Unit magnetic variable online monitoring method and system

Abstract

The invention relates to the technical field of unit detection, in particular to a unit magnetic variable on-line monitoring method and system, and the method comprises the steps: carrying out the multi-source data collection through a sensor, carrying out the correction of the multi-source data through a multi-dimensional calibration mechanism, synchronizing a timestamp through a dual-synchronization system, and carrying out the frequency-band-divided conditioning and standardization processing; environmental noise in the standardized multi-source data is eliminated through an intelligent algorithm, feature vectors are extracted, and purified feature vectors are obtained; screening effective abnormal features through an isolation forest algorithm; and inputting the effective abnormal features into an LSTM prediction model to obtain a corrected prediction value, substituting the corrected prediction value into a logistic regression formula to obtain a fault prediction probability of fault occurrence, calculating a health degree score, and performing graded early warning according to the health degree score. According to the scheme, through multi-dimensional calibration, working condition adaptive feature extraction and graded early warning, the unit monitoring data precision, the fault feature recognition accuracy and the operation and maintenance decision efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unit detection, in particular to a unit magnetic variable online monitoring method and system. BACKGROUND

[0002] In the unit operation state monitoring scene, industrial units often need to collect magnetic variable, temperature, vibration and other parameters through multiple types of sensors to realize state evaluation, but these sensors and collected data have significant heterogeneity problems, mainly reflected in the measurement accuracy difference of different brand sensors can reach more than ± 5%, and the multi-source data deviation of the same parameter can be up to 10%. Since the sensor performance directly determines the data reliability and feature effectiveness, the consistency and stability of heterogeneous data under the same monitoring conditions are greatly different, which leads to high difficulty in precise feature extraction. Temperature, vibration and other parameters to realize state evaluation, but these sensors and collected data have significant heterogeneity problems,

[0003] Traditional monitoring methods mostly use unified data processing flow, without considering the individual characteristic differences of sensors, which is easy to cause feature distortion due to too large error of part of data, and part of data cannot be effectively associated due to uncalibration. At the same time, the working environment of the unit is complex, and factors such as environmental temperature and humidity fluctuation, different position electromagnetic interference intensity difference (such as different interference levels near the motor area and control cabinet area), and equipment vibration, power grid fluctuation will introduce signal noise and parameter measurement deviation, further reducing the data quality.

[0004] In addition, the existing system lacks real-time adaptive ability to dynamic working conditions of the unit, cannot dynamically compensate the feature offset phenomenon caused by load change and temperature drift, and feature extraction mostly depends on fixed model parameters, the response to working condition difference is lagging behind, it is difficult to realize precise extraction of 128-dimensional feature vector and abnormal identification under complex working conditions, which seriously affects the accuracy of unit fault early warning and the timeliness of operation and maintenance decision. SUMMARY

[0005] The present application improves the unit monitoring data precision, fault feature recognition accuracy and operation and maintenance decision efficiency through multi-dimensional calibration, working condition adaptive feature extraction and hierarchical early warning, and guarantees the equipment operation safety and production efficiency.

[0006] The technical scheme provided by the present application is: a unit magnetic variable online monitoring method, the method comprises: Multi-source data acquisition is performed through sensors, multi-dimensional calibration mechanism is used to correct the multi-source data, double synchronization system is used to synchronize the time stamps of the corrected multi-source data, and the synchronized multi-source data is processed by frequency division and standardized. The environmental noise in the standardized processing multi-source data is eliminated by an intelligent algorithm, the feature vector is extracted by dynamically adjusting the long short-term memory network according to different load conditions, the oil quality permeability in the feature vector is corrected by the neural network combining the oil dielectric loss angle and the oil temperature, the magnetic induction intensity in the feature vector is corrected according to the relative permeability of the core and the vibration signal, and the purified feature vector is obtained; The missing data in the purified feature vector is filled in the state mutation scene by using the extended Kalman filter, and the effective abnormal features are screened by using the isolated forest algorithm. The effective abnormal features are input into the LSTM prediction model to obtain the initial prediction value, the initial prediction value is corrected twice by combining the temperature and vibration parameters to obtain the corrected prediction value, the corrected prediction value is substituted into the logistic regression formula to obtain the fault prediction probability of the fault occurrence, the health score is calculated based on the fault prediction probability and the feature deviation degree, and the health score is used for grading early warning.

[0007] Preferably, the specific process of the multi-dimensional calibration mechanism correction is as follows: The automatic zero point calibration is triggered during the unit shutdown period in the early morning, a zero magnetic field environment is generated by short-circuiting the calibration coil, the zero point offset of each sensor is recorded, and the calibrated magnetic field strength is obtained through a real-time compensation formula; a standard magnetic field source is used for three-point calibration every quarter, and the correction coefficient is updated by linear fitting; a temperature sensing branch is added to the signal link of all sensors to realize real-time temperature drift compensation through a hardware circuit; a full-pass filter is used to compensate the phase delay of the integrator, and real-time phase calibration is realized through digital signal processing.

[0008] Preferably, the specific content of the frequency division regulation and the standardization processing is as follows: The analog signals output by the Hall sensor and the fluxgate sensor are gain amplified by the chopper stabilized zero instrument amplifier in the low frequency band, and an 8th order Butterworth low-pass filter is configured to suppress high frequency noise aliasing; the signal converted by the integrator through the wideband amplifier is gain amplified in the high frequency band, a new 2nd order active high-pass filter is added to suppress low frequency noise, and self-adaptive gain control is realized through a digital signal processor to dynamically expand the signal range; a 16-bit multi-channel AD converter is triggered by a synchronous clock generated by the FPGA to synchronously sample the magnetic field signals and the working condition parameters; a GPS time module is used to provide a unified time reference for all sensor nodes, and a 32-bit accurate time stamp is attached to each frame of data.

[0009] Preferably, the specific process of the intelligent algorithm for eliminating noise is as follows: The standardized multi-source data is input into the VAE model, the model parameters are optimized by combining the reconstruction loss and the KL divergence, the model parameter quantity and the inference time are reduced by using the depth separable convolution and the knowledge distillation technology; the wavelet threshold denoising is used for the high-frequency signal, the noise standard deviation is estimated by the median absolute deviation of the wavelet high-frequency coefficients, the adaptive threshold is calculated, the soft threshold truncation processing is performed on the wavelet coefficients, and the high-frequency noise suppression ratio is additionally improved; the pure signal after denoising is synchronously associated with the load coefficient, the rotating speed and the vibration, and is transmitted to the working condition adaptive feature extraction module.

