A method and system for early warning of motor faults in industrial robots

Abstract

This invention relates to the field of industrial motor fault technology, and more particularly to a method and system for early warning of industrial robot motor faults. The method includes the following steps: acquiring industrial robot motor control commands in real time, then judging command value change events to obtain motor load change events; monitoring transient operating condition characteristic data of the motor in real time based on the motor load change events; evaluating the abnormal abrupt response of the flexible sheave based on the transient operating condition characteristic data of the motor, generating a comprehensive response abnormality score for the flexible sheave; performing electromagnetic fault characteristic analysis based on the transient operating condition characteristic data of the motor, obtaining electromagnetic fault characteristic coupling data; and performing flexible sheave health status analysis based on the electromagnetic fault characteristic coupling data and the comprehensive response abnormality score of the flexible sheave, thereby achieving early warning of industrial robot motor faults. This invention, through the detection of operating condition abrupt points and the analysis of the motor flexible sheave health status, can accurately identify motor faults and achieve multi-level fault early warning.

Description

Technical Field

[0001] This invention relates to the field of industrial motor fault technology, and in particular to a method and system for early warning of industrial robot motor faults. Background Technology

[0002] Harmonic reducers are widely used in the joint transmission systems of industrial robots due to their advantages such as high transmission accuracy, small size, light weight, and large transmission ratio. Their core component, the flexible wheel, transmits power through flexible deformation. Its working state directly reflects the motor's load condition. Motor faults, especially common inter-turn short-circuit faults, cause fluctuations in the motor's output torque, which in turn affects the deformation pattern of the flexible wheel. Therefore, monitoring the deformation state of the flexible wheel can reflect the motor's health status. However, traditional fault warning methods for industrial robot motors mainly focus on analyzing signals such as the motor's current, voltage, vibration, and temperature. These methods achieve good results under constant operating conditions, but under complex operating conditions such as variable speed and variable load, signal fluctuations caused by load changes often mask fault characteristics, leading to false alarms or missed alarms. For example, an inter-turn short-circuit fault in the motor windings will cause changes in the motor's current and vibration signals, but under sudden load changes, the current and vibration fluctuations caused by the load change are similar to fault characteristics, making them difficult to distinguish. This makes it difficult to set warning thresholds for traditional fault warning methods when dealing with variable operating conditions, resulting in poor adaptability and failing to meet the high reliability requirements of industrial robots. Summary of the Invention

[0003] Based on this, the present invention provides a method and system for early warning of motor faults in industrial robots to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, an industrial robot motor fault early warning method includes the following steps:

[0005] Step S1: Acquire the industrial robot motor control commands in real time, then judge the command value change event to obtain the motor load change event; monitor the transient operating condition characteristic data of the motor in real time based on the motor load change event.

[0006] Step S2: Perform deformation analysis of the flexible wheel under transient operating conditions based on the transient operating condition characteristic data of the motor, and generate deformation data of the flexible wheel under transient operating conditions; evaluate the abnormality of the flexible wheel's transient response based on the deformation data of the flexible wheel under transient operating conditions, and generate a comprehensive abnormality score of the flexible wheel's response.

[0007] Step S3: Perform electromagnetic fault feature analysis based on the transient operating condition feature data of the motor to obtain the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum respectively; perform multi-channel monitoring fault synchronous coupling on the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum to obtain electromagnetic fault feature coupling data.

[0008] Step S4: Analyze the health status of the motor flexure based on the electromagnetic fault characteristic coupling data and the flexure comprehensive response anomaly score to generate motor fault status data; perform multi-level fault early warning processing based on the motor fault status data to achieve early warning of motor faults in industrial robots.

[0009] Preferably, the present invention also provides an industrial robot motor fault early warning system, which executes the industrial robot motor fault early warning method described above. The industrial robot motor fault early warning system includes:

[0010] The transient operating condition monitoring module is used to acquire industrial robot motor control commands in real time, then judge command value change events to obtain motor load change events; and monitor the transient operating condition characteristic data of the motor in real time based on the motor load change events.

[0011] The flexible wheel response analysis module is used to perform flexible wheel deformation analysis under transient operating conditions based on the transient operating condition characteristic data of the motor, and generate flexible wheel deformation data under transient operating conditions; and to evaluate the abnormality of the flexible wheel's sudden response based on the flexible wheel deformation data under transient operating conditions, and generate a comprehensive abnormality score of the flexible wheel response.

[0012] The electromagnetic feature coupling module is used to perform electromagnetic fault feature analysis based on the transient operating condition feature data of the motor, and obtain the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum respectively; the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum are coupled synchronously with multi-channel monitoring faults to obtain electromagnetic fault feature coupling data.

[0013] The health assessment and early warning module is used to analyze the health status of the motor flexure based on electromagnetic fault characteristic coupling data and flexure comprehensive response anomaly score, and generate motor fault status data; based on the motor fault status data, multi-level fault early warning processing is performed to realize motor fault early warning for industrial robots.

[0014] This invention proactively identifies "events" of load or speed changes by accurately capturing changes in motor control commands, focusing the analysis on "transient operating conditions" during these abrupt changes in motor state. By specifically analyzing the transient characteristic data of the motor in response to command changes, it can more clearly identify abnormal dynamic responses caused by internal faults (such as torque pulsations caused by inter-turn short circuits), distinguishing them from dynamic responses caused by normal load changes. This fundamentally solves the problems of difficult-to-set warning thresholds and frequent false alarms and missed alarms under varying operating conditions, significantly improving the reliability of fault warnings. Flexible sheave deformation directly reflects motor load, and motor faults (especially inter-turn short circuits) cause output torque fluctuations, inevitably altering the deformation pattern of the flexible sheave. Based on this, an in-depth analysis of the flexible sheave deformation data during abrupt changes in motor operating conditions is conducted to assess whether its dynamic response is "abnormal." This method has a unique advantage: it directly focuses on the final link in power transmission affected by faults, namely the actual deformation behavior of the flexible sheave. Compared to simply analyzing the electrical or vibration signals of the motor itself, monitoring the transient response of the flexspline can more intuitively and sensitively capture minute torque anomalies that are sufficient to affect the precision of robot joint movements. Even if the fault is not yet obvious in the motor's electrical signals or is masked by noise, the resulting torque fluctuations can effectively identify the disturbances to the flexspline's precision deformation patterns, thus improving the detection sensitivity for early, hidden faults. Synchronous coupling analysis of multi-dimensional electromagnetic features enhances the robustness and specificity of fault feature extraction. Simultaneously monitoring current and magnetic field distortion—two types of electromagnetic signals closely related to the motor winding state—and generating their respective time-series spectra for feature analysis effectively filters out random noise interference and identifies more unique faults that are harder to mimic under normal operating conditions. For example, inter-turn short circuits not only cause current harmonics at specific frequencies but also lead to specific distortions in the local magnetic field distribution. The synchronous abnormal patterns of these two factors in the transient response further reduce the possibility of misjudgment and improve the accuracy of fault identification. By integrating mechanical response and electromagnetic characteristics for comprehensive health assessment and multi-level early warning, a comprehensive and reliable motor status monitoring system is achieved. Only when a suspected fault characteristic detected at the electromagnetic level simultaneously triggers a corresponding abnormal response at the mechanical transmission level (flexible wheel deformation) will the system provide a highly confident fault diagnosis. This dual verification mechanism significantly enhances the confidence of the final diagnostic results. The motor fault status data generated based on this comprehensive assessment more accurately reflects the motor's true health level. Furthermore, the implemented multi-level fault early warning processing can provide users with different levels of warning information, from "attention" to "emergency shutdown," based on the severity and development trend of the fault. This supports more refined predictive maintenance strategies, avoids catastrophic failures, ensures the high reliability and long-term stability of industrial robots in complex operating environments, and ultimately improves production efficiency and safety.Therefore, the industrial robot motor fault early warning method of the present invention accurately captures motor load change events by acquiring industrial robot motor control commands in real time and judging command value change events, and performs in-depth analysis on the collected transient working condition data, including working condition jump flexible wheel deformation analysis based on ellipse fitting, multi-channel electromagnetic fault feature extraction and coupling, and comprehensive health assessment combining mechanical and electromagnetic features, so as to achieve multi-dimensional and high-precision identification of fault features and effectively improve the accuracy and reliability of fault early warning. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of an industrial robot motor fault early warning method according to the present invention.

[0016] Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S2, which involves analyzing the deformation of the flexible wheel under transient operating conditions based on the transient operating condition characteristic data of the motor.

[0017] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0020] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] To achieve the above objectives, please refer to Figures 1 to 2 This invention provides a method for early warning of motor faults in industrial robots, comprising the following steps:

[0022] Step S1: Acquire the industrial robot motor control commands in real time, then judge the command value change event to obtain the motor load change event; monitor the transient operating condition characteristic data of the motor in real time based on the motor load change event.

[0023] Step S2: Perform deformation analysis of the flexible wheel under transient operating conditions based on the transient operating condition characteristic data of the motor, and generate deformation data of the flexible wheel under transient operating conditions; evaluate the abnormality of the flexible wheel's transient response based on the deformation data of the flexible wheel under transient operating conditions, and generate a comprehensive abnormality score of the flexible wheel's response.

[0024] Step S3: Perform electromagnetic fault feature analysis based on the transient operating condition feature data of the motor to obtain the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum respectively; perform multi-channel monitoring fault synchronous coupling on the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum to obtain electromagnetic fault feature coupling data.

[0025] Step S4: Analyze the health status of the motor flexure based on the electromagnetic fault characteristic coupling data and the flexure comprehensive response anomaly score to generate motor fault status data; perform multi-level fault early warning processing based on the motor fault status data to achieve early warning of motor faults in industrial robots.

[0026] In this embodiment of the invention, the industrial robot motor fault early warning method includes the following steps:

[0027] Step S1: Acquire the industrial robot motor control commands in real time, then judge the command value change event to obtain the motor load change event; monitor the transient operating condition characteristic data of the motor in real time based on the motor load change event.

[0028] In this embodiment of the invention, for example, the third axis (shoulder pitch axis) of an ABB IRB 6700 industrial robot uses an EtherCAT bus-controlled servo motor and a matching harmonic reducer (assuming model CSG-40-120). First, the EtherCAT master protocol stack deployed on an independent IPC subscribes to and parses the target speed command PDO (process data object, e.g., object dictionary 60FFh) broadcast from the servo driver of this axis (e.g., an ABB ACSM1 driver) in real time at 2-millisecond intervals. When the robot performs a gripping action on a heavy workpiece (e.g., weight exceeding 20% ​​of the calibrated normal load), the controller issues a rapid acceleration command to instantly increase the motor speed from 500 rpm to 2500 rpm. The monitoring program processes the received raw speed command sequence on the IPC. The latest sampling point (1678886400.134 seconds, 2500 rpm) and its preceding point (1678886400.132 seconds, 2400 rpm) are extracted, and the sampling time interval is calculated to be 10 milliseconds (meeting the 8-12ms requirement), with a command value difference of 100 rpm. Based on the motor's rated speed of 3000 rpm, the absolute value of the instantaneous change rate of the command is calculated to be |(100 / 3000)×100 / (10ms / 1000ms / s)|=333.3% / s. This value is much greater than the preset command rapid change judgment threshold of 10% per second, therefore, a "motor load change event" is determined to have occurred (time T0 is 1678886400.134 seconds). This event immediately triggers the IPC to send a 5V, 100-microsecond synchronous acquisition trigger pulse signal through the digital output port of the NIPXIe-6363 data acquisition card. The pulse signal synchronously initiates parallel multi-channel monitoring of the installed sensors: three-phase current acquired by a through-hole current sensor (50kHz sampling rate), magnetic field strength at three points near the stator detected by a Hall array (50kHz sampling rate), displacement measured by four eddy current displacement sensors attached to the inner wall of the flexible wheel (50kHz sampling rate), and rotation angle obtained by the motor shaft encoder (synchronous reading). The data acquisition system performs pre-trigger acquisition, automatically saving the data 50 milliseconds before the trigger and continuously acquiring data 200 milliseconds after the trigger. These two data segments are spliced ​​together and packaged together with the event trigger information (timestamp T0, rate of change 333.3% / s, event description "grabbing an overweight workpiece") to generate a complete "motor transient operating condition characteristic data" record.

