Digital twin collaborative control system and method for batch testing of unmanned aerial vehicle steering gears

Abstract

The application relates to a digital twin cooperative control system and method for batch testing of unmanned aerial vehicle rudders, which comprises a multi-channel loading test bench, an excitation twin module, a reference twin module, a twin bus and a test main control module; an excitation twin instance generates a target loading torque sequence according to a flight working condition script and shields the sampling feedback of the measured rudder, a reference twin instance generates an expected response sequence according to a control instruction and sampling feedback, the twin bus aligns the test beats of each measured channel according to a unified flight working condition script progress, and a test main control module completes qualification determination based on the difference between the expected response sequence and the actual measured response sequence. The application has the advantages of resolving the excitation-reference role conflict of the twin, realizing multi-channel horizontal cooperation and protecting the edge qualified measured rudder.

Description

Technical Field

[0001] This application relates to the field of servo control, and in particular to a digital twin collaborative control system and method for batch testing of UAV servo motors. Background Technology

[0002] UAV servos are key actuators in flight control systems, directly affecting the aircraft's attitude response characteristics and mission reliability. Before leaving the factory, servos must undergo full-condition performance verification on a load test bench to screen out individual defective products, identify batch incoming material anomalies, and assess the margin of acceptable parts. With the scaling up of the UAV industry, the demand for batch testing throughput of servos is rapidly increasing. Parallel testing of dozens of units per batch has become a common scenario on production lines, which places new demands on the collaborative control capabilities of test benches.

[0003] Existing loading test methods primarily rely on a preset torque curve being sent to a loading motor, which then applies a mechanical load to the servo motor under test. The servo motor's response is then collected and compared with a nominal baseline to determine its pass / fail status. This method is feasible in single-unit static testing scenarios, but in batch parallel testing scenarios, a core contradiction emerges—the conflict between the excitation source and the judgment benchmark. When the test system attempts to introduce digital twin technology to improve judgment accuracy, the twin instance needs to assume both the role of generating the loading torque as an excitation source and the role of generating the expected response to form a residual benchmark. These two roles have conflicting requirements for updating the twin parameters. The loading curve required for the excitation must remain rigid and stable throughout the entire batch of tests to ensure comparability of the benchmark. However, the reference response needs to continuously track the individual differences of the test device to maintain the accuracy of the residual. If both roles are concentrated in the same twin instance, the adaptive update of the twin parameters will pollute the benchmark, making the individuals tested at different times incomparable. If the adaptive update of parameters is abandoned, the deviation between the twin output and the response of the real test device will mask the gradual degradation characteristics of the test device itself.

[0004] This core contradiction further gives rise to multiple derivative defects in batch testing scenarios. Because the twins in existing solutions correspond one-to-one with the test pieces and are isolated from each other, the test bench cannot perform lateral comparisons of residuals across channels at the batch level, making it difficult to distinguish between incoming material deviations within the same batch and individual part failures. Since the loading excitations for each channel are independent, electromagnetic crosstalk and mechanical coupling between channels can contaminate the residual attribution for a single channel, making it impossible to effectively trace residual abnormalities back to specific fault sources. Because the loading excitation is issued all at once with a nominal amplitude immediately after the test piece is placed on the test bench, for edge pieces with parameters at the acceptable boundary, full-amplitude excitation may directly push them into the damage zone, causing edge pieces that could have been screened out for re-inspection to be destroyed by destructive testing. Due to the dynamic hysteresis and nonlinearity of the loading mechanism itself, there is a mismatch between the ideal excitation calculated by the twin and the actual torque applied by the loading motor, further weakening the fidelity of residual determination. Summary of the Invention

[0005] In order to simultaneously resolve the excitation-reference role conflict of the twin, achieve multi-channel lateral collaboration, and provide protection for edge qualified parts in batch parallel testing scenarios, this application provides a digital twin collaborative control system and related methods, devices, and storage media for batch testing of UAV servo motors.

[0006] Firstly, the digital twin collaborative control system for batch testing of UAV servo motors provided in this application adopts the following technical solution: A digital twin collaborative control system for batch testing of UAV servo motors includes a multi-channel loading test bench, an excitation twin module, a reference twin module, a twin bus, and a test master control module; The multi-channel loading test bench includes multiple independent test channels. Each test channel is equipped with a loading motor, a loading torque sensor, and a servo motor under test fixture. The test channel currently equipped with the servo motor under test is called the test channel, and the test channel not currently equipped with the servo motor under test is called the idle channel. The excitation twin module configures an excitation twin instance for each channel under test. The excitation twin instance generates a target loading torque sequence based on the flight condition script and sends it to the loading motor. The input of the excitation twin instance masks the sampling feedback of the corresponding servo motor under test. The reference twin module configures a reference twin instance for each channel under test. The reference twin instance generates the expected response sequence of the servo under test based on the control command of the servo under test and the sampled feedback of the servo under test. The twin bus connects each stimulus twin instance and carries the operational progress information of each stimulus twin instance, so that each channel under test is aligned with the test cycle according to the unified flight operational script. The test control module determines the pass / fail status of the servo under test based on the difference between the expected response sequence and the actual measured response sequence of the servo under test.

[0007] By adopting the above technical solution, the excitation twin instance and the reference twin instance are separated from each other in the input connection topology. The excitation twin instance only receives the flight condition script to output a stable loading excitation, while the reference twin instance simultaneously receives control commands and feedback from the servo under test to output the expected response adjusted according to individual differences. This decouples the excitation source role and the judgment criterion role of the twin into two independent instances, completely eliminating the risk of excitation curve drift due to the adaptive judgment criterion. The twin bus only exchanges condition progress information between each excitation twin instance without transmitting feedback from the device under test, so that each channel under test is aligned with the test schedule according to a unified flight condition script under the premise of mutual independence at the physical signal layer, realizing the lateral coordination of multiple channels. The combination of the two mechanisms enables the test master control module to obtain laterally comparable residuals based on a stable judgment criterion, providing a unified basis for subsequent individual failure identification, model drift detection, and channel fault attribution.

[0008] Secondly, this application provides a digital twin cooperative control method for batch testing of UAV servo motors, which is implemented based on the aforementioned digital twin cooperative control system for batch testing of UAV servo motors, and includes the following steps: S1. Deploy an excitation twin instance and a reference twin instance for each channel under test of the multi-channel loading test bench. Send the flight scenario script to the input terminal of the excitation twin instance. Send the control commands of the servo under test and the sampled feedback of the servo under test to the input terminal of the reference twin instance at the same time. Shield the excitation twin instance from accessing the sampled feedback of the servo under test.

[0009] S2. By stimulating the twin instance, a target loading torque sequence is generated based on the flight condition script and sent to the loading motor of the corresponding test channel. By referencing the twin instance, the expected response sequence of the tested servo is generated based on the control command of the tested servo and the sampled feedback of the tested servo.

[0010] S3. Exchange operational progress information among various excitation twin instances via the twin bus, and align the test cycle of each test channel according to the progress of the unified flight operational script.

[0011] S4. Determine the pass / fail status of the servo under test based on the difference between the expected response sequence and the actual measured response sequence of the servo under test.

[0012] By adopting the above technical solution, the functional relationships of each functional module of the system are translated into executable steps and sequences, providing a methodological anchor for embedding refined mechanisms such as individualized identification, residual decomposition, and idle channel calibration within each step, while ensuring that the system implementation and the method implementation are strictly equivalent in technical features.

[0013] Optionally, S1 includes sub-steps S11-S13: S11. After the servo under test is mounted, the corresponding twin template is retrieved from the pre-set twin template library according to the model information of the servo under test as the initialization structure of the excitation twin instance and the reference twin instance.

[0014] S12. Apply a preset identification excitation to the servo motor under test, and identify the individual parameters of the servo motor under test based on the identification response of the servo motor under test. The individual parameters include static friction coefficient, dead zone width, load inertia and zero position deviation.

[0015] S13. Synchronously inject the individual parameters into the stimulus twin instance and the reference twin instance, respectively, as the internal parameters of the stimulus twin instance and the reference twin instance; after the injection is completed, freeze the internal parameters of the stimulus twin instance until the test of the servo under test ends.

[0016] By adopting the above technical solution, individualized twin construction is completed through a one-time identification window when each servo under test is mounted, so that the two twin instances can simultaneously obtain the initial values ​​of the internal parameters of the test component; then the internal parameters of the excitation twin instance are frozen to ensure that the judgment benchmark remains rigid and stable throughout the entire testing process of the component.

[0017] Optionally, S4 includes sub-steps S41-S44: S41. Generate individual residual sequences based on the difference between the measured response sequence and the expected response sequence of the tested servo motor.

[0018] S42. Generate the model residual sequence based on the difference between the expected response sequence and the internal reference sequence of the excitation twin instance.

[0019] S43. Generate a channel residual sequence based on the difference between the individual residual sequence and the time-varying mean sequence of all individual residual sequences in the tested channels of this batch.

[0020] S44. The pass / fail status of the tested servo motor is determined based on the individual residual sequence, the model residual sequence, and the channel residual sequence.

[0021] By adopting the above technical solution, a single residual is decomposed into three orthogonal dimensions: individual, model, and channel. This allows the same anomaly to be mapped to different quadrants of the three-dimensional attribution space, upgrading the original one-dimensional binary detection capability to a three-dimensional attributable capability, and providing structured input for subsequent anomaly attribution decisions.

[0022] Optionally, after S4, the following steps are also included: in response to the sliding window mean of the model residual sequence continuously deviating from a preset model baseline range, updating the intrinsic parameters of the reference twin instance at a preset rate limit, thereby encouraging the intrinsic parameters of the twin instance to remain frozen.

[0023] By adopting the above technical solution, the reference twin instance can track the slow degradation of the test piece and the long-term drift of the twin model itself, while the excitation twin instance remains frozen to maintain the stability of the judgment benchmark. This achieves the adaptive capability of the twin model and strictly adheres to the role boundary of the excitation twin.

[0024] Optionally, S43 includes sub-steps S431-S434: S431. Identify an idle channel from the multi-channel loading test bench, and inject a calibration disturbance of known amplitude and known frequency into the loading motor corresponding to the idle channel. The amplitude of the calibration disturbance is significantly smaller than the rated loading torque of the corresponding channel under test, and the frequency of the calibration disturbance is located at an independent frequency point outside the frequency band covered by the flight condition script.