[0010] Preferably, the specific process of the feature vector acquisition is as follows: Based on the denoised multi-source data, the convolution kernel size of the long short-term memory network is dynamically adjusted based on the load coefficient, the key time node feature attention is strengthened through the attention mechanism of the network, and an initial feature sequence containing the time domain, frequency domain and time correlation features of the magnetic field is extracted.

[0011] Preferably, the specific process of the feature vector purification is as follows: The oil medium loss angle collected by the fiber grating sensor and the oil temperature collected by the temperature sensor are input into the magnetic permeability correction model, and the corrected magnetic permeability is output, the magnetic field parameters in the feature sequence affected by the change of oil quality characteristics are compensated; the relative magnetic permeability of the iron core monitored by the magnetic resistance sensor and the vibration signal collected by the vibration sensor are combined, the magnetic induction intensity related features are adjusted through the comprehensive correction formula, the measurement error caused by the change of the magnetic permeability of the iron core and the magnetostriction is suppressed, and the 128-dimensional purified feature vector is obtained.

[0012] Preferably, the effective abnormal feature screening process is as follows: The median filter is used for processing non-normal distribution data of the purified feature vector, and the pulse noise interference is eliminated; for the missing data in the state mutation scene, the dynamic filling is performed through the extended Kalman filter, and the data continuity is ensured; the isolated forest algorithm is applied for abnormal detection, the feature abnormal score is calculated by constructing multiple isolated trees, the pseudo-abnormal features caused by sensor fluctuation and environmental interference are removed, and the real fault correlation features are reserved; combined with the incremental learning mechanism, the model parameters are fine-tuned by the elastic weight consolidation algorithm based on the newly collected fault samples, the noise fingerprint library is expanded, the stability and accuracy of the abnormal feature recognition under different working conditions are improved, and finally the effective abnormal features are output.

[0013] Preferably, the specific process of the hierarchical early warning is as follows: Based on the fault prediction probability and the feature deviation degree, a score is calculated through a health degree scoring formula, and the value range is 0-100 points;According to the health degree score, three levels of early warning thresholds are set, greater than or equal to 80 points and less than 90 points are the first level of early warning corresponding to slight abnormality, prompting to strengthen monitoring;Greater than or equal to 60 points and less than 80 points are the second level of early warning corresponding to moderate abnormality, triggering special inspection;Less than 60 points is the third level of early warning corresponding to serious abnormality, which needs to be stopped for emergency repair;Different levels of early warning signals are synchronously pushed to the operation and maintenance platform, with fault positioning information and historical data trend chart;The first level of early warning is pushed through APP to remind, the second level of early warning is linked to sound and light alarm and generates an inspection work order, and the third level of early warning automatically triggers the emergency shutdown process, and notifies the operation and maintenance team to handle on site.

[0014] The application also provides a unit magnetic variable online monitoring system, which is used for executing the unit magnetic variable online monitoring method.

[0015] The application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the unit magnetic variable online monitoring method.

[0016] The application has the following beneficial effects: 1、Through the multi-dimensional calibration mechanism composed of automatic zero point calibration, periodic precision calibration and temperature compensation, the errors such as sensor zero drift and temperature drift are systematically eliminated, and the intelligent denoising processing of VAE model and wavelet threshold algorithm is combined to significantly reduce signal noise interference. This makes the long-term measurement error of key parameters such as magnetic field and vibration controlled below 0.3%, and the weak fault feature retention rate improved to above 95%, providing a high-fidelity original data basis for subsequent feature extraction and fault prediction, and solving the misjudgment problem caused by insufficient data precision in traditional monitoring. The multi-dimensional calibration mechanism composed of temperature compensation systematically eliminates errors such as sensor zero drift and temperature drift, and the intelligent denoising processing of VAE model and wavelet threshold algorithm significantly reduces signal noise interference. This makes the long-term measurement error of key parameters such as magnetic field and vibration controlled below 0.3%, and the weak fault feature retention rate improved to 95%.

[0017] 2、The convolution kernel size is dynamically adjusted according to different load conditions, the time correlation features are extracted through the LSTM network, and the feature vectors are corrected in combination with multiple parameters such as oil medium loss angle, core magnetic conductivity and vibration signal, so as to effectively eliminate the interference caused by oil quality, core state change and vibration. The 128-dimensional feature vector after purification can accurately reflect the real running state of the unit, and the pseudo-abnormal feature elimination rate is improved to above 90% by cooperating with the multi-unit decoupling and abnormal detection algorithm, which greatly improves the stability and accuracy of fault feature recognition and reduces the occurrence of invalid early warning.

[0018] 3. Based on the fault prediction probability and the feature deviation degree, the health score is calculated, a three-level early warning system is established, and a differentiated response strategy is matched, realizing full-scene coverage from slight abnormalities to serious faults. The early warning of the first level, the special inspection trigger of the second level, and the emergency shutdown disposal of the third level form a progressive operation and maintenance closed loop. In actual application, the average early warning time of the fault is 72 hours, and the fault positioning accuracy is controlled within 0.5m, which not only reduces the risk of unit sudden failure, but also reduces the unplanned downtime, improves the operation and maintenance efficiency and the safety of equipment operation. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A flowchart of a unit magnetic variable online monitoring method of the application; Figure 2 A monitoring process flowchart of a unit magnetic variable online monitoring method of the application. DETAILED DESCRIPTION

[0020] The following description is used to disclose the application so that those skilled in the art can implement the application. The preferred embodiments in the following description are only used as examples, and other obvious modifications can be thought of by those skilled in the art. The basic principles of the application defined in the following description can be applied to other embodiments, modifications, improvements, equivalents and other technical solutions without departing from the spirit and scope of the application.

[0021] It can be understood that the term "one" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of one element can be one, and in another embodiment, the number of the element can be multiple, and the term "one" cannot be understood as a limitation on the number.

[0022] As shown in Figure 1 and Figure 2 , in order to realize the all-round monitoring of the magnetic variable and the running state of the unit, first, the original data acquisition is completed through a high-precision sensor array.

[0023] The radial magnetic field component of the core is collected by the Hall sensor and the axial component . The Hall sensor is arranged equidistantly along the circumference of the stator core of the unit (2 at each end of the phase winding). The sensitivity is set to (0.1 millivolt voltage output corresponding to each millitesla magnetic field change), the linear error is (the maximum deviation ratio of the measured value to the true value within the full range), the response time is (from the magnetic field change to the stable output), the temperature drift coefficient (in the working range of-20-85℃, the output drift caused by the change of 1℃ of temperature), the zero point drift Output drift (output drift without magnetic field input within 24 hours), measured object is magnetic induction intensity (unit: T).

