A fault early warning system for excitation direct current large current and carbon brush vibration coupling characteristics

Abstract

This invention belongs to the field of synchronous motor condition monitoring and fault early warning technology. It discloses a fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration, aiming to solve the core pain points of existing technologies, such as weak early fault identification capability, high false alarm rate, poor adaptability to operating conditions, and lack of fault tracing and trend prediction capabilities. The system includes a monitoring center and a multi-parameter acquisition unit, data processing module, model building module, and fault early warning module connected to it. Through synchronous acquisition and anti-interference processing of multi-source heterogeneous data, it constructs an electromagnet thermal four-field coupling model and a personalized prediction model. Through pre-identification of coupling strength abrupt changes and accurate identification of degradation rate inflection points, it achieves graded fault early warning and shutdown protection triggering. This invention can achieve advanced early warning of carbon brush-slip ring faults, has strong adaptability to all operating conditions, and provides support for the safe and stable operation and condition-based maintenance of generator sets.

Description

Technical Field

[0001] This invention relates to the field of synchronous motor condition monitoring and fault early warning technology, specifically a fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration. Background Technology

[0002] Synchronous generators are the core power equipment of power systems, and the excitation system is a core component of synchronous generators. The carbon brush-slip ring assembly is the only rotating contact component in the excitation system that transmits large DC currents, and its operational reliability directly determines the safe and stable operation of the unit. According to power industry fault statistics, carbon brush-slip ring failures in the excitation system account for more than 20% of unplanned generator outages, making it one of the core pain points in unit operation and maintenance.

[0003] The excitation DC current of large synchronous generator sets can reach thousands of amperes. The carbon brush-slip ring system has strong electromechanical coupling characteristics during operation: changes in the contact state between the carbon brush and the slip ring will cause fluctuations in contact resistance, which will in turn cause uneven distribution of branch current and sudden changes in current ripple; the fluctuation of current will generate alternating electrodynamic force, which will excite radial and axial vibration of the carbon brush; and the vibration of the carbon brush will further change the contact pressure and contact area, exacerbate the fluctuation of contact resistance, and form a positive feedback fault evolution process of "poor contact - current fluctuation - increased vibration - further deterioration of contact".

[0004] Existing monitoring solutions for carbon brush-slip ring failures have the following core shortcomings: 1. The system primarily relies on single-parameter monitoring, resulting in extremely weak early fault identification capabilities. Most existing solutions monitor single parameters such as carbon brush temperature, vibration acceleration, and excitation current RMS value. However, the changes in these single parameters are minimal during the early stages of a fault, completely obscured by fluctuations in unit operating conditions and environmental interference. This makes it impossible to identify early potential problems such as intermittent contact failures. Often, the fault is only detected when it has progressed to the stage of brush burnout or ring fire, missing the optimal time for operation and maintenance.

[0005] 2. Lacking a supporting coupling mechanism, multi-parameter fusion is merely a simple threshold superposition. A few multi-parameter monitoring schemes simply logically superimpose the thresholds of multiple single parameters, failing to reveal the electromechanical coupling fault mechanism between excitation high current and carbon brush vibration. They cannot distinguish between normal operating condition fluctuations and early fault evolution, resulting in extremely high false alarm and false alarm rates under varying operating conditions.

[0006] 3. Fixed threshold warnings have extremely poor adaptability to operating conditions. Most existing solutions use fixed thresholds for warnings. However, the excitation current and vibration level of the generator set vary greatly across the entire operating range from no-load to overload. Fixed thresholds are prone to false alarms under light loads and false alarms under heavy loads, making them unsuitable for the full range of operating conditions of the generator set.

[0007] 4. Lack of fault tracing and trend prediction capabilities. Existing solutions can only provide fault alarms, but cannot accurately locate fault types or provide targeted guidance for operation and maintenance. At the same time, they lack the ability to predict fault evolution trends, cannot achieve true condition-based maintenance, and still rely on the traditional periodic maintenance model, resulting in high operation and maintenance costs and uncontrollable risks. Summary of the Invention

[0008] To address the aforementioned technical problems, the present invention aims to provide a fault early warning system for the coupling characteristics of excitation DC high current and carbon brush vibration, including a monitoring center, which is communicatively connected to a multi-parameter acquisition unit, a data processing module, a model building module, and a fault early warning module. The multi-parameter acquisition unit is used to synchronously acquire all channels of data from the target unit and set the acquisition cycle; The data processing module is used to perform synchronous timing calibration and abnormal timing processing on the full-channel data, as well as noise reduction preprocessing. It identifies and removes transient interference from the noise-reduced full-channel data, reconstructs the waveform of the removed interference segment, and determines the operating conditions of the current acquisition cycle. The model building module is used to build an electromagnetic thermal coupling model, and to build a fault life cycle evolution mapping table and a personalized four-field coupling prediction model for the target unit based on the simulation test data of the electromagnetic thermal coupling model. The fault early warning module is used to output the four-dimensional coupling coefficient and core mechanism state parameters under the current operating conditions based on the personalized four-field coupling prediction model. Based on the four-dimensional coupling coefficient, it performs pre-identification of coupling strength abrupt changes. Based on the pre-identification results, it determines whether to perform accurate identification of degradation rate based on the core mechanism state parameters. Based on the accurate identification results, it determines whether to perform critical inflection point risk assessment or trigger shutdown protection. Based on the critical inflection point risk assessment results, it generates an early warning level.

[0009] Furthermore, the acquisition channels of the multi-parameter acquisition unit include: the total excitation current channel, the branch current channel of each carbon brush, the vibration acceleration channel of each carbon brush holder, the speed / phase pulse channel, the non-contact acquisition channel, and the unit operating condition channel; The full-channel data includes contact-based multidimensional heterogeneous data and contactless redundant data; The contact-type multidimensional heterogeneous data includes the total current of the excitation circuit, the full waveform data of the current of each carbon brush branch, the vibration acceleration signal of each carbon brush holder, the slip ring speed phase pulse signal, and the unit operating parameters. The non-contact redundant data includes the radial vibration of the carbon brush end face, the temperature field distribution of the slip ring surface and the carbon brush contact surface, and the partial discharge ultra-high frequency signal.

