the downward dashed lines... The arrows point to the aforementioned elliptical frame for obtaining the health decline index. Four unidirectional dashed arrows extend outward from the elliptical frame for obtaining the health decline index. These arrows point in opposite directions to the aforementioned elliptical frame for generating dynamic correction instructions, to the elliptical frame containing text for calibrating nonlinear trend parameters, to the aforementioned elliptical frame for issuing speed limit instructions, and to the elliptical frame containing text for cutting off the drive power supply in the lower right corner. In addition, the elliptical frame for calibrating nonlinear trend parameters inside the main rectangle extends beyond the boundary of the main rectangle with a unidirectional arrow pointing to the right, connecting to an external cloud-shaped graphic containing text for cloud server. The elliptical frame for cutting off the drive power supply inside the main rectangle extends beyond the boundary of the main rectangle with a downward-slightly-right connecting line, connecting to an external rectangular graphic with internal double vertical lines on both sides and text for power management module.
[0042] Example 4: Parameter calibration process for individual assembly deviations and sensor background noise in servo control systems. Due to physical differences in the actuator transmission chain, the system collects feedback velocity sequences from a stationary state during the initial operation phase. A 5-second time window was selected to obtain 5000 raw sampling points, and the feedback velocity sequence was calculated. Standard deviation According to the formula Determine the zero-velocity dead zone threshold, where The noise suppression coefficient is given when the measurement system resolution is 0.001 mm. The value is 5, which is why it is determined. To cover 99.9% of random measurement noise fluctuations, the physical time delay parameter... The extraction only points to the reverse physical gap of the transmission chain, and the physical delay parameters are... When generating healthy feature vectors by aligning and mapping the command position sequence in the spatial coordinate system, a transmission ratio-based quantization conversion operation is performed. The signal processing unit extracts the instantaneous velocity value of the command velocity sequence at the end of the zero-velocity observation window. And the numerical value with physical delay parameters Multiply to obtain the nominal spatial lag distance Call the underlying servo driver encoder feedback pulse equivalent Converted nominal spatial lag distance Number of position deviation pulses generated The number of position deviation pulses The instruction location sequence at the corresponding timestamp By stitching together absolute coordinate data, the peak amplitude of the driving current during commutation transients is extracted. A multidimensional numerical array is jointly constructed as the health feature vector input fault prediction model. The peak amplitude of the drive current is acquired in real time by the current loop at the bottom of the servo driver through the shunt resistor connected in series in the motor winding circuit. The acquired analog electrical signal is discretized by the high-frequency analog-to-digital converter inside the driver and then synchronously transmitted to the dedicated register of the signal processing unit through the field communication bus to accurately quantify the dynamic excitation energy consumed by the commutation transient to overcome the reverse mechanical friction.
[0043] The fault prediction model receives input data as follows: The health feature tensor has 10 dimensions, where 10 represents the feature dimension including position residuals, velocity hysteresis, force feedback gradients, and physical delay parameters. To determine the number of sampling points within the sliding time window, the deep learning layer employs a three-layer one-dimensional convolutional neural network structure. The first layer's convolutional kernel size is set to 3, used to extract transient fluctuation features from the runtime sequence data. The convolutional feature map is activated by a linear rectified function and passed to subsequent layers to obtain nonlinear evolution trend parameters. The reinforcement learning layer is connected in series with the deep learning layer, receiving the nonlinear evolution trend parameters as state space input. Historical operating logs of the same model of equipment are retrieved from the cloud server to identify sample labels similar to the current trajectory. The reinforcement learning algorithm uses policy gradient optimization logic, calculating the residual between the predicted health decline index and the actual hardware wear state at the end of each prediction cycle, and generating a reward signal based on this residual. When the reward signal continuously falls below the set weight, the cloud server triggers model parameter reconstruction, adjusting the weight matrix within the neural network to achieve adaptive learning of tissue impedance interference patterns under different environments. The reinforcement learning layer utilizes historical operating logs from the cloud to perform calibration on the nonlinear trend parameters, calculating the actual hardware wear state required for the reinforcement learning algorithm to calculate the residual. The data originates from the physical benchmark calibration procedure during the equipment's standby period. During the self-test cycle, the control system sends an amplitude of the rated maximum continuous torque to the actuator. Low-frequency alternating torque command It also monitors fluctuations in the end-position encoder values and extracts alternating torque commands. Encoder feedback positive limit coordinates under action Reverse limit coordinates The difference is defined as the actual physical clearance. The cloud server will handle the actual physical bandwidth. The residual value is obtained by encapsulating the loss label vector at the corresponding time node and performing a difference operation with the predicted output value of the reinforcement learning layer. If the residual value If the deviation exceeds the preset threshold for five consecutive cycles, the gradient descent update of the weight matrix in the cloud is triggered and synchronized to the local computing unit. By obtaining the pure intrinsic actual physical gap during the self-test standby period when it is not affected by external organizational impedance interference, and using it as the absolute benchmark label, the reinforcement learning model can establish a baseline comparison anchor point in the subsequent calculation of high impedance dynamic evolution scenarios. The feature vectors in the runtime sequence data are differentially aligned using this unloaded anchor point, thereby effectively filtering out the dynamic elastic deformation noise caused by nonlinear loads and ensuring the data purity of dynamic calibration.