[0024] The high-frequency magnetic field radiation signal of the winding is collected by the Rogowski coil sensor. The Rogowski coil sensor is installed outside the busbar of the outgoing line end of the unit, and is arranged in a ring shape in two (orthogonal) directions. The bandwidth is 1 kHz-1 MHz (effective measurement frequency range), the sensitivity is 0.5 mV / A (induced voltage output per ampere of current), the integrator bandwidth is 1 kHz-1 MHz (-3 dB) (the degree of attenuation of the integrator to different frequency signals, -3 dB is the cut-off point of 3 decibel attenuation), the phase error is ≤5° (1 kHz-100 kHz) (in the frequency range of 1 kHz to 100 kHz, the phase difference value of the measured signal and the true signal), and the measured object is high-frequency magnetic induction intensity .

[0025] The external leakage magnetic field distribution is collected by the fluxgate sensor. The fluxgate sensor is installed at the top center and the bottom four corners of the unit shell, with a total of 5 measuring points. The measurement range is ±500 (the minimum and maximum values of the measurable magnetic induction intensity), the resolution is 0.1 (the minimum change in magnetic induction intensity that can be recognized), the long-term stability (the drift range of the measured value per hour), the temperature coefficient is ≤0.5 / ℃ (the change in the measured value caused by a change of 1℃ in temperature), the annual drift rate is ≤0.1% FS / year (the drift proportion relative to the full scale per year), the applicable frequency band is DC-1 kHz (suitable measurement frequency range, DC is direct current), and the measured object is steady-state magnetic induction intensity .

[0026] The real-time current of the unit is collected by the current sensor . The current sensor can be a core-penetrating CT (current transformer) with a range of 0-1000 A (measurable current range) and an accuracy of 0.2 level.

[0027] The oil temperature and the winding temperature are collected by the temperature sensor. The temperature sensor can be a platinum resistance PT100 with a measurement range of -20-150℃ and an accuracy of ±0.5℃. It is arranged at the bottom of the oil tank (to collect the oil temperature , unit: ℃) and the end of the winding (to collect the winding temperature , unit: ℃).

[0028] The real-time speed is collected by the speed sensor The rotating speed sensor can adopt a magneto-rotating speed probe, the measurement range is 0-3000 r / min (the measurable rotating speed range), the precision is ±1 r / min, and the rotating speed sensor is installed at the non-driving end of the rotating shaft of the unit.

[0029] The vibration signal is collected by a vibration sensor The vibration sensor can adopt a piezoelectric acceleration sensor, the range is 0-50 g (the measurable acceleration range, g is the gravity acceleration), the frequency range is 10 Hz-10 kHz, and the vibration sensor is installed on the outer wall of the iron core.

[0030] The oil dielectric loss angle is monitored by a fiber grating sensor (the dimensionless parameter, reflecting the insulation performance of the transformer oil). The measurement precision is ±0.1%.

[0031] The relative permeability of the iron core is monitored by a magnetic resistance sensor (the dimensionless parameter, the ratio of the permeability of the iron core to the permeability of vacuum). The long-term stability is ±0.5% / year.

[0032] According to the real-time current collected by the current sensor , the load coefficient is calculated by the working condition parameter collection formula (0.2≤ ≤2.0), which is used to divide the light load (0.2≤ ≤0.3), full load (0.3≤ ≤0.8) and overload (0.8<0.2≤ , 0.3≤ ≤0.8, 0.8<0.2≤ ).

[0033] Through the cooperative collection of multiple types of sensors, the space coverage rate of the magnetic field signal is improved to 92%, and the completeness of the working condition parameter collection is 100%.

[0034] Based on the raw data collected in the above steps, there are errors such as zero drift and temperature influence, which need to be improved through calibration mechanism to improve the data reliability.

[0035] Automatic zero point calibration: triggered during the unit shutdown period in the early morning every day, a zero magnetic field environment is generated by short-circuiting the calibration coil, the zero point offset of each sensor is recorded , and the real-time compensation formula is: (0.2≤ ≤2.0), which is used to divide the light load (0.2≤ ≤0.3), full load (0.3≤ ≤0.8) and overload (0.8<0.2≤

[0036] ). Regular precision calibration: every quarter, a standard magnetic field source (precision ±0.01%) is used for three-point calibration (0% FS, 50% FS, 100% FS, FS is the full scale of the sensor), the correction coefficient is updated by linear fitting, and the full scale measurement precision is ensured.

[0037] Temperature compensation circuit: All sensor signal links add PT100 temperature sensing branches, and real-time temperature drift compensation is realized through hardware circuit, with compensation accuracy ≥98%.

[0038] Through automatic zero-point calibration, regular precision calibration, and the addition of temperature compensation circuit, the long-term measurement error of the sensor is controlled below 0.3%, and the cumulative error caused by zero-point drift is reduced by 70%.

[0039] For the phase delay problem of the Rogowski coil signal, phase correction is performed based on the calibrated signal. The Rogowski coil signal link has a phase correction module, which uses an all-pass filter (cut-off frequency 1 MHz, phase error ≤2°) to compensate for the phase delay of the integrator. Real-time phase calibration is realized through digital signal processing: The frequency-phase curve fitting is obtained through factory calibration.

[0040] Through the phase correction module, the 1MHz signal phase error is reduced from 15° to 3°, and the partial discharge pulse positioning accuracy is improved by 40%.

[0041] To ensure the time consistency of multi-sensor data, a GPS and IEEE1588 dual synchronization system is constructed. When the GPS signal is normal, it is preferred to be time-synchronized (synchronization accuracy ≤100ns); when the GPS loses lock (no signal for 5s continuously), it automatically switches to the IEEE1588 precise time protocol (synchronization accuracy ≤1 μs), and the time synchronization between nodes is realized through an industrial Ethernet switch, with a synchronization accuracy of ≤1 μs; the local crystal oscillator (frequency stability 1ppm) is used as a transitional backup to ensure that there is no time interruption during the switching process.

[0042] Through the dual synchronization system, the probability of synchronization failure is reduced to 0.01% / day, and the timestamp deviation of multi-sensor is controlled to be within 1 μs.

[0043] Based on the unified time reference provided by the dual synchronization system, the vibration and magnetic field signals are time-aligned. The timestamp accuracy of the vibration sensor is improved to 10 μs, and the sampling frequency is dynamically adjusted (1kHz for low-frequency vibration and 10kHz for high-frequency impact). Through interpolation algorithm, the vibration signal is aligned with the magnetic field signal at sub-microsecond level. The interpolation algorithm is based on the unified time reference axis , and linear interpolation is used for vibration signals in different scenarios (low-frequency stable working condition: where, is the interpolated vibration acceleration, in g; , are the vibration accelerations of adjacent sampling points, in g; , ​For adjacent sampling point timestamps, units ns) and cubic spline interpolation (high-frequency impact conditions), ensure alignment error 50ns.