[0029] Step S2: Perform deformation analysis of the flexible wheel under transient operating conditions based on the transient operating condition characteristic data of the motor, and generate deformation data of the flexible wheel under transient operating conditions; evaluate the abnormality of the flexible wheel's transient response based on the deformation data of the flexible wheel under transient operating conditions, and generate a comprehensive abnormality score of the flexible wheel's response.

[0030] In this embodiment of the invention, utilizing the data from four flexible wheel displacement sensors {di(t)} and the motor rotation angle data θ(t), the major axis a(t) and minor axis b(t) of the flexible wheel cross-section ellipse are calculated in real time at each sampling time t (from T0-50ms to T0+200ms) using a direct least squares ellipse fitting algorithm. This generates "flexible wheel ellipse parameter time series data". Then, based on this data and the steady-state baseline 50ms before the event (e.g., the calculated mean of the major axis baseline μ_a = 19.95mm), the major and minor axis deviation sequences {Δa(t)} and {Δb(t)} within the entire transient window are calculated. The maximum absolute value of the deviation is extracted, for example, max|Δa| = 0.18mm, max|Δb| = 0.15mm. The "flexible wheel comprehensive peak deviation" is calculated as sqrt(0.18^2 + 0.15^2) = 0.234 mm. Finally, the data after the event is analyzed to determine the time for the flexible wheel shape to return to stability. By setting a stable threshold band (e.g., the stable band for the long axis deviation calculated based on the last segment data is [-0.01mm, +0.01mm]), starting the search from T0, it was found that the long and short axis deviation sequences first simultaneously entered and remained within their respective threshold bands at t_stable_start = T0 + 110ms. Therefore, the "flexible wheel deformation stabilization time" is 110 milliseconds. Within the interval [T0, T0 + 110ms], peak detection was performed on the long axis deviation sequence {Δa(t)} to identify oscillation characteristic points. The average time interval between adjacent peaks (or troughs) was calculated to obtain the "average oscillation period" of 18.2 milliseconds. Taking the reciprocal, the "main oscillation frequency" was obtained as 55 Hz. These three indicators (combined peak deviation of 0.234mm, stabilization time of 110ms, and main oscillation frequency of 55Hz) were packaged into "flexible wheel deformation data under operating conditions". An anomaly assessment was then conducted: The 0.234mm deformation was compared to the severity grading standard (assuming 0.15-0.25mm is grade 3 "severe"), resulting in a peak deformation severity of grade 3. The stabilization time of 110ms was compared to the healthy baseline under this condition (assuming a mean of 70ms and a standard deviation of 5ms), and the Z-score was calculated as (110-70) / 5 = 8.0, yielding an anomaly index of 8.0 (significantly prolonged). The main oscillation frequency of 55Hz was compared with the fault frequency database, finding it to fall within the "characteristic frequency range of early cracks in flexible wheel" (45-55Hz), thus classifying it as "suspected flexible wheel crack risk (high)". Finally, the severity (mapped to 8 points), time anomaly index (mapped to 10 points), and frequency risk (mapped to 9 points) are weighted and summed according to their weights (0.4, 0.3, 0.3) to obtain the "flexible wheel comprehensive response anomaly score" = (0.48) + (0.310) + (0.3 × 9) = 3.2 + 3.0 + 2.7 = 8.9 points (out of 10).

[0031] Step S3: Perform electromagnetic fault feature analysis based on the transient operating condition feature data of the motor to obtain the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum respectively; perform multi-channel monitoring fault synchronous coupling on the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum to obtain electromagnetic fault feature coupling data.

[0032] In this embodiment of the invention, the three-phase current and three-point magnetic field strength signals are extracted after the event (T0 to T0+200ms). Using the three-phase current and the synchronously acquired motor electrical angle (calculated from the motor rotation angle and pole pair number), the Clark-Parker transform is applied to calculate the direct-axis current and quadrature-axis current. A fourth-order Butterworth high-pass filter with a cutoff frequency of 1kHz is applied to the direct-axis current and quadrature-axis current respectively to extract the high-frequency ripple signals {Id_ripple(t)} and {Iq_ripple(t)}. Next, a short-time Fourier transform (STFT) is performed on these two ripple signals during the main power surge phase (e.g., T0 to T0+100ms), with parameters set to a 1024-point Hanning window and 50% overlap. The 2kHz to 5kHz frequency band is examined in the generated time-spectrum graph. During the time interval from T0+25ms to T0+40ms, the STFT result of Id_ripple(t) showed a frequency component with an amplitude of 0.7A at approximately 3.5kHz, exceeding the threshold of 0.5A (while the spectral resolution was ensured to be within 0.2A). Therefore, this period was marked as the "current fault characteristic region," and this marked time-frequency spectrum was saved as the "current fault characteristic time series spectrum." Simultaneously, a sliding window (50ms Hanning window, 50% overlap) FFT analysis was performed on one of the magnetic field signals (e.g., Bx(t)). Spectral peak changes were tracked within the 100Hz to 1kHz frequency band. It was found that during the period from T0+20ms to T0+70ms, the amplitude of a spectral peak at approximately 600Hz changed by 7mT relative to its previous window, exceeding the threshold of 5mT. This information on the time windows exhibiting significant distortion was marked on the time-frequency spectrum to generate the "magnetic field distortion characteristic time series spectrum." Finally, the generated "current fault characteristic time series spectrum" (containing abnormal region markers) and "magnetic field distortion characteristic time series spectrum" (containing distortion window markers), along with their common time reference information, are encapsulated into a data structure to form "electromagnetic fault characteristic coupled data".

[0033] Step S4: Analyze the health status of the motor flexure based on the electromagnetic fault characteristic coupling data and the flexure comprehensive response anomaly score to generate motor fault status data; perform multi-level fault early warning processing based on the motor fault status data to achieve early warning of motor faults in industrial robots.

[0034] In this embodiment of the invention, electromagnetic data is evaluated: The current spectrum is examined, and a marked region is found, therefore the "current channel electromagnetic anomaly flag" = 1. The magnetic field spectrum is examined, and a distortion window is found, therefore the "magnetic field channel electromagnetic anomaly flag" = 1. The sub-channel flags {current: 1, magnetic field: 1} are obtained. Next, based on this flag and the flexspline score of 8.9, according to a preset rule (e.g., if the flexspline score >= 7 and (current = 1 or magnetic field = 1), then it is considered high confidence), it is determined to be a "high confidence electromechanical joint anomaly," mapped to a "mechatronic joint anomaly confidence" of 0.9. Then, a time-series correlation assessment is performed: the flexspline peak response period (assuming T0 to T0+20ms) has low overlap with the current anomaly period (T0+25ms to T0+40ms), but is close to the magnetic field distortion initiation period (T0+20ms). The overall judgment is moderate correlation. The confidence level is adjusted (assuming a moderate correlation adjustment factor of 1.0): the adjusted single-event coupling anomaly is 0.9 × 1.0 = 0.9. Next, a risk level analysis is performed: the anomaly of 0.9 is compared with the risk threshold (assuming >= 0.9 is level 3 "immediate action"), and the risk level is determined to be "level 3". Finally, a "Motor Fault Status Data" report is generated, containing: event timestamp, axis number 3, event type "grabbing overweight workpiece", adjusted coupling anomaly of 0.9, risk level "level 3 - immediate action", main basis "flexible wheel deformation severely exceeds the standard, recovery time is extremely long, the main vibration frequency is suspected to have crack characteristics, accompanied by significant magnetic field distortion", and the recommended operation is "immediately stop robot operation, arrange for an emergency inspection of the third axis harmonic reducer, focusing on checking for cracks in the flexible wheel". Based on the motor fault status data of the highest risk level (Level 3), the system executes the highest level of early warning processing: it sends an "emergency stop" command to the robot controller through the industrial network, and at the same time, a red highlighted alarm window pops up on the central monitoring system (SCADA) interface, detailing the fault location (third axis motor / reducer) and risk assessment results, and automatically pushes a high-priority repair work order to the maintenance management system (CMMS), thus achieving accurate and timely early warning of industrial robot motor faults.

[0035] Preferably, the step S1 of acquiring industrial robot motor control commands in real time and then judging command value change events includes:

[0036] The speed or load command signal output in real time by the motor control system of the industrial robot is digitally sampled to obtain the original command sequence value;

[0037] Extract the instruction value of the latest sampling point from the original instruction sequence value, and obtain its corresponding sampling time point to obtain the current industrial robot motor control data;

[0038] Based on the current industrial robot motor control data, the instruction value and its timestamp of the immediately preceding sampling cycle are extracted to obtain the previous instruction time point data;

[0039] The sampling time interval is obtained by calculating the difference between the timestamps of the current industrial robot motor control data and the previous instruction time point data; the instruction value difference is obtained by calculating the difference between the instruction values ​​of the current industrial robot motor control data and the previous instruction time point data; the sampling time interval should be set in the range of 8 to 12 ms.

[0040] The instantaneous rate of change of the instruction is calculated based on the sampling time interval and the difference in instruction value, and then the absolute value is processed to generate the absolute value of the rate of change of the instruction.

[0041] The motor load change event is determined by rapidly changing the absolute value of the command change rate using a preset command threshold. The preset command rapid change threshold is set to 10% per second. When the absolute value of the command change rate is greater than 10% per second, it is determined to be a motor load change event.

[0042] In this embodiment of the invention, for an industrial robot system equipped with a six-axis joint and employing the EtherCAT bus communication protocol, the motor speed control commands for its third axis (typically responsible for arm swing and significant load changes) are acquired in real time. The monitoring system is implemented through a separately deployed industrial control computer (IPC) equipped with an EtherCAT master protocol stack. First, before the robot runs its predetermined pick-and-place operation program normally, the EtherCAT master on the IPC is configured, connecting it as a monitoring node in the network to the robot's EtherCAT fieldbus. Subsequently, through the master configuration tool, the process data objects (PDOs) broadcast by the servo drive unit controlling the third-axis motor are accurately identified and subscribed to. Specifically, the PDO mapping entry containing the "Target Velocity" command value is subscribed to, for example, the parameter with address index 60FFh in the servo drive object dictionary (assuming a standard CiA402 drive specification). After the robot's operation program is started, the master station protocol stack on the IPC receives real-time status PDO messages sent from the servo driver in strict synchronization according to a preset EtherCAT communication cycle (e.g., every 2 milliseconds), and parses out a 32-bit integer value representing the target motor speed. This value is the raw speed command issued by the controller to the motor at the current moment. This process is uninterrupted, ensuring continuous capture of the command stream. A high-precision system clock (e.g., synchronized with the EtherCAT distributed clock or using a microsecond-level timer provided by the operating system) is used to timestamp each received speed command value. Specifically, whenever the EtherCAT master station protocol stack successfully receives and parses the target speed value (e.g., 1500 rpm) from the third-axis servo driver, it immediately calls the system timing function to obtain the current timestamp. The speed value (1500 rpm) and the corresponding timestamp are stored as a data pair. This data pair is sequentially added to a first-in-first-out (FIFO) memory buffer. This buffer is set to a fixed length, for example, storing the most recent 1000 sampling points. As new data pairs are continuously generated and enter the buffer, the oldest data pairs are removed, forming a sliding window-like record of the original instruction sequence values. When a change event needs to be determined (usually executed periodically or triggered by the arrival of new data), the tail of the FIFO buffer (the position of the most recently added element) is accessed. Specifically, the data pair contained in the last element of the buffer is read directly. For example, the current last data pair in the buffer. This speed instruction value, 1550 rpm, will be extracted as the "instruction value of the latest sampling point". Simultaneously, its corresponding timestamp, 1678886400.133456789 seconds, will be extracted as the "corresponding sampling time point".The system backtracks to find the data immediately preceding the previous sample. This is done by accessing the second-to-last element of the FIFO buffer (the element preceding the tail element). This requires at least two data points in the buffer; the data pair at that position is read. For example, the second-to-last data pair might be (1678886400.123456789 seconds, 1500 rpm). The speed command value, 1500 rpm, is extracted as the "command value of the previous sampling period." Simultaneously, its corresponding timestamp, 1678886400.123456789 seconds, is extracted as the "timestamp of the previous sampling period." These two values ​​constitute the "previous command time point data." After obtaining the current and previous data, the following calculation is performed: First, the difference in timestamps is calculated: sampling time interval ΔT = T_current - T_previous = 0.010000000 seconds. This result is then converted to 10.0 milliseconds. Next, a check is performed to confirm whether the sampling time interval (10.0 milliseconds) falls within the preset valid range of 8 to 12 milliseconds. In this example, 10.0 milliseconds meets the condition (8 ≤ 10.0 ≤ 12), therefore the calculation is valid. Then, the difference in command values ​​is calculated: command value difference ΔV = V_current - V_previous = 50 rpm. These two calculation results—the sampling time interval (10.0 milliseconds) and the command value difference (50 rpm)—will be used in the next step to calculate the instantaneous rate of change. If the time interval is not within this range, the calculation is marked as invalid or subject to special handling to avoid incorrect judgments due to abnormal data acquisition timing. The time interval is converted to seconds: ΔT_s = ΔT_ms / 1000.0 = 0.01 seconds. Assume that the rated or maximum operating speed of the third axis motor is V_rated = 3000 rpm. Command instantaneous rate of change Rate = (ΔV / V_rated) × 100 / ΔT_s. Substituting the values: Rate = 166.67% / s. This calculation indicates that the speed command value changed by 1.6667% relative to the rated speed within 0.01 seconds, which translates to a rate of change of 166.67% / s per second. Finally, the absolute value of this instantaneous rate of change is processed to obtain the absolute value of the command change rate |Rate| = |166.67% / s| = 166.67% / s. This is then compared with the preset "command rapid change judgment threshold," which is preset to 10% per second. The basis for setting this threshold is: through statistical analysis of a large amount of normal operating data of this type of industrial robot in its typical working scenarios, it was found that during stable operation, the absolute value of the motor speed command change rate is usually much lower than 10% per second; however, under conditions of rapid acceleration and deceleration, or sudden load changes (such as the instant of grasping or releasing heavy objects) that cause a large impact on the harmonic reducer flexure, this rate of change will increase significantly. Therefore, 10% per second is chosen as the boundary to distinguish between normal fluctuations and potential high-load impact events.The comparison process is as follows: determine if |Rate| is greater than 10% per second. In this example, 166.67% / s > 10% per second. Since the absolute value of the calculated command change rate exceeds the preset threshold, the system determines that a "motor load change event" has occurred.

