Experimental calibration method based on multi-sensor signal fusion processing

Abstract

The invention discloses an experimental calibration method based on multi-sensor signal fusion processing, particularly relates to the field of data analysis, and comprises the steps of a stepped confrontation disturbance protocol, data stream processing and real-time analysis, and multi-stage logic judgment and calibration execution. A stepped confrontation disturbance protocol is created for the first time, potential errors of the sensor are excited in real time and corrected in a closed loop mode under the condition that system operation is not interrupted, and a traditional shutdown calibration mode is thoroughly replaced; a multi-stage disturbance defense system is constructed, and composite interferences such as spectrum aliasing, overload distortion and signal-to-noise ratio attenuation caused by electromechanical thermal multi-physical field strong coupling are effectively suppressed; laser space calibration and dynamic rigidity modulation technologies are fused, and space drift errors caused by deformation of the mounting base and mechanical loosening are accurately diagnosed and compensated. Finally, long-term precision stability and data fusion reliability of the multi-source sensor under complex working conditions are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and more specifically, to an experimental calibration method based on multi-sensor signal fusion processing. Background Art

[0002] In the condition monitoring and fault diagnosis of large rotating machinery (such as steam turbines, compressors, generator sets, etc.), the prior art uses periodic shutdown calibration combined with a single-dimensional compensation strategy for sensor calibration. The standard process is as follows: manually calibrate the sensor with a standard measuring tool in the system's stationary state, or during operation, based on the feedback of a single environmental factor sensor, apply a preset compensation coefficient (such as the temperature drift compensation curve of a thermistor) to correct the sensor output. The calibration process relies on static data acquisition and manual parameter adjustment under fixed working conditions.

[0003] However, the existing methods have serious limitations: First, the shutdown calibration mode disrupts the continuous operation state of the system, cannot capture transient process errors, and has high maintenance costs; Second, the single environmental factor compensation mechanism is difficult to cope with multi-physical field coupling disturbances such as voltage fluctuations, mechanical shocks, electromagnetic interference, and oil pollution, resulting in the failure of compensation parameters under dynamic working conditions; Third, there is a lack of the ability to detect deep coupling errors such as deformation of the sensor mounting base, loosening of mechanical connections, and spatial position drift, allowing compound errors such as vibration transmission path distortion and electromagnetic induction crosstalk to continuously accumulate and be uncorrectable. Summary of the Invention

[0004] To overcome the above-mentioned defects of the prior art, the present invention provides an experimental calibration method based on multi-sensor signal fusion processing, and through the following solutions, to solve the problems raised in the above background art.

[0005] To achieve the above object, the present invention provides the following technical solution: An experimental calibration method based on multi-sensor signal fusion processing, including:

[0006] S1: Stepwise adversarial perturbation protocol: Design a multi-stage electromechanical-thermal coupling perturbation scenario to stimulate potential errors of the sensor;

[0007] S2: Data stream processing and real-time analysis: Generate a dynamic benchmark through optical and material marking techniques to replace traditional manual calibration. The data stream passes through a spatio-temporal alignment pipeline to unify the sampling rate and coordinate system, online residual analysis calculates the deviation between multi-source signals and physical models, and frequency domain coherence monitoring verifies the correlation strength between sensors;

[0008] S3: Multi-level logical judgment and calibration execution: Implement a four-level decision rule base and dynamic confidence weight management to diagnose the sensor state and trigger calibration.

[0009] Preferably, the first stage of the multi-stage electromechanical-thermal coupling disturbance scenario in S1 is the establishment of the base condition and the injection of low-frequency disturbances; the second stage is the load step and high-frequency common-mode interference; the third stage is the transient shock and dynamic stiffness modulation; the fourth stage is the surge boundary limit test.

[0010] Preferably, for the establishment of the base condition and the injection of low-frequency disturbances, first, the equipment runs at the rated speed for ten minutes under no-load conditions, and the reference data of sensors for vibration, temperature, current, and oil parameters are collected. Subsequently, the first-stage disturbance is injected: a ±5% 0.5-times power frequency voltage fluctuation with a frequency of 25 Hz is superimposed on the power input side through a programmable frequency converter for five minutes. The first stage focuses on monitoring the change in the harmonic spectrum of the current transformer and the response delay of the low-frequency component of the bearing housing vibration. The low-frequency component specifically refers to the frequency band below 200 Hz. The temperature distribution uniformity on the surface of the stator core by the infrared thermal imager is synchronously recorded. When the detected vibration phase lag exceeds 15° or the temperature gradient sudden change rate exceeds 8 °C / m, the sensor gain calibration module is triggered.

[0011] Preferably, the sensor gain calibration module first collects the output signal S of the target sensor under the reference condition base , and at the same time, obtains the real physical quantity P at the same position through a laser Doppler vibrometer ref , and constructs a gain correction model based on deviation analysis: , is the calibration gain coefficient, k is the temperature compensation factor, which is calculated in real time by adjacent temperature sensors, represents the signal DC offset, represents the original output signal of the sensor, is the high-frequency noise mean. The execution process adopts a three-step closed-loop strategy: first, a sinusoidal swept-frequency signal with an increasing amplitude in the frequency range of 5 Hz - 500 Hz is injected to establish the sensor transfer function curve; second, the gain-frequency characteristic surface is fitted by the least squares method; finally, the correction coefficient k is written into the sensor transmitter register.