[0044] By dynamically adjusting the sampling frequency and interpolation optimization, the vibration and magnetic field signal phase alignment error is ≤50ns, and the multi-physical field collaborative analysis accuracy is improved to 99.5%.

[0045] According to the calibrated and synchronized signals, the signals are processed in different frequency bands and standardized to provide high-quality input for subsequent algorithms.

[0046] The analog signal (0-5V) output by the Hall / magnetic flux gate sensor is connected to the signal conditioning module, and the output signal of the Rogowski coil is converted into a voltage signal by the integrator and then connected; Low frequency band (DC-1kHz) uses chopping zero instrument amplifier (common mode rejection ratio ≥120dB, temperature drift ≤0.1μV / ℃) to realize 20dB gain ( , where, is the output voltage of the low frequency band, unit V; is the input voltage of the sensor, unit V), which is suitable for monitoring the steady-state magnetic field of the core; High frequency band (1kHz-50kHz) uses wideband amplifier to realize 40dB gain ( , where, is the output voltage of the high frequency band, unit V), which enhances the high frequency signal of partial discharge; The front end is configured with an 8th order Butterworth low pass filter, with a cutoff frequency (signal amplitude attenuation 3dB frequency), attenuation rate ≥80dB / dec (signal attenuation per ten times frequency), effectively suppresses high frequency noise aliasing; The output signal of the Rogowski coil is converted by a passive integrator (RC time constant 10 ) and a new 2nd order active high pass filter (cutoff frequency 1kHz) is added to suppress low frequency noise, and adaptive gain control (AGC) is realized through a digital signal processor (DSP). AGC detects ( is the signal peak value of the first frame, unit V; is the input signal sequence number), the gain coefficient (target amplitude ; is the smoothed signal amplitude, unit V), the maximum gain change rate is limited to ≤6dB / ms, and the dynamic range is expanded to 60dB (1mV-1V input signal); The signal phase spectrum is calculated by FFT transform, and point-by-point correction is performed based on the pre-stored phase compensation curve to ensure that the phase characteristic linearity in the 1kHz-1MHz frequency band is ≥98%. Generated by FPGA Synchronous clock (jitter ≤10 ns), trigger 16-bit multi-channel AD converter (conversion time ), realize the synchronous sampling of magnetic field signal and working condition parameters; Adopt GPS time service module (synchronization accuracy ≤100 ns) to provide unified time reference for all sensor nodes, and add 32-bit accurate time stamp (year+month+day+hour+minute+second+millisecond+microsecond) to each frame of data; Pack the sampling data in frame format, each frame contains time stamp (accuracy 1 ms), three-dimensional magnetic field components / / , current , temperature / , rotating speed , vibration , etc. 32 parameters, transmitted to edge computing nodes through industrial Ethernet (ring network redundancy design).

[0047] Inter-channel synchronization error ≤50 ns, data transmission packet loss rate ≤0.01%, single-channel sampling power consumption ≤0.5 W.

[0048] Adopt low-temperature drift amplifier to reduce thermal noise by 40%, high-frequency signal (1 kHz-1 MHz) distortion rate ≤1%, partial discharge pulse signal SNR improved by 12 dB; Signal synchronization error is controlled below 1 ms, time synchronization error between vibration and magnetic field signal ≤0.1 ms, multi-physical field superposition analysis accuracy improved to 99%.

[0049] Based on the standardized signals after conditioning, noise is eliminated and key features are extracted through intelligent algorithms. Through intelligent denoising algorithm, environmental noise and measurement interference are eliminated, and key fault features are retained, laying a foundation for precise feature extraction.

[0050] The preprocessed signal is first input into the variational autoencoder (VAE) denoising module, which adopts joint optimization of reconstruction loss and KL divergence: , wherein, is the loss function of variational autoencoder, is the expectation operator, is the log likelihood, is the KL divergence, is the approximate posterior distribution, is the prior distribution. The training data adds extreme working condition samples (oil temperature 80-100℃, overload The value range is ) to expand the distribution coverage. ​

[0051] In the VAE model, replace the traditional convolution with a depth separable convolution, reduce the encoder / decoder channel number from 64 / 128 / 256 to 32 / 64 / 128, introduce the knowledge distillation technology (parameter migration from teacher model to student model), reduce the model parameter by 60%, and compress the inference time to 30ms.

[0052] VAE multi-fault generalization optimization: the training set has 2000 groups of multi-fault mixed samples (such as having winding short circuit and core looseness at the same time, and having partial discharge and bearing wear at the same time), the encoder adopts a mixed structure of "depth separable convolution and 1×1 standard convolution" (specific levels: the first layer is 32 channels 3×3 depth separable convolution; the second layer is 32 channels 3×3 depth convolution + 64 channels 1×1 standard convolution; the third layer is 30% standard convolution + 70% depth convolution; the fourth layer is 96 channels 1×1 standard convolution + 128 channels 3×3 depth convolution), which retains 30% standard convolution to improve feature interaction capability, and the F1 score in multi-fault scene is improved to 0.92 and above.

[0053] Compared with traditional denoising algorithms, the signal-to-noise ratio is improved by 15dB, the weak fault feature retention rate is improved to 92%, and the "pseudo denoising" rate under extreme working conditions is reduced to 3%; the recall rate in multi-fault concurrent scene is improved from 75% to 90%, while maintaining performance, the GPU occupancy rate of the edge computing node is reduced from 75% to 35%.

[0054] Signal framing (N=1024 points / frame, overlap rate r=50%); the encoder extracts features through 4 layers of convolution, and the decoder reconstructs the denoised signal; use Adam optimizer (learning rate , iteration E=50 rounds); high-frequency signal (>5kHz) is supplemented by wavelet threshold denoising: is the original wavelet coefficient; is the sign function; is the adaptive threshold, , where, is the noise standard deviation, estimated by the absolute deviation of the median of the wavelet high-frequency coefficient , is the median absolute deviation); the output pure signal is synchronized with the working condition parameters , speed , vibration , and is transmitted to the working condition adaptive feature extraction module. Through this method, the high-frequency noise suppression is additionally improved by 10dB, and the denoising processing time is ≤50ms, meeting the real-time requirement.