[0043] Preferably, the step S1 of real-time monitoring of motor transient operating condition characteristic data based on motor load change events includes:

[0044] Based on the motor load change event, a synchronous trigger signal is sent to obtain the synchronous acquisition trigger pulse signal;

[0045] Parallel multi-channel monitoring is performed based on synchronously acquired trigger pulse signals to generate multi-channel raw sampling point data; among which, the multi-channel raw sampling point data includes raw electromagnetic timing signals and raw flexible wheel displacement and rotation angle data;

[0046] By using a preset pre-event traceback duration threshold, parallel raw digital stream segments are extracted from the raw sampling point data of multiple channels to obtain the pre-event data segment.

[0047] The time point of the motor load change event is taken as the event start time node. Parallel raw digital streams are continuously extracted from the multi-channel raw sampling point data until the preset coverage transient response is reached and the total acquisition time is extended to the new steady state, thus obtaining the data segment acquired after the event.

[0048] By splicing the pre-collected data segment before the event and the data segment collected after the event in chronological order, the complete window's original data is obtained.

[0049] The event trigger attribute is processed by combining the motor load change event with the original data of the complete window, and transient operating condition features are integrated to generate transient operating condition feature data of the motor.

[0050] In this embodiment of the invention, when the previous step determines that a "motor load change event" has occurred (i.e., the absolute value of the command change rate exceeds 10% per second, for example, a change rate of 166.67% / s is detected at time T0), the monitoring main program deployed on the industrial control computer (IPC) immediately executes a preset trigger action. Specifically, this action involves generating a precisely time-synchronized digital pulse signal through the digital output (DO) channel (e.g., Port0 / Line0) of a data acquisition (DAQ) card (e.g., a multi-function I / O card of model NIPXIe-6363) connected to the IPC. This pulse signal is set to a high level (e.g., 5-volt TTL level) for a duration of 100 microseconds. The rising edge of this pulse strictly corresponds to the moment T0 when the load change event is detected (or within a very small, definite delay, such as 10 microseconds). This 100-microsecond high-level pulse with a clearly defined time base, actively sent by the monitoring program to the external hardware device after the internal logic determines that the condition is met, is the "synchronization acquisition trigger pulse signal." The output voltage signal is connected to the three analog input (AI) channels (AI0, AI1, AI2) of the PXIe-6363, with a sampling rate set to 50 kHz and a resolution of 16 bits. Simultaneously, the PXIe-4331 module is used to acquire flexure state data. Several miniature strain gauges (e.g., forming a full-bridge to measure torsional strain) are attached to specific locations on the outer wall of the flexure of the harmonic reducer. The strain gauge bridge is connected to a channel of the PXIe-4331 (e.g., CH0), which performs high-precision strain acquisition at a sampling rate of 25.6 kHz. Furthermore, a high-resolution absolute encoder (e.g., 24-bit resolution) is installed on both the motor shaft and the reducer output. Its signal is connected to an encoder interface module inside the chassis (e.g., connected to the counter / timer resource of the PXIe-6363 for high-speed position tracking) to calculate the real-time relative rotation angle of the flexure (output rotation angle minus the input rotation angle calculated according to the reduction ratio). When the PFI0 pin of the PXIe-6363 detects the rising edge of the trigger pulse signal, the system immediately starts data acquisition for all configured channels simultaneously at a preset sampling rate (50kHz for current, 25.6kHz for strain, and the counter records the position value at the corresponding moment). The system stores the sampled digital values ​​(current value, strain value, encoder count value) and their precise timestamps (provided by the PXIe chassis's synchronization clock) into the onboard buffer (FIFO) of each channel, forming a parallel, time-synchronized "multi-channel raw sampling point data" stream. A "pre-event traceback duration threshold" is set, its value predetermined based on the transient response characteristics analysis of the robot axis under typical operating conditions, aiming to capture the stable or quasi-steady state before the event occurs as a baseline. For example, this threshold is preset to 50 milliseconds (ms).The specific implementation is as follows: Before receiving the trigger command, the data acquisition system continuously acquires data from all channels and temporarily stores it in a circular buffer. This buffer is large enough to hold at least 50ms of data (for example, for a current channel with a sampling rate of 50kHz, 50kHz × 0.050s = 2500 sampling points are needed; for a strain channel with a sampling rate of 25.6kHz, 25.6kHz × 0.050s = 1280 sampling points are needed). When the synchronous acquisition trigger pulse signal (occurring at time T0) is confirmed by the DAQ system, the system automatically extracts and saves the sampled data of all channels from T0-50ms to time T0. The precise time point T0 when the motor load change event is detected (i.e., the effective edge time of the synchronous trigger pulse) is used as the "event start time node". After time T0, the data acquisition system continues to acquire data from all channels uninterruptedly at the set sampling rate (50kHz for current, 25.6kHz for strain, and encoder position tracking). A pre-set "total acquisition duration" is required, sufficient to cover the complete dynamic response process of the motor and flexible wheel caused by load changes (including impact, oscillation, and stabilization) and extend to the point of reaching a new steady-state condition. This value is determined based on research into the robot's dynamic characteristics and typical work cycles, for example, set to 200 milliseconds (ms). Therefore, the data acquisition system will start acquiring data from time T0 and continue acquiring data until time T0+200ms. The parallel raw digital streams of all channels acquired during the time period from T0 to T0+200ms constitute the "post-event acquisition data segment". Specifically, for each monitoring channel (e.g., the U-phase current channel), the sample sequence of its pre-event acquisition data segment (e.g., 2500 points) is directly concatenated with the sample sequence of the post-event acquisition data segment (e.g., 10000 points) in chronological order. It is ensured that the data at the connection point (time T0) is continuous (i.e., the last point of the previous segment is immediately adjacent to the first point of the next segment in time). Perform the same operation on all channels (U / V / W phase current, flexible wheel strain, motor / output angle). The final data is a continuous, time-synchronized multi-channel data record containing all monitoring channels, spanning from T0-50ms to T0+200ms (total duration 250ms). Perform "event trigger attribute processing": Attach the relevant information of the "motor load change event" that triggered this data acquisition (e.g., the precise timestamp T0 of the event, the absolute value of the command change rate that caused the triggering 166.67% / s, the robot axis number where the event occurred: the third axis, the current task stage of the robot: e.g., "the moment of grasping the workpiece") as metadata to this data window.Next, "transient condition feature integration" is performed: a structured data record is created (e.g., a new record is created in a database, or a file in a specific format such as HDF5 or MAT is generated). This record contains two parts: one part is the complete, multi-channel "full window raw data" itself (e.g., stored as a multidimensional array or multiple parallel time series); the other part is additional metadata that details the attributes of the triggering event. This structured data package, which integrates the raw time series data and the context information of the triggering event, is the final "motor transient condition feature data".

[0051] Preferably, the parallel multi-channel monitoring based on the synchronously acquired trigger pulse signal includes:

[0052] The current sensor, magnetic sensor array, and non-contact displacement sensor are precisely installed on the motor input line, the predetermined monitoring point of the stator winding, and the corresponding position on the inner wall of the flexspline, respectively. The signal output terminals of each sensor are physically connected to the corresponding input ports of the data acquisition equipment to obtain the connected sensor port group.

[0053] Based on the synchronous acquisition trigger pulse signal, the current and voltage of the three-phase input line of the industrial robot motor are sampled at high frequency through the connected sensor port group, and the magnetic field strength near the motor housing or motor air gap is detected at multiple points synchronously to obtain the original electromagnetic timing signal.

[0054] Based on the synchronous acquisition of trigger pulse signals, the non-contact multi-point position displacement synchronous measurement of the outer wall of the flexure wheel of the harmonic reducer in the industrial robot motor is performed through the connected sensor port group, and the motor rotation angle signal is acquired simultaneously to obtain the original flexure wheel displacement and rotation angle data.