[0012] Preferably, the load step and high-frequency common-mode interference boost the equipment load step to 80% of the rated value. After the operation stabilizes, a second-stage perturbation is injected: a high-frequency common-mode interference signal with a frequency of 5 kHz and a peak-to-peak value of 50 Vpp is injected into the control loop for 120 seconds. Three operations are synchronously performed in the second stage: using a high-voltage differential probe to capture the transient voltage waveform of the motor terminal box terminals to identify the electromagnetic interference intrusion path; turning on a laser Doppler vibrometer to scan the rotor journal area to quantify the signal-to-noise ratio attenuation curve under high-frequency interference; injecting ISO 4406 grade 16 standard particles with a particle size distribution range of 15 - 25 μm into the lubricating oil path, and verifying the signal fidelity of the pollution condition through the correlation analysis of an electrostatic sensor and a high-speed microcamera. When the measurement error of the current harmonic sensor at the 5 kHz frequency point exceeds ±3%, the anti-aliasing filter reconstruction algorithm is automatically activated.

[0013] Preferably, the anti-aliasing filter reconstruction algorithm adopts a dual-path adaptive filtering architecture: the main path retains the original signal x(n), and the auxiliary path generates the reconstructed signal y(n). The weight vector is dynamically adjusted through an LMS filter: , where M is the filter order with a default value of 128, μ is the convergence factor with a dynamic range of 0.01 - 0.001, e(n) is the error signal, the reference source is the output of an optical fiber current sensor, and the reconstruction process is divided into three stages: in the initialization stage, an out-of-band attenuation model is established with a stopband attenuation > 80 dB, and in the convergence stage, the stopband energy ratio is monitored in real time , is the stopband energy, is the passband energy, and the target value < 10 -4 . In the stable stage, the optimal weight vector is locked, and the final output signal is resampled by 128 times of interpolation to ensure that the Nyquist frequency is extended to four times the original sampling rate, completely eliminating the spectrum aliasing phenomenon.

[0014] Preferably, the transient shock and dynamic stiffness modulation inject a third-stage perturbation under the condition of maintaining 80% load: a current spike with a peak value of 200 A and a pulse width of 100 μs is generated through a preset short-circuit device, and a pneumatic shaker is synchronously triggered to apply a 3 ms half-sine mechanical shock with an acceleration amplitude of 15 times the acceleration of gravity. A high-speed acquisition card with a sampling rate of 100 MS / s is used to capture the sensor overload recovery characteristics, and it is required that the time for the signal to return to the normal fluctuation band does not exceed 50 ms, where the normal fluctuation band is defined as the range of ±2 times the standard deviation. At the same time, the preload of the hydraulic release base is reduced by 50%, so that the natural frequency drops from 500 Hz to 200 Hz. The time taken for the hydraulic release base to reduce the preload is 3 minutes. The displacement of the equipment foundation is monitored in real time through a laser tracker to construct a vibration transfer function matrix: , where represents the new transfer function, represents the bearing vibration spectrum, denotes the basic vibration spectrum, where \(J\) is the imaginary unit and the basis element of the negative number domain. The physical meaning of \(-J\) is phase lag. is the phase angle. When the amplitude change of the transfer function exceeds 20 dB or the phase shift exceeds 30°, it is determined that the sensor installation coupling fails and the recalibration process is started.

[0015] Preferably, the recalibration process first deploys a laser tracker to establish a reference grid with a grid spacing of 50 mm in the equipment base coordinate system, and then performs three-dimensional space mapping, including static calibration, dynamic tracking, and coordinate transformation. The static calibration drives the retroreflective target ball through the robotic arm to traverse each sensor installation point and records the position matrix , , , represents the coordinate values of the sensor in the X, Y, and Z axes of the base coordinate system. The dynamic tracking captures the vibration trajectory \(Q(t)\) of the sensor housing at 2000 frames per second through a CCD high-speed camera under 80% load conditions. The coordinate transformation solves the homogeneous transformation matrix , where \(R\) is the rotation matrix and \(t\) is the translation vector. After the transformation, a spatial compensation vector is generated and written into the spatial coordinate synchronization unit, \(\delta\) x , \(\delta\) y , \(\delta\) z is the linear displacement compensation amount, \(\delta\) θx , \(\delta\) θy , \(\delta\) θz are the rotation compensation angles around the X / Y / Z axes.

[0016] Preferably, for the surge boundary limit test, the outlet valve is closed to the critical opening range of 65% - 70% to create a progressive surge condition: the pressure relief valve is periodically adjusted at a frequency of 0.2 Hz, and its opening changes stepwise between 30% - 80%. The propagation speed of the flow channel pulsation wavefront is captured by a dynamic pressure sensor. The turning point of the axial and radial energy ratio is analyzed using a vibration accelerometer. The temperature entropy value of the inlet guide vane area is calculated based on the infrared thermal imager data: , where \(S\) T is the temperature entropy, and \(p\) i represents the distribution probability of each temperature interval. The fusion system needs to trigger a surge warning when the mutation growth rate of the temperature entropy exceeds 0.8 bits / s and the pressure pulsation amplitude exceeds 25% of the rated value, and at the same time verify that the time alignment error of multiple sensors is less than 100 μs.