[0025] S432. Extract the frequency components of the calibration disturbance from the individual residual sequences of each channel under test, and identify the crosstalk transfer function matrix between channels based on the extraction results.

[0026] S433. Based on the crosstalk transfer function matrix between channels, crosstalk separation processing is performed on the individual residual sequence to obtain the net individual residual sequence after removing the coupling components between channels.

[0027] S434. Subtract the time-varying mean sequence of the net individual residual sequence from the time-varying mean sequence of the net individual residual sequences of all channels under test in this batch to obtain the channel residual sequence.

[0028] By adopting the above technical solution, low-amplitude orthogonal disturbances are injected into the naturally existing idle channels during the testing process to achieve online crosstalk calibration, and the electromagnetic and mechanical coupling between physical channels is separated from the residuals, so that the channel residual sequence truly reflects the abnormal level of the channel itself without being contaminated by interference from neighboring channels.

[0029] Optionally, S44 includes sub-steps S441-S444: S441. Maintain an absolute reference distribution constructed from the individual parameters of the tested servos from historical batches. The absolute reference distribution remains unchanged during the testing of the tested servos.

[0030] S442. The relative distribution of this batch is constructed in real time from the individual parameters of the first few tested servos in this batch. The relative distribution of this batch is dynamically updated as the number of tested servos in the multi-channel loading test bench increases.

[0031] S443. If the amplitude of the individual residual sequence, the amplitude of the model residual sequence, and the amplitude of the channel residual sequence are all lower than the corresponding preset anomaly threshold, and the individual parameters of the tested servo motor fall within the qualified range of the absolute reference distribution, and the individual parameters of the tested servo motor do not belong to the outlier of the relative distribution of this batch, the tested servo motor is marked as qualified; otherwise, the tested servo motor is marked as re-inspection or unqualified.

[0032] S444. Perform statistical testing on the offset of the distribution center of the relative distribution of this batch relative to the distribution center of the absolute reference distribution. If the offset exceeds the preset incoming batch offset threshold, trigger an incoming batch offset warning.

[0033] By adopting the above technical solution, the absolute benchmark distribution is used as the rigid specification boundary across batches, and the relative distribution of this batch is used as the flexible boundary for outlier detection within the same batch. The two distributions form an AND pass / fail criterion, which can distinguish and handle the two situations of the whole batch deviating but passing within the model and outliers within the model but within the batch. At the same time, when the whole batch deviates from the historical benchmark, an incoming batch deviation warning is triggered.

[0034] Optionally, S44 also includes the following parallel branches: in response to the individual residual sequence amplitude exceeding the individual anomaly threshold, and the model residual sequence amplitude being lower than the model anomaly threshold, and the channel residual sequence amplitude being lower than the channel anomaly threshold, the anomaly is attributed to the individual failure of the servo under test; in response to the model residual sequence amplitude exceeding the model anomaly threshold, and the amplitudes of the individual residual sequences of multiple channels under test in the multi-channel loading test bench increasing synchronously, the anomaly is attributed to model drift of the reference twin instance, and parameter correction of the reference twin instance is triggered; in response to the channel residual sequence amplitude of at least one channel under test in the multi-channel loading test bench exceeding the channel anomaly threshold, and the amplitudes of the channel residual sequences of the remaining channels under test in the multi-channel loading test bench being lower than the channel anomaly threshold, the anomaly is attributed to hardware failure of at least one channel under test, and at least one channel under test is suspended from the current batch of tests.

[0035] By adopting the above technical solution, different combinations of ternary residuals are mapped to three distinct causes: individual failure, model drift, and channel failure. This allows anomaly detection results to be directly transformed into actionable measures, shortening the response time from anomaly occurrence to location.

[0036] Optionally, before sending the target loading torque sequence to the corresponding loading motor in the test channel in S2, the following protective processing is performed on the target loading torque sequence: Within the excitation twin instance, using the sampling of the internal parameters of the reference twin instance within the preset tolerance range as initial values, random sampling pre-play is performed on the test cases covered by the target loading torque sequence to generate a pre-play response set of the tested servo motor under the test cases; in response to any sampling pre-play showing that the tested servo motor has entered the preset damage criterion area in the pre-play response set, the target loading torque sequence is decomposed into several step sub-sequences according to the amplitude, and sequentially applied to the loading motor. The machine issues each step of the ladder sequence; after each step of the ladder sequence is issued, the measured response sequence of the tested servo under the ladder sequence is fed back to the reference twin instance, and the internal parameters of the reference twin instance are updated; when the excitation twin instance performs random sampling pre-play of the next step of the ladder sequence, the updated internal parameters of the reference twin instance are used as the sampling center value, and the internal parameters of the excitation twin instance are kept frozen; if both the response and the pre-play response set show that the tested servo has not entered the damage criterion area, the target loading torque sequence is issued to the loading motor at once; after the protective processing is completed, the pass / fail judgment of S4 is entered.

[0037] By adopting the above technical solution, random sampling pre-playing is performed in the twin before the excitation is issued to predict the response of the tested servo under extreme sampling. For the tested servo with damage risk, graded progressive loading is started, and the parameters of the reference twin instance are updated with the measured response after each step. The pre-playing uses the latest parameters of the reference twin instance to improve the pre-playing accuracy, while the internal parameters of the excitation twin instance are always kept frozen so as not to affect the judgment criteria. This achieves protection for the edge qualified tested servo and avoids the full-amplitude excitation directly pushing qualified parts with small margins into the damage zone.

[0038] Optionally, a loading mechanism inverse model compensator is connected in series between the excitation twin instance and the loading motor. The loading mechanism inverse model compensator is configured to perform a pre-transformation on the target loading torque sequence output by the excitation twin instance based on the parameters of the loading mechanism inverse model compensator to obtain a drive command sequence and send it to the loading motor. The parameters of the loading mechanism inverse model compensator are obtained in the following way: after each multi-channel loading test bench is powered on, the first servo motor under test that is mounted after the power-on is taken as the first piece. The frequency sweep excitation of the first piece is performed and the frequency sweep response of the first piece is collected. The parameters of the loading mechanism inverse model compensator are identified based on the frequency sweep response of the first piece. The identified parameters of the loading mechanism inverse model compensator are used in all subsequent tests during the power-on.

[0039] By adopting the above technical solution, the inverse model compensator of the loading mechanism compensates for the inherent dynamic lag and nonlinearity of the loading mechanism, so that the actual output loading torque of the loading motor tracks the target loading torque sequence, eliminating the fidelity loss of the twin excitation in the loading link; and by using the first component sweep frequency identification method at each start-up to track the mechanical attenuation of the loading mechanism itself, so that the compensator remains effective throughout the entire life cycle of the test bench.

[0040] In summary, this application includes at least one of the following beneficial technical effects: 1. By using a dual-track twin architecture, the excitation source role and the judgment benchmark role are assigned to two twin instances that are isolated from each other in the input topology. With the twin bus, the working progress of each excitation twin instance is synchronized, so that the loading excitation in the batch parallel test scenario remains rigid and stable throughout the batch. At the same time, each channel under test is physically independent but logically coordinated, which fundamentally eliminates the excitation-benchmark role conflict and supports multi-channel horizontal comparison.

[0041] 2. By decomposing a single residual into three orthogonal dimensions—individual, model, and channel—and combining online crosstalk calibration of idle channels with dual matching of absolute baseline distribution and relative distribution of the current batch, the testing system can complete individual failure identification, model drift detection, channel hardware fault location, and incoming batch offset warning within the same residual system, significantly improving the determinism of anomaly attribution under complex working conditions.

[0042] 3. By driving graded progressive loading through random sampling pre-simulation within the twin domain, and using the inverse model compensator of the loading mechanism to pre-transform the target loading torque sequence, the testing process can avoid the destructive impact of full-amplitude excitation on edge qualified parts, and eliminate the loss of excitation fidelity due to the dynamic lag of the loading process, thus balancing the stringency of screening and the protection of the test part. Attached Figure Description

[0043] Figure 1 This is a flowchart of a digital twin collaborative control method for batch testing of UAV servo motors in one embodiment of the present invention. Detailed Implementation

[0044] The present application will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the scope of the application.

[0045] This application provides a digital twin collaborative control system and method for batch testing of unmanned aerial vehicle (UAV) servo motors. The system uses a multi-channel loading test bench to support the servo motor under test. An excitation twin instance and a reference twin instance are assigned to each channel under test, one master and one parameter. The excitation twin instance generates loading excitations based on the flight script, thus masking the sampling feedback from the servo motor under test. The reference twin instance simultaneously receives control commands and sampling feedback to generate a judgment criterion. A twin bus synchronizes the flight script progress among the excitation twin instances. The test master control module determines the pass / fail status based on the difference between the measured response and the judgment criterion.

[0046] Before detailing the embodiments of this application, some terms are explained. A UAV servo refers to an electric actuator mounted on a UAV to drive the deflection of aerodynamic control surfaces. Each servo includes a motor, a reduction gear, a position feedback unit, and an internal control loop. Its input is an angle or angular velocity command, and its output is the actual deflection angle of the control surface. A flight scenario script refers to a sequence of loading torque commands organized according to flight phases, covering typical aerodynamic loads in takeoff, climb, cruise, maneuver, and landing phases. It is used to reproduce the torque conditions experienced by the servo during flight on a ground test bench. Both excitation twin instances and reference twin instances refer to servo mathematical model instances running in software. The models include mechanical equations, electrical models, and nonlinear elements such as dead zones, friction, and saturation. The excitation twin instance generates a target loading torque sequence based on the flight scenario script, while the reference twin instance generates an expected response sequence based on control commands and sampled feedback. A twin bus refers to a publish-subscribe data distribution bus in the software layer, used for exchanging flight scenario script progress information between excitation twin instances. Flight condition script progress refers to the segment identifier and time scale within the segment that the script has been executed at a certain moment.

[0047] As the baseline scenario throughout the subsequent embodiments, the multi-channel loading test bench is configured with 8 independent test channels. The rated loading torque of the servo under test is 20 N·m, and the total duration of the flight scenario is 300 seconds, with the takeoff segment corresponding to 0 to 30 seconds, the climb segment to 30 to 80 seconds, the cruise segment to 80 to 200 seconds, the maneuver segment to 200 to 260 seconds, and the landing segment to 260 to 300 seconds. Eight servos were tested in this batch, corresponding to the eight test channels. The components of the system and method are described in detail below.