[0055] Adjust the feature extraction strategy dynamically according to different load working conditions, solve the feature aliasing problem under complex working conditions, and accurately extract fault-related features. ​

[0056] Correlation-based load coefficient Dynamic adjustment of convolution kernel size: ( For the convolution kernel size; For the minimum value of 2.0), limit the maximum size to 7x7. The average accuracy of feature extraction under different working conditions is improved by 18%, effectively solving the feature aliasing problem and avoiding the convolution kernel size exceeding the limit. LSTM network for time series hidden state

[0057] Calculate attention weight: ( For the attention weight at the moment; For the energy function; For the sequence length), where ( , , For the weight matrix; For the bias term), combined with multiple source mutation signals Add bias.

[0058] LSTM sequence length optimization: based on information entropy analysis to determine the optimal sequence length: by calculating the feature information entropy ( For the feature probability distribution; For the information entropy, which measures the uncertainty of the feature), select The smallest stable window (256 points for light load, 384 points for heavy load), which reduces the original design by 20% of the calculation amount. Use a bidirectional LSTM structure (hidden layer dimension 128+128, number of layers 2) to reduce 30% of the parameters, and use TensorRT quantization acceleration (INT8 precision) to improve the inference speed by 2 times.

[0059] Adaptive attention bias: dynamic calculation of bias value: ( For the rated speed, unit r / min; For the attention bias term (dimensionless); For the current speed, unit r / min; For the speed at the previous moment, unit r / min; ), small capacity motor ( =1500r / min) and large capacity motor ( =3000r / min) to avoid bias in empirical values. This approach increases feature focus at key time nodes by 35%, reduces weight allocation errors caused by single sensor errors to 2%, achieves ≥95% retention of long-term dependent features, and improves computational efficiency by 15%. Model inference latency is ≤20ms, meeting the real-time matching requirement of 100 times / second, and increases fault feature recognition to 90% during dynamic transitions.

[0060] The output D = 128-dimensional feature vector contains the time domain, frequency domain and time correlation characteristics of the magnetic field, which is passed to the environmental interference elimination module for further purification.

[0061] Based on the extracted 128-dimensional feature vector, it is necessary to eliminate the impact of changes in the internal medium characteristics of the unit and external interference on the signal to ensure that the signal truly reflects the unit status.

[0062] Eliminate the impact of changes in the characteristics of the unit's internal media (such as transformer oil, iron core) on the magnetic field signal to ensure that the signal reflects the actual unit status.

[0063] Clarify the relationship between physical quantities: magnetic induction intensity ( is the magnetic induction intensity, unit is T; is the magnetic permeability, unit is H / m; is the magnetic intensity, unit is A / m; , is the relative magnetic permeability, is the vacuum permeability), the sensor directly measures ,pass Convert to obtain magnetic field strength .

[0064] Oil quality nonlinear compensation: Using BP neural network to establish magnetic permeability correction model ,in, is the corrected magnetic permeability, unit is H / m; The network input is the oil dielectric loss angle collected by the fiber Bragg grating sensor. and oil temperature , the output is the corrected permeability .

[0065] Neural network structure parameters: Oil quality compensation BP network adopts 2 inputs ( )-3 hidden layers (16,32,16 neurons)-1 output ( ) structure, the activation functions are ReLU (hidden layer) and Linear (output layer), the optimizer uses Adam (learning rate 5e-4), iterates for 200 rounds, and the validation set loss is ≤ 0.001.

[0066] BP network extrapolation constraint: add boundary limit in output layer: when 0.1 or 80℃, start linear extrapolation constraint (slope ≤ 0.02 / ℃), and trigger out-of-range warning. After testing, =0.2, the compensation error is controlled within 2.0%.

[0067] Thus, the network goodness-of-fit ( is the coefficient of determination, which measures the degree of model fitting), the nonlinear compensation error of oil quality is stable at 1.5%, the compensation accuracy is maintained at ≥98% under out-of-range conditions, the model failure is avoided, and the signal stability is significantly improved.

[0068] Synchronously monitor the relative permeability of the core through the magnetic resistance sensor (the reference value , combined with the vibration signal to establish a comprehensive correction formula ( is the corrected magnetic induction intensity, unit T; is the measured magnetic induction intensity, unit T; is the real-time relative permeability; is the reference relative permeability; is the real-time vibration acceleration, unit g; is the normal vibration reference value, unit g, obtained by statistical analysis of historical data), when (magnetic permeability deviation exceeds 10%) or (vibration deviation exceeds 20%) start correction. The measurement error of magnetic induction intensity caused by the change of core permeability is controlled within 1.2%, the magnetostriction influence suppression rate is 80%, and the physical quantity consistency is ensured.

[0069] The corrected signal has greatly eliminated internal medium interference, and the external unit interference is transmitted to the multi-unit magnetic interaction decoupling module for processing.

[0070] To separate the magnetic field interference of adjacent units and extract the true magnetic field characteristics of the target unit, solve the problem of multi-unit cooperative operation

[0071] An 8-element uniform linear array (spacing d=0.5m) is used, and the minimum description length criterion (MDL) is introduced to automatically estimate the number of interference sources : ( is the minimum description length value; is the probability of signal when a given number of interference sources; is the number of array elements, is the number of snapshots; is the number of interference sources). The spatial spectrum function of the MUSIC algorithm: (azimuth angle for spatial spectrum power; for azimuth angle; for array manifold vector; for noise subspace matrix; superscript for conjugate transpose), azimuth angle .

[0072] Fast spatial spectrum estimation by FFT replaces direct matrix inversion, reducing spatial spectrum calculation complexity from to ( for snapshots), sliding window accumulation (10 frames / cumulation) is introduced to reduce calculation frequency, and dimension reduction PCA (principal component retention rate 95%) is used for array element data preprocessing.

[0073] Array element position calibration mechanism: the array element spacing is calibrated by laser range finder (accuracy ±1mm) every month, a position error database is established, and an error compensation term is introduced in the MUSIC algorithm: ( for compensated array manifold vector; for original array manifold vector; for correction vector of spacing deviation d). The estimation accuracy of the number of interference sources reaches 98%, the single calculation time of the 8-element MUSIC algorithm is reduced from 50ms to 8ms, and the real-time performance is improved by 6 times in the multi-machine group scene; when the ground settlement causes a deviation of ±5cm, the positioning error is still controlled within 0.5m, and the anti-installation deviation ability is significantly enhanced.

[0074] Based on the reference signal generated by the interference source parameters, the variable step size LMS adaptive filter is used to update the coefficients ( for filter coefficients at the th moment; for variable step size factor; for error signal at the th moment; for input signal at the th moment), step size ( 0.01, =1e-6), to improve the convergence speed.