[0017] Preferably, the data stream processing of S2 is based on the acquisition of the original signals of the sensor array. All data are timestamped based on the IEEE 1588 Precision Time Protocol. The original signals first enter a circular buffer for temporary storage. The buffer depth is set to twice the maximum expected delay. Subsequently, the spatio-temporal normalization processing is performed by a resampling engine: the sampling rates of each sensor are unified to the 100 kHz benchmark through the cubic spline interpolation algorithm, and at the same time, the microsecond-level time delay caused by the difference in the length of the transmission cable is compensated. In the coordinate system conversion stage, the laser tracker scans the key reference points of the device every thirty seconds to construct the spatial transformation matrix in real time: , where Δx, Δy, and Δz are the position offsets, and θ is the installation deflection angle. The data after conversion form a spatio-temporal synchronization cube and are input into the fusion core for multi-dimensional analysis.

[0018] Preferably, the core of the real-time analysis in S2 is the residual monitoring engine, which calculates the multi-source residual matrix every 200 milliseconds: , where i represents the sensor number and j corresponds to the physical model type. represents the real-time measurement value of the i-th sensor. represents the predicted value of the j-th type of physical model. The vibration residual analysis uses the Campbell diagram comparison technology to match the measured acceleration spectrum with the critical speed trajectory output by the rotor dynamics model; the temperature residual generates the theoretical temperature field by solving the three-dimensional heat conduction equation , where k is the thermal conductivity, T is the temperature, ρ is the density, c p is the specific heat capacity, t is the time, is the gradient operator, is the partial derivative symbol. When the kurtosis value of the residual exceeds 3.0 or five consecutive sampling points exceed the ±3σ boundary, an anomaly flag is triggered in real time and the root cause analysis thread is started.

[0019] Preferably, the S2 uses the frequency domain coherence monitoring as an auxiliary verification means to track the transfer characteristics of current and vibration: , is the cross-power spectral density of current and vibration. and are the auto-power spectra respectively. In the characteristic frequency band, when the coherence coefficient is lower than 0.7 and lasts for more than three seconds, it is determined that the sensor is out of calibration.

[0020] Preferably, the pipeline architecture is adopted in the processing process to ensure real-time performance: the first-level FPGA completes the timestamp alignment and filtering, the second-level GPU cluster executes the model calculation and residual analysis, and the third-level CPU realizes the logical decision-making.

[0021] Preferably, the multi-level logical judgment of S3 starts from the input of raw data. First, single-sensor self-check is performed: preliminary screening is achieved by analyzing that the standard deviation of the signal baseline stability is < 0.1% of the full scale, the signal-to-noise ratio is > 60 dB, and the power supply ripple is < 5 mV for 10 consecutive minutes. The sensors passing the self-check enter the physical consistency check stage. At this time, the system calls the preset physical model library and compares the measured values of the sensors with the predicted values of the model in real time: the vibration signal needs to satisfy , where represents the frequency-domain amplitude of the measured vibration acceleration signal, and σ represents the standard deviation, where is the theoretical value calculated through the transfer function and the electromagnetic force , F em represents the electromagnetic excitation force, k t represents the motor torque constant, r represents the equivalent radius of the rotor, represents the current harmonic component, and the temperature distribution needs to conform to the numerical solution boundary of the heat conduction equation .

[0022] Preferably, when the physical consistency check fails, the system automatically activates the model parameter correction module, updates the transfer function or the material thermal conductivity k based on the residual gradient descent method. The data passing the check enters the multi-sensor cross-verification link, and the complex coherence function of the current harmonic and the vibration spectrum is cross-compared : If and the vibration residual exceeds the standard, it is determined that the sensor has drifted. If and the vibration residual exceeds the standard, the device protection mechanism is triggered.

[0023] Preferably, the sensors that fail the cross-verification will start the fourth-level anti-interference test verification: reproduce the step disturbance under controlled working conditions. If the fault characteristics reappear, it is confirmed that the sensor fails, and the data is automatically isolated and the replacement instruction is triggered; if not, it is determined as a false alarm, and the confidence weight W i (t) of the sensor is adjusted. The confidence management adopts a dynamic weighting model: , R recent is the residual compliance rate in the last five minutes, and R history is calculated based on the exponential decay function of the historical fault times. E env is the environmental severity factor. When the temperature > 80 °C, E env takes the value of 0.8. The data of the sensors with the final confidence level lower than 0.6 will be isolated.

[0024] Preferably, in the calibration execution stage of S3, the drift mode is first identified: for the bias error, compensation is performed, and μ residual represents the statistical mean of the residual sequence; for the gain drift, is used, and S rawIt represents the original sensing signal, and the k value is calculated by backpropagation from adjacent sensors; for non-linear distortion, a pre-trained neural network corrector is activated. Immediately after calibration, a 0.5 Hz sine sweep signal is injected to verify the phase consistency, and the 24-hour Allan variance monitoring is started. To ensure long-term stability, a requirement of < 1e-6 is imposed. It represents the duration of data grouping. It represents the arithmetic mean of the sensor output values within the k-th time window. It represents the arithmetic mean of the adjacent next time window. Finally, the theoretical value and the measured value of the bearing reaction force are compared through a laser Doppler vibrometer to complete the closed-loop verification.

[0025] The technical effects and advantages of the present invention are as follows:

[0026] 1. First, based on the stepped adversarial perturbation protocol and the online closed-loop correction mechanism, various potential errors can be excited and captured in real time under continuous operation, completely replacing the traditional shutdown calibration mode, and significantly improving the system availability and calibration timeliness.

