A method and system for online monitoring of unit magnetic variables

Abstract

This invention relates to the field of unit monitoring technology, and particularly to a method and system for online monitoring of unit magnetic variables. The method includes: acquiring multi-source data through sensors; correcting the multi-source data using a multi-dimensional calibration mechanism; synchronizing timestamps using a dual-synchronization system; and performing frequency band conditioning and standardization processing. It also includes: eliminating environmental noise in the standardized multi-source data using intelligent algorithms; extracting feature vectors to obtain purified feature vectors; screening effective abnormal features using an isolated forest algorithm; inputting the effective abnormal features into an LSTM prediction model to obtain corrected predicted values; substituting these values ​​into a logistic regression formula to obtain the predicted probability of a fault occurrence; calculating a health score; and providing graded early warnings based on the health score. This solution improves the accuracy of unit monitoring data, the accuracy of fault feature identification, and the efficiency of operation and maintenance decision-making through multi-dimensional calibration, adaptive feature extraction based on operating conditions, and graded early warnings.

Description

Technical Field

[0001] This invention relates to the field of generator unit testing technology, and in particular to a method and system for online monitoring of generator unit magnetic variables. Background Technology

[0002] In the scenario of monitoring the operating status of industrial power units, it is often necessary to collect parameters such as magnetic variables, temperature, and vibration using multiple types of sensors to achieve condition assessment. However, these sensors and the collected data exhibit significant heterogeneity, primarily manifested in the measurement accuracy differences between different brands of sensors reaching ±5% or more, and the deviation of multi-source data for the same parameter reaching up to 10%. Since sensor performance directly determines data reliability and feature validity, the consistency and stability of heterogeneous data vary greatly under the same monitoring conditions, resulting in extremely high difficulty in accurate feature extraction.

[0003] Traditional monitoring methods often employ a uniform data processing workflow, failing to consider the individual characteristics of sensors. This can easily lead to issues such as distorted features due to excessive errors in some data and ineffective correlation of other data due to lack of calibration. Furthermore, the complex operating environment of the generating unit, including fluctuations in ambient temperature and humidity, differences in electromagnetic interference intensity at different locations (e.g., different interference levels near the motor area versus the control cabinet area), equipment vibration, and power grid fluctuations, introduces signal noise and parameter measurement deviations, further degrading data quality.

[0004] In addition, the existing system lacks the ability to adapt to the dynamic operating conditions of the unit in real time, and cannot dynamically compensate for the feature shift caused by load changes and temperature drift. Furthermore, feature extraction relies heavily on fixed model parameters, resulting in a lag in response to differences in operating conditions. It is difficult to achieve accurate extraction and anomaly identification of 128-dimensional feature vectors under complex operating conditions, which seriously affects the accuracy of unit fault early warning and the timeliness of operation and maintenance decisions. Summary of the Invention

[0005] This invention improves the accuracy of unit monitoring data, the accuracy of fault feature identification, and the efficiency of operation and maintenance decision-making through multi-dimensional calibration, adaptive feature extraction under operating conditions, and hierarchical early warning, thereby ensuring equipment operation safety and production efficiency.

[0006] The technical solution proposed in this invention is: a method for online monitoring of magnetic variables in a generator unit, the method comprising:

[0007] Multi-source data is acquired through sensors, and the multi-dimensional calibration mechanism is used to correct the multi-source data. The timestamps of the corrected multi-source data are synchronized using a dual synchronization system, and the synchronized multi-source data is subjected to frequency band conditioning and standardization.

[0008] The system eliminates environmental noise in standardized multi-source data through intelligent algorithms, dynamically adjusts the long short-term memory network to extract feature vectors for different load conditions, and corrects the oil permeability in the feature vectors through neural networks by combining oil loss angle and oil temperature. The system also corrects the magnetic induction intensity in the feature vectors based on the relative permeability of the iron core and vibration signals to obtain the purified feature vectors.

[0009] Extended Kalman filter is used to fill in missing data in the state change scenario on the purified feature vector, and effective abnormal features are screened by the isolated forest algorithm.

[0010] The effective abnormal features are input into the LSTM prediction model to obtain the initial prediction value. The initial prediction value is then corrected by combining temperature and vibration parameters to obtain the corrected prediction value. The corrected prediction value is then substituted into the logistic regression formula to obtain the fault prediction probability. The health score is calculated based on the fault prediction probability and feature deviation. The health score is then used for graded early warning.

[0011] Preferably, the specific process of the multidimensional calibration mechanism correction is as follows:

[0012] Automatic zero-point calibration is triggered during the daily early morning unit shutdown period. A zero magnetic field environment is generated by short-circuiting the calibration coil, and the zero-point offset of each sensor is recorded. The calibrated magnetic field strength is obtained through a real-time compensation formula. Three-point calibration is performed quarterly using a standard magnetic field source, and the correction coefficients are updated through linear fitting. Temperature sensing branches are added to all sensor signal links, and real-time temperature drift compensation is achieved through hardware circuitry. A full-pass filter is used to compensate for the phase delay of the integrator, and real-time phase calibration is achieved through digital signal processing.

[0013] Preferably, the specific details of the frequency band conditioning and the standardization process are as follows:

[0014] In the low-frequency band, a chopper-stabilized instrumentation amplifier is used to 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 wideband amplifier is used to amplify the signal after conversion by the Rogowski coil and integrator, and a new 2nd-order active high-pass filter is added to suppress low-frequency noise. Adaptive gain control is implemented through a digital signal processor to dynamically expand the signal range. A synchronous clock generated by the FPGA triggers a 16-bit multi-channel AD converter to synchronously sample the magnetic field signal and operating parameters. A GPS timing module is used to provide a unified time reference for all sensor nodes, and a 32-bit precise timestamp is added to each frame of data.

[0015] Preferably, the specific process of the intelligent algorithm to eliminate noise is as follows:

[0016] Standardized multi-source data is input into the VAE model, and reconstruction loss and KL divergence are used to jointly optimize the model parameters. The number of model parameters and inference time are reduced by depthwise separable convolution and knowledge distillation techniques. Wavelet thresholding is used to denoise high-frequency signals. The noise standard deviation is estimated by the absolute deviation of the median of the wavelet high-frequency coefficients, and an adaptive threshold is calculated. The wavelet coefficients are then truncated with a soft threshold to further improve the high-frequency noise suppression ratio. The denoised clean signal is synchronously correlated with the load coefficient, rotational speed, and vibration, and then input into the working condition adaptive feature extraction module.