[0010] Furthermore, the process of synchronizing timing calibration and handling abnormal timing data across all channels includes: An absolute time base is established for all channel data, and full-channel delay calibration is performed on all acquisition channels. A full-link delay compensation table is established. Based on the absolute time base and the full-link delay compensation table, synchronous timing calibration is performed on the full channel data for each acquisition cycle. Sampling loss points are determined for the full channel data after synchronous timing calibration. Abnormal data segments are marked based on the sampling loss point determination results. Abnormal data segments include a first abnormal segment and a second abnormal segment. The first abnormal segment is interpolated and completed by adjacent normal data segments, and the second abnormal segment is removed.

[0011] Furthermore, the process of noise reduction preprocessing and transient interference identification and removal for the entire channel data includes: Establish a typical disturbance feature library for the excitation system, which includes multidimensional instantaneous features corresponding to different typical disturbances; Steady-state periodic interference and random noise reduction preprocessing are performed on the contact-type multidimensional heterogeneous data in the full channel data to obtain high signal-to-noise ratio basic signals, including excitation current signals and carbon brush vibration acceleration signals. A sliding detection window is set up, and multi-dimensional transient features are detected on the high signal-to-noise ratio basic signal within the sliding detection window to obtain multi-dimensional transient features, including amplitude change rate, rising edge slope, modulus maxima, and time-frequency domain energy entropy change. A transient threshold corresponding to each transient feature in the multi-dimensional transient features is preset. If any transient feature is greater than the corresponding transient threshold, the detection window is determined to be a transient event. When a transient event occurs, the instantaneous coherence coefficients of the excitation current signal and the carbon brush vibration acceleration signal within the detection window are obtained. The instantaneous coherence coefficients are compared with the preset instantaneous coherence coefficient threshold. If the instantaneous coherence coefficient is greater than the preset instantaneous coherence coefficient threshold, the transient event is determined to be a real fault transient and no rejection processing is performed. If the instantaneous coherence coefficient is less than or equal to the preset instantaneous coherence coefficient threshold, the multi-dimensional transient features are matched with the typical interference feature library of the excitation system using cosine similarity to obtain the matching degree between the multi-dimensional transient features and each typical interference. If there is a typical interference whose matching degree is greater than the preset matching degree threshold, the transient event is determined to be the interference interval corresponding to the typical interference. The high signal-to-noise ratio basic signal in the interference interval is removed and the waveform is reconstructed. If there is no typical interference and the matching degree is greater than the preset matching degree threshold, the corner position of the transient event is detected based on the slip ring speed phase pulse signal. It is determined whether the transient event repeats at the same corner position within 10 consecutive frequency cycles. If it repeats, it is determined to be a real fault transient and no rejection process is performed. If it does not repeat, the transient event is determined to be an interference range, the high signal-to-noise ratio basic signal in the interference range is removed, and the waveform is reconstructed.

[0012] Furthermore, the waveform reconstruction process includes: The start and end timestamps of the interference interval are selected as the reference, and the normal waveform reference segment is extracted. The normal waveform reference segment is divided into a forward reference segment and a backward reference segment. Differential fitting is performed on the excitation current signal and carbon brush vibration acceleration signal in the normal waveform reference segment. The forward reference segment trend term and the backward reference segment trend term are extracted respectively. The forward reference segment trend term is subtracted from the high signal-to-noise ratio base signal of the forward reference segment to obtain the forward reference segment detail term. The backward reference segment detail term is obtained by subtracting the backward reference segment trend term from the high signal-to-noise ratio base signal of the backward reference segment. Based on the trend terms of the forward and backward reference segments, a continuous trend baseline is constructed within the interference interval. The cubic spline interpolation curve is superimposed on the continuous trend baseline to generate a reconstructed waveform. The detail component features of the detail terms of the forward and backward reference segments are extracted to obtain the detail components. Based on the slip ring rotation speed phase pulse signal, the normal detail component of the corresponding phase is matched for each sampling point within the interference interval to construct a detail compensation sequence for the interference interval. The detail compensation sequence is superimposed on the reconstructed waveform to generate the final reconstructed waveform.

[0013] Furthermore, the process of determining the operating conditions of the current data acquisition cycle includes: The effective value of excitation current and the active power of the unit in the contact-type multidimensional heterogeneous data are selected as the operating condition characteristic indicators. An operating condition interval mapping table is preset. The operating condition interval mapping table includes the range of operating condition characteristic indicators and the transition zone corresponding to different operating conditions. The operating condition of the current collection cycle is determined according to the operating condition characteristic indicators and the operating condition interval mapping table.

[0014] Furthermore, the process of constructing an electromagnetic-mechanical thermal coupling model, and building a fault life-cycle evolution mapping table and a personalized four-field coupling prediction model for the target unit based on the simulation test data of the electromagnetic-mechanical thermal coupling model, includes:

[0015] The electromagnetic-mechanical thermal coupling model includes a circuit field model, an electromagnetic field model, a mechanical field model, and a temperature field model.

[0016] Simulation tests were conducted on the electromagnetic motor thermal coupling model under different fault types and operating conditions to obtain simulation test data of the electromagnetic motor thermal coupling model under different fault types and operating conditions.

[0017] A fault lifecycle evolution mapping table is constructed based on simulation test data. The fault lifecycle evolution mapping table includes the mechanism core state parameter ranges of different evolution stages. The evolution stages include nonlinear degradation region, critical inflection point region and failure deterioration region.

[0018] A four-field coupling prediction model is constructed, and simulation test data is used as training data to train the four-field coupling prediction model. The trained four-field coupling prediction model is obtained. Based on the equipment characteristics of the target unit, the individual parameters of the four-field coupling prediction model are calibrated to generate a personalized four-field coupling prediction model for the target unit.

[0019] For each operating condition range, full-channel data of the target unit that has not generated a warning level within 72 hours is collected as healthy operation data. The healthy operation data is input into the personalized four-field coupling prediction model of the target unit, and a health benchmark library is constructed based on the output of the personalized four-field coupling prediction model.