[0027] 2. Second, through the multi-level perturbation scenario design and the multi-sensor cross-verification logic, a comprehensive defense system against strong coupling interference in the electro-mechanical-thermal multi-physical fields is constructed for the first time, effectively suppressing complex errors such as spectral aliasing caused by voltage fluctuations, overload distortion caused by mechanical shocks, and signal-to-noise ratio attenuation caused by oil contamination.

[0028] 3. Finally, by creatively integrating the spatial position calibration and the coupling error decoupling technology, spatial drift errors such as deformation of the installation base and loosening of mechanical connections are accurately diagnosed using laser grid positioning and transfer function analysis, and the autonomous reconstruction of the measurement reference is achieved through dynamic stiffness modulation, fundamentally ensuring the spatial consistency and long-term stability of multi-source data fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the overall structure of the present invention.

[0030] Figure 2 It is a schematic diagram of the stepped adversarial perturbation protocol structure of the present invention.

[0031] Figure 3 It is a schematic diagram of the data stream processing and real-time analysis structure of the present invention.

[0032] Figure 4 It is a schematic diagram of the multi-level logic judgment and calibration execution structure of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0033] Next, in combination with the accompanying drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0034] Reference Figures 1-4 An experimental calibration method based on multi-sensor signal fusion processing is shown as follows, including:

[0035] S1: Stepwise adversarial perturbation protocol: Design a multi-stage electromechanical-thermal coupling perturbation scenario to stimulate potential errors of the sensor.

[0036] In the S1, the first stage of the multi-stage electromechanical-thermal coupling perturbation scenario is the establishment of the basic working condition and the injection of low-frequency perturbation; the second stage is the load step and high-frequency common-mode interference; the third stage is the transient impact and dynamic stiffness modulation; the fourth stage is the surge boundary limit test.

[0037] For the establishment of the basic working condition and the injection of low-frequency perturbation, first, the equipment runs at the rated speed for ten minutes under the no-load state, and the sensor reference data of vibration, temperature, current, and oil fluid parameters are collected. Subsequently, the first-stage perturbation is injected: a ±5% 0.5-times power frequency voltage fluctuation is superimposed on the power input side through a programmable frequency converter, the fluctuation frequency is 25 Hz, and it lasts for five minutes. The first stage focuses on monitoring the change of the harmonic spectrum of the current transformer and the response delay of the low-frequency component of the bearing housing vibration. The low-frequency component is specifically the frequency band below 200 Hz. The temperature distribution uniformity on the surface of the stator core is synchronously recorded by an infrared thermal imager. When the detected vibration phase lag exceeds 15° or the temperature gradient sudden change rate exceeds 8 °C / m, the sensor gain calibration module is triggered.

[0038] The sensor gain calibration module first collects the output signal S of the target sensor under the reference working condition base , and at the same time obtains the real physical quantity P at the same position through a laser Doppler vibrometer ref , and constructs a gain correction model based on deviation analysis: , G cal is the calibration gain coefficient, k is the temperature compensation factor, which is calculated in real time by adjacent temperature sensors, μ offset represents the signal DC offset, S raw represents the original output signal of the sensor, μ noise is the high-frequency noise mean. The execution process adopts a three-step closed-loop strategy: first, a sinusoidal swept-frequency signal with an increasing amplitude is injected, the frequency range is 5 Hz - 500 Hz, and the sensor transfer function curve is established; second, the gain-frequency characteristic surface is fitted by the least squares method; finally, the correction coefficient k is written into the sensor transmitter register.

[0039] The load step and high-frequency common-mode interference increase the equipment load step to 80% of the rated value. After stable operation, a second-stage perturbation is injected: a high-frequency common-mode interference signal with a frequency of 5 kHz and a peak-to-peak value of 50 Vpp is injected into the control loop for 120 seconds. Three operations are synchronously performed in the second stage: using a high-voltage differential probe to capture the transient voltage waveform of the motor terminal box terminals to identify the electromagnetic interference intrusion path; turning on a laser Doppler vibrometer to scan the rotor journal area to quantify the signal-to-noise ratio attenuation curve under high-frequency interference; injecting ISO 4406 grade 16 standard particles with a particle size distribution range of 15 - 25 μm into the lubricating oil circuit, and verifying the signal fidelity of the pollution condition through the correlation analysis of an electrostatic sensor and a high-speed microcamera. When the measurement error of the current harmonic sensor at the 5 kHz frequency point exceeds ±3%, the anti-aliasing filter reconstruction algorithm is automatically activated.

[0040] The anti-aliasing filter reconstruction algorithm adopts a two-path adaptive filtering architecture: the main path retains the original signal x(n), and the auxiliary path generates the reconstructed signal y(n). The weight vector is dynamically adjusted through an LMS filter: , where M is the filter order with a default value of 128, μ is the convergence factor with a dynamic range of 0.01 - 0.001, e(n) is the error signal, and the reference source is the output of an optical fiber current sensor. The reconstruction process is divided into three stages: in the initialization stage, an out-of-band attenuation model is established with a stopband attenuation > 80 dB. In the convergence stage, the stopband energy ratio , E stopband is the stopband energy, E passband is the passband energy, and the target value < 10 -4 . In the stable stage, the optimal weight vector is locked, and the final output signal is interpolated and resampled by 128 times to ensure that the Nyquist frequency is extended to four times the original sampling rate, completely eliminating the spectral aliasing phenomenon.