[0017] Preferably, the specific process for obtaining the feature vector is as follows:

[0018] 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 of key time node features is enhanced through the network's attention mechanism, and an initial feature sequence containing magnetic field time domain, frequency domain and time-related features is extracted.

[0019] Preferably, the specific process of feature vector purification is as follows:

[0020] The oil dielectric loss angle collected by the fiber optic grating sensor and the oil temperature collected by the temperature sensor are input into the permeability correction model, and the corrected 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 permeability of the iron core monitored by the magnetoresistive sensor and the vibration signal collected by the vibration sensor, the magnetic induction intensity related features are adjusted through a comprehensive correction formula to suppress the measurement error caused by changes in iron core permeability and magnetostriction, and obtain a 128-dimensional purified feature vector.

[0021] Preferably, the effective anomaly feature screening process is as follows:

[0022] Median filtering is applied to the purified feature vectors to process non-normally distributed data and eliminate impulse noise interference. For missing data in scenarios with abrupt changes in state, extended Kalman filtering is used for dynamic imputation to ensure data continuity. Anomaly detection is performed using the isolated forest algorithm. By constructing multiple isolated trees, feature anomaly scores are calculated to eliminate false anomaly features caused by sensor fluctuations and environmental interference, while retaining true fault-related features. Combined with an incremental learning mechanism, the model parameters are fine-tuned based on newly collected fault samples using an elastic weight consolidation algorithm. This expands the noise fingerprint database and improves the stability and accuracy of anomaly feature identification under different operating conditions, ultimately outputting valid anomaly features.

[0023] Preferably, the specific process of the tiered early warning is as follows:

[0024] Based on the probability of fault prediction and the deviation of characteristics, a health score is calculated using a health scoring formula, with a value range of 0-100. Three levels of warning thresholds are defined according to the health score: a score greater than or equal to 80 and less than 90 corresponds to a Level 1 warning, indicating a minor anomaly and prompting enhanced monitoring; a score greater than or equal to 60 and less than 80 corresponds to a Level 2 warning, indicating a moderate anomaly and triggering a special inspection; and a score less than 60 corresponds to a Level 3 warning, indicating a severe anomaly and requiring emergency shutdown for maintenance. Warning signals of different levels are simultaneously pushed to the operations and maintenance platform, along with fault location information and historical data trend charts. Level 1 warnings are pushed via the APP, Level 2 warnings trigger audible and visual alarms and generate inspection work orders, and Level 3 warnings automatically trigger the emergency shutdown process and simultaneously notify the operations and maintenance team for on-site handling.

[0025] The present invention also provides an online monitoring system for generator magnetic variables, the system being used to execute the aforementioned online monitoring method for generator magnetic variables.

[0026] The present invention also provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the aforementioned method for online monitoring of unit magnetic variables.

[0027] The beneficial effects of this invention are:

[0028] 1. A multi-dimensional calibration mechanism consisting of automatic zero-point calibration, periodic accuracy calibration, and temperature compensation systematically eliminates errors such as sensor zero-point drift and temperature drift. Combined with intelligent denoising processing using a VAE model and wavelet thresholding algorithm, signal noise interference is significantly reduced. This allows the long-term measurement error of key parameters such as magnetic field and vibration to be controlled below 0.3%, and the retention rate of weak fault features to over 95%. This provides a high-fidelity raw data foundation for subsequent feature extraction and fault prediction, solving the misjudgment problem caused by insufficient data accuracy in traditional monitoring. The multi-dimensional calibration mechanism consisting of temperature compensation systematically eliminates errors such as sensor zero-point drift and temperature drift. Combined with intelligent denoising processing using a VAE model and wavelet thresholding algorithm, signal noise interference is significantly reduced. This allows the long-term measurement error of key parameters such as magnetic field and vibration to be controlled below 0.3%, and the retention rate of weak fault features to be increased to 95%.

[0029] 2. The convolution kernel size is dynamically adjusted for different load conditions. Temporal correlation features are extracted using an LSTM network, and the feature vector is corrected by combining multiple parameters such as oil loss angle, core permeability, and vibration signals. This effectively eliminates interference from changes in oil quality, core condition, and vibration. The purified 128-dimensional feature vector accurately reflects the actual operating status of the unit. Combined with multi-unit decoupling and anomaly detection algorithms, the false anomaly feature removal rate is increased to over 90%, significantly improving the stability and accuracy of fault feature identification and reducing the occurrence of invalid warnings.

[0030] 3. Based on fault prediction probability and feature deviation, a health score is calculated, establishing a three-tiered early warning system and matching differentiated response strategies, achieving full-scenario coverage from minor anomalies to severe faults. Level 1 early warning provides advance monitoring and alerts, Level 2 early warning triggers specialized inspections, and Level 3 early warning prompts emergency shutdowns, forming a progressive closed-loop operation and maintenance system. In practical applications, the average early warning time for faults reaches 72 hours, and the fault location accuracy is controlled within 0.5 meters. This not only reduces the risk of sudden unit failures but also reduces unplanned downtime, improving operation and maintenance efficiency and equipment operational safety. Attached Figure Description

[0031] Figure 1 This is a flowchart of a method for online monitoring of unit magnetic variables according to the present invention;

[0032] Figure 2 This is a flowchart illustrating the monitoring process of an online monitoring method for unit magnetic variables according to the present invention. Detailed Implementation

[0033] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0034] It is understood that the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0035] like Figure 1 and Figure 2 As shown, in order to achieve comprehensive monitoring of the unit's magnetic variables and operating status, the raw data is first collected through a high-precision sensor array.

[0036] The radial magnetic field component of the iron core is collected using a Hall sensor. and axial component Six Hall effect sensors are evenly spaced along the circumference of the unit's stator core (two at the end of each phase winding). The sensitivity is set to... (Each millitalas change in magnetic field corresponds to a 0.1 millivolt voltage output), with a linearity error of... (The maximum deviation between the measured value and the true value within the full-scale range), response time is (Time from magnetic field change to stable output), temperature drift coefficient (Output drift caused by a 1°C change in temperature within the operating range of -20°C to 85°C), zero-point drift (Output drift when there is no magnetic field input within 24 hours), the object of measurement is magnetic induction intensity. (Unit: T).