[0020] Furthermore, the process of pre-identifying abrupt changes in coupling strength based on four-dimensional coupling coefficients includes:

[0021] The full-channel data of the target unit in the current acquisition cycle is input into the personalized four-field coupling prediction model. Based on the personalized four-field coupling prediction model, the four-dimensional coupling coefficient and core mechanism state parameters in the next few acquisition cycles under the current operating conditions are generated. The comprehensive coupling intensity increase and the duration of the increase of the four-dimensional coupling coefficient in the next few acquisition cycles are obtained. The coupling intensity increase threshold and time threshold are preset. If the comprehensive coupling intensity increase is greater than the coupling intensity increase threshold and the increase duration is greater than the time threshold, the degradation rate is accurately identified.

[0022] Furthermore, the process of accurately identifying the degradation rate based on the core state parameters of the mechanism includes:

[0023] Extract the mean health benchmark of the core mechanism state parameters under the current working condition from the health benchmark library, compare the core mechanism state parameters with the corresponding mean health benchmark in several future acquisition cycles, obtain the comprehensive degradation degree in each future acquisition cycle, perform least squares fitting on the comprehensive degradation degree in each future acquisition cycle, generate degradation degree time series curve, and obtain the first and second derivatives of the degradation degree time series curve.

[0024] Determine whether the second derivative changes from negative to positive and remains positive. If it does, mark the target unit as a nonlinear deterioration zone and perform a critical inflection point risk assessment.

[0025] If the second derivative peaks and then falls back, but the first derivative continues to grow exponentially, the target unit is marked as a failure deterioration zone, triggering shutdown protection.

[0026] Furthermore, the process of conducting a critical inflection point risk assessment and generating an early warning level based on the assessment results includes:

[0027] Based on the fault lifecycle evolution mapping table, the critical benchmarks of the core mechanism state parameters in the critical inflection point region are obtained. Based on the average values ​​of the core mechanism state parameters, critical benchmarks, and health benchmarks in each future acquisition cycle, the safety margins in each future acquisition cycle are obtained. The warning levels corresponding to different safety margin intervals are preset. Based on the safety margins in each future acquisition cycle and the warning levels corresponding to different safety margin intervals, the warning levels in each future acquisition cycle are determined, and the preset operation and maintenance strategies corresponding to the warning levels are executed.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] I. This invention completely abandons the limitations of existing technologies that only monitor the surface phenomena of faults. For the first time, it starts from the physical essence of carbon brush-slip ring fault evolution and constructs a complete multi-physics field coupled quantitative model. It fully reveals the positive feedback closed-loop evolution mechanism of faults from initial deterioration to final failure, breaking the core constraint of existing technologies that "lack mechanistic support and rely solely on the simple superposition of multiple parameter thresholds." All monitoring, analysis, and early warning logics of this invention are rigidly constrained by the inherent physical laws of faults, completely solving the core problem of existing technologies that cannot distinguish between normal operating condition fluctuations and early fault evolution. At the same time, it avoids the black-box defects of pure data-driven models, upgrading early warning decisions from "unfounded empirical threshold judgments" to "scientific quantitative decisions based on fault evolution laws," providing a solid and interpretable underlying mechanistic support for the entire early warning system.

[0030] Second, this invention breaks through the core limitations of existing single-parameter monitoring technologies by constructing a multi-physics, multi-dimensional heterogeneous sensing system. It retains the core coupling characteristics of excitation current and carbon brush vibration while supplementing it with multi-dimensional non-contact redundant sensing parameters. This enables it to capture early, weak fault signals that existing technologies cannot identify, especially early-stage problems such as intermittent poor contact without significant steady-state parameter changes. It completely changes the passive mode of existing technologies, where faults are only detected when they reach severe stages like brush burnout or ring fire. It enables accurate identification at the fault's nascent stage, providing ample time for maintenance and repair, and mitigating equipment damage and unit operation risks caused by the continuous deterioration of faults. This truly achieves the goal of proactive early warning, preventing problems before they occur.

[0031] Third, this invention addresses the full operating characteristics of generator sets by accurately dividing operating ranges based on the core operating parameters of the generator set. It also constructs a matching health benchmark library and adaptive early warning logic for different operating ranges, completely abandoning the rigid early warning mode with fixed thresholds in existing technologies.

[0032] IV. This invention addresses the harsh operating environment of excitation systems, characterized by strong electromagnetic fields and vibrations. It constructs a full-link signal quality assurance system, starting from multi-channel timing synchronization calibration to ensure the homogeneity and timing accuracy of multi-source heterogeneous data. Simultaneously, through progressive noise reduction processing, precise identification of transient interference, and waveform fidelity reconstruction techniques, it achieves accurate separation of interference signals from actual fault characteristics. This invention effectively eliminates various steady-state and transient interference signals while completely preserving the subtle characteristics of actual faults, thoroughly solving the problem of false early warnings caused by signal distortion in complex field environments. This allows the early warning system to operate stably and reliably for extended periods in the harsh industrial environment of power plants, demonstrating strong practicality for engineering applications and avoiding the shortcomings of existing technologies that perform well in the laboratory but poorly in field applications. Attached Figure Description

[0033] Figure 1 This is a flowchart of a fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration, according to an embodiment of this application. Detailed Implementation

[0034] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] like Figure 1 As shown, a fault early warning system for the coupling characteristics of excitation DC high current and carbon brush vibration includes a monitoring center, which is connected to a multi-parameter acquisition unit, a data processing module, a model building module and a fault early warning module.

[0036] The multi-parameter acquisition unit is a microsecond-level synchronous acquisition unit, used to synchronously acquire all channels of data of the target unit and set the acquisition period (5us).

[0037] The data processing module is used to perform synchronous timing calibration and abnormal timing processing on the full-channel data, as well as noise reduction preprocessing. It identifies and removes transient interference from the noise-reduced full-channel data, reconstructs the waveform of the removed interference segment, and determines the operating conditions of the current acquisition cycle.

[0038] The model building module is used to build an electromagnetic thermal coupling model, and to build a fault life cycle evolution mapping table and a personalized four-field coupling prediction model for the target unit based on the simulation test data of the electromagnetic thermal coupling model.