[0044] Before being transmitted to the drive compensation module, the acquired health decline index is first de-identified by a temporary register in memory, retaining the floating-point value representing the physical attributes while removing the device's unique identifier. The drive compensation module then calculates the position compensation amount based on the offset determined by the health decline index. and will The instruction register of the underlying controller is written with 16-bit precision, and a dynamic correction instruction is generated by arithmetic superposition with the original target coordinate set. Specifically, when the drive compensation module is calculating, it first extracts the nominal spatial lag distance from the system's factory calibration, performs a differential operation between the currently acquired health degradation index and a preset safety threshold to obtain a dimensionless difference, and then directly multiplies this dimensionless difference with the nominal spatial lag distance. The resulting product is mapped to the actual mechanical clearance increment under the current wear state, and this increment is directly used as the position compensation amount. This method is effective when the wear is within the degradation range of 0.05mm to 0.15mm. Maintaining the position compensation frequency at 100Hz offsets over 85% of the trajectory hysteresis deviation caused by transmission backlash. When the system detects a health degradation index reaching the failure threshold of 0.95, the signal processing unit sends a level transition signal to the power management module, cutting off the drive power and locking the actuator's degrees of freedom to avoid surgical risks caused by accumulated physical wear. In a working condition that classifies and judges abnormal actuator states, the reinforcement learning layer determines the failure mode characteristics corresponding to the current loss by analyzing the nonlinear evolution trend parameters extracted by the deep learning layer. The fault prediction model classifies the physical loss state into transmission chain wear states. Sensor zero-point drift state Related to motor rotor overheating The system acquires normalized probability distribution vectors for different failure mode characteristics within each sampling period. ,in, Pointing to the physical delay parameter Monotonically increasing pattern probabilities are compared by the signal processing unit with the alignment vector. The maximum probability in, when Greater than Furthermore, when the position residual fluctuation characteristics match the wear characteristics in the discretized mapping table, the system outputs a classification result for the potential fault types of physical aging of the transmission chain.
[0045] Example 5: In the case of precision calibration and benchmark feature library construction for a servo control system, the system uses a standard organization model with a known viscoelastic coefficient to generate an impedance gradient of 50N to 200N. A laser displacement meter is used to measure the displacement deviation of the transmission chain in step increments of 0.01mm to 0.20mm, and command speed sequences under different load levels are collected. With feedback velocity sequence Calculate physical delay parameters This leads to the determination of physical wear. and A discretized mapping table between tissues was created and input into the fault prediction model as the initial weights for the deep learning layer. For tissue models of different hardness levels, 1000 reciprocating commutation actions were performed, and the standard deviation of the sampling points was collected. The statistical distribution characteristics were determined by calculating the statistical distribution characteristics under different working conditions. The mean drift is used to determine a preset safety threshold for a specific hardware batch, enabling the system to acquire characteristic data on the transmission chain loss pattern before accessing real-time data, and ensuring that the command speed sequence value exceeds... The preset step size is used in the adjustment process of linearly increasing the sampling time of the zero-velocity observation window. The values were established by calibrating the mechanical inertia decay response characteristics of the actuator. During the initial commissioning phase, a frequency sweep speed command sequence covering the entire operating speed range was issued to the actuator, and the actual physical time for the feedback speed to return to zero was measured at different commutation speed nodes. Regarding the commutation speed and actual physical time consumption Perform a first-order linear regression fit on the corresponding data point set to obtain the slope parameter of the straight line. The slope parameter of the line With the system's basic control cycle Multiply and round down to the nearest integer multiple of the milliseconds to define the preset step size. The data is stored in memory for the monitoring system to retrieve and calculate. The aforementioned adjustment threshold of 50 mm / s is determined based on the electromechanical braking characteristics of the drive system. When the operating speed of the actuator is lower than this value, the residual kinetic energy of the rotor can be completely absorbed and dissipated within the normal window period by the electromagnetic reverse braking inside the servo driver. However, once the speed limit is exceeded, the system will generate a non-negligible mechanical sliding wake due to the surge in inertia. If the observation window is not extended according to the speed gradient, some hysteresis features will be outside the sampling range.