[0041] The transient shock and dynamic stiffness modulation inject a third-stage perturbation under the condition of maintaining 80% load: a current spike with a peak value of 200 A and a pulse width of 100 μs is generated through a preset short-circuit device, and a pneumatic shaker is synchronously triggered to apply a 3 ms half-sine mechanical shock with an acceleration amplitude of 15 times the acceleration of gravity. A high-speed acquisition card with a sampling rate of 100 MS / s is used to capture the sensor overload recovery characteristics, and it is required that the time for the signal to return to the normal fluctuation band does not exceed 50 ms, where the normal fluctuation band is defined as the range of ±2 times the standard deviation. At the same time, the preload of the hydraulic release base is reduced by 50%, so that the natural frequency drops from 500 Hz to 200 Hz. The gradual change process of the preload of the hydraulic release base takes 3 minutes. The displacement of the equipment foundation is monitored in real time through a laser tracker to construct a vibration transfer function matrix: , where represents the new transfer function, Vib bearingRepresents the bearing vibration spectrum, Represents the foundation vibration spectrum. J is the imaginary unit and the basis element of the negative number field. The physical meaning of -J is phase lag, is the phase angle. When the amplitude change of the transfer function exceeds 20 dB or the phase shift exceeds 30°, it is determined that the sensor installation coupling fails and the recalibration process is started.

[0042] The recalibration process first deploys a laser tracker to establish a reference grid with a grid spacing of 50 mm in the equipment foundation coordinate system, and then performs three-dimensional space mapping, including static calibration, dynamic tracking, and coordinate transformation. The static calibration drives the reflective target ball through the robotic arm to traverse each sensor installation point and records the position matrix , x i , y i , z i represents the coordinate values of the sensor in the X, Y, and Z axes of the foundation coordinate system. The dynamic tracking captures the vibration trajectory Q(t) of the sensor housing at 2000 frames per second through a CCD high-speed camera under 80% load conditions. The coordinate transformation solves the homogeneous transformation matrix , where R is the rotation matrix and t is the translation vector. After the transformation, a space compensation vector is generated and written into the space coordinate synchronization unit, δ x a, δ y b, δ z cis the linear displacement compensation amount, δ θx α, δ θy β, δ θz γis the rotation compensation angle around the X / Y / Z axis.

[0043] For the surge boundary limit test, close the outlet valve to the critical opening range of 65% - 70% to create a progressive surge condition: periodically adjust the pressure relief valve at a frequency of 0.2 Hz, and its opening changes stepwise between 30% - 80%; capture the propagation speed of the flow channel pulsation wavefront through a dynamic pressure sensor; analyze the turning point of the axial and radial energy ratio using a vibration accelerometer; calculate the temperature entropy value of the inlet guide vane area based on the infrared thermal imager data: , where S T is the temperature entropy, p i represents the distribution probability of each temperature interval. The fusion system needs to trigger a surge warning when the mutation growth rate of the temperature entropy exceeds 0.8 bits / s and the pressure pulsation amplitude exceeds 25% of the rated value, and at the same time verify that the time alignment error of multiple sensors is less than 100 μs.

[0044] S2: Data stream processing and real-time analysis: Generate a dynamic reference through optical and material marking techniques to replace traditional manual calibration. The data stream unifies the sampling rate and coordinate system through a spatio-temporal alignment pipeline, calculates the deviation between multi-source signals and physical models through online residual analysis, and verifies the correlation strength between sensors through frequency domain coherence monitoring.

[0045] The data stream processing of S2 is based on the acquisition of the original signals of the sensor array. All data are timestamped based on the IEEE 1588 Precision Time Protocol. The original signals first enter a circular buffer for temporary storage, and the buffer depth is set to twice the maximum expected delay. Subsequently, the spatio-temporal normalization process is performed by a resampling engine: the sampling rates of each sensor are unified to the 100 kHz benchmark through the cubic spline interpolation algorithm, and at the same time, the microsecond-level time delay caused by the difference in the length of the transmission cables is compensated. In the coordinate system transformation stage, the laser tracker scans the key reference points of the device every thirty seconds to construct the spatial transformation matrix in real time: , where Δx, Δy, and Δz are the position offsets, and θ is the installation deflection angle. The data after transformation form a spatio-temporal synchronization cube and are input into the fusion core for multi-dimensional analysis.

[0046] The core of the real-time analysis in S2 is the residual monitoring engine, which calculates the multi-source residual matrix every 200 milliseconds: , where i represents the sensor number and j corresponds to the physical model type, represents the real-time measurement value of the ith sensor, represents the predicted value of the jth type of physical model. The vibration residual analysis uses the Campbell diagram comparison technique to match the measured acceleration spectrum with the critical speed trajectory output by the rotor dynamics model; the temperature residual generates the theoretical temperature field by solving the three-dimensional heat conduction equation , where k is the thermal conductivity, T is the temperature, ρ is the density, c p is the specific heat capacity, t is the time, is the gradient operator, is the partial derivative symbol. When the kurtosis value of the residual exceeds 3.0 or five consecutive sampling points exceed the ±3σ boundary, an anomaly flag is triggered in real time and the root cause analysis thread is started.

[0047] S2 uses frequency-domain coherence monitoring as an auxiliary verification means to track the transfer characteristics of current and vibration: , is the cross-power spectral density of current and vibration, and are the auto-power spectra respectively. In the characteristic frequency band, when the coherence coefficient is lower than 0.7 and lasts for more than three seconds, it is determined that the sensor is misaligned.

[0048] The processing process adopts a pipeline architecture to ensure real-time performance: the first-level FPGA completes timestamp alignment and filtering, the second-level GPU cluster performs model calculation and residual analysis, and the third-level CPU realizes logical decision-making.