[0048] The digital twin collaborative control system for batch testing of UAV servos includes a multi-channel loading test bench, an excitation twin module, a reference twin module, a twin bus, and a test master control module. The multi-channel loading test bench includes multiple independent test channels, each equipped with a loading motor, a loading torque sensor, and a fixture for the servo under test (DUT). Channels currently equipped with the DUT are called "on-test channels," while those without are called "idle channels." The excitation twin module configures an excitation twin instance for each on-test channel. The excitation twin instance generates a target loading torque sequence based on the flight script and sends it to the loading motor. The input of the excitation twin instance masks the sampling feedback from the corresponding DUT. The reference twin module configures a reference twin instance for each on-test channel. The reference twin instance generates the expected response sequence of the DUT based on the control commands and sampling feedback from the DUT. The twin bus connects each excitation twin instance and carries the operational progress information of each excitation twin instance. The test control module determines the pass / fail status of the servo under test based on the difference between the expected response sequence and the actual measured response sequence of the servo under test.

[0049] As an example, both the excitation twin instance and the reference twin instance are hosted in independent software processes, each running a complete servo motor mathematical model. The servo motor mathematical model includes mechanical equations describing the servo motor's output shaft moment of inertia, damping, and transmission stiffness; an electrical model describing the conversion of motor armature voltage to electromagnetic torque; and compensation terms describing nonlinear elements such as static friction, dead zone, saturation, and backlash. The twin bus is implemented using a publish-subscribe data distribution service. Each excitation twin instance acts as a publisher, publishing its current flight script progress to the twin bus, while simultaneously receiving progress information from other excitation twin instances as a subscriber. The test master control module runs on the master computer, and each instance of the excitation and reference twin modules runs on the corresponding real-time computing node of the tested channel. These nodes are connected via an industrial Ethernet network.

[0050] The multi-channel loading test bench has eight test channels arranged in parallel. Each test channel is independently configured with a loading motor, a loading torque sensor, and a fixture for the servo under test. The excitation twin module runs an excitation twin instance for each channel under test. The output of the excitation twin instance is connected to the loading motor of the corresponding channel under test, and the input of the excitation twin instance is only connected to the flight script and not to the sampling feedback of the servo under test. The reference twin module runs a reference twin instance for each channel under test. The reference twin instance is connected to both the control command channel and the sampling feedback channel of the servo under test. The output of the reference twin instance sends the expected response sequence to the test master control module.

[0051] The flight scenario script is sent to the input terminal of the excitation twin instance. The output terminal of the excitation twin instance outputs the target loading torque sequence to the loading motor, which applies physical torque to the servo under test on the servo under test fixture. The control command channel and the sampling feedback channel of the servo under test are simultaneously sent to the control command input terminal and the feedback input terminal of the reference twin instance. The output terminal of the reference twin instance outputs the expected response sequence to the test master control module. The path from the sampling feedback of the servo under test to the input terminal of the excitation twin instance is represented by a cut-off symbol, indicating that the path is explicitly masked at both the physical and software layers, and any sampling feedback from the servo under test does not enter the excitation twin instance.

[0052] Through the aforementioned dual-track connection topology, the excitation twin instance proceeds solely according to the flight scenario script during execution, without adjusting its output based on real-time feedback from the tested servo. Even if the tested servo experiences feedback drift at a certain moment due to individual anomalies, the target loading torque sequence output by the excitation twin instance remains rigidly consistent for the same tested servo throughout the entire batch of tests. Therefore, the expected response sequence obtained by the test master control module from the reference twin instance, serving as the reference value for the matching criterion, originates from the combined effect of control commands and sampled feedback. The physical excitation applied by the loading motor, on the other hand, comes from the output driven purely by the scenario script. These two functions correspond to the "tracking function of the matching criterion" and the "criterion function of excitation generation," respectively, and are functionally independent. Different tested servos receive the same physical excitation at the same flight scenario script moment, thus enabling cross-channel lateral comparison of the measured response sequences obtained from each channel. The channel residual sequences in the subsequent ternary residuals can be generated based on this.

[0053] A loading mechanism inverse model compensator is connected in series between the excitation twin instance and the loading motor. The loading mechanism inverse model compensator is configured to perform a pre-transformation on the target loading torque sequence output by the excitation twin instance based on its parameters, thereby obtaining a drive command sequence and sending it to the loading motor. The parameters of the loading mechanism inverse model compensator are obtained as follows: After each multi-channel loading test bench is powered on, the first servo motor under test mounted after the power-on is taken as the first piece. A frequency sweep excitation is performed on the first piece, and the frequency sweep response of the first piece is collected. The parameters of the loading mechanism inverse model compensator are identified based on the frequency sweep response of the first piece, and the identified parameters of the loading mechanism inverse model compensator are used throughout all subsequent tests during the power-on period.

[0054] As an example, the loading mechanism inverse model compensator is implemented in the form of a finite impulse response filter, and the filter coefficients are the parameters of the loading mechanism inverse model compensator. The loading mechanism refers to the complete mechanical and electrical link from the armature winding of the loading motor, through the electromagnetic torque element, reduction transmission, output shaft to the servo motor under test. The inherent dynamics of this link can be approximated as a second-order linear element superimposed with Coulomb friction and backlash, exhibiting overall phase lag and amplitude attenuation. The loading mechanism inverse model compensator constructs the inverse of the loading mechanism's forward dynamics using a forward transformation, making the transfer function between the target loading torque sequence output by the excitation twin instance and the physical torque actually acting on the servo motor under test via the loading mechanism approximately unit transfer.

[0055] In the pre-transformed signal stream, the target loading torque sequence first enters the input of the loading mechanism inverse model compensator. After the loading mechanism inverse model compensator performs finite impulse response convolution operation according to its parameters, it outputs a drive command sequence. The drive command sequence is converted into armature voltage by the loading motor's digital-to-analog conversion, driving the loading motor to output electromagnetic torque, which acts on the tested servo motor fixture through the transmission chain. Compared with the target loading torque sequence in the time domain, the drive command sequence is advanced by several sampling periods in phase and moderately boosted in amplitude at high frequencies to compensate for the loading mechanism's own phase lag and amplitude attenuation.

[0056] As an example, during the frequency sweep identification process after each power-on of the multi-channel loading test bench, a linear frequency sweep excitation is performed on the first servo under test (SUT) mounted in this power-on session. The frequency sweep covers a bandwidth from 0.1 Hz to 200 Hz, and the sweep duration is approximately 30 seconds. The sweep amplitude is 10% of the rated loading torque. The test main control module synchronously acquires the frequency sweep response of the first component. Based on the ratio of the frequency sweep excitation to the frequency sweep response spectrum, the forward transfer function of the loading mechanism is identified, and its stable inverse is used as the parameter of the loading mechanism inverse model compensator. The frequency sweep excitation of the first component is performed in parallel with the flight condition script excitation used in normal testing. The first component itself also participates in the qualification assessment as a regular SUT and is not exempted from testing because it undertakes the identification task.

[0057] The parameters of the loading mechanism inverse model compensator are refreshed only once during the first-piece identification at each startup. All subsequent tests of the tested servos under test (SUTs) using these parameters during this startup are independent of individual SUT differences. This design decouples the slow degradation of the loading mechanism itself, such as lubrication degradation, reduction gear wear, and motor armature aging, from the individual differences of the SUTs. The former changes over the timescale of each startup and is captured by the first-piece identification; the latter is identified and captured when each SUT is mounted and is carried by the internal parameters of the excitation twin instance.

[0058] Through the aforementioned pre-transformation, the inherent phase lag and amplitude attenuation of the loading mechanism are compensated at the drive command sequence level, enabling the torque actually applied by the loading motor to the tested servo motor to track the target loading torque sequence output by the excitation twin instance. Therefore, the model residual sequence subsequently generated in S42 reflects the mismatch between the reference twin instance and the excitation twin instance at the model level, without adding the dynamic deviation of the loading mechanism itself; the orthogonality of the ternary residuals is thus physically guaranteed.

[0059] Reference Figure 1 The digital twin collaborative control method for batch testing of UAV servo motors is implemented based on the above system, including steps S1-S4.

[0060] S1. Deploy an excitation twin instance and a reference twin instance for each channel under test of the multi-channel loading test bench. Send the flight scenario script to the input terminal of the excitation twin instance. Send the control commands of the servo under test and the sampled feedback of the servo under test to the input terminal of the reference twin instance at the same time. Shield the excitation twin instance from accessing the sampled feedback of the servo under test.

[0061] Specifically, S1 includes sub-steps S11-S13.

[0062] S11. After the servo under test is mounted, the corresponding twin template is retrieved from the pre-set twin template library according to the model information of the servo under test as the initialization structure of the excitation twin instance and the reference twin instance.

[0063] S12. Apply a preset identification excitation to the servo motor under test, and identify the individual parameters of the servo motor under test based on the identification response of the servo motor under test. The individual parameters include static friction coefficient, dead zone width, load inertia and zero position deviation.

[0064] S13. Synchronously inject the individual parameters into the stimulus twin instance and the reference twin instance, respectively, as the internal parameters of the stimulus twin instance and the reference twin instance; after the injection is completed, freeze the internal parameters of the stimulus twin instance until the test of the servo under test ends.

[0065] As an example, the twin template library organizes templates according to the model of the servo motor under test. Each template contains the nominal mechanical structure, electrical parameters, and nonlinear term structure of that model. The templates do not contain specific individual parameter values, but rather reserve placeholders for the parameters, which are then filled in by S12 after identification. The meanings of each parameter in the individual parameters are as follows: the static friction coefficient represents the static damping torque of the output shaft of the servo motor under test at the moment of startup; the dead zone width represents the non-response angle range of the servo motor under test under small signal control commands; the load inertia represents the combined rotational inertia of the output shaft of the servo motor under test and the tooling; and the zero-position deviation represents the steady-state angular deviation of the servo motor under test under zero control command input.