[0046] When the system is applied to the debugging condition after replacing the transmission components, the signal processing unit starts a 10-second self-test program before the actual cutting action. The drive component generates a simple harmonic reciprocating motion with a frequency of 2Hz and an amplitude of 5mm in space, and calculates the command speed sequence. The dwell time amplitude when passing through the zero axis determines the initial physical delay parameters of the current hardware system. ,use The judgment benchmark issued by the cloud server is zero-point shifted and corrected, and the corrected feature vector is used as the initial state space input of the reinforcement learning layer. In the first 5 cycles of the cutting action, the system adjusts the reward signal weight factor inside the model according to the measured force feedback gradient. The local benchmark is verified by monitoring the fluctuation of the health decline index in the preceding stage. If the index fluctuation difference is less than 0.01 for 3 consecutive cycles, the parameter adjustment is completed and the system is switched to the prediction monitoring mode. At this time, the position compensation calculation logic of the driving component points to the deviation caused by component wear.
[0047] In scenarios where quantitative assessment of equipment wear and tear is required, the fault prediction model relies on the acquired physical delay parameters. The initial physical delay parameters determined during the initial self-test phase Calculate the health decline index that characterizes the trajectory of the evolution of the operating state. The specific calculation logic is to calculate the current delay increment. According to the formula Determine the value of the health decline index, where, The preset exponential decay constant is calibrated to 0.12 based on the displacement deviation sensitivity of the transmission chain at the failure critical point. The resulting value... The index is limited to the range of 0 to 1, so as to characterize the evolution of the health status of the control system from standard operating conditions to failure conditions.
[0048] Example 6: In the adaptive parameter matrix calibration procedure for actuators from different manufacturing batches, the system faces the problem of position hysteresis deviation caused by differences in physical wear characteristics. To establish the health degradation index... The quantization mapping relationship between the control gain and the underlying loop is defined by dividing the system into multiple logical intervals according to a standardized process. When the value is in the range of 0.7 to 0.9, the drive chain is determined to be in a state of slight wear. At this time, the drive compensation module reads the feedforward gain coefficient from the underlying control loop. And calculate the updated feedforward gain coefficient. The specific calculation formula is as follows: ,in, This is a correction factor, the value of which is based on the physical time delay parameter. The sensitivity test data for position deviation was set to 1.5. This adjustment was injected into the underlying control loop via the controller's instruction register to compensate for the effects caused by the physical gap by adjusting the drive current response speed during commutation transients.
[0049] During the maintenance procedure of remote model iteration and parameter synchronization, the remote computing unit obtains historical operation logs stored on the cloud server, identifies characteristic patterns that match the current equipment wear trajectory, and the system calculates physical delay parameters. Mean shift within the preset sampling window To determine whether the weight matrix of the fault prediction model needs to be reconstructed, when... Exceeding the benchmark value When the trend shows a monotonically increasing trend, the cloud server initiates a parameter update process, using labeled sample data to correct the optimization strategy of the reinforcement learning layer, and transmits the generated updated weight packet to the signal processing unit via an encrypted link. The system completes the overwriting of local model parameters during the interval of the uninterrupted switching operation. When the health decline index is monitored... When the failure threshold of 0.95 is reached, the signal processing unit sends a level reversal signal to the power management module, cuts off the power supply circuit of the drive motor and triggers mechanical self-locking, so that the operating state of the actuator is maintained within the safety envelope.
[0050] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.
Claims
1. A method for predicting operational faults in a needle-knife robot, implemented in a monitoring system comprising a signal processing unit and a remote computing unit, characterized in that, Includes the following steps: Step 101: Collect runtime sequence data, which includes instruction speed sequence, feedback speed sequence, and instruction position sequence; Step 102: Calculate the derivative of the command velocity sequence with respect to time to obtain the command acceleration value; Step 103: Monitor the polarity switching state of the command speed sequence, and determine the zero-crossing trigger point of the reversal by combining the command acceleration value. The zero-crossing trigger point of the reversal corresponds to the transient zero speed point when the motion direction of the actuator reverses. Step 104: In response to the zero-crossing trigger point of the reversal, open the zero-velocity observation window of a preset duration and collect the real-time velocity values of the feedback velocity sequence within the zero-velocity observation window; Step 105: Determine the duration during which the absolute value of the real-time speed is lower than the preset zero-speed dead zone threshold, in order to define the physical time delay parameter characterizing the backlash of the actuator transmission chain. Step 106: Align and map the physical delay parameters with the command position sequence in the spatial coordinate system to generate a health feature vector; Step 107: Input the health feature vector into the fault prediction model and obtain the health degradation index, which represents the evolution trajectory of the system's operating state, output by the fault prediction model. Step 108: When the health decline index exceeds the preset safety threshold, calculate the position compensation amount based on the offset determined by the health decline index, and superimpose the position compensation amount onto the instruction position sequence to generate a dynamic correction instruction.