[0049] S3: Multi-level logical judgment and calibration execution: Implement a four-level decision rule library and dynamic confidence weight management to diagnose the sensor status and trigger calibration.

[0050] The multi-level logical judgment of S3 starts from the input of raw data. First, it performs a single-sensor self-check: through analyzing that the standard deviation of the signal baseline stability is < 0.1% of the full scale, the signal-to-noise ratio > 60 dB, and the power supply ripple < 5 mV for 10 consecutive minutes for preliminary screening. The sensors passing the self-check enter the physical consistency check stage. At this time, the system calls the preset physical model library and compares the measured values of the sensors with the predicted values of the model in real time: the vibration signal needs to satisfy , represents the frequency-domain amplitude of the measured vibration acceleration signal, σ represents the standard deviation, where is the theoretical value calculated through the transfer function and the electromagnetic force The theoretical value calculated by em represents the electromagnetic excitation force, k t represents the motor torque constant, r represents the equivalent radius of the rotor, represents the current harmonic component, and the temperature distribution needs to conform to the numerical solution boundary of the heat conduction equation ;

[0051] When the physical consistency check fails, the system automatically activates the model parameter correction module and updates the transfer function or the material thermal conductivity k based on the residual gradient descent method. The data passing the check enters the multi-sensor cross-verification link, and cross-compares the complex coherence function of the current harmonic and the vibration spectrum : If and the vibration residual exceeds the standard, it is determined that the sensor drifts. If and the vibration residual exceeds the standard, the device protection mechanism is triggered;

[0052] The sensors that fail the cross-verification will start the fourth-level adversarial test verification: reproduce the step disturbance under controlled working conditions. If the fault characteristics reappear, it is confirmed that the sensor fails, and the data is automatically isolated and the replacement instruction is triggered; if not, it is determined as a false alarm, and the confidence weight of the sensor is adjusted , and the confidence management adopts a dynamic weighting model: , R recent is the residual compliance rate in the last five minutes, and R history is calculated based on the exponential decay function of the historical fault times. E env is the environmental severity factor. When the temperature > 80 °C, E env takes the value of 0.8. Finally, the sensor data with a final confidence level lower than 0.6 will be isolated;

[0053] In the calibration execution stage of S3, it first identifies the drift mode: for the bias error, it performs compensation, μ residual represents the statistical mean of the residual sequence; for the gain drift, it uses , S rawDenote the original sensing signal, and the k value is calculated by backtracking from adjacent sensors; for non-linear distortion, activate the pre-trained neural network corrector. Immediately after calibration, inject a 0.5 Hz sine sweep signal to verify the phase consistency, and start the 24-hour Allan variance monitoring. To ensure long-term stability, a requirement of < 1e-6 is imposed. Denote the duration of data grouping. Denote the arithmetic mean of the sensor output values within the k-th time window. Denote the arithmetic mean of the adjacent next time window. Finally, compare the theoretical value and the measured value of the bearing reaction force through a laser Doppler vibrometer to complete the closed-loop verification.

[0054] The present invention first executes a stepped adversarial perturbation protocol, and designs a multi-stage electromechanical-thermal coupling perturbation scenario including four stages to stimulate potential errors of sensors; the first stage is the establishment of the base condition and the injection of low-frequency perturbations. After running at the no-load rated speed, apply a specific voltage fluctuation, and mainly monitor the change of the current spectrum and the delay of the low-frequency vibration response. When a phase lag or a sudden change in the temperature gradient is detected, trigger the sensor gain calibration module for closed-loop correction; the second stage is the load step and high-frequency common-mode interference. After increasing the load, inject a high-frequency interference signal, and simultaneously execute electromagnetic interference path identification, signal-to-noise ratio attenuation quantification, and oil contamination condition verification. If the current measurement error exceeds the standard, activate the anti-aliasing filter reconstruction algorithm to use dual-path adaptive filtering to eliminate spectrum aliasing; the third stage is the transient shock and dynamic stiffness modulation. Under the condition of maintaining the load, inject current spikes and mechanical shocks, capture the overload recovery characteristics of the sensors and change the natural frequency of the system through hydraulic release. When the change of the transfer function exceeds the limit, determine that the installation coupling fails and start the re-calibration process. This process establishes a reference grid through a laser tracker and combines robotic arm positioning and high-speed camera dynamic tracking to complete spatial compensation; the fourth stage is the surge boundary limit test. Create a surge condition by periodically adjusting the valve, and use the fused data of a dynamic pressure sensor, a vibration accelerometer, and an infrared thermal imager to trigger an alarm; subsequently, perform data stream processing and real-time analysis. After aligning the time stamps of the original signals based on the precision time protocol, temporarily store them in a circular buffer, unify the sampling rate through a resampling engine and compensate for the transmission delay, and achieve spatio-temporal synchronization by combining the spatial transformation of the laser tracker. After inputting into the fusion core, the residual monitoring engine periodically calculates the multi-source residual matrix for anomaly detection, and at the same time uses frequency-domain coherence monitoring to verify the sensor correlation strength; finally, implement multi-level logical judgment and calibration execution, sequentially execute four-level decision-making of single-sensor self-check, physical consistency check, multi-sensor cross-verification, and adversarial test verification, isolate the low-confidence sensor data through dynamic confidence weight management, and in the calibration execution stage, after identifying the drift mode, perform offset compensation, gain correction, or neural network correction respectively. After calibration, verify the phase consistency through a sine sweep signal and start long-term stability monitoring. Finally, complete the closed-loop verification by comparing the theoretical value and the measured value through a laser Doppler vibrometer.