[0066] For the eight servos under test in the baseline scenario, S12 applies a 5-second identification excitation to each servo after it is mounted. The identification excitation consists of a small-amplitude sinusoidal sweep signal superimposed with a pseudo-random binary sequence. The sweep amplitude is set to 5% of the rated loading torque to avoid entering the saturation or damage zone of the servo under test. The test main control module synchronously acquires the identification response of each servo under test and solves the individual parameters from the correspondence between the identification excitation and the identification response using a least-squares identification algorithm. Typical identification results are a static friction coefficient of approximately 0.08 N·m, a dead zone width of approximately 0.3 degrees, and a load inertia of approximately 3.5 × 10⁻⁶. -3 The identification results of each tested servo motor exhibit reasonable dispersion in various dimensions, with a zero-position deviation of approximately 0.1 degrees and a kilogram-square meter measurement. The dispersion range is consistent with the manufacturing tolerance of the tested servo motor.

[0067] During the injection process in S13, the stimulus twin instance and the reference twin instance synchronously receive the same set of identification results, ensuring that the two instances have a completely consistent starting point at the time of injection completion. Immediately after injection, the internal parameters of the stimulus twin instance are frozen by setting its parameter update interface to read-only, preventing any subsequent data stream during this test from modifying these parameters. The parameter update interface of the reference twin instance remains writable, allowing subsequent limited updates according to the parameter correction mechanism after S4 and the graded progressive loading mechanism before S2.

[0068] That is, after S13 is completed, the internal parameters of the excitation twin instance are locked at the values ​​at the identification moment throughout the entire test cycle of the servo under test, and the excitation twin instance thus becomes a "rigid mirror image of the servo under test at the identification moment"; the reference twin instance allows for slow tracking adjustments based on measured feedback in subsequent processes. The two instances overlap at the moment of identification completion, and gradually separate in the parameter space as the test progresses. The distance that separates is the physical quantity that the model residual sequence aims to capture.

[0069] If two twin instances share a nominal template without individual identification, the manufacturing tolerance differences between each tested servo will be directly presented as residuals, making it impossible for inspectors to distinguish between "individual tolerances" and "real faults." After individual identification in S11-S13, the residuals no longer contain individual tolerance components. The remaining part corresponds to the mismatch between the state changes of the tested servo after the identification time and the twin model, providing a high signal-to-noise ratio criterion after the noise floor is raised for the pass / fail judgment in S4.

[0070] Before proceeding to step S2, in some embodiments, the following protective processing is performed on the target loading torque sequence: random sampling pre-simulation is performed on the test cases covered by the target loading torque sequence to obtain a pre-simulation response set. If any sample in the pre-simulation response set enters the damage criterion region, the target loading torque sequence is decomposed into several stepped sub-sequences according to amplitude and sequentially distributed, with the internal parameters of the reference twin instance updated after each step. If none of the pre-simulation response sets enter the damage criterion region, the target loading torque sequence is distributed all at once. After the protective processing is completed, the pass / fail determination in S4 is performed. The above process is described in detail below.

[0071] Within the stimulus twin instance, using samples of the reference twin instance's internal parameters within a preset tolerance range as initial values, random sampling pre-simulation is performed on the test cases covered by the target loading torque sequence, generating a pre-simulation response set for the servo under test under the test cases. If any sampled pre-simulation in the pre-simulation response set indicates that the servo under test has entered a preset damage criterion zone, the target loading torque sequence is decomposed into several step sub-sequences by amplitude, and each step sub-sequence is sequentially sent to the loading motor. After each step sub-sequence is sent, the measured response sequence of the servo under test under the step sub-sequence is fed back to the reference twin instance, updating the reference twin instance's internal parameters. When the stimulus twin instance performs random sampling pre-simulation for the next step sub-sequence, the updated internal parameters of the reference twin instance are used as the sampling center value, and the stimulus twin instance's internal parameters remain frozen. If all pre-simulation response sets indicate that the servo under test has not entered the damage criterion zone, the target loading torque sequence is sent to the loading motor all at once. After protective processing is completed, the process proceeds to the pass / fail determination in S4.

[0072] As an example, the damage criterion zone consists of a combination of multiple quantified boundaries, including: the absolute value of the loading torque borne by the tested servo exceeding 120% of the rated loading torque; the duration of the loading torque exceeding 110% of the rated loading torque exceeding 1 second; the current of the tested servo exceeding the thermal overload threshold; and the angular rate of the tested servo exceeding the mechanical allowable limit of 600 degrees per second. Touching any of these boundaries constitutes entry into the damage criterion zone. The preset tolerance range corresponds to a reasonable engineering discreteness of individual parameters, with each dimension taking ±3σ or ±10% of the identified value. The tolerances for the static friction coefficient and dead zone width are relatively narrow, while the tolerance for load inertia is relatively wide.

[0073] For a specific servo motor under test in the baseline scenario, using the internal parameters of the reference twin instance as the sampling center and ±3σ as the tolerance range, 100 random sampling pre-plays are performed on the target loading torque sequence of the maneuver segment from 200 to 260 seconds within the excitation twin instance. If the pre-play results show that in 5 out of the 100 samples, the servo motor angle rate reaches 620 degrees per second around the 230th second, touching the angle rate boundary of the damage criterion zone, it is determined that "there is a sample entering the damage criterion zone in the pre-play response set," and graded progressive loading is initiated. Graded progressive loading decomposes the target loading torque sequence of the maneuver segment into three stepped sub-sequences according to the amplitude, corresponding to 60%, 80%, and 100% of the rated torque, respectively, and sends them to the loading motor sequentially.

[0074] After the first step (60%) is completed, the test control module collects the measured response sequence of the servo under test at this step and feeds it back to the reference twin instance. The reference twin instance updates its internal parameters based on the measured response sequence with a limited amplitude. The updated internal parameters are slightly adjusted according to the feedback direction, with each adjustment not exceeding 1% of the current parameter value. When the excitation twin instance performs the pre-rehearsal of the second step (80%), it obtains the updated internal parameters from the reference twin instance as the sampling center value and re-executes 100 random sampling pre-rehearsals. If the pre-rehearsal result shows that the servo under test does not enter the damage criterion area at the second step, the loading torque sequence of the second step is issued, and then the update and pre-rehearsal cycle is repeated until all three steps are completed. Throughout the process, the internal parameters of the excitation twin instance itself remain frozen and do not change due to the measured feedback of each step, thus ensuring that the judgment criterion for the subsequent S4 determination is still locked at the initial value at the identification time of S13.

[0075] In other embodiments, the aforementioned random sampling pre-simulation can be implemented using the Monte Carlo method. The number of Monte Carlo samplings, N, ranges from 100 to 1000. The sampling distribution is a multidimensional normal distribution centered on the internal parameters of the reference twin instance, and the variance of each dimension is determined by the identification covariance matrix of the individual parameters. This alternative method differs slightly from the aforementioned default random sampling pre-simulation in its sampling algorithm, but it is functionally equivalent in its purpose of "predicting the response distribution of the tested servo motor within the parameter tolerance range before the excitation is issued," and it also belongs to the protective processing implementation method for the target loading torque sequence.

[0076] Furthermore, if no samples from the pre-simulated response set enter the damage criterion area, it indicates that the tested servo motor has sufficient margin for this test case within the current parameter tolerance range. The target loading torque sequence is then directly sent to the loading motor in one go, without initiating graded progressive loading to avoid unnecessary test cycle overhead. After the protective treatment is completed, the pass / fail judgment in S4 begins. At this point, the measured response sequence on which S4 is based is either the response under a single, one-time loading or the response after three-stage cumulative loading. The judgment logic for S4 is the same in both cases.

[0077] For servos under test whose individual parameters fall within the acceptable range, directly applying the full-amplitude maneuvering load from the flight script could cause instantaneous peak values ​​that could lead the servo into a short-term damage zone, resulting in damage during testing and loss of the opportunity for retesting. This protective measure uses the measured response feedback to correct the parameters of the reference twin instance after each step, enabling the subsequent step's simulation to more accurately predict the servo's true response boundary. If the simulation reveals that a certain step is close to the damage zone, stopping at that step completes the screening, preserving the servo's complete state. Engineers can then decide whether to retest or release it at a lower specification.

[0078] S2. By stimulating the twin instance, a target loading torque sequence is generated based on the flight condition script and sent to the loading motor of the corresponding test channel. By referencing the twin instance, the expected response sequence of the tested servo is generated based on the control command of the tested servo and the sampled feedback of the tested servo.

[0079] S3. Exchange operational progress information among various excitation twin instances via the twin bus, and align the test cycle of each test channel according to the progress of the unified flight operational script.

[0080] As an example, the progress information includes three fields: current scenario segment identifier, time scale within the current segment, and percentage of completed scenario points within the current segment. The scenario segment identifier can be one of the following: takeoff segment, climb segment, cruise segment, maneuver segment, or landing segment. The time scale within the current segment is in seconds, indicating the progress time from the start point to the current moment. The percentage of completed scenario points within the current segment is the ratio of the number of sampling points issued within the current scenario segment to the total number of sampling points in that segment. The control commands for the tested servo correspond to the angle or angular velocity commands issued by the flight control computer to the servo. In this test environment, these commands are synchronously generated by the test master control module based on the flight scenario script and sent to the control command input terminal of the reference twin instance and the control command channel of the tested servo. The sampling feedback of the tested servo corresponds to the current angular position measured by the encoder or rotary transformer inside the tested servo, and is transmitted back to the feedback input terminal of the reference twin instance in real time.

[0081] During the execution of S2, the output of the excitation twin instance is transformed into a drive command sequence by the inverse model compensator of the loading mechanism and then sent to the loading motor. The loading motor converts the drive command sequence into physical torque acting on the servo under test fixture. At the same time, the reference twin instance takes the control command of the servo under test as input and the sampled feedback of the servo under test as the internal state correction source, and deduces according to its own servo mathematical model to output the expected response sequence to the test master control module. The two signal chains of S2 are executed in parallel. The excitation chain drives the physical loading, and the reference chain generates the judgment benchmark. The two chains converge at the test master control module to perform residual calculation in S4.