2. The method for predicting operational faults of a needle knife robot according to claim 1, characterized in that, The process of determining the zero-crossing trigger point in step 103 includes the following sub-steps: Step 1031: Compare adjacent sampling points of the command velocity sequence in real time and determine the velocity polarity change interval where the product of adjacent sampling points is less than 0; Step 1032: Extract the command acceleration value within the velocity polarity change interval, and determine the time corresponding to the velocity polarity change interval as the zero-crossing trigger point when the absolute value of the command acceleration value is greater than the preset acceleration threshold.
3. The method for predicting operational faults of a needle knife robot according to claim 1, characterized in that, The fault prediction model consists of a deep learning layer and a reinforcement learning layer connected in series. The deep learning layer is used to extract the nonlinear trend parameters of the health feature vector, and the reinforcement learning layer is used to calibrate the nonlinear trend parameters based on historical operation logs in the cloud.
4. The method for predicting operational faults of a needle knife robot according to claim 1, characterized in that, Step 106 further includes the following calibration process: Step 1061: Obtain the reference time delay sequence of the actuator during the fault-free initial operation phase; Step 1062: Use the reference time delay sequence to perform mean centering processing on the physical time delay parameters to eliminate the influence of the inherent assembly deviation of the transmission chain on the wear trend prediction.
5. The method for predicting operational faults of a needle knife robot according to claim 1, characterized in that, In addition to collecting real-time velocity values, the process also includes adjusting the zero-velocity observation window: monitoring changes in the values in the command velocity sequence; and linearly increasing the sampling duration of the zero-velocity observation window by a preset step size when the values in the command velocity sequence exceed 50 mm / s.
6. The method for predicting operational faults of a needle knife robot according to claim 1, characterized in that, Step 108 specifically includes: Step 1081: Compare the health decline index with multiple preset risk ranges; Step 1082: Match the corresponding step compensation gain according to the risk level determined by the comparison results.
7. The method for predicting operational faults of a needle knife robot according to claim 6, characterized in that, The stepped compensation gain includes adjusting the feedforward gain coefficient in the dynamic correction command when the health decline index is in the first risk range, in order to offset the position lag caused by the reverse backlash.
8. The method for predicting operational faults of a needle knife robot according to claim 6, characterized in that, The stepped compensation gain includes: when the health decline index is in the second risk range, issuing a speed limit command to the drive circuit to reduce the commutation speed of the actuator.
9. The method for predicting operational faults of a needle knife robot according to claim 1, characterized in that, The method also includes an alarm locking step: when the health degradation index reaches the failure threshold of 0.95, an early warning signal for physical wear of the transmission chain is generated and the drive power is cut off, forcibly locking the motion degree of freedom of the actuator.
10. A fault prediction system for a needle knife robot, used to implement the fault prediction method for a needle knife robot as described in claim 1, characterized in that, The system includes: a data acquisition module, a signal processing unit, a remote computing unit, and a drive compensation module; The data acquisition module is used to collect runtime sequence data, which includes the issued instruction speed sequence, feedback speed sequence, and instruction position sequence. The signal processing unit is used to calculate the derivative of the command velocity sequence with respect to time to obtain the command acceleration value, and to determine the reversal zero-crossing trigger point corresponding to the reversal of the motion direction of the controlled mechanism by combining the polarity switching state of the command velocity sequence. The signal processing unit is also used to open the zero-velocity observation window in response to the commutation zero-crossing trigger point, and to acquire the real-time velocity value of the feedback velocity sequence within the zero-velocity observation window in order to determine the physical time delay parameter; The remote computing unit, connected to the signal processing unit, is used to align and map physical delay parameters with instruction position sequences to generate health feature vectors and obtain the health degradation index output by the fault prediction model. The drive compensation module is used to calculate the position compensation amount based on the offset determined by the health decline index when the health decline index exceeds a preset safety threshold, and send the position compensation amount to the controlled mechanism for deviation correction.