[0055] Secondly, in the accompanying drawings of the disclosed embodiments of the present invention, only the structures related to the disclosed embodiments are involved. For other structures, reference can be made to the general design. Without conflict, the same embodiment and different embodiments of the present invention can be combined with each other;

[0056] Finally, the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An experimental calibration method based on multi-sensor signal fusion processing, characterized in that, Including: S1: Step - type counter - measure disturbance protocol: Design multi - stage electromechanical - thermal coupling disturbance scenarios to stimulate potential errors in sensors; S2: Data stream processing and real - time analysis: Generate dynamic benchmarks through optical and material marking technologies to replace traditional manual calibration. The data stream passes through a spatio - temporal alignment pipeline to unify the sampling rate and coordinate system. Online residual analysis calculates the deviation between multi - source signals and physical models, and frequency - domain coherence monitoring verifies the correlation strength between sensors; S3: Multi - level logical judgment and calibration execution: Implement a four - level decision rule base and dynamic confidence weight management for diagnosing sensor status and triggering calibration.

2. The experimental calibration method based on multi-sensor signal fusion processing according to claim 1, characterized in that: In the multi - stage electromechanical - thermal coupling disturbance scenario in S1, the first stage is the establishment of the basic working condition and the injection of low - frequency disturbance; the second stage is load step - change and high - frequency common - mode interference; the third stage is transient shock and dynamic stiffness modulation; the fourth stage is the surge boundary limit test.

3. An experimental calibration method based on multi-sensor signal fusion processing according to claim 2, characterized in that: For the establishment of the basic working condition and the injection of low - frequency disturbance, first, the equipment runs at the rated speed for ten minutes under no - load conditions, and the sensor reference data of vibration, temperature, current, and oil - fluid parameters are collected. Subsequently, the first - stage disturbance is injected: A ±5% 0.5 - times power - frequency voltage fluctuation with a frequency of 25 Hz is superimposed on the power input side through a programmable frequency converter for five minutes. In the first stage, the change in the harmonic spectrum of the current transformer and the response delay of the low - frequency component of the bearing housing vibration are mainly monitored. The low - frequency component is specifically the frequency band below 200 Hz. The temperature distribution uniformity on the surface of the stator core by the infrared thermal imager is synchronously recorded. When the detected vibration phase lag exceeds 15° or the sudden change rate of the temperature gradient exceeds 8℃ / m, the sensor gain calibration module is triggered.

4. An experimental calibration method based on multi-sensor signal fusion processing according to claim 2, characterized in that: The load step and high-frequency common-mode interference increase the device load step to 80% of the rated value. After the operation stabilizes, a second-stage perturbation is injected: a high-frequency common-mode interference signal with a frequency of 5 kHz and a peak-to-peak value of is injected into the control loop for 120 seconds. Three operations are synchronously performed in the second stage: using a high-voltage differential probe to capture the transient voltage waveform of the motor terminal box terminals to identify the electromagnetic interference intrusion path; turning on a laser Doppler vibrometer to scan the rotor journal area to quantify the signal-to-noise ratio attenuation curve under high-frequency interference; injecting ISO 4406 grade 16 standard particles with a particle size distribution range of 15 - 25 μm into the lubricating oil path, and verifying the signal fidelity of the contamination condition through the correlation analysis of an electrostatic sensor and a high-speed microcamera. When the measurement error of the current harmonic sensor at the 5 kHz frequency point exceeds ±3%, the anti-aliasing filter reconstruction algorithm is automatically activated.

5. An experimental calibration method based on multi-sensor signal fusion processing according to claim 2, characterized in that: The transient shock and dynamic stiffness modulation inject a third-level disturbance under the condition of maintaining 80% load: a current spike with a peak value of 200 A and a pulse width of 100 μs is generated by presetting a short-circuit device, and a pneumatic shaker is synchronously triggered to apply a 3-ms half-sine mechanical shock with an acceleration amplitude of 15 times the acceleration of gravity. A high-speed acquisition card with a sampling rate of 100 MS / s is used to capture the overload recovery characteristics of the sensor. It is required that the time for the signal to recover to the normal fluctuation band does not exceed 50 ms, where the normal fluctuation band is defined as the range of ±2 times the standard deviation. At the same time, the pre-tightening force of the hydraulic release base is reduced by 50%, so that the natural frequency drops from 500 Hz to 200 Hz. It takes 3 minutes to release the pre-tightening force of the hydraulic release base. The displacement of the equipment foundation is monitored in real time by a laser tracker, and a vibration transfer function matrix is constructed: , where represents the new transfer function, Vib bearing represents the bearing vibration spectrum, Vib base represents the foundation vibration spectrum, J is the imaginary unit and is the basis element of the negative number domain. The physical meaning of -J is phase lag, is the phase angle. When the amplitude change of the transfer function exceeds 20 dB or the phase shift exceeds 30°, it is determined that the sensor installation coupling fails and the re-calibration process is started.