[0082] During the execution of S3, each stimulus twin instance publishes its current operational progress information to the twin bus at fixed intervals, such as every 10 milliseconds, while simultaneously subscribing to the progress information of other stimulus twin instances from the twin bus. The test master module calculates the progress deviation between channels based on the progress information of each channel and makes fine adjustments to channels whose progress deviation exceeds a preset allowable value. The physical clocks of the channels under test are not forced to be synchronized, allowing the real-time computing nodes of each channel to run on independent master clocks. However, the flight operational script advancement of each stimulus twin instance is aligned according to a unified progress negotiation mechanism, and the script progress deviation of each channel does not exceed the preset allowable value at any given time.

[0083] For the eight channels under test in the baseline scenario, the preset allowable script progress deviation is 1 second. For example, when the flight scenario execution reaches the 230th second of the maneuver phase, the actual script progress of the eight channels are 229.5 seconds, 230.2 seconds, 229.8 seconds, 230.1 seconds, 230.0 seconds, 229.7 seconds, 230.3 seconds, and 229.9 seconds, respectively, with a maximum deviation of 0.8 seconds, which is within the allowable value, and the test master module does not issue a fine-tuning command. If at a certain moment the script progress of a certain channel is 228.0 seconds while the other channels are concentrated around 230.0 seconds, the maximum deviation reaches 2.0 seconds, which exceeds the allowable value, and the test master module issues an acceleration synchronization command to the excitation twin instance of that channel, so that the channel catches up with the progress of the other channels in the next negotiation cycle. This fine-tuning mechanism has a negotiation cycle of 50 milliseconds, and the magnitude of a single fine-tuning does not exceed 200 milliseconds, avoiding a large-scale one-time alignment that introduces an excitation step.

[0084] It should be understood that the only information exchanged between the various excitation twin instances via the twin bus is operational progress information, excluding the sampled feedback of the servo under test, individual parameters, and the internal reference sequence of the excitation twin instance. That is, the twin bus only performs timing negotiation functions among the multi-channel excitation twin instances, and does not perform data sharing functions; the physical signal paths of each channel under test are completely isolated, and abnormal feedback from any channel's servo under test will not propagate to other channels' excitation twin instances via the twin bus.

[0085] Through multi-channel virtual progress synchronization in S3, each channel processes the same segment of the flight scenario script at the same physical moment, providing a temporal basis for cross-channel lateral comparison of the ternary residual sequences in S4. The calculation of "taking the time-varying mean of the individual residual sequences of all channels under test in this batch" in S43 is only meaningful under the premise of temporal alignment. If the script progress of each channel is not aligned, the time-varying mean will mix residuals from different script segments at each time point, losing the semantics of lateral comparison. This simultaneous synchronization mechanism, while achieving script progress alignment of each channel, still keeps the physical signal paths of each channel independent, providing both a temporal basis for multi-channel collaborative comparison and physical isolation conditions for fault isolation.

[0086] S4. Determine the pass / fail status of the servo under test based on the difference between the expected response sequence and the actual measured response sequence of the servo under test.

[0087] Specifically, S4 includes sub-steps S41-S44.

[0088] S41. Generate individual residual sequences based on the difference between the measured response sequence and the expected response sequence of the tested servo motor.

[0089] S42. Generate the model residual sequence based on the difference between the expected response sequence and the internal reference sequence of the excitation twin instance.

[0090] S43. Generate a channel residual sequence based on the difference between the individual residual sequence and the time-varying mean sequence of all individual residual sequences in the tested channels of this batch.

[0091] S44. The pass / fail status of the tested servo motor is determined based on the individual residual sequence, the model residual sequence, and the channel residual sequence.

[0092] As an example, the internal reference sequence of the excitation twin instance refers to the "response sequence that the tested servo should exhibit under the condition of no mismatch other than individual differences," which is derived from the internal parameters frozen in S13 during the execution of S2, while outputting the target loading torque sequence. The internal reference sequence is not sent to the loading motor, nor does it participate in the generation of physical excitation; it is only sent to the test master control module as the minuend of S42 to subtract from the expected response sequence output by the reference twin instance. Since the internal parameters of the excitation twin instance are frozen at the injection time of S13, the internal reference sequence reflects the "ideal response of the tested servo at the identification time," while the expected response sequence output by the reference twin instance reflects the "response of the tested servo after feedback tracking at the current time." The difference between the two is the model residual sequence required by S42.

[0093] The ternary residuals are all time series synchronized with the flight script in terms of data structure, with the sampling rate consistent with the flight script sampling rate. Each sampling moment corresponds to a scalar deviation value, with the dimension consistent with the response quantity. When the response is taken as an angle, the dimension of the ternary residuals is degrees; when the response is taken as an angular velocity, the dimension of the ternary residuals is degrees per second. The individual residual sequence in S41 is obtained by subtracting the expected response sequence point by point from the measured response sequence. The model residual sequence in S42 is obtained by subtracting the internal reference sequence point by point from the expected response sequence. The channel residual sequence in S43 is obtained by first taking the arithmetic mean of the individual residual sequences of all channels under test in this batch at each moment to obtain the time-varying mean sequence, and then subtracting the time-varying mean sequence point by point from the individual residual sequence of that channel.

[0094] For the servo under test in channel 3 of the baseline scenario, at the 230-second mark of the maneuver phase during the flight scenario execution, the measured response angle is 12.3 degrees, the expected response angle output by the reference twin instance is 12.0 degrees, and the angle value of the internal reference sequence output by the excitation twin instance at that moment is 11.9 degrees. Simultaneously, the arithmetic mean (i.e., the time-varying mean sequence) of the individual residuals of the eight channels under test in this batch at that moment is 0.25 degrees, with the mean of the individual residuals of the other seven channels under test (excluding channel 3) being approximately 0.243 degrees. Therefore, the individual residual of channel 3 at that moment is 12.3 - 12.0 = 0.3 degrees, the model residual is 12.0 - 11.9 = 0.1 degrees, and the channel residual is 0.3 - 0.25 = 0.05 degrees. These three residuals respectively reflect the deviation of the servo under test in channel 3 from the expected response of the reference twin instance, the deviation of the reference twin instance from the internal reference of the excitation twin instance, and the deviation of channel 3 from the average of this batch.

[0095] In other words, the individual residual sequence generally reflects the comprehensive deviation between the actual response of the tested servo and the judgment benchmark, which includes three components: the individual differences and degradation of the tested servo in this channel, the mismatch of the reference twin model itself, and the channel differences of this channel relative to other channels. The model residual sequence separately reflects the parameter drift of the reference twin instance relative to the excitation twin instance. Since the parameters of the two twin instances are exactly the same when injected in S13, and the parameters of the excitation twin instance are frozen, the deviation reflected by the model residual sequence only comes from the parameter offset generated by the reference twin instance tracking the feedback of the tested servo during operation. This offset is precisely the mismatch of the reference twin model relative to the ideal model. The channel residual sequence deducts the components common to all channels in this batch from the individual residual sequence, and the remaining part only reflects the deviation specific to that channel.

[0096] Mathematically, individual residuals are approximately equal to "real fault signal + model mismatch + channel-specific components + batch-specific noise", model residuals are approximately equal to "model mismatch", and channel residuals are approximately equal to "channel-specific components". These three constitute an approximately orthogonal residual basis. Based on this residual basis, the three scalar values ​​at the same time constitute a point in three-dimensional space, and different locations of the abnormal state in this three-dimensional space correspond to different physical causes.

[0097] If two twin instances share a nominal template without individual identification, the manufacturing tolerance differences between each tested servo will be directly presented as residuals, with the residual signal overlapping between the "individual tolerance component" and the "real fault component". After individual identification in S11-S13, the residuals no longer contain the individual tolerance component. The remaining part corresponds to the state change of the tested servo after the identification time and the mismatch between the twin model, providing a high signal-to-noise ratio criterion after the noise floor is raised for the pass / fail judgment in S4.

[0098] This ternary residual decomposition upgrades the original binary judgment of "pass or fail" based on a single residual to a three-dimensional attribution that can precisely pinpoint an anomaly to one of the individual, model, or channel based on the amplitude combination of the three residuals. This attribution capability provides structured input for the S44's pass / fail determination and subsequent attribution decisions, enabling the test bench to provide clearly physical and meaningful handling suggestions when an anomaly occurs, rather than simply marking it as fail.

[0099] Furthermore, S43 includes sub-steps S431-S434.

[0100] S431. Identify an idle channel from the multi-channel loading test bench, and inject a calibration disturbance of known amplitude and known frequency into the loading motor corresponding to the idle channel. The amplitude of the calibration disturbance is significantly smaller than the rated loading torque of the corresponding channel under test, and the frequency of the calibration disturbance is located at an independent frequency point outside the frequency band covered by the flight condition script.

[0101] S432. Extract the frequency components of the calibration disturbance from the individual residual sequences of each channel under test, and identify the crosstalk transfer function matrix between channels based on the extraction results.

[0102] S433. Based on the crosstalk transfer function matrix between channels, crosstalk separation processing is performed on the individual residual sequence to obtain the net individual residual sequence after removing the coupling components between channels.

[0103] S434. Subtract the time-varying mean sequence of the net individual residual sequence from the time-varying mean sequence of the net individual residual sequences of all channels under test in this batch to obtain the channel residual sequence.

[0104] As an example, in this batch of tests involving six servos, channels 7 and 8 of the 8-channel loading test bench were idle. At the 100-second mark of the cruise phase during the flight scenario, a single-frequency sinusoidal disturbance with an amplitude of 0.06 N·m (three-thousandths of the rated loading torque) was injected through the loading motor corresponding to idle channel 7. The disturbance frequency was set to 377 Hz. The main response frequency band of the servos under test in the flight scenario was 0 Hz to 50 Hz. 377 Hz was an independent frequency point outside this main band and would not overlap with any response components excited by the flight scenario in the frequency domain. After the calibration disturbance was applied through the loading motor in the idle channel, it was transmitted to other channels under test via the mechanical connection structure and electromagnetic coupling path of the test bench, appearing as a weak component of 377 Hz in the individual residual sequences of each channel under test.

[0105] In S432, the test master control module performs bandpass filtering and phase-locked amplification on the individual residual sequences of each channel under test, extracting the complex amplitude of the 377 Hz component, i.e., the amplitude and phase of this component. For the six channels under test in this embodiment, six sets of complex amplitude values ​​of 377 Hz are obtained respectively. These six sets of values ​​are the coupling coefficients from the unit perturbation injected from the idle channel 7 to each channel under test. If the 8th channel is simultaneously activated to inject independent perturbations at different frequencies, such as 413 Hz, the coupling coefficients from the 8th channel to each channel under test can also be identified synchronously. The two sets of results are combined to form a 2-column, 6-row inter-channel crosstalk transfer function matrix; if only one idle channel participates in the injection, the crosstalk transfer function matrix degenerates into a 1-column, 6-row column vector.