[0055] In this embodiment of the invention, three LEMLA55-P type Hall effect current sensors are selected, employing a through-hole structure, and are precisely clamped and fixed to the three-phase (U, V, W) power input cables of the motor driving the third axis (assuming it is the monitoring target axis), ensuring that the cables pass completely centered through the sensor core window. Next, an array module containing three Melexis MLX91208 programmable linear Hall sensors is selected and precisely installed at predetermined monitoring points near the end windings of the motor stator using a custom-designed high-temperature resistant non-metallic mounting bracket. For example, these sensors are evenly spaced at 120-degree intervals in the circumferential direction and 3 mm radially from the winding surface, used to detect the leakage magnetic field inside the motor. Furthermore, four Micro-Epsilon eddyNCDT3005 eddy current displacement sensors were selected. Using specially designed mounting brackets that could withstand the internal environment of the reducer, the probes were precisely fixed on the stationary components of the harmonic reducer (such as the rigid wheel or base). Each probe's measuring surface was aligned with a specific circumferential position on the inner wall of the flex wheel (e.g., evenly distributed at 0°, 90°, 180°, and 270° along the expected paths of the major and minor axes of the flex wheel's deformed ellipse). The initial gap between the probe and the inner wall of the flex wheel was adjusted to 1 mm. After all sensor bodies were installed, the signal output lines of each sensor (±10V voltage signal from the current sensor, 0-5V voltage signal from the Hall sensor array, and 0-10V voltage signal from the eddy current displacement sensor) were physically connected to the designated analog input (AI) ports of the NIPXIe-6363 multifunction data acquisition card using highly shielded twisted-pair cables and BNC or equivalent connectors. Upon receiving the synchronous acquisition trigger pulse signal (a 5V TTL level pulse lasting 100 microseconds), this signal is input to the PFI0 (Programmable Function Interface) trigger input of the NIPXIe-6363 data acquisition card. The data acquisition system (controlled by a LabVIEW program running on the PXI controller) is configured to immediately initiate a multi-channel synchronous analog acquisition task upon detecting a rising edge at the PFI0 port. This task operates on the channels assigned to electromagnetic signals in the "Connected Sensor Port Group": synchronously converting the voltage signals output from the three-phase input current sensors to digital signals at a rate of 50,000 samples per second (50kHz) and 16-bit resolution via ports AI0, AI1, and AI2. Simultaneously, if the system configuration includes voltage monitoring, the three-phase input line voltages are also synchronously acquired at a rate of 50kHz and 16-bit resolution via a separately configured voltage transformer or high-voltage differential probe (e.g., connected to AI10, AI11, AI12). Furthermore, the voltage signals (representing local magnetic field strength) output by three Hall sensors installed near the stator winding are simultaneously sampled through the AI3, AI4, and AI5 ports at the same 50kHz rate and 16-bit resolution.All sampling actions across these channels are strictly synchronized and driven by the same high-precision onboard clock (e.g., 100MHz time base) on the PXIe-6363 card, ensuring that all electromagnetically related physical quantities are measured simultaneously at each sampling point. The acquired continuous, time-accurately aligned digital sequences (including U-phase current value sequences, V-phase current value sequences, W-phase current value sequences, U-phase voltage value sequences, V-phase voltage value sequences, W-phase voltage value sequences, and magnetic field strength value sequences at the three measurement points) together constitute the "raw electromagnetic timing signal." The data acquisition task for the flexible wheel condition monitoring channel is initiated. Specifically, through ports AI6, AI7, AI8, and AI9 in the "connected sensor port group," the voltage signals output by the four eddy current displacement sensors installed inside the harmonic reducer are simultaneously converted from analog to digital at a rate of 50,000 samples per second (50kHz) and 16-bit resolution. These signals directly reflect the real-time distance (displacement) of the flexible wheel's inner wall relative to the sensors at four fixed circumferential positions. Simultaneously, the system acquires the motor rotation angle signal. This operation is achieved through a high-resolution absolute encoder (e.g., a 24-bit resolution BiSS-C interface encoder) connected to the motor shaft. The encoder signal is input to a dedicated encoder interface module configured within the PXI chassis (or utilizing the high-speed counter / timer resources on the PXIe-6363 card). At the moment the trigger signal arrives and for each subsequent sampling clock cycle, the system precisely records the current absolute angle value reported by the encoder. The synchronous sampling mechanism of the PXIe-6363 card ensures precise time alignment between the sampling of the four displacement channels and the reading of the encoder angle values ​​(with errors on the nanosecond level). The acquired continuous, time-synchronized digital sequence (including the flexspline displacement value sequence at the four measuring points and the motor rotation angle value sequence) together constitutes the "raw flexspline displacement and rotation angle data".

[0056] As an example of the present invention, reference is made to... Figure 2 As shown, Figure 1 A detailed flowchart illustrating the implementation steps of step S2, which involves analyzing the deformation of a flexible wheel under transient operating conditions based on the transient operating condition characteristic data of the motor. In this example, step S2, which involves analyzing the deformation of a flexible wheel under transient operating conditions based on the transient operating condition characteristic data of the motor, includes:

[0057] S21: Based on the transient operating condition feature data of the motor, perform elliptical geometric fitting of the cross-sectional profile of the flexible wheel, and then extract the time series of the elliptical parameters of the flexible wheel to generate the time series data of the elliptical parameters of the flexible wheel.

[0058] In this embodiment of the invention, at each synchronous sampling time point t, the motor rotation angle θ(t) and the displacement values ​​d1(t), d2(t), d3(t), and d4(t) of the inner wall of the flexure measured by four eddy current displacement sensors are extracted. The fixed installation angles of the four sensors are known to be 0 degrees, 90 degrees, 180 degrees, and 270 degrees (relative to a certain reference coordinate system). Combined with the motor rotation angle θ(t), the instantaneous angles of these four measurement points on the flexure relative to the main axis of the wave generator at that moment can be calculated. Based on these four spatial position points (the angles are known; the radial position is calculated from the displacement sensor readings d_i(t) combined with the sensor installation position and the nominal radius of the flexure), the "direct least squares ellipse fitting" algorithm is adopted. This algorithm finds an optimal ellipse equation Ax^2 + Bxy + Cy^2 + Dx + Ey + F = 0 by solving a constrained optimization problem to fit these four (or more, if the number of sensors is greater) data points. This fitting process is repeated for each sampling time point in the "complete window of original data" (time span T0-50ms to T0+200ms). From the ellipse equation coefficients (A, B, C, D, E, F) obtained from each fitting, the geometric parameters of the flexible wheel cross-section at that moment are calculated: the coordinates of the ellipse center (xc(t), yc(t)), the length of the major semi-axis a(t), the length of the minor semi-axis b(t), and the orientation angle φ(t) of the ellipse major axis relative to the reference coordinate system. These time-varying parameters (a(t), b(t), xc(t), yc(t), φ(t)) are collected to form a time series, which is the "flexible wheel ellipse parameter time series data".

[0059] S22: Perform pre-event steady-state baseline statistics based on the transient operating condition characteristic data of the motor, and generate pre-event steady-state baseline characteristic data;

[0060] In this embodiment of the invention, statistical calculations are performed on the multi-channel raw data within this time period and the flexure ellipse parameters (e.g., the length of the major semi-axis a(t)) within the corresponding time period to establish a steady-state baseline. Specifically, all N sampled values ​​of the flexure major semi-axis length a(t) within the time window T0-50ms to T0 are selected (e.g., N = 2500 points for a 50kHz sampling rate and a 50ms window). The arithmetic mean μ_a = (1 / N) × Σa(t) and the standard deviation σ_a = sqrt[(1 / (N-1)) × Σ(a(t) - μ_a)^2] are calculated. Similarly, the mean and standard deviation of other key signals such as the effective value of the three-phase motor current (e.g., the effective value of the U-phase current I_U_rms) and the motor speed (which can be obtained by differentiating the rotation angle data) within this time period are also calculated. For example, the mean μ_I_U_rms and standard deviation σ_I_U_rms of the effective value of the U-phase current within this window are calculated. These calculated statistics (such as μ_a, σ_a, μ_I_U_rms, σ_I_U_rms, etc.) are collected together to form the "pre-event steady-state baseline characteristic data". This dataset quantifies the average level and fluctuation range of various indicators of the robot motor and flexible wheel in a relatively stable state before the occurrence of a load change event.

[0061] S23: Perform post-event dynamic response analysis based on the transient operating condition characteristic data of the motor, and generate post-event dynamic response characteristic data;

[0062] In this embodiment of the invention, dynamic characteristic analysis is performed on the multi-channel raw data and the corresponding time-series data of the flexure ellipse parameters (e.g., the length of the major semi-axis a(t)) within this time period. The specific operation is as follows: For the sequence of the length of the major semi-axis a(t) of the flexure, first determine its maximum peak value a_peak after time T0 and its occurrence time t_peak. Then, determine the new steady-state value a_final that the signal enters after the transient process ends (which can be obtained by calculating the average value within the time period from T0+150ms to T0+200ms). Calculate the overshoot: Overshoot = (a_peak - a_final) / a_final × 100%. Define a stability criterion, for example, if the signal enters and remains within the error band of a_final ± 2%, find the time point t_settling where the signal last entered this error band, then the settling time is Ts = t_settling - T0. Fast Fourier Transform (FFT) is applied to the data of a(t) during the period from T0 to t_settling to identify the frequency component with the largest amplitude, which is taken as the main oscillation frequency f_osc. Similar analysis is performed on other key signals such as motor current and speed to extract their dynamic features such as peak value, settling time, and oscillation frequency. All these dynamic indicators extracted from the post-event data segment (such as a_peak, Ts_a, f_osc_a, I_peak, Ts_I, f_osc_I, etc.) are integrated to generate "post-event dynamic response feature data".

[0063] S24: Based on the steady-state baseline characteristic data before the event and the dynamic response characteristic data after the event, perform a working condition jump deformation analysis on the time series data of the flexible wheel ellipse parameters to generate working condition jump deformation data. The working condition jump deformation data includes the comprehensive peak deviation of the flexible wheel, the stabilization time of the flexible wheel deformation, and the main oscillation frequency.

[0064] In this embodiment of the invention, an index comprehensively reflecting the severity of deformation is defined. For example, the peak deviation of the major axis Δa_peak = a_peak - μ_a and the peak deviation of the minor axis Δb_peak = b_peak - μ_b are calculated (note that the peak value of the minor axis can be the minimum value, in which case the deviation is μ_b - b_min). A comprehensive deviation can be Dev_peak_comprehensive = sqrt((Δa_peak)^2 + (Δb_peak)^2), or the peak deviation of the major axis, Dev_peak_comprehensive = |Δa_peak|, which better reflects stress concentration, can be directly used. Taking Dev_peak_comprehensive = |a_peak - μ_a| as an example, this value (for example, calculated to be 0.05 mm) is recorded as the "comprehensive peak deviation of the flexible wheel". The adjustment time Ts_a of the flexible wheel geometric parameters (such as the major axis a(t)) obtained from the post-event dynamic response analysis is used. For example, if Ts_a = 85 milliseconds is calculated, this value is recorded as the "flexible wheel deformation stabilization time". This represents the time required for the flexure shape to recover to a new steady state (or near-steady state) after being impacted. The dominant oscillation frequency f_osc_a is obtained using FFT analysis of the flexure geometry parameters (such as the major axis a(t)) in post-event dynamic response analysis. For example, if the FFT analysis shows that the dominant oscillation frequency of a(t) is 120 Hz during the transient process, this value is recorded as the "dominant oscillation frequency". These three specifically calculated values ​​(flexure composite peak deviation = 0.05 mm, flexure deformation stabilization time = 85 ms, dominant oscillation frequency = 120 Hz) are packaged into a "flexure deformation data under sudden load change" record. This data quantifies the dynamic deformation characteristics of the flexure under a specific load change event.

[0065] Preferably, the analysis of flexible wheel deformation under sudden changes in operating conditions based on the time series data of the flexible wheel ellipse parameters using pre-event steady-state baseline characteristic data and post-event dynamic response characteristic data includes:

[0066] Based on the time series data of the flex wheel ellipse parameters and the steady-state baseline characteristic data before the event, the deviations of the major axis and minor axis relative to their respective baseline values ​​at each sampling time within the preset transient window are calculated, and the flex wheel major axis deviation sequence and flex wheel minor axis deviation sequence are obtained respectively.

[0067] Take the absolute values ​​of the flex wheel major axis deviation sequence and the flex wheel minor axis deviation sequence respectively, and extract their respective maximum values ​​within a preset transient window to obtain the maximum absolute value of the major and minor axis deviations;

[0068] The overall peak value of the flexible wheel is processed based on the absolute value of the maximum major and minor axis deviations to obtain the overall peak value deviation of the flexible wheel;

[0069] Extract the last segment of the preset transient window data based on the time series data of the flexible wheel ellipse parameters, and perform steady-state deviation judgment threshold analysis of the long and short axes based on the long axis deviation sequence and the short axis deviation sequence of the flexible wheel to generate a deviation stability judgment threshold band.

[0070] Extract event trigger times from transient operating condition characteristic data of motor;

[0071] Based on the deviation stability determination threshold band, the starting time of deformation stability is obtained by searching backward from the event trigger time point for the first time when the major and minor axis deviation sequences simultaneously enter and remain within their respective threshold bands.

[0072] Analyze the deformation stabilization time of the flexible wheel based on the initial moment of deformation stabilization;

[0073] Based on the stabilization time of flexible wheel deformation, the oscillation behavior of the dynamic response characteristic data after the event is identified, and a sequence of deviation oscillation characteristic points is generated.

[0074] Calculate the average oscillation period based on the sequence of characteristic points of the deviation oscillation;

[0075] The main oscillation frequency is obtained by taking the reciprocal of the average oscillation period.