[0039] The fault early warning module is used to output the four-dimensional coupling coefficient and core mechanism state parameters under the current operating conditions based on the personalized four-field coupling prediction model. Based on the four-dimensional coupling coefficient, it performs pre-identification of coupling strength abrupt changes. Based on the pre-identification results, it determines whether to perform accurate identification of degradation rate based on the core mechanism state parameters. Based on the accurate identification results, it determines whether to perform critical inflection point risk assessment or trigger shutdown protection. Based on the critical inflection point risk assessment results, it generates an early warning level.

[0040] It should be further explained that, in the specific implementation process, the acquisition channels of the multi-parameter acquisition unit (all signal channels that are synchronously acquired in the system) include: the total excitation current channel (1 channel each for positive and negative buses), the branch current channel for each carbon brush (if a single-pole slip ring is equipped with 24 carbon brushes, 24 carbon brushes = 24 channels), the vibration acceleration channel for each carbon brush holder (24 = 24 channels), the speed / phase pulse channel (1 channel), the non-contact acquisition channel (3 channels, namely the laser Doppler vibration meter channel, the infrared thermal imaging array channel, and the ultra-high frequency (UHF) antenna channel), and the unit operating condition channel (DCS acquisition: active power, reactive power, and excitation voltage).

[0041] The full-channel data includes contact-based multidimensional heterogeneous data and contactless redundant data;

[0042] The contact-type multidimensional heterogeneous data includes the total current of the excitation circuit, the full waveform data of the current of each carbon brush branch, the vibration acceleration signal of each carbon brush holder, the slip ring speed phase pulse signal, and the unit operating parameters (excitation bus voltage, ambient temperature, cumulative carbon brush running time, unit active power, etc.), and the sampling synchronization accuracy of the current and vibration signals is ≤ ±1μs.

[0043] Non-contact redundant data includes radial vibration of the carbon brush end face, temperature field distribution of the slip ring's entire surface and the contact surface with the carbon brush, and ultra-high frequency partial discharge signals.

[0044] The multi-parameter acquisition unit includes contact sensors and non-contact sensors;

[0045] Contact sensors are:

[0046] A high-precision Hall current sensor group, including the main sensor of the total excitation circuit and the branch sensor of each carbon brush branch, with a sensor bandwidth ≥200kHz, measurement accuracy ≥0.2 grade, and sampling rate ≥100kHz / channel;

[0047] IEPE type accelerometer group is installed on each carbon brush holder body. The sensor bandwidth is ≥50kHz, the range is ≥±50g, and the sampling rate is completely synchronized with the current sensor.

[0048] A photoelectric speed encoder, mounted on the coaxial end of the slip ring, with a resolution of ≥1024 lines, is used to provide the slip ring frequency reference and phase reference;

[0049] The multi-channel synchronous acquisition card has a resolution of 16 bits or higher ADC, supports full-channel synchronous sampling, and has a synchronous triggering accuracy of ≤±1μs. It also supports the interface with the unit's DCS system to collect active power, reactive power, and excitation voltage operating data.

[0050] Non-contact sensors are:

[0051] Laser Doppler Vibration Measurement (LDV): Non-contact measurement of radial vibration of carbon brush end face, directly capturing micron-level runout at the contact interface between carbon brush and slip ring, avoiding installation stress and lifespan issues of contact sensors;

[0052] Infrared thermal imaging array: A non-contact measurement array of the temperature field distribution between the entire surface of the slip ring and the carbon brush contact surface, capturing local hot spots caused by poor contact, with a resolution of up to 0.1℃;

[0053] Ultra-high frequency (UHF) antenna: It collects ultra-high frequency signals of partial discharge caused by poor carbon brush contact, and captures the transient characteristics of contact discharge at the nanosecond level. It is the most sensitive characteristic parameter of early contact failure.

[0054] It should be further explained that, in the specific implementation process, the process of synchronizing timing calibration and handling abnormal timing of data across all channels includes:

[0055] An absolute time base is established for all channel data (using the Z-phase zero-position pulse of the 1024-line photoelectric encoder coaxially mounted on the slip ring as the system's sole absolute time base; this pulse is output once per revolution, corresponding to a fixed mechanical phase of the slip ring, unaffected by electromagnetic interference or load changes, providing a unified time anchor point for all channel data). Full-channel delay calibration is performed on all acquisition channels. Specifically, during system power-on initialization, a standard synchronous trigger pulse is output through the synchronous acquisition card to calibrate the end-to-end delay (including cable transmission delay, sensor response delay, and acquisition card channel quantization delay) of each current sensor and vibration sensor. A 100ns-level precision end-to-end delay compensation table is established to eliminate inherent delay differences caused by different sensor types and cable lengths. Based on the absolute time base and the end-to-end delay compensation table, the end-to-end delay of each acquisition cycle is calibrated. Synchronous timing calibration is performed to ensure that the timing synchronization error between the processed current signal and vibration signal is stable at ≤±1μs, thus fundamentally avoiding the distortion of coupling characteristic calculation caused by timing mismatch. After synchronous timing calibration, sampling point loss is determined for all channel data. Specifically, for any two adjacent valid sampling points in any channel, the actual timestamp difference is calculated. If the actual timestamp difference is greater than 7.5µs, the latter of the two valid sampling points is determined as a single-point sampling loss. Then, based on the sampling point loss determination result, abnormal data segments are marked. Abnormal data segments include a first abnormal segment and a second abnormal segment. Specifically, after obtaining the first single-point sampling loss point n1, the effective sampling point n-end with an interval of 5µs between it and the previous valid sampling point is obtained. The time interval between the first single-point sampling loss point n1 and the effective sampling point n-end is also obtained. ,like If the time exceeds 10ms, all valid sampling points between the single-point sampling lost point n1 and the valid sampling point n-end are marked as the second abnormal segment. Otherwise, all valid sampling points between the single-point sampling lost point n1 and the valid sampling point n-end are marked as the first abnormal segment. The first abnormal segment is then filled by interpolation of adjacent normal data segments, and the second abnormal segment is removed.

[0056] It should be further explained that, in the specific implementation process, the process of noise reduction preprocessing of the entire channel data and the identification and removal of transient interference includes:

[0057] Establish a typical interference feature library for the excitation system (rectifier commutation ripple, plant power switching impulse, lightning transient interference). The typical interference feature library for the excitation system includes multi-dimensional transient features corresponding to different typical interferences.