[0075] LMS convergence optimization: the step size formula is improved to (introducing momentum term 0.2), the convergence time is shortened from 20ms to 12ms in dynamic scenes. The same frequency interference suppression ratio is maintained at 35dB and above in dynamic scenes, the interference suppression ratio is maintained at 30dB and above when the load suddenly changes, and the purity of the target signal extraction is improved by 40%.

[0076] The purified target magnetic field feature vector has greatly eliminated internal and external interference, and the incoming signal verification module performs anomaly detection.

[0077] Based on the purified feature vector, the robust algorithm detects anomalies and updates the model to ensure reliable extraction of fault features. The dynamic threshold determines the abnormality of the features, eliminates false interference, and selects real fault features to provide reliable input for fault diagnosis.

[0078] Median filtering is used instead of 3σ criterion to process non-normal distribution data, and the window size is dynamically adjusted (3-7 points); Extended Kalman filter (EKF) is used instead of traditional Kalman filter to handle missing data filling in state mutation scenarios is the state transition function; is the state vector at the time; is the control input; is the state prediction increment; is the load nonlinear correction term, is the load coefficient); Pulse feature is retained: set pulse detection threshold (greater than 5 times the root mean square value), mark suspected partial discharge pulses and skip the elimination process.

[0079] Through this optimization method, the non-normal feature false elimination rate is reduced from more than 5% to less than 1%, the missing data filling accuracy is maintained at more than 92% in state mutation, and the data integrity is improved to 99.9%.

[0080] Incremental learning module is introduced, new collected normal working condition data (1000 samples) is used to fine-tune VAE and LSTM model parameters every week, and elastic weight consolidation (EWC) algorithm is used to protect key feature extraction capability is the total loss function; is the new data loss; is the regularization strength; is the parameter importance matrix; is the current parameter; is the old parameter reference value).

[0081] Incremental learning triggers optimization: clear KL divergence calculation rules: use the last 30 days of data, minimum sample size 2000 (if insufficient, extend the collection period), set double trigger conditions (KL divergence > 0.1 and continuous 3 days over limit), false trigger rate reduced to 1 / quarter and below. The performance of the model under un-covered working conditions is reduced by 30% to 8%; the triggering accuracy of new unit commissioning is improved to 95%, avoiding ineffective retraining. ​​

[0082] Construct a noise fingerprint library (including 20 typical noise patterns such as lightning and equipment start-stop), and use SVM classifier to identify signal type (fault / noise) in real time. Start enhanced noise reduction (wavelet threshold increased by 20%) for suspected noise signals.

[0083] Noise fingerprint library expansion mechanism: Establish a quarterly update mechanism, add 10 new noise samples (such as frequency converter harmonics and 5G base station interference), use incremental SVM algorithm (support online sample addition), and maintain noise recognition accuracy ≥97%.

[0084] Fault misjudgment rate under strong noise environment decreased from 15% to ≤3%; misjudgment rate of uncoated noise types decreased from 12% to ≤4%.

[0085] Based on effective abnormal features, through trend prediction and health assessment, realize hierarchical early warning and support operation and maintenance decision-making. Receive decoupled feature vectors, calculate dynamic threshold interval based on normal data in the past 3 months: Upper limit Lower limit Update mean and variance every day through sliding window N=1000), where is the upper threshold; is the lower threshold; is the mean; is the standard deviation.

[0086] Threshold adaptive update reduces false positive rate to 1% and false negative rate to 0.5%.

[0087] Label abnormal features that exceed the threshold, and remove interference residues through the Isolation Forest algorithm.

[0088] Isolation Forest parameter optimization: Determine the optimal parameters through grid search: subsample size s=512, tree number t=200, and abnormality detection F1 score improves to 0.93.

[0089] Through the Isolation Forest algorithm, the interference residue removal rate can reach 98%; the stability of interference residue removal is improved, and the accuracy of abnormal feature recognition reaches 99%, ensuring that the effective abnormal features are passed to the next link.

[0090] Effective abnormal features are transmitted to the dynamic working condition pattern matching module for fault type identification and trend prediction.

[0091] Identify fault types and assess development trends through feature matching and trend prediction to achieve early fault warning.

[0092] Extreme condition detection enhancement: Detect load sudden change through sliding window variance (window size 50ms) Change rate>0.5 / s) or temperature rise suddenly ( Change rate>5℃ / min), triggering the temporary reset mechanism of model parameters and loading the pre-trained parameters of the corresponding working conditions.

[0093] Improved parameter reset system: The pre-trained parameter library is stored by unit model (e.g. 8 categories such as 10kV / 200kW, 35kV / 1000kW, etc.), and contains parameters for 10 typical operating conditions such as no-load, full-load, and 120% overload. Trigger threshold is customized according to the unit: load sudden change threshold Rate of change = 0.3 + 0.2 × ( / 1000)( is the rated power (in kW), the temperature rise threshold Rate of change = 3 + 0.02 × ); Reset process: Parameter switching is completed within 50ms after triggering, and cached data transition is used during switching.

[0094] By resetting parameters, the false trigger rate of extreme working condition parameter reset is reduced to ≤0.5 times / month, and the model stability is improved by 40%.

[0095] Using the DTW algorithm with slope constraint ( and ) Calculate the similarity between real-time features and atlas samples: ,in, is the similarity ( Determine if the match is successful); is the dynamic time warping distance; is the characteristic sequence length; is the maximum value of the feature.

[0096] The DTW algorithm with slope constraints can improve the fault type identification accuracy to 95%, and the matching speed can reach 100 times / second or more.

[0097] The LSTM prediction model inputs the feature sequence of the past 30 days and uses the ReLU activation function and MAE loss function (target error ≤ 5%) predict future trends, where is the hidden layer output; 、 is the weight matrix; is bias; is the sequence length; is the true value; is the initial prediction value. The initial prediction value of the fault trend is obtained by calculating the output layer of the LSTM network ,in, is the output layer weight matrix; is the output layer bias term. Multi-factor trend correction: combine temperature and vibration parameters to make a secondary correction to the prediction result, the correction formula is: wherein, is the corrected prediction value; is the initial prediction value; is the temperature correction coefficient; is the vibration correction coefficient; temperature correction term , , , , ; vibration correction term , is the indicator function. The trend prediction error is further reduced to 4.5% and below under extreme working conditions.

[0098] Substitute the corrected prediction value into the logistic regression formula to obtain the failure prediction probability of failure occurrence , and the logistic regression formula is: wherein, is the failure prediction probability (dimensionless, value range 0-1), the value closer to 1 indicates the higher the risk of failure occurrence; is the failure trend prediction value at the time output by the LSTM model; , is the logistic regression coefficient (obtained by training historical failure data, is the weight matrix, is the bias term).