[0037] High-frequency magnetic field radiation signals from the windings are acquired using Rogowski coil sensors. Two Rogowski coil sensors are installed on the outside of the unit's outgoing busbar, arranged in a circular pattern (orthogonal). The bandwidth is 1kHz-1MHz (the effective measurement frequency range), the sensitivity is 0.5mV / A (the induced voltage output per ampere of current), the integrator bandwidth is 1kHz-1MHz (-3dB) (the integrator's attenuation of signals at different frequencies; -3dB is the cutoff point with 3 dB attenuation), and the phase error is ≤5° (1kHz-100kHz) (the phase difference between the measured signal and the true signal within the 1kHz to 100kHz frequency range). The measured object is high-frequency magnetic flux density. .

[0038] The distribution of the external leakage magnetic field is collected using fluxgate sensors. The fluxgate sensors are installed at five points: the center of the top and the four corners of the bottom of the unit casing. The measurement range is ±500°. (Measurable minimum and maximum magnetic flux density), with a resolution of 0.1. (Minimum detectable change in magnetic flux density), long-term stability (Drift range of measured values ​​within one hour), temperature coefficient ≤ 0.5 / ℃ (change in measured value per 1℃ change in temperature), annual drift rate ≤0.1%FS / year (per year relative to full scale drift), applicable frequency range DC-1kHz (suitable frequency range for measurement, DC refers to direct current), the object of measurement is steady-state magnetic flux density. .

[0039] Real-time current of the unit is collected via a current sensor. The current sensor can be a through-core CT (current transformer), with a range of 0-1000A (measurable current range) and an accuracy of 0.2 class.

[0040] Oil temperature is collected by a temperature sensor. and winding temperature The temperature sensor can be a PT100 platinum resistance thermometer, with a measurement range of -20-150℃ and an accuracy of ±0.5℃. It is positioned at the bottom of the oil tank to collect oil temperature data. (unit: °C) and winding ends (for collecting winding temperature) (unit: °C).

[0041] Real-time rotation speed is collected by a speed sensor. The speed sensor can be a magnetoelectric speed probe with a measurement range of 0-3000 r / min (measurable speed range) and an accuracy of ±1 r / min. It is installed on the non-drive end of the unit's shaft.

[0042] Vibration signals are collected using vibration sensors. The vibration sensor can be a piezoelectric accelerometer with a range of 0-50g (the measurable acceleration range, where g is the acceleration due to gravity) and a frequency range of 10Hz-10kHz, and is installed on the outer wall of the iron core.

[0043] Monitoring oil loss angle using a fiber optic grating sensor (Dimensionless parameter, reflecting the insulation performance of transformer oil). Measurement accuracy is ±0.1%.

[0044] Monitoring the relative magnetic permeability of the iron core using a magnetoresistive sensor (Dimensionless, ratio of core permeability to vacuum permeability). Long-term stability is ±0.5% / year.

[0045] Based on the real-time current collected by the current sensor The load factor is calculated using the formula obtained from the operating condition parameters. ( For real-time current, For rated current, 0.2≤ ≤2.0), used to classify light loads ( ), full load (0.3≤ ≤0.8), overload ( Operating conditions.

[0046] Through collaborative acquisition by multiple types of sensors, the spatial coverage of magnetic field signals has been increased to 92%, and the completeness of the acquisition of operating parameters has reached 100%.

[0047] The raw data collected based on the above steps contains errors such as zero-point drift and temperature effects, and a calibration mechanism is needed to improve data reliability.

[0048] Automatic zero-point calibration: Triggered daily during the unit shutdown period in the early morning, a zero magnetic field environment is generated by short-circuiting the calibration coil, and the zero-point offset of each sensor is recorded. Real-time compensation formula: ( To calibrate the magnetic field strength, (The original measurement value).

[0049] Regular accuracy calibration: Three-point calibration (0%FS, 50%FS, 100%FS, where FS is the full scale of the sensor) is performed quarterly using a standard magnetic field source (accuracy ±0.01%). The correction coefficients are updated through linear fitting to ensure measurement accuracy across the entire range.

[0050] Temperature compensation circuit: A PT100 temperature sensing branch is added to all sensor signal links to achieve real-time temperature drift compensation through hardware circuitry, with a compensation accuracy of ≥98%.

[0051] By implementing automatic zero-point calibration, periodic accuracy calibration, and adding temperature compensation circuitry, the long-term measurement error of the sensor is controlled to below 0.3%, and the cumulative error caused by zero-point drift is reduced by 70%.

[0052] To address the phase delay issue of Rogowski coil signals, phase correction is performed based on the calibrated signal. The Rogowski coil signal link incorporates a phase correction module, employing an all-pass filter (cutoff frequency 1MHz, phase error ≤2°) to compensate for the integrator's phase delay, and real-time phase calibration is achieved through digital signal processing. ( (Frequency-phase curve fitting was obtained through factory calibration).

[0053] The phase correction module reduces the phase error of the 1MHz signal from 15° to 3°, improving the positioning accuracy of the partial discharge pulse by 40%.

[0054] To ensure time consistency of multi-sensor data, a dual synchronization system of GPS and IEEE 1588 is constructed. When GPS signal is normal, time synchronization is prioritized (synchronization accuracy ≤100ns); when GPS loses lock (no signal for 5 consecutive seconds), it automatically switches to the IEEE 1588 precise time protocol (synchronization accuracy ≤1). Time synchronization between nodes is achieved through industrial Ethernet switches, with a synchronization accuracy of ≤1. A local crystal oscillator (frequency stability 1ppm) serves as a backup to ensure uninterrupted switching.

[0055] By employing a dual synchronization system, the probability of synchronization failure is reduced to 0.01% / day, and the timestamp deviation of multiple sensors is controlled within 1. the following.

[0056] Based on the unified time reference provided by the dual synchronization system, vibration and magnetic field signals are time-aligned. The timestamp accuracy of the vibration sensor is improved to 10. The sampling frequency is dynamically adjusted (1kHz for low-frequency vibration and 10kHz for high-frequency impact), and sub-microsecond alignment with the magnetic field signal is achieved through an interpolation algorithm. The interpolation algorithm is based on a unified time reference axis. Linear interpolation is used for vibration signals in different scenarios (low-frequency stable operating condition: ,in, The interpolated vibration acceleration is expressed in g. , Vibration acceleration at adjacent sampling points, in g; , For adjacent sampling point timestamps (in nanoseconds) and cubic spline interpolation (for high-frequency impact conditions), ensure alignment error. 50ns.