[0058] Steady-state periodic interference and random noise reduction preprocessing are performed on the contact-type multidimensional heterogeneous data in the full channel data to obtain high signal-to-noise ratio basic signals, including excitation current signals and carbon brush vibration acceleration signals.

[0059] The steady-state periodic disturbance noise reduction preprocessing includes:

[0060] A second-order IIR adaptive notch filter bank is adopted. The notch center frequency is dynamically adjusted based on the real-time acquisition of the generator terminal voltage frequency via the unit's operating condition channel, covering core interference frequencies such as 50Hz, 100Hz, 150Hz, 200Hz, and 300Hz. The notch depth is ≥60dB, and the passband ripple is ≤0.5dB. This eliminates steady-state periodic interference (50Hz fundamental frequency from the power grid, 6-pulse / 12-pulse rectifier ripple from the excitation rectifier bridge, and power frequency vibration crosstalk from the generator unit itself) that causes early fault coupling characteristics in the submerged low-frequency band.

[0061] Adaptive adjustment of operating conditions: When the unit is in no-load / start-stop conditions, the filter order is automatically increased to enhance the ability to suppress rectified ripple; when in rated load conditions, the order is automatically reduced to reduce damage to fault characteristics, balancing noise reduction effect and real-time performance.

[0062] Random noise reduction preprocessing includes:

[0063] Based on the sampling rate and slip ring frequency under the current operating conditions, the number of modes K and the penalty factor α of the variational mode decomposition (VMD) are adaptively set; K=8 and α=2000 under rated operating conditions, and K=10 and α=2500 under no-load operating conditions. For the K intrinsic mode functions (IMFs) obtained from the variational mode decomposition, three core screening criteria are calculated: ① correlation coefficient with the original signal ≥0.1; ② kurtosis value ≥3 (including impulse characteristics); ③ normalized energy entropy ≤0.7. Pure noise spurious components that do not meet the conditions are removed. For the effective IMF components after screening, a 3-level wavelet decomposition is performed using the db4 wavelet basis, and an adaptive soft thresholding function is used to process the high-frequency detail coefficients. The threshold formula is as follows: , where σ is the noise standard deviation of the median absolute deviation estimate of detail coefficients, and N is the number of sampling points; the effective IMF component after noise reduction is superimposed with the low-frequency trend component to reconstruct the high signal-to-noise ratio basic signal.

[0064] Steady-state periodic disturbances and random noise denoising preprocessing cannot solve strong electromagnetic transient disturbances at the μs level, which are characterized by large amplitude, short duration, and non-periodicity.

[0065] A sliding detection window (window length 100 ns, step size 10 μs) is set, and multi-dimensional transient feature detection is performed on the high signal-to-noise ratio fundamental signal within the sliding detection window, including:

[0066] For each carbon brush branch current and each carbon brush vibration acceleration, calculate the amplitude abrupt change rate and rise time slope separately. The formula for the amplitude abrupt change rate is: The formula for the rising slope is: ,in, The value is the current (A) or vibration (m / s²) at the current sampling point. This represents the valid value of the channel within the normal 10ms window. The amplitude change rate of the current sampling point is denoted by m, which is the slope. The number of sampling point intervals is calculated and is set to 3. Ts is the sampling period.

[0067] The excitation current waveform and the carbon brush vibration waveform were decomposed into 4-level db6 wavelet decompositions using the db6 wavelet basis.

[0068] The energy entropy was calculated for the time-frequency spectrum of the excitation current and the time-frequency spectrum of the carbon brush vibration, and the abrupt change in energy entropy in the time-frequency domain was statistically analyzed. , ,in, Energy entropy within the first 10ms The energy entropy within the current sliding detection window;

[0069] Acquire multi-dimensional transient features, including amplitude mutation rate, rising edge slope, modulus maxima, and time-frequency domain energy entropy mutation. Preset transient thresholds for each transient feature in the multi-dimensional transient features. If any transient feature is greater than the corresponding transient threshold, the detection window is determined to be a transient event.

[0070] Transient events are detection windows that meet any of the following triggering conditions:

[0071] 1. Amplitude mutation rate ≥5. For current signals ≥1× A / s, for vibration signals ≥1× g / s;

[0072] 2. Wavelet modulus maxima alignment across layers: Aligned modulus maxima appear across three consecutive decomposition scales, and the modulus maxima are greater than a threshold. For example, high-frequency details in layer 1. The modulus maxima appear, level 2 detail. Simultaneous occurrence of modulo maxima, level 3 details Simultaneous occurrence of modulo maxima, and Maximum modulus ≥ 0.1 × rated current Maximum modulus ≥ 0.08 × rated current; Maximum modulus ≥ 0.05 × rated current;

[0073] 3. Sudden change in time-frequency energy entropy ≥2.

[0074] When a transient event occurs, the instantaneous coherence coefficients of the excitation current signal and the carbon brush vibration acceleration signal within the detection window are obtained (calculated using the Pearson correlation coefficient). The instantaneous coherence coefficients are compared with the preset instantaneous coherence coefficient threshold (0.8). If the instantaneous coherence coefficient is greater than the preset instantaneous coherence coefficient threshold, the transient event is determined to be a real fault transient. The waveform of the high signal-to-noise ratio basic signal is completely preserved without any rejection processing.

[0075] If the instantaneous coherence coefficient is less than or equal to the preset instantaneous coherence coefficient threshold, the multi-dimensional transient features are matched with the typical interference feature library of the excitation system using cosine similarity to obtain the matching degree between the multi-dimensional transient features and each typical interference. If there is a typical interference whose matching degree is greater than the preset matching degree threshold (85%), the transient event is determined to be the interference interval corresponding to the typical interference. The high signal-to-noise ratio basic signal in the interference interval is removed and the waveform is reconstructed.

[0076] If there is no typical interference and the matching degree is greater than the preset matching degree threshold, the corner position of the transient event is detected based on the slip ring speed phase pulse signal. It is determined whether the transient event repeats at the same corner position within 10 consecutive frequency cycles. If it repeats, it is determined to be a real fault transient. The waveform of the high signal-to-noise ratio basic signal is completely preserved without any rejection processing. If it does not repeat, the transient event is determined to be an interference interval. The high signal-to-noise ratio basic signal in the interference interval is removed and the waveform is reconstructed.