[0106] In S433, the test master control module performs crosstalk separation processing on the individual residual sequences of each channel under test based on the crosstalk transfer function matrix. Specifically, the calibration disturbance injected into the idle channel is multiplied by the corresponding crosstalk transfer function to obtain the expected component of the disturbance in the individual residual sequence of each channel under test. Then, the corresponding expected component is subtracted from the individual residual sequence of each channel under test to obtain the net individual residual sequence after removing the inter-channel coupling components. In the frequency domain, this processing is manifested as estimating the mutual leakage between channels according to the crosstalk matrix and subtracting the leakage part from the original individual residual; in the time domain, it is manifested as crosstalk separation filtering, so that the net individual residual sequence only reflects the deviation generated by the servo under test itself within the channel.

[0107] In S434, the channel residual sequence is recalculated using the net individual residual sequence instead of the individual residual sequence defined in S43: first, the arithmetic mean of the net individual residual sequences of all channels under test in this batch is taken at each time point to obtain a new time-varying mean sequence; then, the channel residual sequence is obtained by subtracting this time-varying mean sequence from the net individual residual sequence of the channel point by point. After the crosstalk separation processing of S431-S434, the channel residual sequence no longer contains interference leaked from adjacent channels, and only reflects the abnormal level of the channel itself.

[0108] In other embodiments, crosstalk calibration between channels can also be performed using a power-on initialization scan method instead of online calibration of idle channels. Specifically, after each multi-channel loaded test bench is powered on and before the servo under test is mounted, a low-amplitude test disturbance is injected into each channel individually, while the remaining channels remain stationary. The crosstalk transfer function from each channel to other channels is measured sequentially, and the crosstalk transfer function matrix identified at power-on is stored in the test main control module. All tests after this power-on process use this matrix for crosstalk separation. Compared with online calibration of idle channels, the power-on initialization scan method does not depend on the existence of idle channels in this batch and can be used in full-load batch scenarios where all channels are mounted. Its cost is that it cannot track crosstalk changes during the test process, such as coupling changes caused by loosening of mechanical connections during the test. However, it is effective enough for slowly changing structural crosstalk and also achieves the purpose of removing inter-channel coupling components to generate a net individual residual sequence.

[0109] The reason for setting the amplitude of the calibration disturbance to between one-thousandth and one-hundredth is for engineering balance: if the amplitude is further reduced, the signal-to-noise ratio of the 377 Hz component extracted from the individual residuals of the tested channel decreases, resulting in poorer identification accuracy; if the amplitude is further increased, the disturbance torque output by the motor in the idle channel may cause significant mechanical vibration when transmitted to the tested channel through the mechanical structure, which will have a visible impact on the test accuracy of the tested servo motor. The range of one-thousandth to one-hundredth further includes sub-steps S441-S444 in S444.

[0110] S441. Maintain an absolute reference distribution constructed from the individual parameters of the tested servos from historical batches. The absolute reference distribution remains unchanged during the testing of the tested servos.

[0111] S442. The relative distribution of this batch is constructed in real time from the individual parameters of the first few tested servos in this batch. The relative distribution of this batch is dynamically updated as the number of tested servos in the multi-channel loading test bench increases.

[0112] S443. If the amplitude of the individual residual sequence, the amplitude of the model residual sequence, and the amplitude of the channel residual sequence are all lower than the corresponding preset anomaly threshold, and the individual parameters of the tested servo motor fall within the qualified range of the absolute reference distribution, and the individual parameters of the tested servo motor do not belong to the outlier of the relative distribution of this batch, the tested servo motor is marked as qualified; otherwise, the tested servo motor is marked as re-inspection or unqualified.

[0113] S444. Perform statistical testing on the offset of the distribution center of the relative distribution of this batch relative to the distribution center of the absolute reference distribution. If the offset exceeds the preset incoming batch offset threshold, trigger an incoming batch offset warning.

[0114] As an example, the absolute baseline distribution is constructed from the accumulated individual parameters of several batches of tested servos over the past. Specifically, individual parameters of 500 servos of the same model are extracted from the most recent 10 historical batches. The mean and standard deviation are calculated for four dimensions: static friction coefficient, dead zone width, load inertia, and zero-position deviation. The absolute baseline distribution is calculated by adding or subtracting three times the standard deviation from the mean within the acceptable range of each dimension. The absolute baseline distribution does not change during the testing of the tested servos regardless of the number of tested servos in the current batch. Even if the distribution of individual parameters in the current batch shows an overall shift, the mean and standard deviation of the absolute baseline distribution remain unchanged from the historical baseline until a sufficient number of historical batches are available for periodic recalculation by the maintenance process.

[0115] The construction of the relative distribution for this batch begins with the M0th servo motor in this batch that has completed testing. M0 is set to 3 to ensure that the initial statistics have basic reliability. After each servo motor is tested, its individual parameters are included in the relative distribution for this batch, and the mean and standard deviation of each dimension are dynamically updated. The outlier determination interval for the relative distribution of this batch is the mean plus or minus 2 times the standard deviation, and the threshold is stricter than that of the absolute benchmark distribution to highlight outliers within the same batch.

[0116] For example, in the baseline scenario, the static friction coefficient of a tested servo motor is identified as 0.12 N·m. The absolute baseline distribution has a mean of 0.08 N·m and a standard deviation of 0.015 N·m along the static friction coefficient dimension, with a pass / fail range of [0.035, 0.125] N·m. The tested servo motor's 0.12 N·m falls within this pass / fail range and is determined by the absolute baseline distribution. However, the static friction coefficient samples of the first 5 tested servos in this batch are 0.080, 0.085, 0.090, 0.082, and 0.088 N·m. The calculated mean of the relative distribution of this batch along this dimension is 0.085 N·m, with a standard deviation of approximately 0.004 N·m. The outlier range is outside of [0.077, 0.093] N·m. The tested servo motor's 0.12 N·m is far outside this outlier range and is therefore identified as an outlier in this batch. If the amplitude of the three-dimensional residual sequence of the tested servo is lower than the corresponding anomaly threshold at this time, the AND criterion of S443 fails as a whole because the "outlier in this batch" condition is not met. The tested servo is marked as "re-inspected" instead of "qualified" and the engineer makes a second decision.

[0117] During the execution of S443, the three conditions—ternary residual amplitude, absolute reference distribution qualification, and batch relative distribution non-outlier—are combined using an AND relationship. Only when all conditions are met is the device marked as qualified. If any condition is not met, the marking—"re-inspection" or "unqualified"—is determined based on the nature of the specific condition: if only the batch outlier condition is not met, it is marked as re-inspection; if the absolute reference distribution qualification interval condition is not met or the ternary residual amplitude exceeds the limit, it is marked as unqualified. Servos marked as re-inspection enter a manual secondary decision-making process, where engineers make a release or rejection decision based on historical process fluctuations. Servos marked as unqualified directly enter the rework or scrapping process.

[0118] In S444, the test control module performs a statistical test on the offset between the distribution center of the relative distribution of this batch and the distribution center of the absolute reference distribution. Specifically, for each parameter dimension, the difference between the mean of the relative distribution of this batch and the mean of the absolute reference distribution is taken, and divided by the standard deviation of the absolute reference distribution to obtain the normalized offset. When the absolute value of the normalized offset of any dimension exceeds the preset incoming batch offset threshold, an incoming batch offset warning is triggered. The preset incoming batch offset threshold is, for example, 1.5 times. As an example of another batch, after a subsequent batch has completed the testing of 5 tested servos, the mean of the relative distribution of this batch in the static friction coefficient dimension is 0.105 N·m, the mean of the absolute reference distribution in this dimension is 0.08 N·m, and the standard deviation is 0.015 N·m. The normalized offset is (0.105-0.080) / 0.015≈1.67, which exceeds the preset threshold of 1.5, triggering an incoming batch offset warning and indicating that there may be batch-wide process drift in the upstream incoming material process. After the warning is triggered, the S443 determination of the tested servo motors in this batch will still be handled on an individual basis. The warning for the entire batch will be archived as a batch-level quality evaluation record for engineers to determine whether to suspend the testing of this batch and to work with upstream suppliers to investigate.

[0119] Through the dual-distribution criteria of S441-S444, the pass / fail criteria simultaneously possess two dimensions: "cross-batch specification stability" and "outlier identification within the same batch." The absolute baseline distribution corresponds to the specification boundaries of the model, ensuring consistency in cross-batch criteria; the relative distribution within the batch corresponds to outlier detection within the batch, identifying individuals within the same batch that are abnormal compared to other tested servos. After the two distributions form the AND criterion, tested servos that are qualified within the model but outlier relative to the batch will not be released, and tested servos that deviate from the overall batch but are all within the specifications will not be misjudged as the entire batch being unqualified; the batch deviation warning of S444 further provides an upstream feedback entry point for process closure at the batch level, so that the output of the test bench is not only a single-piece pass / fail judgment, but also includes a batch-level quality status evaluation.

[0120] While performing the S44 pass / fail assessment, the system also identifies the causes of abnormal states in the test bench based on the combination patterns of the ternary residual sequences. If the amplitude of an individual residual sequence exceeds the individual abnormality threshold, and the amplitude of the model residual sequence is below the model abnormality threshold, and the amplitude of the channel residual sequence is below the channel abnormality threshold, the abnormality is attributed to an individual failure of the servo under test. If the amplitude of the model residual sequence exceeds the model abnormality threshold, and the amplitudes of the individual residual sequences in multiple channels under test in the multi-channel loading test bench increase synchronously, the abnormality is attributed to model drift of the reference twin instance, triggering parameter correction for the reference twin instance. If the amplitude of the channel residual sequence in at least one channel under test in the multi-channel loading test bench exceeds the channel abnormality threshold, and the amplitudes of the channel residual sequences in the remaining channels under test in the multi-channel loading test bench are all below the channel abnormality threshold, the abnormality is attributed to a hardware failure in at least one channel under test, and at least one channel under test is suspended from this batch of tests.