[0057] By dynamically adjusting the sampling frequency and optimizing the interpolation, the phase alignment error between vibration and magnetic field signals is reduced to ≤50ns, and the accuracy of multi-physics collaborative analysis is improved to 99.5%.

[0058] Based on the calibrated and synchronized signals, frequency band conditioning and standardization are performed to provide high-quality input for subsequent algorithms.

[0059] The analog signal (0-5V) output from the Hall / fluxgate sensor is connected to the signal conditioning module, and the Rogowski coil output signal is converted into a voltage signal by an integrator before being connected.

[0060] In the low-frequency range (DC-1kHz), a chopper-stabilized instrumentation amplifier (common-mode rejection ratio ≥120dB, temperature drift ≤0.1μV / ℃) is used to achieve a 20dB gain. ,in, This refers to the low-frequency output voltage, in volts (V). (This refers to the sensor input voltage, in V), adapted for monitoring the steady-state magnetic field of the iron core; a wideband amplifier is used to achieve a 40dB gain in the high-frequency range (1kHz-50kHz). ,in, (High-frequency output voltage, unit: V) to enhance the high-frequency signal of partial discharge;

[0061] The front end is configured with an 8th-order Butterworth low-pass filter with a cutoff frequency of (Frequency at which the signal amplitude attenuates by 3dB), attenuation rate ≥80dB / dec (signal attenuation per decade), effectively suppressing high-frequency noise aliasing;

[0062] The Rogowski coil output signal is passed through a passive integrator (RC time constant 10). After conversion, a second-order active high-pass filter (cutoff frequency 1kHz) is added to suppress low-frequency noise, and adaptive gain control (AGC) is implemented through a digital signal processor (DSP). AGC is achieved through sliding window peak detection. ( For the first Peak value of frame signal, in V; (Input signal number), and the gain coefficient is calculated after smoothing. (Target amplitude) ; The signal amplitude is smoothed (in V). The maximum gain change rate is limited to ≤6dB / ms, and the dynamic range is extended to 60dB (1mV-1V input signal).

[0063] The signal phase spectrum is calculated by FFT transformation, and point-by-point correction is performed based on the pre-stored phase compensation curve to ensure that the phase characteristic linearity is ≥98% in the 1kHz-1MHz frequency band.

[0064] Generated by FPGA Synchronous clock (jitter ≤ 10ns) triggers 16-bit multi-channel AD converter (conversion time) This enables synchronous sampling of magnetic field signals and operating parameters;

[0065] A GPS timing module (synchronization accuracy ≤100ns) is used to provide a unified time reference for all sensor nodes, and a 32-bit precise timestamp is appended to each frame of data. =Year + Month + Day + Hour + Minute + Second + Millisecond + Microsecond);

[0066] The sampled data is encapsulated in frame format, with each frame containing a timestamp. (Accuracy 1ms), three-dimensional magnetic field components / / Current ,temperature / Rotation speed ,vibration 32 parameters are transmitted to the edge computing node via industrial Ethernet (ring network redundancy design).

[0067] Inter-channel synchronization error ≤50ns, data transmission packet loss rate ≤0.01%, single-channel sampling power consumption ≤0.5W.

[0068] The use of a low-temperature drift amplifier reduces thermal noise by 40%, the distortion rate of high-frequency signals (1kHz-1MHz) is ≤1%, and the signal-to-noise ratio of partial discharge pulse signals is improved by 12dB; the signal synchronization error is controlled below 1ms, the time synchronization error of vibration and magnetic field signals is ≤0.1ms, and the accuracy of multi-physics superposition analysis is improved to 99%.

[0069] Based on the conditioned, standardized signal, noise is eliminated and key features are extracted using intelligent algorithms. Intelligent denoising algorithms eliminate environmental noise and measurement interference while retaining key fault characteristics, laying the foundation for accurate feature extraction.

[0070] The preprocessed signal is first input into the variational autoencoder (VAE) denoising module, where it is jointly optimized using reconstruction loss and KL divergence. ,in, Let be the loss function of the variational autoencoder. For expectation operator, For log-likelihood, Let KL divergence be the KL divergence. For approximate posterior distribution, The prior distribution is used. The training data includes additional samples under extreme operating conditions (oil temperature 80-100℃, overload). The range of values ​​is Expand the coverage area.

[0071] In the VAE model, depthwise separable convolution is used to replace traditional convolution, reducing the number of encoder / decoder channels from 64 / 128 / 256 to 32 / 64 / 128. Knowledge distillation (parameter transfer from teacher model to student model) is introduced, reducing the number of model parameters by 60% and compressing the inference time to 30ms.

[0072] VAE Multi-Fault Generalization Optimization: The training set contains 2000 mixed samples with multiple faults (such as simultaneous winding short circuit and core loosening, and simultaneous partial discharge and bearing wear). The encoder adopts a hybrid structure of "depth-separable convolution and 1×1 standard convolution" (specific layers: layer 1: 32-channel 3×3 depth-separable convolution; layer 2: 32-channel 3×3 depth convolution + 64-channel 1×1 standard convolution; layer 3: 30% standard convolution + 70% depth convolution; layer 4: 96-channel 1×1 standard convolution + 128-channel 3×3 depth convolution). 30% of the standard convolution is retained to improve feature interaction capabilities, and the F1 score in multi-fault scenarios is improved to 0.92 and above.

[0073] Compared to traditional denoising algorithms, the signal-to-noise ratio is improved by 15dB, the retention rate of weak fault features is increased to 92%, and the "pseudo-denoising" rate is reduced to 3% under extreme conditions; the recall rate in multi-fault concurrent scenarios is increased from 75% to 90%, and while maintaining performance, the GPU utilization rate of edge computing nodes is reduced from 75% to 35%.