[0077] It should be further explained that, in the specific implementation process, the waveform reconstruction process includes:

[0078] Using the start and end timestamps of the interference interval as a reference, a normal waveform reference segment is extracted. The normal waveform reference segment is divided into a forward reference segment (the high signal-to-noise ratio basic signal before the interference interval) and a backward reference segment (the high signal-to-noise ratio basic signal after the interference interval). For the extracted forward and backward reference segments, two rigid checks are performed: the amplitude gradient abrupt change rate is less than 3; there are no consecutive single-point sampling loss points; if the check fails, the reference segment is extended in a direction away from the interference interval until the check passes. The minimum length of the forward / backward reference segment is 100 valid sampling points.

[0079] Differential fitting is performed on the excitation current signal and carbon brush vibration acceleration signal in the normal waveform reference segment. For the excitation current signal: weighted least squares linear fitting is used, with the amplitude of the sampling point in the reference segment as input and the distance from the interference interval as weight (the closer to the interference interval, the higher the weight) to fit the linear trend term.

[0080] For carbon brush vibration signals: a moving average polynomial fitting method is used to fit the frequency trend term of the vibration signal based on the slip ring rotation speed phase pulse signal;

[0081] After fitting, the forward reference segment trend term T1(n) and the backward reference segment trend term T2(n) are extracted respectively. The forward reference segment trend term is subtracted from the high signal-to-noise ratio basic signal X1(n) of the forward reference segment to obtain the forward reference segment detail term D1(n), where D1(n) = X1(n) - T1(n). The backward reference segment trend term is subtracted from the high signal-to-noise ratio basic signal X2(n) of the backward reference segment to obtain the backward reference segment detail term D2(n), where D2(n) = X2(n) - T2(n).

[0082] Based on the trend terms of the forward and backward reference segments, a continuous trend baseline is constructed within the interference interval. Using the trend values ​​at the start and end timestamps of the interference interval as boundaries, linear interpolation is employed to construct the trend baseline within the interference interval. This ensures that the trend baseline is completely continuous with the trends of the preceding and following normal signals at the start and end timestamps, without any abrupt changes. Cubic spline interpolation is then performed on the interference interval to obtain the cubic spline interpolation curve within the interference window, including:

[0083] Set the left boundary (start timestamp): The first derivative of the interpolation function is equal to the average slope of the last 10 sampling points of the forward reference segment, and the second derivative is equal to the average curvature of the last 10 sampling points of the forward reference segment.

[0084] Set the right boundary (end timestamp): The first derivative of the interpolation function is equal to the average slope of the first 10 sampling points of the backward reference segment, and the second derivative is equal to the average curvature of the first 10 sampling points of the backward reference segment.

[0085] Interpolation node construction: Using the start and end timestamps of the interference interval as fixed boundary nodes, interpolation nodes are set at equal intervals within the interference window. The node density is adaptively adjusted according to the interference duration: when the interference duration is ≤50μs, the node interval is 2 sampling points; when the interference duration is >50μs, the node interval is 5 sampling points, balancing interpolation accuracy and calculation speed.

[0086] Based on boundary conditions and interpolation nodes, the three bending moment equations are solved to obtain the cubic spline interpolation curve within the disturbance window;

[0087] The cubic spline interpolation curve is superimposed on the continuous trend baseline to generate a reconstructed waveform. The detail component features of the forward and backward reference segments are extracted to obtain the detail components. Based on the slip ring rotation speed phase pulse signal, the normal detail component of the corresponding phase is matched for each sampling point in the interference interval to construct the detail compensation sequence of the interference interval. The detail compensation sequence is superimposed on the reconstructed waveform to generate the final reconstructed waveform.

[0088] The feature extraction of the reference segment details includes:

[0089] For the excitation current signal: the fundamental frequency, amplitude, phase, and normal high-frequency harmonic components of the rectified ripple are extracted by Fast Fourier Transform (FFT);

[0090] For carbon brush vibration signals: through order spectrum analysis, the amplitude and phase of the slip ring frequency from 1st to 10th harmonics, as well as the vibration components at the natural frequency of the carbon brush holder, are extracted;

[0091] Simultaneously calculate the average energy, kurtosis, and peak factor of the detail components.

[0092] It should be further explained that, in the specific implementation process, the process of determining the operating conditions of the current data collection cycle includes:

[0093] The effective value of excitation current and the active power of the unit in the contact-type multidimensional heterogeneous data are selected as the operating condition characteristic indicators. An operating condition interval mapping table is preset. The operating condition interval mapping table includes the range of operating condition characteristic indicators and the transition band corresponding to different operating conditions. Taking the actual operating characteristics of a 600MW thermal power unit as an example, four operating condition intervals are calibrated, and a transition band is set to avoid the reference jump caused by frequent switching of operating conditions, as shown in Table 1 below:

[0094] Table 1

[0095]

[0096] The operating condition of the current acquisition cycle is determined based on the operating condition characteristic indicators and the operating condition interval mapping table. Specifically, the average value of the two operating condition characteristic indicators is calculated in real time using a 1s sliding window and a 0.2s step size. When the average value of three consecutive windows falls into the same operating condition interval and does not exceed the transition zone, the operating condition interval of the current acquisition cycle is determined.

[0097] It should be further explained that, in the specific implementation process, the process of constructing an electromagnetic motor thermal coupling model, and constructing a fault life-cycle evolution mapping table and a personalized four-field coupling prediction model for the target unit based on the simulation test data of the electromagnetic motor thermal coupling model includes:

[0098] A thermal coupling model of an electromagnet was constructed using COMSOL Multiphysic finite element software. The thermal coupling model of the electromagnet includes a circuit field model, an electromagnetic field model, a mechanical field model, and a temperature field model.