[0121] As an example, the individual anomaly threshold, model anomaly threshold, and channel anomaly threshold are set independently, each corresponding to the statistical characteristics of the ternary residual sequence under normal testing conditions. The individual anomaly threshold is taken as three times the standard deviation of the amplitude of the individual residual of the tested servo under the absolute benchmark distribution, approximately 0.45 degrees for the benchmark scenario; the model anomaly threshold is taken as two times the standard deviation of the reasonable deviation between the reference sequences of two twin instances, approximately 0.15 degrees for the benchmark scenario; and the channel anomaly threshold is taken as three times the standard deviation of the root mean square value of the channel residuals under long-term operation, approximately 0.20 degrees for the benchmark scenario. The three thresholds are independent of each other and independently determine whether an anomaly has occurred from the three residual dimensions.

[0122] Three scenarios are constructed to correspond to the implementation process of the three-branch attribution.

[0123] Scenario 1 is an individual failure attribution. At the 275-second mark of the landing phase during flight scenario execution, the individual residual for channel 5 reached 0.60 degrees, the model residual was 0.10 degrees, and the channel residual was 0.08 degrees. The individual residual for channel 5 exceeded the individual anomaly threshold of 0.45 degrees, the model residual was below the model anomaly threshold of 0.15 degrees, and the channel residual was below the channel anomaly threshold of 0.20 degrees. At this point, the combination pattern of the three residuals satisfies "individual exceeding limits, model normal, channel normal," and the anomaly is attributed to the individual failure of the tested servo in channel 5. The test master control module marks the tested servo in channel 5 as unqualified; the tests of the remaining channels in this batch are unaffected.

[0124] Scenario 2 involves model drift attribution for the reference twin instance. At the same time, the model residuals for channels 1, 2, 4, and 5 all reach 0.22 degrees, 0.20 degrees, 0.24 degrees, and 0.21 degrees respectively, simultaneously exceeding the model anomaly threshold of 0.15 degrees. Furthermore, the corresponding individual residuals rise synchronously from the previous level of 0.10 degrees to the range of 0.35 to 0.40 degrees, while the channel residuals for each channel are all below the channel anomaly threshold. At this point, the combination pattern of the ternary residuals satisfies "model exceeding limits, synchronous increase in individual residuals across multiple channels," and the anomaly is attributed to model drift in the reference twin instance. The test control module triggers parameter correction for the reference twin instance; the specific correction process is executed by the reference twin instance parameter correction handling described later.

[0125] Scenario 3 is attributed to channel hardware failure. At the same time, the channel residual of channel 7 reaches 0.45 degrees, exceeding the channel anomaly threshold of 0.20 degrees. The channel residuals of the other five channels under test are 0.06 degrees, 0.08 degrees, 0.05 degrees, 0.07 degrees, and 0.06 degrees, respectively, all below the channel anomaly threshold. Simultaneously, the individual residual of channel 7 increases, for example, reaching 0.50 degrees, while the model residual remains within the normal range. At this point, the combination mode of the ternary residuals satisfies "a certain channel's residual exceeds the limit, while the residuals of the other channels are normal," and the anomaly is attributed to a hardware failure in channel 7. The test control module suspends channel 7 from this batch of tests, and engineers go online to check whether there is electrical loosening or mechanical damage to the loading motor windings, tooling clamping mechanism, and sensor cables of channel 7. The remaining channels in this batch continue normal testing, and the tested servo motor of channel 7 is considered incomplete and transferred to the next batch for retesting.

[0126] The three branches are independent of each other in terms of response conditions and are not mutually exclusive; multiple branches may be triggered at the same time. For example, when the tested servo fails in a certain channel and is accompanied by a hardware failure in that channel, the first and third branches are triggered simultaneously, each executing its own handling action: the tested servo is marked as unqualified, and the channel is suspended from this batch of tests. The time granularity of the attribution decision is consistent with the time granularity of the ternary residual sequence generation, and the judgment is made according to the time series; to avoid misattribution caused by instantaneous interference, the triggering of each branch adopts the sliding window criterion, and attribution is formally triggered only when the condition is continuously met for multiple consecutive sampling times; instantaneous single over-limit does not trigger directly.

[0127] Through a three-branch attribution decision, the ternary residual is upgraded from a "detection signal" to a "response signal." During operation, the test bench autonomously completes a closed loop of "detecting anomalies, locating root causes, and executing remedial actions." The individual failure branch provides a remedial action by marking the servo under test; the model drift branch provides a remedial action by triggering parameter correction for the reference twin instance; and the channel hardware failure branch provides a remedial action by suspending the channel and dispatching a maintenance work order. All three types of remedial actions can be automatically executed by the test master module, allowing the entire batch of tests to continue even when some channels or some servos under test are abnormal, without the need for engineer intervention at the moment an anomaly occurs.

[0128] Following S4, parameter correction processing for the reference twin instance is also included. In response to the sliding window mean of the model residual sequence continuously deviating from the preset model baseline range, the intrinsic parameters of the reference twin instance are updated at a preset rate and limited to keep the intrinsic parameters of the twin instance frozen.

[0129] As an example, the sliding window length is taken as a complete flight scenario phase, corresponding to a cruise segment of 120 seconds and a maneuver segment of 60 seconds; the window sliding step size is 10 seconds, and the window mean is recalculated every 10 seconds. The model baseline range is taken as the confidence interval of the reasonable deviation between the two twin instances at the injection time of S13 identification. In the baseline scenario, this range is set as the mean of the model residuals being in the range of [-0.05, +0.05]. The preset rate limit corresponds to the maximum change amplitude of each parameter update, which is taken as 1% of the current parameter value of the reference twin instance, that is, the change of parameters after each update does not exceed 1%; the purpose of the rate limit is to prevent the parameter jitter from causing discontinuity in the expected response sequence output by the reference twin instance in the time domain, thereby polluting the individual residual calculation in S41.

[0130] Parameter correction is triggered only if two conditions are met simultaneously: first, the mean of the sliding window of the model residual sequence exceeds the model baseline range; second, this excess state persists for multiple consecutive windows, such as three consecutive windows corresponding to 30 seconds. A single window exceeding the limit does not trigger correction to avoid unnecessary parameter adjustments caused by transient interference; only when multiple consecutive windows exceed the limit is it considered a genuine signal of model drift, and the correction process is officially initiated.

[0131] For the reference twin instance in the baseline scenario, if the mean model residual reaches 0.08 degrees, 0.09 degrees, and 0.10 degrees respectively within three consecutive sliding windows of the cruise phase (100-120 seconds, 110-130 seconds, and 120-140 seconds), all exceeding the baseline model range of [-0.05, +0.05] degrees and persisting for three windows, then parameter correction is triggered at 140 seconds. The correction direction is determined by the sign of the mean model residual: a positive mean model residual indicates that the expected response of the reference twin instance is generally too small and differs positively from the internal reference of the excitation twin instance. The parameters of the corresponding dimension of the reference twin instance need to be adjusted in the direction that increases the expected response, with the adjustment magnitude not exceeding 1% of the current value of the parameter. After correction, monitoring continues for the next window. If the mean model residual still exceeds the baseline range, a limiting update continues in the next triggering cycle to gradually converge the mean model residual to the baseline range.

[0132] During the correction process, the internal parameters of the excitation twin instance remain frozen and do not participate in the correction. The parameter correction of the reference twin instance only applies to the reference twin instance itself. The corrected reference twin instance continues to generate the expected response sequence with the new parameters. This expected response sequence is sent to S4 to participate in the subsequent calculation of individual residuals and model residuals. Since the parameters of the excitation twin instance remain unchanged, its internal reference sequence still reflects the ideal response at the identification time. Therefore, the new round of model residuals in S42 will reflect the difference between the "corrected reference twin instance" and the "excitation twin instance at the identification time", rather than a further expansion of model drift.

[0133] The parameter correction process for the reference twin instance is clearly separated from the parameter updates in the aforementioned hierarchical progressive loading process in terms of timing. Hierarchical progressive loading occurs during the pre-protective processing phase before entering S2, updating the reference twin instance's parameters after each step of the subsequence is completed to provide more accurate sampling center values ​​for the random sampling pre-simulation of the next step. The parameter correction described in this section occurs after the normal judgment phase in S4, continuously tracking slowly changing model drift that occurs during the main test. The two phases are located at different points in the test flow, and their processing timescales differ. The former corresponds to the loading window of each step, i.e., a second-level scale, while the latter corresponds to the sliding window of the entire flight scenario script, i.e., a minute-level scale. Neither will trigger simultaneously to avoid conflicting updates to the reference twin instance's parameters by two different mechanisms.

[0134] Through this parameter correction mechanism, the reference twin instance can track two types of slowly changing factors during the long-term operation of the entire batch of tests: first, the gradual degradation of the tested servo itself, including lubrication wear and seal aging; second, the long-term modeling error of the twin model relative to the real servo. This error is not significant at the time of identification, but gradually emerges during the full-condition progression covered by the flight scenario script. The excitation twin instance is always kept frozen to ensure that the judgment criterion does not change with the running signal, while the limited update of the reference twin instance ensures that the residual accuracy does not significantly decrease at the end of the test cycle due to model aging. Together, they constitute a dual-track adaptive mechanism in the main test process, maintaining the tracking accuracy of residual calculation while ensuring the rigidity and stability of the judgment criterion.