6. An experimental calibration method based on multi-sensor signal fusion processing according to claim 2, characterized in that: The surge boundary limit test closes the outlet valve to the critical opening range of 65% - 70% to create a progressive surge condition: the pressure relief valve is periodically adjusted at a frequency of 0.2 Hz, and its opening changes stepwise between 30% - 80%; the propagation speed of the flow channel pulsation wavefront is captured by a dynamic pressure sensor; the turning point of the axial and radial energy ratio is analyzed using a vibration accelerometer; the temperature entropy value of the inlet guide vane area is calculated based on the infrared thermal imager data: , where S T is the temperature entropy, and p i represents the distribution probability of each temperature interval. The fusion system needs to trigger a surge warning when the mutation growth rate of the temperature entropy exceeds 0.8 bits / s and the pressure pulsation amplitude exceeds 25% of the rated value, and at the same time verify that the time alignment error of the multi-sensor is less than 100 μs.

7. An experimental calibration method based on multi-sensor signal fusion processing according to claim 1, characterized in that: The data stream processing of S2 is based on the acquisition of the original signals of the sensor array. All data are timestamped based on the IEEE 1588 Precision Time Protocol. The original signals first enter the circular buffer for temporary storage. The buffer depth is set to twice the maximum expected delay. Subsequently, the spatio-temporal normalization process is performed by the resampling engine: the sampling rates of each sensor are unified to the 100 kHz benchmark through the cubic spline interpolation algorithm, and at the same time, the microsecond-level time delay caused by the difference in the length of the transmission cables is compensated. In the coordinate system transformation stage, the laser tracker scans the key reference points of the device every thirty seconds to construct the spatial transformation matrix in real time: , where Δx, Δy, and Δz are the position offsets, and θ is the installation deflection angle. The data after transformation form a spatio-temporal synchronization cube and are input into the fusion core for multi-dimensional analysis.

8. An experimental calibration method based on multi-sensor signal fusion processing according to claim 1, characterized in that: The core of the real-time analysis in S2 is the residual monitoring engine, which calculates the multi-source residual matrix every 200 milliseconds: , where i represents the sensor number and j corresponds to the physical model type, represents the real-time measurement value of the i-th sensor, represents the predicted value of the j-th type of physical model. Vibration residual analysis uses Campbell diagram comparison technology to match the measured acceleration spectrum with the critical speed trajectory output by the rotor dynamics model; the temperature residual generates the theoretical temperature field by solving the three-dimensional heat conduction equation , where k is the thermal conductivity, T is the temperature, ρ is the density, c p is the specific heat capacity, t is the time, is the gradient operator, is the partial derivative symbol. When the residual kurtosis value exceeds 3.0 or five consecutive sampling points exceed the ±3σ boundary, an anomaly flag is triggered in real-time and the root cause analysis thread is started.

9. An experimental calibration method based on multi - sensor signal fusion processing according to claim 1, wherein: The multi-level logic judgment of S3 starts from the input of raw data. First, a single-sensor self-check is performed: preliminary screening is achieved by analyzing that the standard deviation of the signal baseline stability is < 0.1% of the full scale, the signal-to-noise ratio > 60 dB, and the power supply ripple < 5 mV for 10 consecutive minutes. The sensors that pass the self-check enter the physical consistency check stage. At this time, the system calls the preset physical model library and compares the measured values of the sensors with the predicted values of the models in real time: the vibration signal needs to satisfy , represents the frequency-domain amplitude of the measured vibration acceleration signal, σ represents the standard deviation, where is the theoretical value calculated through the transfer function and the electromagnetic force , F em represents the electromagnetic excitation force, k t represents the motor torque constant, r represents the equivalent radius of the rotor, represents the current harmonic component, and the temperature distribution needs to conform to the numerical solution boundary of the heat conduction equation . When the physical consistency check fails, the system automatically activates the model parameter correction module and updates the transfer function based on the residual gradient descent method Or the material thermal conductivity k. The data obtained through the check enters the multi-sensor cross-verification process, and the complex coherence function of the current harmonic and the vibration spectrum is cross-compared If And the vibration residual exceeds the standard, it is determined that the sensor has drifted. If And the vibration residual exceeds the standard, the device protection mechanism is triggered; Sensors that fail the cross-verification will initiate the fourth-level countermeasure test for verification: reproduce the step disturbance under controlled conditions. If the fault characteristics reappear, it is confirmed that the sensor has failed, and the data is automatically isolated and a replacement instruction is triggered; if they do not reappear, it is determined as a false alarm, and the confidence weight of the sensor is adjusted. , and confidence management uses a dynamic weighting model: , R recent is the residual compliance rate in the last five minutes, and R history is calculated based on an exponentially decaying function of the historical number of faults, and E env is the environmental severity factor. When the temperature > 80 °C, E env takes a value of 0.8, and sensor data with a final confidence level lower than 0.6 will be isolated.

10. An experimental calibration method based on multi-sensor signal fusion processing according to claim 1, characterized in that: In the calibration execution stage of S3, the drift mode is first identified: for the bias error, compensation is performed, where μ residual represents the statistical mean of the residual sequence; for the gain drift, , S raw represents the original sensing signal, and the k value is calculated by backtracking from adjacent sensors; for the non-linear distortion, a pre-trained neural network corrector is activated. Immediately after calibration, a 0.5 Hz sine sweep signal is injected to verify the phase consistency, and a 24-hour Allan variance monitoring is started to ensure long-term stability, with the requirement < 1e-6, represents the duration of the data grouping, represents the arithmetic mean of the sensor output values within the k-th time window, represents the arithmetic mean of the adjacent next time window. Finally, the theoretical value and the measured value of the bearing reaction force are compared by a laser Doppler vibrometer to complete the closed-loop verification.

Images