[0074] The signal is framed (N=1024 points / frame, overlap rate r=50%); the encoder extracts features through 4 layers of convolution, and the decoder reconstructs the denoised signal; the Adam optimizer (learning rate) is used. (Iteration E=50 rounds); High-frequency signals (>5kHz) supplemented with wavelet threshold denoising: ( These are the original wavelet coefficients; It is a symbolic function; For adaptive threshold, ,in, The noise standard deviation is estimated using the absolute deviation of the median of the wavelet high-frequency coefficients. , (Median absolute deviation); Output pure signal synchronous correlation operating parameters Rotation speed ,vibration The data is then fed into the adaptive feature extraction module. This method improves the high-frequency noise suppression ratio by an additional 10dB, and the denoising processing time is ≤50ms, meeting real-time requirements.

[0075] The feature extraction strategy is dynamically adjusted for different load conditions to solve the problem of feature aliasing under complex conditions and accurately extract fault-related features.

[0076] Load coefficient based on association Dynamically adjust the kernel size: ( The kernel size; To obtain (Minimum value of 2.0), limiting the maximum size to 7×7. This improves the feature extraction accuracy by an average of 18% under different working conditions, effectively solves the feature aliasing problem, and avoids exceeding the convolution kernel size limit.

[0077] LSTM network for hiding time series states Calculate attention weights: ( For the first Moment-by-moment attention weighting; It is an energy function; (where is the sequence length), ( , This is the weight matrix; (as a bias term), combined with multi-source mutation signals to Add bias.

[0078] LSTM Sequence Length Optimization: Determining the Optimal Sequence Length Based on Information Entropy Analysis: Calculating the Feature Information Entropy under Different Windows ( For characteristic probability distribution; (For information entropy, a measure of feature uncertainty), select A stable minimum window (256 points for light load, 384 points for heavy load) reduces computation by 20% compared to the original design. A bidirectional LSTM structure (128+128 hidden layer dimensions, 2 layers) reduces parameters by 30%, and TensorRT quantization acceleration (INT8 precision) doubles the inference speed.

[0079] Adaptive attention bias: Bias value is dynamically calculated. ( Rated speed, in r / min; This is the attention bias term (dimensionless). The rotational speed at the current moment, in r / min; The speed at the previous moment, in r / min;), small capacity motor ( =1500r / min) and large capacity motor ( =3000r / min) respectively adapted to avoid empirical value deviation. In this way, the attention of key time node features is increased by 35%, the error rate of weight allocation caused by single sensor error is reduced to 2%; the long-term feature retention rate is ≥95%, the computational efficiency is improved by 15%; the model inference latency is ≤20ms, meeting the real-time matching requirement of 100 times / second, and the fault feature recognition rate is improved to 90% during dynamic transition.

[0080] The output D=128-dimensional feature vector contains the time domain, frequency domain, and time-related features of the magnetic field, and is passed to the environmental interference elimination module for further purification.

[0081] Based on the extracted 128-dimensional feature vector, it is necessary to eliminate the influence 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's status.

[0082] Eliminate the influence of changes in the characteristics of the internal medium (such as transformer oil and iron core) on the magnetic field signal to ensure that the signal reflects the true state of the unit.

[0083] Clarify the relationship between physical quantities: magnetic induction intensity ( Magnetic flux density, in tons (T). Permeability, in H / m; Magnetic intensity, unit: A / m; , Relative permeability (vacuum permeability), directly measured by the sensor. ,pass Convert to obtain magnetic field strength .

[0084] Oil quality nonlinearity compensation: A permeability correction model is established using a BP neural network. ,in, This is the corrected permeability, in H / m; This refers to the oil loss angle. The network input is the oil loss angle acquired by the fiber Bragg grating sensor. and oil temperature The output is the corrected permeability. .

[0085] Neural network structure parameters: The oil quality compensation BP network uses 2 inputs ( -3 hidden layers (16, 32, 16 neurons) -1 output ( The structure is as follows: the activation functions are ReLU (hidden layer) and Linear (output layer), the optimizer is Adam (learning rate 5e-4), the iteration is 200 rounds, and the validation set loss is ≤0.001.

[0086] BP network extrapolation constraints: Add boundary constraints to the output layer: when 0.1 or At 80℃, linear extrapolation constraints are activated (slope ≤ 0.02 / ℃), triggering an out-of-range warning. Testing shows... When the value is 0.2, the compensation error is controlled within 2.0%.

[0087] This improves the network fit. ( The coefficient of determination (to measure the model fit) keeps the nonlinear compensation error of oil quality stable at 1.5%, and the compensation accuracy remains ≥98% under out-of-range conditions, avoiding model failure and significantly improving signal stability.

[0088] Simultaneous monitoring of the relative permeability of the iron core using a magnetoresistive sensor (benchmark value) Combined with vibration signals Establish a comprehensive correction formula ( The corrected magnetic flux density is expressed in tons (T). The unit for measuring magnetic flux density is T. Real-time relative permeability; The relative permeability is used as a reference. Real-time vibration acceleration, in g; (This is the normal vibration reference value, in g, obtained through historical data statistics). (Permeability deviation exceeds 10%) or Correction is initiated when the vibration deviation exceeds 20%. This ensures that the measurement error of magnetic induction intensity caused by changes in the core permeability is controlled within 1.2%, the magnetostriction effect is suppressed by 80%, and the consistency of physical quantities is guaranteed.

[0089] The corrected signal has significantly reduced internal medium interference and is then fed into the multi-unit magnetic interaction decoupling module to handle external unit interference.

[0090] To isolate magnetic field interference from adjacent units, the true magnetic field characteristics of the target unit are extracted to address the challenges of multi-unit cooperative operation.

[0091] 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. : ( This is the minimum description length value; For a given Signal from one interference source The probability of; For the number of array elements, This refers to the number of snapshots; (Number of interference sources). MUSIC algorithm spatial spectrum function: ( Spatial spectral power; It is the azimuth angle; For array manifold vectors; The noise subspace matrix; superscript (For conjugate transpose), azimuth angle is obtained by searching for peak values. .

[0092] Using FFT for fast spatial spectrum estimation instead of direct matrix inversion reduces the computational complexity of spatial spectrum estimation from... Down to ( To reduce the number of snapshots, a sliding window accumulation (10 frames / accumulation) is introduced to reduce the computation frequency. Array data preprocessing adopts dimensionality reduction PCA (principal component retention rate of 95%).