[0076] In this embodiment of the invention, "flexible ellipse parameter time series data" (including major axis sequence a(t) and minor axis sequence b(t), with a time range of T0-50ms to T0+200ms) and "pre-event steady-state baseline characteristic data" (including major axis baseline average value μ_a and minor axis baseline average value μ_b) are used. For each sampling time point t within the "preset transient window" (i.e., the complete data window from T0-50ms to T0+200ms), deviation calculation is performed. Specifically, the corresponding steady-state baseline value μ_a is subtracted from the major axis sequence a(t) to obtain the major axis deviation Δa(t) = a(t) - μ_a at that moment. Similarly, the steady-state baseline value μ_b is subtracted from the minor axis sequence b(t) to obtain the minor axis deviation Δb(t) = b(t) - μ_b at that moment. This calculation is repeated for all time points within the window, generating two new time series: one is the "flexible wheel major axis deviation series" {Δa(t)}, and the other is the "flexible wheel minor axis deviation series" {Δb(t)}. The absolute value of each value in these two deviation series is processed to generate the "major axis deviation absolute value series" {|Δa(t)|} and the "minor axis deviation absolute value series" {|Δb(t)|}. Then, within the "preset transient window" (T0-50ms to T0+200ms), the maximum value of each of these two absolute value series is searched. Specifically, all values ​​in the {|Δa(t)|} series are iterated through, and the maximum value is found and denoted as max|Δa|. Similarly, all values ​​in the {|Δb(t)|} series are iterated through, and the maximum value is found and denoted as max|Δb|. For example, if max|Δa| = 0.08 mm and max|Δb| = 0.06 mm are calculated. These two values ​​represent the maximum absolute deviations of the flexure's major and minor axes from their respective steady-state baselines during the entire transient process, and are recorded as the "maximum absolute values ​​of major and minor axis deviations." The Euclidean norm is used as the synthesis method: the calculation formula is flexure composite peak deviation = sqrt((max|Δa|)^2 + (max|Δb|)^2), which is the "flexure composite peak deviation" for this transient event, comprehensively considering the maximum deformation in both the major and minor axis directions. For example, data from T0+150ms to T0+200ms is selected, considered to represent the system entering a new steady state or approaching a steady state. The corresponding "flexure major axis deviation sequence" Δa(t) and "flexure minor axis deviation sequence" Δb(t) are also used. First, the average value μ_Δa_final and standard deviation σ_Δa_final of Δa(t) in this final segment (T0+150ms to T0+200ms) are calculated. Similarly, calculate the mean μ_Δb_final and standard deviation σ_Δb_final of Δb(t) at this final segment. Then, define the stability threshold band based on these statistics.A decision coefficient k is set (e.g., k = 3, this coefficient is predetermined based on the assessment of noise level and required stability confidence). The stability threshold band for the major axis deviation is then defined as [μ_Δa_final - k × σ_Δa_final, μ_Δa_final + k × σ_Δa_final]. Similarly, the stability threshold band for the minor axis deviation is [μ_Δb_final - k × σ_Δb_final, μ_Δb_final + k × σ_Δb_final]. These two ranges constitute the "deviation stability decision threshold band." The precise timestamp of the recorded "motor load change event" is located. The time point when the "absolute value of the command change rate" first exceeds the preset threshold (10% per second) is precisely recorded by the real-time monitoring system and used to generate the synchronous acquisition trigger pulse. The deviation values ​​Δa(t) and Δb(t) of each sampling point are checked in chronological order. The goal of the search is to find the first time point t_stable_start that satisfies the following conditions: for all time points t' from t_stable_start to the end of the transient window (T0+200ms), Δa(t') must fall within the major axis threshold band [0.004, 0.016] and Δb(t') must fall within the minor axis threshold band [-0.008, -0.002]. In practice, we can start iterating from T0. Once a point t is found to satisfy both conditions, it is tentatively designated as t_stable_start. The search continues until the end of the window. If any point fails to meet the conditions, the current t is discarded, and the search continues for a new candidate t_stable_start. The first time point that passes the verification of all subsequent points is the desired time point. For example, if the search finds that t = T0+85ms is the first time this sustained stability condition is met, then t_stable_start = T0+85ms, which is the "deformation stabilization start time". Based on the "deformation stabilization start time" t_stable_start and the "event trigger time" T0, the time difference between the two yields the duration from the impact to the stable shape of the flexible wheel, which is the quantified "flexible wheel deformation stabilization time". For example, a "flexible wheel major axis deviation sequence" {Δa(t)}, which typically has a large deviation amplitude, can be selected. The analysis time range is from the "event trigger time" T0 to the time t_stable_start corresponding to the "flexible wheel deformation stabilization time" (i.e., the interval [T0, t_stable_start]). Within this interval, a peak detection algorithm (e.g., based on the zero point of the first derivative and the sign of the second derivative, or a simple neighborhood comparison method) is applied to identify all local maxima (peaks) and local minima (troughs) in the {Δa(t)} sequence. The timestamps corresponding to each peak and trough are recorded and arranged in chronological order.For example, within this 85ms interval, peaks are identified at T0+15ms, T0+35ms, T0+55ms, and T0+75ms, while troughs occur at T0+25ms, T0+45ms, and T0+65ms. Using the peak timestamps from the example above: the first period P1 = (T0+35ms) - (T0+15ms) = 20ms; the second period P2 = (T0+55ms) - (T0+35ms) = 20ms; the third period P3 = (T0+75ms) - (T0+55ms) = 20ms. Then, the arithmetic mean of these periods is calculated. In this example, the average oscillation period = (P1 + P2 + P3) / 3 = (20ms + 20ms + 20ms) / 3 = 20 milliseconds. If we use the trough time points: P1' = (T0 + 45ms) - (T0 + 25ms) = 20ms; P2' = (T0 + 65ms) - (T0 + 45ms) = 20ms. The average period is (P1' + P2') / 2 = 20ms. Therefore, the calculated "average oscillation period" is 20 milliseconds. Taking the reciprocal of the average oscillation period (which needs to be converted to seconds) gives the dominant oscillation frequency of the flexible wheel during the transient response. Using the result from the example above, the average oscillation period = 20 milliseconds = 0.020 seconds. Calculating its reciprocal: dominant oscillation frequency = 1 / average oscillation period = 1 / 0.020s = 50 Hz.

[0077] Most importantly, the step of extracting the last segment of data from the preset transient window based on the time series data of the flexible wheel ellipse parameters, and performing steady-state deviation judgment threshold analysis of the major and minor axes based on the major axis deviation sequence and the minor axis deviation sequence of the flexible wheel, specifically involves:

[0078] Based on the timestamp information of the flexible wheel ellipse parameter time series data, the last time period of the preset transient window is determined, and the data points of the two sequences, the flexible wheel major axis deviation sequence and the flexible wheel minor axis deviation sequence, are extracted in the last time period to obtain the major axis deviation last segment sequence and the minor axis deviation last segment sequence, respectively.

[0079] Calculate the arithmetic mean of the data points in the major axis deviation last segment sequence and the minor axis deviation last segment sequence respectively to estimate the average deviation level after reaching the new steady state, and obtain the estimated value of the new steady state center deviation of the major and minor axes.

[0080] The tolerance range of the new steady-state deviation mean of the major and minor axes and the comprehensive peak deviation of the flexible wheel are determined by a small tolerance value of 5% of the peak deviation, and the stability boundary of the major and minor axis deviation is generated.

[0081] By combining the estimated deviation of the new steady-state center of the major and minor axes with the deviation stability boundary of the major and minor axes, the deviation stability judgment threshold band is obtained.

[0082] In this embodiment of the invention, based on the total duration of the preset transient window (250ms) and the set length of the final time period (e.g., the last 50 milliseconds), the specific time range of the final time period is determined to be from T0+150ms to T0+200ms. Next, a programming script (e.g., using the Python Pandas library to process time series data) is used to traverse the timestamp sequence of the flexible elliptic parameter time series data, filtering out all time point indices falling within the interval [T0+150ms, T0+200ms]. Using these indices, a subset of corresponding data points is extracted from the complete {Δa(t)} sequence to form the "major axis deviation final segment sequence". Similarly, the same indices are used to extract corresponding data points from the {Δb(t)} sequence to form the "minor axis deviation final segment sequence". Specifically, all 2500 data points in the "major axis deviation final segment sequence" are summed, and then the sum is divided by the number of data points (2500). The calculation formula is μ_Δa_final = (1 / 2500) × ΣΔa_end(t), where the summation iterates through all time points t of the final segment sequence. For example, μ_Δa_final is calculated to be 0.01 mm. Similarly, the same calculation is performed on the "minor axis deviation final segment sequence": μ_Δb_final = (1 / 2500) × ΣΔb_end(t). For example, μ_Δb_final is calculated to be -0.005 mm. These two calculated averages (0.01 mm and -0.005 mm) represent the center deviation levels of the major and minor axes that tend to stabilize after the event, i.e., the "estimates of the new steady-state center deviation of the major and minor axes". The tolerance is defined using the "5% small tolerance value of peak deviation" rule. First, this tolerance value is calculated: Tolerance = 0.05 × flexspline composite peak deviation = 0.05 × 0.1 mm = 0.005 mm. Then, based on this tolerance value, the new steady-state center deviation of each axis is extended upwards and downwards to generate stable boundaries. The upper boundary of the major axis (Upper_Bound_a) is μ_Δa_final + Tolerance. The lower boundary of the major axis (Lower_Bound_a) is μ_Δa_final - Tolerance. Similarly, the upper boundary of the minor axis (Upper_Bound_b) is μ_Δb_final + Tolerance. The lower boundary of the minor axis (Lower_Bound_b) is μ_Δb_final - Tolerance. These four calculated values ​​(0.015mm, 0.005mm, 0mm, -0.010mm) are the "stable boundaries of the major and minor axis deviations". Specifically, the lower boundary (0.005mm) and upper boundary (0.015mm) of the major axis are combined to define the stability threshold band of the major axis as a closed interval: Threshold_Band_a = [Lower_Bound_a, Upper_Bound_a].Similarly, by combining the lower boundary (-0.010 mm) and upper boundary (0 mm) of the minor axis, a stability threshold band for the minor axis is defined: Threshold_Band_b = [Lower_Bound_b, Upper_Bound_b]. These two clearly defined intervals together constitute the final output "deviation stability threshold band".

[0083] Preferably, the abnormal assessment of the flexible wheel's sudden change response based on the deformation data of the flexible wheel under sudden change in operating conditions in step S2 includes:

[0084] The severity of peak deformation is assessed based on the comprehensive peak deviation of the flexible wheel.

[0085] The abnormal lengthening or shortening of the deformation stabilization time of the flexible wheel is calculated by using a preset statistical time baseline under the same working condition jump, and the abnormal deformation recovery time index is obtained.

[0086] Fault mode matching is performed based on the comprehensive peak deviation of the flexible wheel to obtain the risk label of the flexible wheel vibration frequency;

[0087] Based on the peak deformation severity, abnormal deformation recovery time index, and flexure vibration frequency risk marker, a weighted assessment of flexure abrupt response anomalies is performed to generate a comprehensive flexure response anomaly score.