[0099] After 10 units for 6 months of actual measurement, the failure development trend prediction error is controlled at 5% and below, and the early warning time is averagely 72h.

[0100] The output failure type and probability are transmitted to the warning module to trigger the corresponding level of warning mechanism.

[0101] Based on the failure severity, a graded warning is issued to ensure that the operation and maintenance personnel respond in time and avoid the expansion of the failure.

[0102] Based on the failure prediction probability and the characteristic deviation degree ( ), the health score is calculated: wherein, is the health score (0-100 points); is the real-time feature; is the normal feature; , is the feature threshold upper and lower limit.

[0103] Health score dynamic weight: weight dynamically adjusted according to fault type: winding fault (feature deviation weight 15, probability weight 60), abnormal oil temperature (feature deviation weight 5, probability weight 30), through fault influence factor matrix Realize differentiated scoring.

[0104] The health state quantification accuracy rate reaches 97%, and by adjusting the denominator, negative scores are avoided, providing a scientific basis for grading early warning; the fault severity distinction accuracy rate is improved to 98%, and the rationality of early warning grading is significantly enhanced.

[0105] According to the score grading early warning: first level (80-90 points) push APP notification (slight abnormality, with feature deviation details and trend chart), second level (60-80 points) trigger sound and light alarm (potential fault, automatically generate preliminary maintenance suggestion), third level (<60 points) automatically cut off the control circuit (emergency fault, synchronously push fault positioning coordinates and emergency treatment plan), synchronously associated fault position positioning information (accuracy ≤0.5m).

[0106] The early warning response time is shortened to ≤3s, and through industrial field verification, the fault handling efficiency is improved by 40%, and the third level early warning accuracy rate reaches 100%.

[0107] For example, according to the raw data collected by the sensor array, first calculate the core working condition parameters. Taking a certain 35kV / 1000kW unit as an example, the key parameter collection results are as follows: the Hall sensor collects the radial magnetic field component of the core , the axial component , synchronously records the zero point offset axial component (daily automatic calibration to obtain); the current sensor collects the real-time current , the rated current is known , the load coefficient is calculated (determined as full load working condition); the temperature sensor collects the oil temperature Winding temperature ; the vibration sensor collects the vibration signal (normal vibration reference value , determined by historical data statistics); the magnetic resistance sensor collects the relative magnetic permeability of the core (reference ).

[0108] According to the raw data obtained by the signal collection steps, calibration and time alignment processing are carried out to eliminate system error. The Hall sensor has temperature drift characteristics (temperature drift coefficient 50ppm / ℃), based on the current oil temperature (Reference ambient temperature 25℃), calculate the temperature drift error: temperature drift error = 50ppm / ℃x(55-25)℃=1500ppm=0.15%, after PT100 temperature compensation circuit processing, the actual error is reduced to 0.03%.

[0109] Interpolation alignment of vibration signal based on unified time reference provided by GPS and IEEE1588 dual synchronization system. The interpolation result at , , , needs to be calculated , .

[0110] According to the calibrated and synchronized signal, the signal quality is improved by frequency division amplification, filtering and denoising. Low frequency band (DC-1kHz): the output voltage of Hall sensor , after processing by chopper stabilized instrument amplifier (20dB gain): . High frequency band (1kHz-50kHz): partial discharge signal collected by Rogowski coil , after processing by wideband amplifier (40dB gain): .

[0111] Supplement denoising for high frequency signal (>5kHz), signal frame length , median absolute deviation of wavelet high frequency coefficient , calculate noise standard deviation: , adaptive threshold: . After denoising for certain wavelet coefficient , .

[0112] According to the pure signal after conditioning, dynamically adjust the feature extraction parameters and eliminate the medium interference. Based on the load coefficient , optimize the convolution kernel size: (integral number), adapt to full load working condition feature extraction.

[0113] According to the magnetic reluctance sensor monitored and vibration signal , start comprehensive correction:

[0114] Magnetic permeability deviation: (no trigger magnetic permeability correction).

[0115] Vibration deviation: (trigger vibration correction).

[0116] Corrected magnetic induction intensity: .

[0117] According to the feature vector after interference elimination, fault trend prediction and health state scoring are performed. The LSTM model inputs the feature sequence of the last 30 days to obtain the corrected prediction value , and substitutes it into the logistic regression formula (known , ) .

[0118] The real-time value of a certain feature , the normal reference value , the upper threshold , and the lower threshold , (triggers a three-level early warning).

[0119] The processes described above with reference to the flowcharts can be implemented as computer software programs in accordance with embodiments of the present disclosure. Embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer readable medium, the computer program comprising program code for performing the methods illustrated by the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication section, and / or installed from a detachable medium. When the computer program is executed by a central processing unit (CPU), the above-described functions defined in the methods of the present application are performed. It should be noted that the computer readable medium of the present application can be a computer readable signal medium or a computer readable storage medium or any combination of the two. The computer readable storage medium may, for example, but not limited to, be an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any suitable combination of the above. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, the computer readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus or device. In the present application, the computer readable signal medium can include a data signal carried in a baseband or as part of a carrier wave, in which the computer readable program code is carried. Such a propagated data signal can take many forms, including but not limited to, an electromagnetic signal, an optical signal, or any suitable combination of the above. The computer readable signal medium can also be any computer readable medium that can send, propagate or transfer a program for use by or in connection with an instruction execution system, apparatus or device. The program code contained on the computer readable medium can be transmitted by any suitable medium, including but not limited to, wireless, wire, optical cable, RF, or any suitable combination of the above.

[0120] The computer program product of the present application can be a computer program product comprising a computer-readable medium bearing computer program code embodied therein for use with a computer. The computer program product can be propagated signal on a carrier wave carried by a transmission medium (for example, on a network or in a transmission medium). A propagated signal can be an example of a carrier wave, which can comprise one or more exemplifications of combinations of signals. The computer program product can or can not be encrypted. The computer program code can be downloaded from the Internet and / or used in conjunction with a software development program. The computer program product can be embodied in one or more computer- readable media, which can be non-transitory computer-readable media, or transitory computer- readable media (for example, electromagnetic signals).

[0121] Those skilled in the art will understand that the application described above and illustrated in the accompanying drawings is presented by way of example only and not limitation. It is therefore intended that the application's full patent rights and the equivalents thereof be determined by the following claims.