[0093] Element position calibration mechanism: The element spacing is calibrated monthly using a laser rangefinder (accuracy ±1mm), a position error database is established, and an error compensation term is introduced into the MUSIC algorithm. ( The compensated array manifold vector; The vector of the original array manifold; (This is the correction vector for the spacing deviation d). This improves the accuracy of interference source number estimation to 98%, reduces the single calculation time of the 8-element MUSIC algorithm from 50ms to 8ms, and improves real-time performance by 6 times in multi-unit scenarios. Even when the foundation settlement causes a deviation of ±5cm, the positioning error is still controlled within 0.5m, and the resistance to installation deviation is significantly enhanced.

[0094] A reference signal is generated based on the interference source parameters, and the coefficients are updated using variable step size LMS adaptive filtering. ( For the first Time-major filter coefficients; This is a variable step size factor; For the first Timing error signal; For the first (input signal at any time), step size ( 0.01, =1e-6), which improves the convergence speed.

[0095] LMS convergence optimization: Step size formula improved to (Introducing the momentum term) (0.2), the convergence time in dynamic scenarios is reduced from 20ms to 12ms. This ensures that the co-channel interference suppression ratio remains at 35dB or higher in dynamic scenarios, and at 30dB or higher when the load changes abruptly, while improving the purity of the target signal extraction by 40%.

[0096] The purified target magnetic field feature vector has been significantly freed from internal and external interference and is then fed into the signal verification module for anomaly detection.

[0097] Based on the purified feature vectors, a robust algorithm is used to detect anomalies and update the model, ensuring reliable extraction of fault features. Dynamic thresholds are used to determine feature anomalies, eliminating false interference and filtering out genuine fault features, providing reliable input for fault diagnosis.

[0098] Median filtering is used to replace the 3σ criterion for processing non-normally distributed data, and the window size is dynamically adjusted (3-7 points).

[0099] An extended Kalman filter (EKF) is used instead of a traditional Kalman filter to handle missing data imputation in scenarios with abrupt state changes, and a nonlinear term is introduced into the state equation. ( This is the state transition function; For the first Time-state vector; For control input; Predict increments for the state; This is a load nonlinearity correction term. (load factor).

[0100] Preserve pulse characteristics: Set a pulse detection threshold (greater than 5 times the root mean square value) and skip the rejection process after marking suspected partial discharge pulses.

[0101] This optimization method reduces the false rejection rate of nonnormal features from over 5% to below 1%, maintains the accuracy of missing data imputation at the time of state change above 92%, and improves data integrity to 99.9%.

[0102] An incremental learning module is introduced, and the parameters of the VAE and LSTM models are fine-tuned weekly using newly collected normal operating condition data (1000 samples). The Elastic Weight Consolidation (EWC) algorithm is used to protect the key feature extraction capabilities. ( This is the total loss function; Loss due to new data; The regularization strength; This is the parameter importance matrix; For the current parameter; (Based on the old parameter values).

[0103] Incremental learning trigger optimization: Clearly define the KL divergence calculation rules: use data from the most recent 30 days, with a minimum sample size of 2000 (extend the collection period if insufficient), and set dual trigger conditions (KL divergence > 0.1 and exceeding the limit for 3 consecutive days), reducing the false trigger rate to 1 time per quarter or less. The model performance degradation under uncovered operating conditions is controlled from 30% to 8%; trigger accuracy is improved to 95% in the initial stage of new unit commissioning, avoiding ineffective retraining.

[0104] A noise fingerprint database (containing 20 typical noise patterns such as lightning strikes and equipment start-up and shutdown) is constructed. The signal type (fault / noise) is identified in real time by an SVM classifier. Enhanced denoising is initiated for suspected noise signals (the wavelet threshold is increased by 20%).

[0105] Noise fingerprint database expansion mechanism: Establish a quarterly update mechanism, add 10 new types of noise samples (such as inverter harmonics and 5G base station interference), adopt incremental SVM algorithm (supporting online sample addition), and maintain noise identification accuracy of ≥97%.

[0106] In high-noise environments, the fault misjudgment rate decreased from 15% to ≤3%; the misjudgment rate for uncovered noise types decreased from 12% to ≤4%.

[0107] Based on valid anomaly features, tiered early warnings are implemented through trend prediction and health assessment to support operational and maintenance decisions. The system receives the decoupled feature vector and calculates a dynamic threshold range based on nearly three months of normal data.

[0108] upper limit lower limit (The mean is updated daily via a sliding window of N=1000) and variance ),in, Upper threshold; The lower threshold; The mean; The standard deviation is denoted as .

[0109] Adaptive threshold updates reduce the false alarm rate to 1% and the false negative rate to 0.5%.

[0110] Abnormal features exceeding the threshold are marked, and interference residues are removed using the Isolation Forest algorithm.

[0111] Isolation forest parameter optimization: The optimal parameters were determined by grid search: subsample size s=512, number of trees t=200, and the anomaly detection F1 score was improved to 0.93.

[0112] The Isolation Forest algorithm can achieve a 98% rejection rate for interference residues; the stability of interference residue rejection is improved, and the accuracy of abnormal feature identification reaches 99%, ensuring that the features passed to the next stage are valid abnormal features.

[0113] Effective abnormal features are input into the dynamic operating condition map matching module for fault type identification and trend prediction.

[0114] By identifying fault types and assessing development trends through feature matching and trend prediction, early fault warnings can be achieved.

[0115] Enhanced extreme condition detection: Detects sudden load changes using sliding window variance (window size 50ms). (rate of change > 0.5 / s) or sudden temperature rise ( If the rate of change is >5℃ / min, the model parameter temporary reset mechanism is triggered, and the corresponding pre-trained parameters for the working condition are loaded.

[0116] The parameter reset system is comprehensive:

[0117] The pre-training parameter library is stored in categories according to the unit model (e.g., 8 categories such as 10kV / 200kW, 35kV / 1000kW, etc.), and includes 10 typical operating condition parameters such as no-load, full-load, and 120% overload.

[0118] Trigger thresholds are customized for each unit: sudden load change thresholds Rate of change = 0.3 + 0.2 × ( / 1000) (Rated power, unit kW), temperature surge threshold Rate of change = 3 + 0.02 × );

[0119] Reset process: Parameter switching is completed within 50ms after triggering, and cached data is used for transition during the switching period.

[0120] By resetting parameters, the false trigger rate of parameter reset under extreme conditions is reduced to ≤0.5 times / month, and the model stability is improved by 40%.