[0088] In this embodiment of the invention, the standard is formulated for a specific model of harmonic reducer (e.g., CSG-25-100) by combining the relationship between stress level and deformation predicted by finite element analysis (FEA) and the statistical distribution of peak deviation (e.g., the upper limit of the 99% confidence interval) measured when a large number of this model of reducer under healthy conditions undergo various typical operating condition transitions (such as rapid start-stop under different loads, collision recovery, etc.). This standard divides the peak deviation into several levels, for example: Level 0 (normal): deviation < 0.05 mm; Level 1 (slight): 0.05 mm ≤ deviation < 0.09 mm; Level 2 (moderate): 0.09 mm ≤ deviation < 0.15 mm; Level 3 (severe): 0.15 mm ≤ deviation < 0.25 mm; Level 4 (critical): deviation ≥ 0.25 mm. The input comprehensive peak deviation of the flexible wheel (0.1 mm) is compared with this classification standard to determine the range it falls into. In this example, 0.1 mm falls within the "moderate" (level 2) range. Therefore, the assessment result is "Peak Deformation Severity" = Level 2. The specific type of "condition jump" that triggered this data acquisition is identified (this information is stored as metadata in "Motor Transient Condition Characteristics Data," such as "Load suddenly increases from no-load to 5 kg"). Then, an entry matching this type is searched from a pre-defined "Basic Database of Condition Jump Duration." This database stores statistical baselines of flexspline deformation stabilization time calibrated in numerous healthy operating experiments for each defined typical condition jump type, including the mean (μ_Ts_base) and standard deviation (σ_Ts_base). For example, the database shows that for a jump of "Load suddenly increases from no-load to 5 kg," the healthy baseline is μ_Ts_base = 70 ms and σ_Ts_base = 5 ms. Next, the standardized deviation (Z-score) of the currently measured duration (85 ms) relative to the baseline mean is calculated. Check if the peak deformation severity exceeds a warning threshold (e.g., set to the upper limit of 0.09 mm for Level 1 "Slight"). Since 0.1 mm > 0.09 mm, the deformation amplitude is noteworthy, requiring further analysis of the vibration frequency. Next, the measured main oscillation frequency (50 Hz) is compared with a pre-built "Harmonic Gearbox Fault Characteristic Frequency Library". This library is based on fault injection experiments (such as simulating flexure cracks, bearing wear, lubrication failure, etc.) and dynamic simulation analysis of this type of gearbox, and includes typical ranges of characteristic vibration frequencies under different fault modes. For example, the library records: flexure early crack characteristic frequency range: 45-55 Hz; input bearing outer ring fault frequency (BPFO): approximately 180 Hz; output bearing inner ring fault frequency (BPFI): approximately 125 Hz. Match the measured 50 Hz with the entry in the library. In this example, 50 Hz falls within the "flexure early crack characteristic frequency range" (45-55 Hz).Therefore, the system generates a "flexible wheel vibration frequency risk marker" equal to "suspected flexible wheel crack risk (high)". If the frequency does not match any known failure mode, it is marked as "not matching known failure frequency (low)". These different types of indicators are quantified and normalized, mapped to a unified scoring range (e.g., 0 to 10 points). Severity level 2 is mapped to a score of 6 points; a Z-score of 3.0 (indicating a significant deviation from normal) is mapped to a score of 8 points; and the risk marker "suspected flexible wheel crack risk (high)" is mapped to a risk score of 9 points. Then, preset weights (W_severity, W_time, W_frequency) are assigned to these three quantified indicators. These weights are determined based on expert experience, historical failure data analysis, or the importance of each factor to the overall health in a specific application scenario, and their sum is 1. For example, W_severity = 0.4, W_time = 0.3, and W_frequency = 0.3. Finally, the weighted average is calculated as the overall anomaly score: Flexible wheel overall response anomaly score = (W_severity × Score_severity) + (W_time × Score_time) + (W_frequency × Score_frequency). This final score (7.5 out of 10) is the "flexible wheel overall response anomaly score".

[0089] Preferably, step S3 includes the following steps:

[0090] After the event, the instantaneous electromagnetic signal was extracted from the transient operating condition characteristic data of the motor to obtain the transient current-voltage data and the magnetic field data of the motor.

[0091] Calculate the instantaneous input active power of the motor based on the transient current-voltage data of the motor;

[0092] The transient power fluctuation intensity is calculated based on the instantaneous input active power of the motor, and the power impact response index is obtained.

[0093] Based on the transient current-voltage data of the motor, the direct-to-quadrature axis current of the motor is converted to obtain the direct-to-quadrature axis current component data of the motor;

[0094] Extract high-frequency ripple signals of the motor's direct- and quadrature-axis current components from the motor's direct- and quadrature-axis current component data;

[0095] Based on the power impulse response index, the current fault characteristics of the motor's direct-to-quadrature shaft high-frequency ripple signal are analyzed to obtain the current fault characteristic time spectrum; if the amplitude of any frequency component in the range of 2kHz to 5kHz exceeds 0.5A, then the time period is marked as the current fault characteristic region, and the spectral resolution is controlled within 0.2A.

[0096] Local distortion assessment of motor magnetic field data is performed using a 50ms window for fast Fourier transform. Magnetic field amplitude fluctuations are detected in the 100Hz to 1kHz frequency band. When the amplitude change of any spectral peak in this frequency band exceeds 5mT, a magnetic field distortion characteristic time series spectrum is generated.

[0097] In this embodiment of the invention, a signal sequence related to the electromagnetic behavior of the motor is located and extracted from the data record, selecting only the portion within a time range from the "event trigger time point" T0 (e.g., 1678886400.150 seconds) to the end of the "post-event data acquisition segment" (e.g., T0 + 200 milliseconds = 1678886400.350 seconds). The extracted signals include: a three-phase current signal sequence (I_u(t), I_v(t), I_w(t, sampling rate 50kHz) acquired by a previously installed Hall current sensor and a NIPXIe-6363 card; if the system configuration includes voltage acquisition, a synchronously acquired three-phase voltage signal sequence (U_u(t), U_v(t), U_w(t, sampling rate 50kHz)) is extracted. These two sets of signals together constitute the "motor transient current-voltage data". Simultaneously, three magnetic field strength signal sequences (e.g., B_x(t), By(t), B_z(t), with a sampling rate of 50kHz) are extracted from the Melexis MLX91208 magnetic sensor array installed near the stator windings and simultaneously acquired. This set of signals constitutes the "motor magnetic field data". For each synchronous sampling time point t within this data segment (time range T0 to T0+200ms), the standard formula for calculating three-phase instantaneous active power is applied: P(t)=U_u(t)×I_u(t)+U_v(t)×I_v(t)+U_w(t)×I_w(t). The specific calculation process is executed on a computer running a data analysis script (e.g., using Python's NumPy library). The script reads the three voltage values ​​and three current values ​​(all floating-point numbers, in volts and amperes respectively) at the corresponding time point, performs the above multiplication and addition operations, and obtains the instantaneous input active power value P(t) (in watts) at that time. This calculation is repeated for all N sampling points within the data segment (e.g., N = 10,000 points for 200ms data and a 50kHz sampling rate), ultimately generating a time series {P(t)} of length N for "motor instantaneous input active power". To quantify the severity of the power surge during a sudden change in operating conditions (a specific time period starting from T0, e.g., selecting the main transient response phase from T0 to T0+100ms), the standard deviation of the power series within this time period is calculated. Specifically, all power values ​​in {P(t)} with timestamps in the interval [T0, T0+100ms] are selected. Let this subsequence contain M data points (e.g., M = 5,000 points). First, the average power of this subsequence, μ_P = (1 / M) × ΣP(t), is calculated. Then, the standard deviation is calculated: σ_P = sqrt[(1 / (M-1)) × Σ(P(t) - μ_P)^2]. The obtained standard deviation σ_P (for example, a calculated result of 650 watts) is defined as the "power impulse response index".This single value reflects the overall fluctuation range of the input power of the motor when responding to sudden load changes; the larger the value, the stronger the power surge. The extracted data includes the three-phase current sequence (I_u(t), I_v(t), I_w(t)) from the "motor transient current-voltage data" and the synchronous motor angle signal θ_m(t) obtained from the "raw flexspline displacement and angle data". First, the number of pole pairs p of the target motor needs to be known (e.g., p = 4). The motor electrical angle θ_e(t) = p × θ_m(t) at each sampling time t is calculated. Then, the standard Park transform is applied to convert the currents (abc) in the three-phase stationary coordinate system to the direct-axis (d) and quadrature-axis (q) currents in the synchronous rotating coordinate system. The specific calculations are implemented in the data analysis script, typically in two steps: 1. Clark transform: converting I_u(t), I_v(t), I_w(t) to I_α(t) and I_β(t) in the two-phase stationary coordinate system. For example, I_α(t)=sqrt(2 / 3)×(I_u(t)-0.5×I_v(t)-0.5×I_w(t)), I_β(t)=sqrt(2 / 3)×(sqrt(3) / 2×I_v(t)-sqrt(3) / 2×I_w(t)). 2. Parker transformation: Using the calculated electrical angle θ_e(t), I_α(t) and I_β(t) are rotated to the dq coordinate system: I_d(t)=I_α(t)×cos(θ_e(t))+I_β(t)×sin(θ_e(t)); I_q(t)=-I_α(t)×sin(θ_e(t))+I_β(t)×cos(θ_e(t)). This transformation is performed on all sampling points for the entire post-event period (T0 to T0+200ms), generating two new time series: "motor direct-axis current component data" {I_d(t)} and "motor quadrature-axis current component data" {I_q(t)}. Based on the "motor direct--quadrature axis current component data" {I_d(t)} and {I_q(t)}, a digital high-pass filter is applied to each of these two series to separate the high-frequency components typically caused by inverter switching, dead-zone effects, or potential electrical faults. A fourth-order Butterworth high-pass filter is selected, with its cutoff frequency set to 1 kHz. This cutoff frequency is chosen because it is higher than the motor fundamental frequency and its lower harmonics (typically below several hundred hertz), but lower than the fault characteristic frequency range of focus (2 kHz to 5 kHz). The filtering operation is implemented using signal processing library functions (such as signal.butter and signal.filtfilt from the SciPy library in Python) in the data analysis script. Inputting {I_d(t)} into the filter yields the output sequence {I_d_ripple(t)}, which is the "high-frequency ripple signal of the motor's direct shaft".Similarly, inputting {I_q(t)} into the same filter yields the output sequence {I_q_ripple(t)}, which is the "motor cross-axis high-frequency ripple signal". Short-time Fourier Transform (STFT) is used for time-frequency analysis of the ripple signal. STFT parameters are set as follows: a Hanning window is selected, with a window length of 1024 sampling points (corresponding to approximately 20.5 milliseconds at a 50kHz sampling rate, providing a frequency resolution of approximately 48.8Hz), and a window overlap rate of 50% (i.e., sliding 512 points at a time). STFT is performed on {I_d_ripple(t)} and {I_q_ripple(t)} within the time interval [T0, T0+100ms] (the interval used for calculating the power impulse response index), yielding their respective time-frequency spectra (time-frequency-amplitude). Next, all time-frequency units within the frequency range of 2kHz to 5kHz in these two time-frequency spectra are examined. At each time slice (corresponding to the center of an STFT window), all frequency points between 2kHz and 5kHz are traversed, and their amplitudes (in amperes) are recorded. An amplitude threshold of 0.5 amperes is set. If the amplitude at any frequency point within this range exceeds 0.5 amperes, the time slice is marked as a "current fault characteristic region." To ensure the spectral resolution (amplitude accuracy) is controlled within 0.2 amperes, the accuracy of the STFT calculation and amplitude estimation methods must be guaranteed, for example, by selecting an appropriate number of FFT points and amplitude correction factors. Finally, the time-spectrum graphs (usually two, one on the d-axis and one on the q-axis, or a composite) with the marked regions (or only containing marked region information) are saved as the "current fault characteristic time series spectrum." The window length is set to 50 milliseconds (corresponding to 2500 sampling points), the window uses a Hanning window, and the overlap rate is 50% (sliding 25 milliseconds). This sliding window FFT is performed on the B_x(t) sequence with a time range from T0 to T0+200ms. Within the spectrum calculated for each window, the frequency band of interest is the range of 100 Hz to 1 kHz. For each significant spectral peak within this band (e.g., a peak whose amplitude exceeds a certain threshold of background noise level), its frequency and amplitude are recorded. Then, the amplitude of the spectral peak in the current window is compared with the amplitude of a peak in the previous window (temporally adjacent) at the same (or very close) frequency. The absolute value of the difference between the two amplitudes is calculated. A change threshold of 5 millitas (mT) is set. If, within the 100 Hz to 1 kHz frequency band, the amplitude of any spectral peak changes by more than 5 mT compared to the previous window, a significant magnetic field distortion is considered to have occurred within that time window. The time windows and time-frequency information of significant distortions are recorded. Finally, a "magnetic field distortion characteristic time series spectrum" annotated with these distortion events is generated.

[0098] Preferably, the motor flexure health status analysis based on electromagnetic fault characteristic coupling data and flexure integrated response anomaly score in step S4 includes:

[0099] Electromagnetic fault assessment of each channel is performed based on electromagnetic fault characteristic coupling data to obtain sub-channel electromagnetic anomaly indicators.