[0099] Among them, the circuit field model is , ;

[0100] in, Excitation bus voltage; : No. The branch current of the carbon brush is acquired in real time by the multi-parameter acquisition unit; : No. The body resistance of the carbon brush; : No. Dynamic contact resistance of carbon brushes; Dynamic contact resistance; The resistivity of carbon brush and slip ring materials; : Brinell hardness of the contact material; : The contact pressure between the carbon brush and the slip ring; Temperature correction factor, which takes into account the effect of temperature rise in the contact area on resistivity and hardness; : Contact gap correction factor, taking into account instantaneous contact separation caused by carbon brush vibration. The contact gap is output from the mechanical field model;

[0101] Electromagnetic field model is ;

[0102] in, The instantaneous electromagnetic force experienced by the carbon brush; The magnetic induction intensity of the excitation circuit at the carbon brush mounting location was calibrated using finite element simulation. : Effective current-carrying length of carbon brush; The instantaneous current of the carbon brush branch is output by the circuit field model; The dynamic electrodynamic coefficient is determined by the inductance parameters of the carbon brush-brush holder circuit. Rate of change of current, in units .

[0103] Mechanical field model is , ;

[0104] In the formula: These represent the mass matrix, damping matrix, and stiffness matrix of the carbon brush-brush holder system, where the stiffness... Determined by the spring stiffness, it is corrected for thermal decay as the temperature field rises; Radial / axial displacement vector of carbon brush For speed, For acceleration; : The contact pressure between the carbon brush and the slip ring; Carbon brush wear volume, unit ; The wear coefficient is related to the carbon brush material, slip ring surface condition, and temperature. : Linear velocity of slip ring surface, unit The data is acquired in real time by a multi-parameter acquisition unit.

[0105] The temperature field model is ;

[0106] In the formula: These represent the material's density, specific heat capacity, and thermal conductivity, respectively, all of which change with temperature. Nonlinear change; Instantaneous temperature of the contact area; Joule heat flux density generated by contact resistance, in units The output from the circuit field model calculate; Heat flux density generated by the sliding friction between the carbon brush and the slip ring, in units of The contact pressure is calculated from the sliding speed and the output of the mechanical field model.

[0107] The circuit field model, electromagnetic field model, mechanical field model, and temperature field model are bidirectionally coupled through core output variables, forming a complete closed-loop solution system. The coupling interfaces are shown in Table 2 below:

[0108] Table 2

[0109]

[0110] The inputs for the circuit field model, electromagnetic field model, mechanical field model, and temperature field model are full-channel data.

[0111] Simulation tests were conducted on the electromagnetic motor thermal coupling model under different fault types and operating conditions to obtain simulation test data (including full-channel data and core output variables of the four field models) of the electromagnetic motor thermal coupling model under different fault types and operating conditions.

[0112] Based on simulation test data, the entire life cycle of carbon brush-slip ring failure is divided into four quantitative stages, and a failure life cycle evolution mapping table is constructed. The failure life cycle evolution mapping table includes the mechanism core state parameter intervals of different evolution stages. The evolution stages include the nonlinear degradation region, the critical inflection point region, and the failure deterioration region, as shown in Table 3 below:

[0113] Table 3

[0114]

[0115] A four-field coupling prediction model was constructed, and simulation data was used as training data to train the model, resulting in a trained model. The network structure of the four-field coupling prediction model consists of three fully connected hidden layers, with 32-16-8 neurons per layer. The activation function is Tanh, and the overall parameter count is ≤2000, significantly smaller than that of conventional deep learning models, making it suitable for edge computing power. The circuit field model, electromagnetic field model, mechanical field model, and temperature field model were used as regularization terms in the loss function, and the loss function formula is as follows: ,in To compensate for the data fitting loss, we train the model using simulation test data to ensure the model's fitting accuracy. To compensate for the loss in the physical equations, the output of the constraint model must conform to the physical laws of the four field models to avoid non-physical interpretations.

[0116] Based on the equipment characteristics of the target unit, the four-field coupled prediction model is calibrated with individual parameters, including:

[0117] 1. Basic parameter calibration: Based on the unit design drawings and maintenance survey data, input the inherent parameters such as carbon brush / slip ring material parameters, dimensions, spring stiffness, brush holder installation parameters, and excitation circuit inductance / resistance;

[0118] 2. Operating parameter calibration: Collect 72 hours of fault-free healthy operation data of the unit, and use the recursive least squares method to calibrate slowly changing parameters such as wear coefficient, thermal attenuation coefficient, and contact resistance correction coefficient;

[0119] Generate a personalized four-field coupled prediction model for the target unit;

[0120] For each operating condition interval, all channel data of the target unit that has not generated a warning level within 72 hours are collected as healthy operation data. The healthy operation data is input into the personalized four-field coupling prediction model of the target unit. A health benchmark library is constructed based on the output of the personalized four-field coupling prediction model (a Gaussian mixture model (GMM) health benchmark is constructed with 5 model components to fit the multidimensional probability distribution of features under healthy conditions, and the 95% confidence interval boundary is calibrated).

[0121] It should be further explained that, in the specific implementation process, the process of pre-identifying abrupt changes in coupling strength based on the four-dimensional coupling coefficient includes:

[0122] The full-channel data of the target unit in the current acquisition cycle is input into the personalized four-field coupling prediction model. Based on the personalized four-field coupling prediction model, the four-dimensional coupling coefficient and core mechanism state parameters (dynamic contact resistance) for several future acquisition cycles under the current operating conditions are generated. Alternating electromagnetic excitation force Contact pressure (Contact area temperature T), the four-dimensional coupling coefficient includes the frequency domain coherence coefficient of the current-vibration signal. Pearson correlation coefficient between current and temperature rise Pearson correlation coefficient between vibration and temperature rise Phase synchronization rate of current-discharge pulses To obtain the increase in the overall coupling strength of the four-dimensional coupling coefficient over several future acquisition cycles (by...) , , , The summation and averaging process generates a comprehensive coupling strength. The increase in comprehensive coupling strength within 10 minutes is statistically analyzed, and the increase in comprehensive coupling strength and the duration of the increase (the cumulative time during which comprehensive coupling strength continues to rise without any downward trend) are obtained. A preset coupling strength increase threshold (200%) and a time threshold (30 minutes) are set. If the increase in comprehensive coupling strength is greater than the coupling strength increase threshold and the duration of the increase is greater than the time threshold, then the degradation rate is accurately identified.