[0121] The DTW algorithm with slope constraints is adopted. and Calculate the similarity between real-time features and samples in the spectral library: ,in, For similarity ( Match successful); For dynamic time-normalized distance; The length of the feature sequence; This represents the maximum value of the characteristic.

[0122] The DTW algorithm with slope constraints improves the accuracy of fault type identification to 95% and the matching speed to 100 times / second or more.

[0123] The LSTM prediction model takes a feature sequence of the past 30 days as input and applies it through the ReLU activation function. and MAE loss function (Target error ≤ 5%) Predict future trends, where, Output for hidden layer; , This is the weight matrix; For bias; The sequence length; The actual value; These are the initial predicted values. The initial predicted values ​​of the fault trend are obtained through calculation by the output layer of the LSTM network. ,in, This is the output layer weight matrix; This is the output layer bias term. Multi-factor trend correction: The prediction results are then corrected a second time by incorporating temperature and vibration parameters. The correction formula is as follows: ,in, This is the corrected predicted value; These are the initial predicted values; This is a temperature correction factor; Vibration correction factor; temperature correction term , , , , Vibration correction term , This is the indicator function. Under extreme operating conditions, the trend prediction error is further reduced to 4.5% or below.

[0124] Corrected predictions Substituting into the logistic regression formula, we obtain the predicted probability of the fault occurring. The logistic regression formula is: ,in, The probability of failure prediction (dimensionless, ranging from 0 to 1), the closer the value is to 1, the higher the risk of failure. The first output of the LSTM model Real-time fault trend prediction value; , The logistic regression coefficients (obtained through training with historical fault data) This is the weight matrix. (This is a bias term).

[0125] After six months of actual testing on 10 units, the error in predicting the fault development trend was controlled at 5% or less, and the average early warning time reached 72 hours.

[0126] The output fault type and probability are sent to the early warning module to trigger the corresponding level of early warning mechanism.

[0127] Tiered warnings are issued based on the severity of the fault to ensure that maintenance personnel respond in a timely manner and prevent the fault from escalating.

[0128] Based on fault prediction probability and feature deviation ( Calculate health score: ,in, Rate your health (0-100 points); For real-time features; These are normal characteristics; , These are the upper and lower limits of the feature threshold.

[0129] Health score dynamic weighting: Weights are dynamically adjusted according to fault type: winding fault (characteristic deviation weight 15, probability weight 60), abnormal oil temperature (characteristic deviation weight 5, probability weight 30), based on the fault influence factor matrix. To achieve differentiated scoring.

[0130] The accuracy rate of health status quantification reached 97%, and negative scores were avoided by adjusting the denominator, providing a scientific basis for graded early warning; the accuracy rate of fault severity differentiation was improved to 98%, and the rationality of early warning grading was significantly enhanced.

[0131] Early warning is based on a scoring system: Level 1 (80-90 points) pushes an APP notification (minor anomaly, with details of feature deviation and trend chart); Level 2 (60-80 points) triggers an audible and visual alarm (potential fault, automatically generates preliminary maintenance suggestions); Level 3 (<60 points) automatically cuts off the control circuit (emergency fault, simultaneously pushes fault location coordinates and emergency response plan), and simultaneously associates fault location information (accuracy ≤0.5m).

[0132] The early warning response time has been shortened to ≤3s. Industrial field verification has shown that fault handling efficiency has been improved by 40%, and the accuracy of the Level 3 early warning has reached 100%.

[0133] For example, based on the raw data collected by the sensor array, the core operating parameters are first calculated. Taking a 35kV / 1000kW unit as an example, the key parameter acquisition results are as follows: The Hall sensor acquires the radial magnetic field component of the iron core. axial component Synchronously record the axial component of the zero-point offset (Automatic calibration is performed daily at midnight); Real-time current is acquired by a current sensor. Given the rated current Calculate the load factor (Determined to be under full load); oil temperature is collected by a temperature sensor. Winding temperature Vibration sensors collect vibration signals. (Normal vibration reference value) (Determined through historical data statistics); the magnetoresistive sensor collects the relative magnetic permeability of the iron core. (benchmark) ).

[0134] Based on the raw data obtained from the signal acquisition steps, calibration and time alignment are performed to eliminate system errors. Hall effect sensors exhibit temperature drift characteristics (temperature drift coefficient 50ppm / ℃), based on the current oil temperature. (Referencing ambient temperature 25℃), the temperature drift error is calculated as follows: Temperature drift error = 50ppm / ℃ × (55-25)℃ = 1500ppm = 0.15%. After processing by the PT100 temperature compensation circuit, the actual error is reduced to 0.03%.

[0135] The vibration signal is interpolated and aligned based on a unified time reference provided by the GPS and IEEE 1588 dual synchronization system. The adjacent sampling points of the vibration sensor are known. , , , , needs to be calculated Interpolation results at: .

[0136] Based on the calibrated and synchronized signal, frequency band amplification, filtering, and noise reduction are performed to improve signal quality. Low frequency band (DC-1kHz): Hall sensor output voltage. After processing by a chopper-stabilized instrumentation amplifier (20dB gain): High-frequency band (1kHz-50kHz): Rogowski coils acquire partial discharge signals. After processing by a wideband amplifier (40dB gain): .

[0137] Additional noise reduction is performed on high-frequency signals (>5kHz), with a signal frame length of [missing information]. absolute deviation of the median of wavelet high-frequency coefficients Calculate the noise standard deviation: Adaptive threshold: For a certain wavelet coefficient After noise reduction: .

[0138] Based on the conditioned pure signal, feature extraction parameters are dynamically adjusted and media interference is eliminated. This is based on the load factor. Optimize the kernel size: (Round to integer), adapted for full-load operating condition feature extraction.

[0139] Based on monitoring by magnetoresistive sensors With vibration signal Initiate comprehensive correction:

[0140] Permeability deviation: (Permeability correction not triggered).

[0141] Vibration deviation: (Triggering vibration correction).

[0142] Corrected magnetic field strength: .

[0143] Based on the feature vectors after interference removal, fault trend prediction and health status scoring are performed. The LSTM model is input with a feature sequence of the past 30 days to obtain corrected prediction values. Substitute into the logistic regression formula (given) , ) .