[0100] Electromechanical anomaly determination is performed based on the sub-channel electromagnetic anomaly flag and the flexible wheel integrated response anomaly score, and an electromechanical joint anomaly confidence level is generated.

[0101] Based on the electromagnetic fault characteristic coupling data and the flexible wheel deformation data under sudden change conditions, the abnormal response time sequence correlation during the sudden change event is evaluated. Then, the single event anomaly degree is corrected for the electromechanical joint anomaly confidence degree to obtain the corrected single event coupling anomaly degree.

[0102] The risk level of the flexible wheel is analyzed to correct the coupling anomaly degree of a single event, and then the motor health status is processed to generate motor fault status data.

[0103] In this embodiment of the invention, the time-series spectrum data structure is scanned to check if any time slices are marked as "current fault characteristic regions" (i.e., frequency components with amplitudes exceeding 0.5 amperes exist in the 2 kHz to 5 kHz frequency band). If at least one such marked region is detected, the "current channel electromagnetic anomaly flag" is set to 1 (indicating an anomaly); if no such flag is found in the entire time-series spectrum, it is set to 0 (indicating normal). Next, the magnetic field distortion characteristic time-series spectrum is processed independently: it is checked if any time window is marked as having experienced significant magnetic field distortion (i.e., any spectral peak amplitude change exceeding 5 millitalas in the 100 Hz to 1 kHz frequency band). If at least one such distortion window is detected, the "magnetic field channel electromagnetic anomaly flag" is set to 1; otherwise, it is set to 0. First, a rule set is defined, pre-set based on historical data and expert knowledge, to combine anomaly information from both mechanical (flexible wheel) and electrical (electromagnetic) aspects. For example, set the following rules: Rule 1: If the flex wheel score >= 7 and (current flag = 1 or magnetic field flag = 1), then it is judged as "high confidence electromechanical anomaly". Rule 2: If 5 <= flex wheel score < 7 and (current flag = 1 or magnetic field flag = 1), then it is judged as "medium confidence electromechanical anomaly". Rule 3: If the flex wheel score >= 7 and current flag = 0 and magnetic field flag = 0, then it is judged as "medium confidence flex wheel anomaly dominant". Rule 4: If the flex wheel score < 5 and (current flag = 1 or magnetic field flag = 1), then it is judged as "medium confidence electromagnetic anomaly dominant". Rule 5: Other cases are judged as "low confidence anomaly". Substitute the input value (flex wheel score 7.5, current flag 1, magnetic field flag 0) into the rule set for matching. According to Rule 1 (7.5 >= 7 and (1 = 1 or 0 = 1)), it is judged as "high confidence electromechanical anomaly". This determination result is then mapped to a numerical "electromechanical joint anomaly confidence level" between 0 and 1. For example, "high confidence level" is mapped to 0.9. Therefore, the "electromechanical joint anomaly confidence level" output in this embodiment is 0.9. First, a time-series correlation assessment is performed: the time period with the most severe flex wheel response is identified, for example, from the event trigger point T0 to the moment the flex wheel peak deviation occurs, T_peak (which can be extracted from the flex wheel elliptic parameter time-series data, assumed to be T0+20 ms). Simultaneously, the main time intervals for electromagnetic anomalies (current fault characteristic regions or magnetic field distortion windows) are identified (for example, the current fault characteristic region is concentrated between T0+15 ms and T0+30 ms). The overlap between these two time intervals is calculated, for example, using the Jaccard Index or a simple time overlap ratio. If the overlap is high (e.g., Jaccard Index > 0.6), a strong time-series correlation is considered to exist. In this example, the peak response period of the flexspline (assumed to be T0 to T0+20ms) and the current anomaly period (T0+15ms to T0+30ms) significantly overlap. Next, the confidence level is adjusted based on the correlation strength.Preset correction rules: For strong correlation, multiply the confidence level by a correction factor of 1.1; for moderate correlation, multiply by 1.0; for weak or no correlation, multiply by 0.9. Due to strong correlation, the corrected anomaly is calculated as follows: Corrected single-event coupling anomaly = Electromechanical joint anomaly confidence level × correction factor = 0.9 × 1.1 = 0.99. This 0.99 is the output "corrected single-event coupling anomaly". Perform risk level analysis: Compare this anomaly with a pre-set risk classification threshold. This threshold system is based on Failure Mode, Effects and Consequences Analysis (FMEA) and historical maintenance data statistics for this model of harmonic reducer. For example, the thresholds are set as follows: Level 0 (Healthy): Anomaly < 0.4; Level 1 (Observe Carefully): 0.4 ≤ Anomaly < 0.7; Level 2 (Inspection Recommended): 0.7 ≤ Anomaly < 0.9; Level 3 (Immediate Action): Anomaly ≥ 0.9. The input anomaly score of 0.99 is compared with the threshold. Since it is ≥0.9, the risk level is determined to be "Level 3 (Immediate Action)". Next, motor health status processing is performed: based on the determined risk level, a structured "Motor Fault Status Data" report is generated. This report includes: the timestamp of the event, the robot axis number, the trigger event type, the calculated corrected single event coupling anomaly score (0.99), the determined risk level (Level 3), a textual description of the risk level ("Immediate Action"), a summary of the main factors leading to the high risk (e.g., "severe peak deformation of the flexspline, long recovery time, abnormal current ripple and strong correlation with the flexspline response"), and suggested operations generated by the system according to the preset maintenance strategy (e.g., "trigger an emergency stop signal, notify maintenance personnel to immediately perform a detailed endoscopic inspection of the third axis harmonic reducer, and prepare to replace the flexspline spare part").

[0104] Most importantly, the electromagnetic fault assessment of each channel based on electromagnetic fault characteristic coupling data specifically involves:

[0105] Based on the current fault characteristic time sequence spectrum, the time points of the current fault characteristic region and the maximum amplitude of the corresponding over-limit frequency component are extracted to obtain the current characteristic over-limit amplitude sequence.

[0106] The cumulative number of occurrences or total duration of over-limit events are calculated based on the current characteristic over-limit amplitude sequence, and the monitoring channel status is determined to obtain preliminary channel status determination data.

[0107] Based on the time series spectrum of magnetic field distortion characteristics, the time points where the amplitude change of the detected spectral peak exceeds 5 mT and the corresponding maximum amplitude change are extracted to obtain the magnetic field distortion peak change sequence.

[0108] The magnetic field distortion anomaly is judged based on the peak change sequence of magnetic field distortion, and the preliminary judgment data of channel status is corrected by monitoring the channel status to generate sub-channel electromagnetic anomaly indicators.

[0109] In this embodiment of the invention, by parsing the time-series spectrum data structure, all time slices marked as "current fault characteristic regions" (i.e., STFT windows with frequency components exceeding 0.5A in the 2kHz to 5kHz frequency band) are identified. For each marked time slice, its center timestamp t is recorded, and the actual maximum amplitude Amax(t) (which must be greater than 0.5A) is further searched within the 2kHz to 5kHz frequency range of that time slice. For example, when analyzing data on events involving the capture of overweight workpieces, an STFT window with timestamp T0+27ms was found to be marked, and a maximum amplitude of 0.7A was detected at the 3.5kHz frequency point of that window; a window with timestamp T0+35ms was also marked, with a maximum amplitude of 0.6A at 4.1kHz. All such information is collected to form a sequence containing timestamps and corresponding maximum over-limit amplitudes. The "cumulative occurrence count" is used as a quantification indicator. The total number of data points contained in the sequence is calculated. In this example, the sequence contains 3 data points, so the cumulative occurrence count is 3. Next, this number of times is compared with a preset "current channel status judgment threshold". This threshold (for example, set to 2 times) is determined based on the analysis of a large amount of historical data, distinguishing between sporadic over-limits caused by noise or accidental interference under healthy conditions and continuous or frequent over-limits caused by potential faults. Comparison result: 3 times > 2 times. Based on this comparison, "preliminary channel status judgment data" is generated, the content of which is: {monitoring channel: "current", preliminary status: "abnormal", judgment basis: "cumulative over-limit times 3 times exceeding the threshold 2 times"}. If the calculated number of times does not exceed the threshold, the preliminary status is judged as "normal". The operation program scans the time series spectrum and locates all time windows marked as having significant magnetic field distortion (i.e., in the 100Hz to 1kHz frequency band, there is a window where the spectral peak amplitude changes by more than 5mT compared to the previous window). For each marked time window, the program records its center timestamp t' and extracts the maximum spectral peak amplitude change |ΔBmax(t')| (this value must be greater than 5mT) observed in the 100Hz to 1kHz frequency band within that window. For example, the program identifies and marks a window with a center timestamp of T0+50ms, where the maximum amplitude change of the 600Hz spectral peak is 7mT; a window with a timestamp of T0+75ms is also marked, where the maximum amplitude change of the 450Hz spectral peak is 5.5mT, forming a sequence containing timestamps and corresponding maximum amplitude changes. Anomaly detection is performed on the magnetic field channel: the "Magnetic Field Distortion Peak Change Sequence" is checked for emptiness. Since the sequence is not empty (containing two data points), it indicates an anomaly in the magnetic field channel. Therefore, the anomaly flag for the magnetic field channel is set to 1. Next, the initial state of the current channel is corrected based on the magnetic field channel's detection result (this correction logic aims to improve the reliability of the detection by utilizing multi-source information).Preset correction rules: Rule 1: If the initial current state is "abnormal" and the magnetic field state is also determined to be "abnormal," then the current channel state is confirmed as "abnormal" (flag = 1). Rule 2: If the initial current state is "normal" but the magnetic field state is determined to be "abnormal," then the current channel state is corrected to "suspected abnormal" or kept "normal," according to the preset strategy (here, it is assumed to be kept "normal," flag = 0, unless there is stronger evidence). Rule 3: If the magnetic field state is determined to be "normal," then the initial state of the current channel is the final state. In this example, the initial current state is "abnormal," and the magnetic field state is also determined to be "abnormal," which conforms to Rule 1. Therefore, the current channel state is confirmed as "abnormal," and its abnormal flag is 1. Finally, the judgment results of the two channels are integrated to generate "sub-channel electromagnetic abnormality flag" data.

[0110] Preferably, the present invention also provides an industrial robot motor fault early warning system, which executes the industrial robot motor fault early warning method described above. The industrial robot motor fault early warning system includes:

[0111] The transient operating condition monitoring module is used to acquire industrial robot motor control commands in real time, then judge command value change events to obtain motor load change events; and monitor the transient operating condition characteristic data of the motor in real time based on the motor load change events.

[0112] The flexible wheel response analysis module is used to perform flexible wheel deformation analysis under transient operating conditions based on the transient operating condition characteristic data of the motor, and generate flexible wheel deformation data under transient operating conditions; and to evaluate the abnormality of the flexible wheel's sudden response based on the flexible wheel deformation data under transient operating conditions, and generate a comprehensive abnormality score of the flexible wheel response.

[0113] The electromagnetic feature coupling module is used to perform electromagnetic fault feature analysis based on the transient operating condition feature data of the motor, and obtain the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum respectively; the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum are coupled synchronously with multi-channel monitoring faults to obtain electromagnetic fault feature coupling data.

[0114] The health assessment and early warning module is used to analyze the health status of the motor flexure based on electromagnetic fault characteristic coupling data and flexure comprehensive response anomaly score, and generate motor fault status data; based on the motor fault status data, multi-level fault early warning processing is performed to realize motor fault early warning for industrial robots.

[0115] This invention proactively identifies "events" of load or speed changes by monitoring real-time changes in industrial robot motor control commands, focusing the analysis on "transient operating conditions" during these abrupt changes in motor state. This event-triggered monitoring method can more accurately capture transient characteristic data of the motor in response to command changes, thus more clearly identifying abnormal dynamic responses caused by internal faults (such as torque pulsations caused by inter-turn short circuits) and distinguishing them from dynamic responses caused by normal load changes. This fundamentally solves the problems of difficult-to-set warning thresholds and frequent false alarms and missed alarms under varying operating conditions, significantly improving the reliability of fault warnings.