Claims

1. A method for online monitoring of magnetic variables of a unit, characterized in that: The method comprises: Multi-source data is collected through sensors, and the multi-source data is corrected using a multi-dimensional calibration mechanism. The corrected multi-source data timestamps are synchronized using a dual synchronization system, and the synchronized multi-source data is frequency-band conditioned and standardized. An intelligent algorithm eliminates environmental noise from standardized multi-source data, dynamically adjusts the long-short-term memory network to extract feature vectors for different load conditions, and uses a neural network to correct the oil permeability in the feature vector based on the oil dielectric loss angle and oil temperature. The magnetic induction intensity in the feature vector is corrected based on the relative permeability of the iron core and the vibration signal to obtain a purified feature vector. The extended Kalman filter is used to fill the missing data in the state mutation scenario for the purified feature vector, and the isolation forest algorithm is used to screen effective abnormal features; The effective abnormal features are input into the LSTM prediction model to obtain the initial prediction value. The initial prediction value is then corrected twice based on the temperature and vibration parameters to obtain the corrected prediction value. The corrected prediction value is substituted into the logistic regression formula to obtain the fault prediction probability of the fault. The health score is calculated based on the fault prediction probability and feature deviation, and a graded warning is issued according to the health score.

2. The method for online monitoring of magnetic variables of a unit according to claim 1, characterized in that: The specific process of multi-dimensional calibration mechanism correction is as follows: Automatic zero-point calibration is triggered during the unit shutdown period in the early morning every day. A zero magnetic field environment is generated by short-circuiting the calibration coil, recording the zero-point offset of each sensor, and obtaining the calibrated magnetic field strength through a real-time compensation formula. A three-point calibration is performed quarterly using a standard magnetic field source, and the correction coefficient is updated through linear fitting. A temperature sensing branch is added to all sensor signal chains, and real-time temperature drift compensation is achieved through hardware circuits. An all-pass filter is used to compensate for the integrator phase delay, and real-time phase calibration is achieved through digital signal processing.

3. The method for online monitoring of magnetic variables of a unit according to claim 2, characterized in that: The specific contents of the frequency band conditioning and the standardization processing are as follows: In the low-frequency band, a chopper-stabilized instrumentation amplifier is used to gain-amplify the analog signals output by the Hall sensor and fluxgate sensor, and an 8th-order Butterworth low-pass filter is configured to suppress high-frequency noise aliasing. In the high-frequency band, a broadband amplifier is used to gain-amplify the signal converted by the Rogowski coil through the integrator, and a new 2nd-order active high-pass filter is added to suppress low-frequency noise. Adaptive gain control is achieved through a digital signal processor to dynamically expand the signal range. The FPGA generates a synchronous clock to trigger the 16-bit multi-channel AD converter, which synchronously samples the magnetic field signal and operating parameters. The GPS timing module is used to provide a unified time base for all sensor nodes, and a 32-bit precise timestamp is attached to each frame of data.

4. The method for online monitoring of magnetic variables of a unit according to claim 3, characterized in that: The specific process of noise elimination by the intelligent algorithm is as follows: The standardized multi-source data is input into the VAE model, and the model parameters are jointly optimized using reconstruction loss and KL divergence. The number of model parameters and inference time are reduced through deep separable convolution and knowledge distillation technology. Wavelet threshold denoising is used for high-frequency signals. The noise standard deviation is estimated by the absolute deviation of the median of the wavelet high-frequency coefficients, and the adaptive threshold is calculated. The wavelet coefficients are soft-thresholded to further improve the high-frequency noise suppression ratio. The denoised pure signal is synchronously associated with the load factor, speed, and vibration, and transmitted to the working condition adaptive feature extraction module.

5. The method for online monitoring of magnetic variables of a unit according to claim 4, characterized in that: The specific process of obtaining the feature vector is as follows: Based on the denoised multi-source data and the load coefficient, the convolution kernel size of the long short-term memory network is dynamically adjusted. The attention mechanism of the network is used to strengthen the focus on key time node features, and the initial feature sequence containing the time domain, frequency domain and time correlation features of the magnetic field is extracted.

6. The method for online monitoring of magnetic variables of a unit according to claim 5, characterized in that: The specific process of feature vector purification is as follows: The oil dielectric loss angle acquired by the fiber Bragg grating sensor and the oil temperature acquired by the temperature sensor are input into the magnetic permeability correction model, and the corrected magnetic permeability is output to compensate for the magnetic field parameters in the feature sequence affected by changes in oil quality characteristics. Combined with the relative magnetic permeability of the iron core monitored by the magnetoresistive sensor and the vibration signal acquired by the vibration sensor, the magnetic induction intensity-related characteristics are adjusted through a comprehensive correction formula to suppress the measurement errors caused by changes in the iron core magnetic permeability and magnetostriction, and obtain a 128-dimensional purified feature vector.

7. The method for online monitoring of magnetic variables of a unit according to claim 6, characterized in that: The effective abnormal feature screening process is as follows: The purified feature vector is subjected to median filtering to process non-normally distributed data and eliminate impulse noise interference. For missing data in state mutation scenarios, the extended Kalman filter is used to dynamically fill in the missing data to ensure data continuity. The isolation forest algorithm is applied for anomaly detection. By constructing multiple isolated trees to calculate feature anomaly scores, pseudo-abnormal features caused by sensor fluctuations and environmental interference are eliminated, and true fault-related features are retained. Combined with the incremental learning mechanism, the elastic weight consolidation algorithm is used to fine-tune the model parameters based on newly collected fault samples, expand the noise fingerprint library, improve the stability and accuracy of abnormal feature recognition under different working conditions, and finally output effective abnormal features.

8. The method for online monitoring of magnetic variables of a unit according to claim 7, characterized in that: The specific process of the graded warning is as follows: Based on the fault prediction probability and feature deviation, a score is calculated using the health score formula, with a value range of 0-100. Three levels of warning thresholds are defined based on the health score: a score of 80 or more and less than 90 is a level one warning, corresponding to a minor anomaly, prompting enhanced monitoring; a score of 60 or more and less than 80 is a level two warning, corresponding to a moderate anomaly, triggering a special inspection. Less than 60 points corresponds to a third-level warning, which indicates a serious abnormality and requires emergency shutdown and maintenance; Different levels of warning signals are simultaneously pushed to the operation and maintenance platform, along with fault location information and historical data trend charts; The first-level warning is pushed through the APP to remind you, the second-level warning is linked to the sound and light alarm and generates an inspection work order, and the third-level warning automatically triggers the emergency shutdown process and notifies the operation and maintenance team to deal with it on site.

9. A unit magnetic variable online monitoring system, characterized in that: The system is used to execute the method for online monitoring of magnetic variables of a unit as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement the method for online monitoring of magnetic variables of a unit as described in any one of claims 1 to 8.

Images