[0135] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0136] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A digital twin collaborative control system for batch testing of unmanned aerial vehicle (UAV) servo motors, characterized in that, It includes a multi-channel loading test bench, an excitation twin module, a reference twin module, a twin bus, and a test master control module; The multi-channel loading test bench includes multiple independent test channels. Each test channel is equipped with a loading motor, a loading torque sensor, and a servo motor under test fixture. The test channel currently equipped with the servo motor under test is called the test channel, and the test channel not currently equipped with the servo motor under test is called the idle channel. The excitation twin module configures an excitation twin instance for each of the channels under test. The excitation twin instance generates a target loading torque sequence based on the flight condition script and sends it to the loading motor. The input of the excitation twin instance shields the sampling feedback of the corresponding servo motor under test. The reference twin module configures a reference twin instance for each channel under test. The reference twin instance generates the expected response sequence of the servo under test based on the control command of the servo under test and the sampled feedback of the servo under test. The twin bus connects each of the excitation twin instances and carries the operational progress information of each of the excitation twin instances, so that each of the channels under test is aligned with the test rhythm according to the unified flight operational script. After the servo under test is mounted, the test main control module retrieves the corresponding twin template from the preset twin template library according to the model information of the servo under test as the initialization structure of the excitation twin instance and the reference twin instance. The test control module applies a preset identification stimulus to the servo under test, and identifies the individual parameters of the servo under test based on the identification response of the servo under test. The individual parameters include static friction coefficient, dead zone width, load inertia, and zero-position deviation. The test control module synchronously injects the individual parameters into the stimulus twin instance and the reference twin instance, respectively, as the internal parameters of the stimulus twin instance and the reference twin instance. After the injection is completed, the internal parameters of the stimulus twin instance are frozen until the test of the servo under test ends. The test control module determines the pass / fail status of the servo under test based on the difference between the expected response sequence and the actual response sequence of the servo under test. The test master control module generates an individual residual sequence based on the difference between the measured response sequence of the tested servo motor and the expected response sequence, generates a model residual sequence based on the difference between the expected response sequence and the internal reference sequence of the excitation twin instance, and generates a channel residual sequence based on the difference between the individual residual sequence and the time-varying mean sequence of the individual residual sequences of all channels under test in this batch. The test master control module maintains an absolute reference distribution constructed from the individual parameters of the tested servos in historical batches, and the absolute reference distribution remains unchanged during the testing of the tested servos; the test master control module constructs a relative distribution of the current batch in real time from the individual parameters of the first few tested servos in the current batch, and the relative distribution of the current batch is dynamically updated as the number of tested servos in the multi-channel loading test bench increases. The test control module responds when the amplitudes of the individual residual sequence, the model residual sequence, and the channel residual sequence are all lower than the corresponding preset anomaly thresholds, and the individual parameters of the tested servo fall within the qualified range of the absolute reference distribution, and the individual parameters of the tested servo do not belong to the outliers of the relative distribution of this batch, by marking the tested servo as qualified; otherwise, the tested servo is marked as re-inspected or unqualified. The test control module performs statistical verification on the offset of the distribution center of the relative distribution of this batch relative to the distribution center of the absolute reference distribution, and triggers an incoming batch offset warning when the offset exceeds a preset incoming batch offset threshold.

2. A digital twin cooperative control method for batch testing of UAV servo motors, characterized in that, The digital twin collaborative control system for batch testing of UAV servo motors as described in claim 1 is implemented, and the method includes the following steps: S1. Deploy an excitation twin instance and a reference twin instance for each channel under test of the multi-channel loading test bench. Send the flight scenario script to the input terminal of the excitation twin instance, and simultaneously send the control commands of the servo under test and the sampled feedback of the servo under test to the input terminal of the reference twin instance. Shield the excitation twin instance from accessing the sampled feedback of the servo under test. S1 includes the following sub-steps: S11. After the servo under test is mounted, the corresponding twin template is retrieved from the preset twin template library according to the model information of the servo under test as the initialization structure of the excitation twin instance and the reference twin instance. S12. Apply a preset identification excitation to the servo motor under test, and identify the individual parameters of the servo motor under test based on the identification response of the servo motor under test. The individual parameters include static friction coefficient, dead zone width, load inertia and zero position deviation. S13. The individual parameters are synchronously injected into the excitation twin instance and the reference twin instance, respectively, as the internal parameters of the excitation twin instance and the internal parameters of the reference twin instance; after the injection is completed, the internal parameters of the excitation twin instance are frozen until the test of the servo under test ends; S2. The target loading torque sequence is generated based on the flight condition script through the excitation twin instance and sent to the loading motor of the corresponding test channel. The expected response sequence of the tested servo is generated based on the control command of the tested servo and the sampled feedback of the tested servo through the reference twin instance. S3. Exchange operational progress information among the various excitation twin instances via twin bus, and align the test cycle of each of the test channels according to the unified flight operational script. S4. Determine the pass / fail status of the servo under test based on the difference between the expected response sequence and the measured response sequence of the servo under test; S4 includes the following sub-steps: S41. Generate an individual residual sequence based on the difference between the measured response sequence and the expected response sequence of the tested servo motor; S42. Generate a model residual sequence based on the difference between the expected response sequence and the internal reference sequence of the excitation twin instance; S43. Generate a channel residual sequence based on the difference between the individual residual sequence and the time-varying mean sequence of all individual residual sequences in the tested channels of this batch; S44. Determine the pass / fail status of the tested servo motor based on the individual residual sequence, the model residual sequence, and the channel residual sequence; S44 includes the following sub-steps: S441. Maintain an absolute reference distribution constructed from the individual parameters of the tested servo motors from historical batches, the absolute reference distribution remaining unchanged during the testing of the tested servo motors; S442. The relative distribution of this batch is constructed in real time from the individual parameters of the first few tested servos in this batch that have completed the test. The relative distribution of this batch is dynamically updated as the number of tested servos in the multi-channel loading test bench increases. S443. In response to the fact that the amplitude of the individual residual sequence, the amplitude of the model residual sequence, and the amplitude of the channel residual sequence are all lower than the corresponding preset anomaly threshold, and the individual parameter of the tested servo falls within the qualified range of the absolute reference distribution, and the individual parameter of the tested servo does not belong to the outlier point of the relative distribution of this batch, the tested servo is marked as qualified; otherwise, the tested servo is marked as re-inspection or unqualified. S444. Perform a statistical test on the offset of the distribution center of the relative distribution of the current batch relative to the distribution center of the absolute reference distribution, and trigger an incoming batch offset warning in response to the offset exceeding a preset incoming batch offset threshold.

3. The digital twin cooperative control method for batch testing of UAV servo motors according to claim 2, characterized in that, Following S4, the method further includes: in response to the sliding window mean of the model residual sequence continuously deviating from a preset model baseline range, updating the internal parameters of the reference twin instance at a preset rate limit, while keeping the internal parameters of the excitation twin instance frozen.

4. The digital twin cooperative control method for batch testing of UAV servo motors according to claim 2, characterized in that, S43 includes the following sub-steps: S431. Identify the idle channel from the multi-channel loading test bench, and inject a calibration disturbance of known amplitude and known frequency through the loading motor corresponding to the idle channel. The amplitude of the calibration disturbance is significantly smaller than the rated loading torque of the corresponding channel under test, and the frequency of the calibration disturbance is located at an independent frequency point outside the frequency band covered by the flight condition script. S432. Extract the frequency components of the calibration disturbance from the individual residual sequences of each of the measured channels, and identify the inter-channel crosstalk transfer function matrix based on the extraction results; S433. Based on the inter-channel crosstalk transfer function matrix, perform crosstalk separation processing on the individual residual sequence to obtain a net individual residual sequence with the inter-channel coupling components removed; S434. Subtract the time-varying mean sequence of the net individual residual sequence from the time-varying mean sequence of the net individual residual sequences of all channels under test in this batch to obtain the channel residual sequence.

5. The digital twin cooperative control method for batch testing of UAV servo motors according to claim 2, characterized in that, The S44 also includes the following parallel branches: In response to the individual residual sequence amplitude exceeding the individual anomaly threshold, the model residual sequence amplitude being lower than the model anomaly threshold, and the channel residual sequence amplitude being lower than the channel anomaly threshold, the anomaly is attributed to the individual failure of the tested servo motor. In response to the model residual sequence amplitude exceeding the model anomaly threshold, and the amplitude of the individual residual sequences in multiple channels under test in the multi-channel loading test bench increasing synchronously, the anomaly is attributed to model drift of the reference twin instance, and parameter correction of the reference twin instance is triggered. In response to the fact that the amplitude of the channel residual sequence of at least one channel under test in the multi-channel loading test bench exceeds the channel anomaly threshold, and the amplitude of the channel residual sequence of the remaining channels under test in the multi-channel loading test bench is lower than the channel anomaly threshold, the anomaly is attributed to a hardware failure of the at least one channel under test, and the at least one channel under test is suspended from the current batch of tests.

6. The digital twin cooperative control method for batch testing of UAV servo motors according to claim 2, characterized in that, Before sending the target loading torque sequence to the loading motor corresponding to the measured channel in step S2, the following protective processing is performed on the target loading torque sequence: Within the excitation twin instance, using the sampling of the internal parameters of the reference twin instance within a preset tolerance range as initial values, random sampling pre-simulation is performed on the test cases covered by the target loading torque sequence to generate a pre-simulation response set of the servo under test under the test cases; In response to any sampling pre-simulation showing that the tested servo motor has entered a preset damage criterion zone in the pre-simulation response set, the target loading torque sequence is decomposed into several ladder sub-sequences according to amplitude, and each ladder sub-sequence is sent to the loading motor in sequence; after each ladder sub-sequence is sent, the measured response sequence of the tested servo motor under the ladder sub-sequence is fed back to the reference twin instance, and the internal parameters of the reference twin instance are updated; when the excitation twin instance executes the random sampling pre-simulation of the next ladder sub-sequence, it uses the updated internal parameters of the reference twin instance as the sampling center value, and the internal parameters of the excitation twin instance remain frozen; If the pre-simulation response set shows that the tested servo motor has not entered the damage criterion area, the target loading torque sequence is sent to the loading motor all at once; After the protective process is completed, the qualification determination in step S4 is performed.

7. The digital twin cooperative control method for batch testing of UAV servo motors according to claim 6, characterized in that, The excitation twin instance is connected in series with the loading mechanism inverse model compensator. The loading mechanism inverse model compensator is configured to perform a pre-transformation on the target loading torque sequence output by the excitation twin instance according to the parameters of the loading mechanism inverse model compensator, so as to obtain a drive command sequence and send it to the loading motor. The parameters of the loading mechanism inverse model compensator are obtained in the following way: after each power-on of the multi-channel loading test bench, the first servo motor under test that is mounted after the power-on is taken as the first piece. The first piece is subjected to frequency sweep excitation and the frequency sweep response of the first piece is collected. The parameters of the loading mechanism inverse model compensator are identified based on the frequency sweep response of the first piece. The identified parameters of the loading mechanism inverse model compensator are used in all subsequent tests during the power-on.

Images