[0144] Real-time value of a certain feature Normal baseline value upper limit of threshold lower limit , (Triggered Level 3 warning).

[0145] The processes described above with reference to the flowcharts in the embodiments disclosed in this invention can be implemented as computer software programs. The embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0146] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0147] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the stated principles, the implementation of the present invention may have any variations or modifications.

Claims

1. A method for online monitoring of magnetic variables in a generator unit, characterized in that, The method includes: Multi-source data is acquired through sensors, and the multi-dimensional calibration mechanism is used to correct the multi-source data. The timestamps of the corrected multi-source data are synchronized using a dual synchronization system, and the synchronized multi-source data is subjected to frequency band conditioning and standardization. The system eliminates environmental noise in standardized multi-source data through intelligent algorithms, dynamically adjusts the long short-term memory network to extract feature vectors for different load conditions, and corrects the oil permeability in the feature vectors through neural networks by combining oil loss angle and oil temperature. The system also corrects the magnetic induction intensity in the feature vectors based on the relative permeability of the iron core and vibration signals to obtain the purified feature vectors. Extended Kalman filter is used to fill in missing data in the state change scenario on the purified feature vector, and effective abnormal features are screened by 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 then corrected by combining temperature and vibration parameters to obtain the corrected prediction value. The corrected prediction value is then 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 feature deviation. The graded early warning is then carried out according to the health score. The specific process of the multidimensional calibration mechanism correction is as follows: Automatic zero-point calibration is triggered during the daily early morning unit shutdown period. A zero magnetic field environment is generated by short-circuiting the calibration coil, and the zero-point offset of each sensor is recorded. The calibrated magnetic field strength is obtained through a real-time compensation formula. Three-point calibration is performed quarterly using a standard magnetic field source, and the correction coefficients are updated through linear fitting. Temperature sensing branches are added to all sensor signal links, and real-time temperature drift compensation is achieved through hardware circuitry. A full-pass filter is used to compensate for the phase delay of the integrator, and real-time phase calibration is achieved through digital signal processing.

2. The method for online monitoring of unit magnetic variables according to claim 1, characterized in that, The specific details of the frequency band conditioning and the standardization process are as follows: In the low-frequency band, a chopper-stabilized instrumentation amplifier is used to 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 wideband amplifier is used to amplify the signal after conversion by the Rogowski coil and integrator, and a new 2nd-order active high-pass filter is added to suppress low-frequency noise. Adaptive gain control is implemented through a digital signal processor to dynamically expand the signal range. A synchronous clock generated by the FPGA triggers a 16-bit multi-channel AD converter to synchronously sample the magnetic field signal and operating parameters. A GPS timing module is used to provide a unified time reference for all sensor nodes, and a 32-bit precise timestamp is added to each frame of data.

3. The method for online monitoring of unit magnetic variables according to claim 2, characterized in that, The specific process of noise elimination by the intelligent algorithm is as follows: Standardized multi-source data is input into the VAE model, and reconstruction loss and KL divergence are used to jointly optimize the model parameters. The number of model parameters and inference time are reduced by depthwise separable convolution and knowledge distillation techniques. Wavelet thresholding is used to denoise high-frequency signals. The noise standard deviation is estimated by the absolute deviation of the median of the wavelet high-frequency coefficients, and an adaptive threshold is calculated. The wavelet coefficients are then truncated with a soft threshold to further improve the high-frequency noise suppression ratio. The denoised clean signal is synchronously correlated with the load coefficient, rotational speed, and vibration, and then input into the working condition adaptive feature extraction module.

4. The method for online monitoring of unit magnetic variables according to claim 3, characterized in that, The specific process for 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 of key time node features is enhanced through the network's attention mechanism, and an initial feature sequence containing magnetic field time domain, frequency domain and time-related features is extracted.

5. The method for online monitoring of unit magnetic variables according to claim 4, characterized in that, The specific process of feature vector purification is as follows: The oil dielectric loss angle collected by the fiber optic grating sensor and the oil temperature collected by the temperature sensor are input into the permeability correction model, and the corrected 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 permeability of the iron core monitored by the magnetoresistive sensor and the vibration signal collected by the vibration sensor, the magnetic induction intensity related features are adjusted through a comprehensive correction formula to suppress the measurement error caused by changes in iron core permeability and magnetostriction, and obtain a 128-dimensional purified feature vector.

6. The method for online monitoring of unit magnetic variables according to claim 5, characterized in that, The effective anomaly feature screening process is as follows: Median filtering is applied to the purified feature vectors to process non-normally distributed data and eliminate impulse noise interference. For missing data in scenarios with abrupt changes in state, extended Kalman filtering is used for dynamic imputation to ensure data continuity. Anomaly detection is performed using the isolated forest algorithm. By constructing multiple isolated trees, feature anomaly scores are calculated to eliminate false anomaly features caused by sensor fluctuations and environmental interference, while retaining true fault-related features. Combined with an incremental learning mechanism, the model parameters are fine-tuned based on newly collected fault samples using an elastic weight consolidation algorithm. This expands the noise fingerprint database and improves the stability and accuracy of anomaly feature identification under different operating conditions, ultimately outputting valid anomaly features.

7. The method for online monitoring of unit magnetic variables according to claim 6, characterized in that, The specific process of the tiered early warning is as follows: Based on the fault prediction probability and feature deviation, a health score is calculated using a health score formula, with a value range of 0-100. Three levels of warning thresholds are defined according to the health score: a score greater than or equal to 80 and less than 90 is a level 1 warning corresponding to minor anomalies, prompting enhanced monitoring; a score greater than or equal to 60 and less than 80 is a level 2 warning corresponding to moderate anomalies, triggering special inspections. A score below 60 indicates a Level 3 warning, which corresponds to a serious anomaly and requires emergency shutdown and maintenance. Different levels of early warning signals are simultaneously pushed to the operation and maintenance platform, along with fault location information and historical data trend charts; Level 1 warnings are pushed to the app; Level 2 warnings trigger audible and visual alarms and generate inspection work orders; Level 3 warnings automatically trigger emergency shutdown procedures and notify the maintenance team for on-site handling.

8. A unit magnetic variable online monitoring system, characterized in that, The system is used to execute the online monitoring method for unit magnetic variables as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is executed by a processor to implement the online monitoring method for unit magnetic variables as described in any one of claims 1-7.

Images