[0116] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0117] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for early warning of motor faults in industrial robots, characterized in that, Includes the following steps: Step S1: Acquire the industrial robot motor control commands in real time, and then judge the command value change event to obtain the motor load change event; Real-time monitoring of transient operating condition characteristic data of motor based on motor load change events; Step S2: Perform deformation analysis of the flexible wheel under transient operating conditions based on the transient operating condition characteristic data of the motor, and generate deformation data of the flexible wheel under transient operating conditions; evaluate the abnormality of the flexible wheel's transient response based on the deformation data of the flexible wheel under transient operating conditions, and generate a comprehensive abnormality score of the flexible wheel's response. Step S3: Perform electromagnetic fault characteristic analysis based on the transient operating condition characteristic data of the motor to obtain the time series spectrum of current fault characteristics and the time series spectrum of magnetic field distortion characteristics. By performing multi-channel monitoring and fault synchronous coupling of the current fault characteristic time spectrum and the magnetic field distortion characteristic time spectrum, electromagnetic fault characteristic coupling data is obtained. Step S4: Perform motor flexure health status analysis based on electromagnetic fault characteristic coupling data and flexure integrated response anomaly score to generate motor fault status data; Multi-level fault warning processing is performed based on motor fault status data to achieve early warning of motor faults in industrial robots.

2. The industrial robot motor fault early warning method according to claim 1, characterized in that, Step S1, which involves acquiring industrial robot motor control commands in real time and then determining command value change events, includes: The speed or load command signal output in real time by the motor control system of the industrial robot is digitally sampled to obtain the original command sequence value; Extract the instruction value of the latest sampling point from the original instruction sequence value, and obtain its corresponding sampling time point to obtain the current industrial robot motor control data; Based on the current industrial robot motor control data, the instruction value and its timestamp of the immediately preceding sampling cycle are extracted to obtain the previous instruction time point data; The sampling time interval is obtained by calculating the difference between the timestamps of the current industrial robot motor control data and the previous instruction time point data; the instruction value difference is obtained by calculating the difference between the instruction values ​​of the current industrial robot motor control data and the previous instruction time point data; the sampling time interval should be set in the range of 8 to 12 ms. The instantaneous rate of change of the instruction is calculated based on the sampling time interval and the difference in instruction value, and then the absolute value is processed to generate the absolute value of the rate of change of the instruction. The motor load change event is determined by rapidly changing the absolute value of the command change rate using a preset command threshold. The preset command rapid change threshold is set to 10% per second. When the absolute value of the command change rate is greater than 10% per second, it is determined to be a motor load change event.

3. The industrial robot motor fault early warning method according to claim 1, characterized in that, The step S1, which involves real-time monitoring of the transient operating condition characteristic data of the motor based on motor load change events, includes: Based on the motor load change event, a synchronous trigger signal is sent to obtain the synchronous acquisition trigger pulse signal; Parallel multi-channel monitoring is performed based on synchronously acquired trigger pulse signals to generate multi-channel raw sampling point data; among which, the multi-channel raw sampling point data includes raw electromagnetic timing signals and raw flexible wheel displacement and rotation angle data; By using a preset pre-event traceback duration threshold, parallel raw digital stream segments are extracted from the raw sampling point data of multiple channels to obtain the pre-event data segment. The time point of the motor load change event is taken as the event start time node. Parallel raw digital streams are continuously extracted from the multi-channel raw sampling point data until the preset coverage transient response is reached and the total acquisition time is extended to the new steady state, thus obtaining the data segment acquired after the event. By splicing the pre-collected data segment before the event and the data segment collected after the event in chronological order, the complete window's original data is obtained. The event trigger attribute is processed by combining the motor load change event with the original data of the complete window, and transient operating condition features are integrated to generate transient operating condition feature data of the motor.

4. The industrial robot motor fault early warning method according to claim 3, characterized in that, The parallel multi-channel monitoring based on synchronously acquired trigger pulse signals includes: The current sensor, magnetic sensor array, and non-contact displacement sensor are precisely installed on the motor input line, the predetermined monitoring point of the stator winding, and the corresponding position on the inner wall of the flexspline, respectively. The signal output terminals of each sensor are physically connected to the corresponding input ports of the data acquisition equipment to obtain the connected sensor port group. Based on the synchronous acquisition trigger pulse signal, the current and voltage of the three-phase input line of the industrial robot motor are sampled at high frequency through the connected sensor port group, and the magnetic field strength near the motor housing or motor air gap is detected at multiple points synchronously to obtain the original electromagnetic timing signal. Based on the synchronous acquisition of trigger pulse signals, the non-contact multi-point position displacement synchronous measurement of the outer wall of the flexure wheel of the harmonic reducer in the industrial robot motor is performed through the connected sensor port group, and the motor rotation angle signal is acquired simultaneously to obtain the original flexure wheel displacement and rotation angle data.

5. The industrial robot motor fault early warning method according to claim 1, characterized in that, Step S2, which involves performing deformation analysis of the flexible wheel under transient operating conditions based on the transient operating condition characteristic data of the motor, includes: Based on the transient operating condition characteristic data of the motor, the elliptical geometric fitting of the cross-sectional profile of the flexible wheel is performed, and then the time series extraction of the elliptical parameters of the flexible wheel is performed to generate the time series data of the elliptical parameters of the flexible wheel. Based on the transient operating condition characteristic data of the motor, the steady-state baseline before the event is statistically analyzed to generate the steady-state baseline characteristic data before the event; Post-event dynamic response analysis is performed based on the transient operating condition characteristic data of the motor to generate post-event dynamic response characteristic data. Based on the steady-state baseline characteristic data before the event and the dynamic response characteristic data after the event, the time series data of the flexible wheel ellipse parameters are analyzed for the deformation of the flexible wheel under the working condition jump, and the deformation data of the flexible wheel under the working condition jump includes the comprehensive peak deviation of the flexible wheel, the deformation stabilization time of the flexible wheel, and the main oscillation frequency.

6. The industrial robot motor fault early warning method according to claim 5, characterized in that, The analysis of flexible wheel deformation under sudden changes in working conditions based on the time series data of the flexible wheel ellipse parameters using pre-event steady-state baseline characteristic data and post-event dynamic response characteristic data includes: Based on the time series data of the flex wheel ellipse parameters and the steady-state baseline characteristic data before the event, the deviations of the major axis and minor axis relative to their respective baseline values ​​at each sampling time within the preset transient window are calculated, and the flex wheel major axis deviation sequence and flex wheel minor axis deviation sequence are obtained respectively. Take the absolute values ​​of the flex wheel major axis deviation sequence and the flex wheel minor axis deviation sequence respectively, and extract their respective maximum values ​​within a preset transient window to obtain the maximum absolute value of the major and minor axis deviations; The overall peak value of the flexible wheel is processed based on the absolute value of the maximum major and minor axis deviations to obtain the overall peak value deviation of the flexible wheel; Extract the last segment of the preset transient window data based on the time series data of the flexible wheel ellipse parameters, and perform steady-state deviation judgment threshold analysis of the long and short axes based on the long axis deviation sequence and the short axis deviation sequence of the flexible wheel to generate a deviation stability judgment threshold band. Extract event trigger times from transient operating condition characteristic data of motor; Based on the deviation stability determination threshold band, the starting time of deformation stability is obtained by searching backward from the event trigger time point for the first time when the major and minor axis deviation sequences simultaneously enter and remain within their respective threshold bands. Analyze the deformation stabilization time of the flexible wheel based on the initial moment of deformation stabilization; Based on the stabilization time of flexible wheel deformation, the oscillation behavior of the dynamic response characteristic data after the event is identified, and a sequence of deviation oscillation characteristic points is generated. Calculate the average oscillation period based on the sequence of characteristic points of the deviation oscillation; The main oscillation frequency is obtained by taking the reciprocal of the average oscillation period.

7. The industrial robot motor fault early warning method according to claim 1, characterized in that, Step S2, which involves evaluating the abnormal response of the flexible wheel based on the deformation data of the flexible wheel under the operating condition, includes: The severity of peak deformation is assessed based on the comprehensive peak deviation of the flexible wheel. The abnormal lengthening or shortening of the deformation stabilization time of the flexible wheel is calculated by using a preset statistical time baseline under the same working condition jump, and the abnormal deformation recovery time index is obtained. Fault mode matching is performed based on the comprehensive peak deviation of the flexible wheel to obtain the risk label of the flexible wheel vibration frequency; Based on the peak deformation severity, abnormal deformation recovery time index, and flexure vibration frequency risk marker, a weighted assessment of flexure abrupt response anomalies is performed to generate a comprehensive flexure response anomaly score.

8. The industrial robot motor fault early warning method according to claim 1, characterized in that, Step S3 includes the following steps: After the event, the instantaneous electromagnetic signal was extracted from the transient operating condition characteristic data of the motor to obtain the transient current-voltage data and the magnetic field data of the motor. Calculate the instantaneous input active power of the motor based on the transient current-voltage data of the motor; The transient power fluctuation intensity is calculated based on the instantaneous input active power of the motor, and the power impact response index is obtained. Based on the transient current-voltage data of the motor, the direct-to-quadrature axis current of the motor is converted to obtain the direct-to-quadrature axis current component data of the motor; Extract high-frequency ripple signals of the motor's direct- and quadrature-axis current components from the motor's direct- and quadrature-axis current component data; Based on the power impulse response index, the current fault characteristics of the motor's direct-to-quadrature shaft high-frequency ripple signal are analyzed to obtain the current fault characteristic time spectrum. If the current amplitude of any frequency component in the range of 2kHz to 5kHz exceeds 0.5A, then the time period is marked as the current fault characteristic region, and the spectrum amplitude accuracy is controlled within 0.2A. Local distortion assessment of motor magnetic field data is performed using a 50ms window for fast Fourier transform. Magnetic field amplitude fluctuations are detected in the 100Hz to 1kHz frequency band. When the amplitude change of any spectral peak in this frequency band exceeds 5mT, a magnetic field distortion characteristic time series spectrum is generated.

9. The industrial robot motor fault early warning method according to claim 1, characterized in that, Step S4, which involves analyzing the health status of the motor flexure based on electromagnetic fault characteristic coupling data and the flexure's comprehensive response anomaly score, includes: Electromagnetic fault assessment of each channel is performed based on electromagnetic fault characteristic coupling data to obtain sub-channel electromagnetic anomaly indicators. Electromechanical anomaly determination is performed based on the sub-channel electromagnetic anomaly flag and the flexible wheel integrated response anomaly score, and an electromechanical joint anomaly confidence level is generated. Based on the electromagnetic fault characteristic coupling data and the flexible wheel deformation data under sudden change conditions, the abnormal response time sequence correlation during the sudden change event is evaluated. Then, the single event anomaly degree is corrected for the electromechanical joint anomaly confidence degree to obtain the corrected single event coupling anomaly degree. The risk level of the flexible wheel is analyzed to correct the coupling anomaly degree of a single event, and then the motor health status is processed to generate motor fault status data.

10. An industrial robot motor fault early warning system, characterized in that, For performing the industrial robot motor fault early warning method as described in claim 1, the industrial robot motor fault early warning system includes: The transient operating condition monitoring module is used to acquire industrial robot motor control commands in real time, then judge command value change events to obtain motor load change events; and monitor the transient operating condition characteristic data of the motor in real time based on the motor load change events. The flexible wheel response analysis module is used to perform flexible wheel deformation analysis under transient operating conditions based on the transient operating condition characteristic data of the motor, and generate flexible wheel deformation data under transient operating conditions; and to evaluate the abnormality of the flexible wheel's sudden response based on the flexible wheel deformation data under transient operating conditions, and generate a comprehensive abnormality score of the flexible wheel response. The electromagnetic feature coupling module is used to perform electromagnetic fault feature analysis based on the transient operating condition feature data of the motor, and obtain the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum respectively; the current fault feature time series spectrum and the magnetic field distortion feature time series spectrum are coupled synchronously with multi-channel monitoring faults to obtain electromagnetic fault feature coupling data. The health assessment and early warning module is used to analyze the health status of the motor flexure based on electromagnetic fault characteristic coupling data and flexure comprehensive response anomaly score, and generate motor fault status data; based on the motor fault status data, multi-level fault early warning processing is performed to realize motor fault early warning for industrial robots.

Images