[0123] It should be further explained that, in the specific implementation process, the process of accurately identifying the degradation rate based on the core state parameters of the mechanism includes:

[0124] The mean health benchmark values ​​of the core mechanism state parameters under the current operating conditions are extracted from the health benchmark library. The core mechanism state parameters for several future acquisition cycles are compared with the corresponding mean health benchmark values ​​to obtain the overall degradation degree D for each future acquisition cycle (calculated using Mahalanobis distance; the greater the deviation between the core mechanism state parameters and the mean health benchmark, the greater the overall degradation degree). The overall degradation degree for each future acquisition cycle is then fitted using the least squares method to generate a degradation degree time series curve, and the first derivative of the degradation degree time series curve is obtained. (Deterioration rate) and second derivative (Deterioration acceleration);

[0125] Determine whether the second derivative changes from negative to positive and remains positive. If it does, mark the target unit as a nonlinear deterioration zone and perform a critical inflection point risk assessment.

[0126] Determine if the second derivative peaks and then falls back, but the first derivative continues to grow exponentially. If so, mark the target unit as the failure deterioration zone (has exceeded the critical inflection point), trigger shutdown protection, immediately reduce load, and arrange emergency shutdown for maintenance.

[0127] The core characteristic of carbon brush-slip ring fire accidents is "long-term slow brewing and instantaneous outbreak": before breaking through the critical inflection point, the deterioration rate is extremely slow and the increase in the overall deterioration degree D is very small; once the critical inflection point is broken, the deterioration will enter an exponential acceleration stage, which can take as little as 8 hours, leaving a very short window for operation and maintenance to deal with it.

[0128] Current technology relies solely on D for early warning, and alarms are only triggered when D rises to a fixed threshold. By this time, the problem is often already close to a fire, or even in an irreversible deterioration phase. There is simply not enough time to schedule maintenance, leaving only emergency shutdowns and causing economic losses in the millions. Furthermore, the health baseline of the unit is not static: as the years of operation increase, normal wear on the slip ring surface, aging of the carbon brush holder, and drift of excitation circuit parameters will cause the baseline D under healthy conditions to change.

[0129] Identifying the critical inflection point of degradation rate is based on "a sudden change in trend" rather than a fixed numerical threshold: no matter how much the unit ages, as long as the second derivative of degradation turns from negative to positive, it means that the system has changed from self-stabilization to self-deterioration. There is no need to frequently calibrate the threshold, and the unit can operate stably throughout its entire life cycle.

[0130] When the second derivative < 0: the degradation rate continues to decrease, the system is in a negative feedback self-stabilizing state, the degradation is controllable and recoverable, and even if the value of D is temporarily high, there is no urgent risk.

[0131] The second derivative turns from negative to positive and remains positive: the rate of degradation changes from decelerating to accelerating, the system breaks through the critical inflection point, enters a positive feedback self-deterioration state, and the degradation is irreversible. Even if the value of D has not yet reached the red alert threshold, it must be dealt with immediately.

[0132] This directly addresses the core pain point that D alone cannot distinguish between "normal fluctuations / recoverable degradation" and "irreversible failures," eliminating false alarms and missed alarms at the source.

[0133] After the critical inflection point appears, the system has entered the positive feedback accelerated deterioration stage. By measuring the deterioration acceleration, it can be accurately predicted that D will break through the critical failure threshold within 3 to 7 days, thus locking in the risk of fire in advance. This extends the early warning and handling window from the original "only 8 hours left for emergency shutdown" based solely on D's alarm to "scheduled maintenance 7 days in advance," completely avoiding the million-level economic losses caused by unplanned shutdowns and fully meeting the power industry's safety requirement of "prevention is better than cure."

[0134] It should be further explained that, in the specific implementation process, the process of conducting a critical inflection point risk assessment and generating an early warning level based on the assessment results includes:

[0135] Based on the fault lifecycle evolution mapping table, the critical benchmarks of the core mechanism state parameters in the critical inflection point region are obtained. Based on the average values ​​of the core mechanism state parameters, critical benchmarks, and health benchmarks in each future acquisition cycle, the safety margins in each future acquisition cycle are obtained. The warning levels corresponding to different safety margin intervals are preset. Based on the safety margins in each future acquisition cycle and the warning levels corresponding to different safety margin intervals, the warning levels in each future acquisition cycle are determined, and the preset operation and maintenance strategies corresponding to the warning levels are executed.

[0136] The specific process for obtaining the safety margin is as follows:

[0137] Based on the fault lifecycle evolution mapping table, the relative distance between the current feature vector and the critical failure state is calculated using Mahalanobis distance, and normalized to obtain a safety margin SM of 0~100%.

[0138] ;

[0139] in, These are the core state parameters of the mechanism under the current operating conditions. It serves as the critical benchmark for the core state parameters of the mechanism. is the health baseline mean of the core state parameters of the mechanism, and MD is the Mahalanobis distance calculation function;

[0140] When SM≥50%: No warning level; 20%≤SM<50%: Warning level is yellow, the inflection point is approaching, and monitoring should be strengthened; SM<20%: Warning level is red, the critical inflection point zone has been entered, and maintenance should be arranged immediately.

[0141] By providing a precise "intervention window" for proactive intervention and control, a fundamental upgrade from "passive alarm" to "proactive prevention and control" is achieved:

[0142] Intervention window: Before the inflection point occurs, the system is still in a negative feedback self-stabilizing state. At this time, by adjusting the PI parameters of the excitation regulator, reducing the reactive load, and reducing the excitation current, the potential positive feedback loop can be broken, the equipment can be restored to a healthy state, and downtime can be completely avoided.

[0143] Uninterventional window: After the inflection point appears, the positive feedback loop has been fully triggered. At this time, online intervention will not only fail to suppress the degradation, but may also exacerbate the electrodynamic excitation due to current fluctuations, thus accelerating the deterioration of the fault. The only option is to shut down for maintenance.

[0144] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration, characterized in that, This includes a monitoring center, which has communication connections to a multi-parameter acquisition unit, a data processing module, a model building module, and a fault early warning module. The multi-parameter acquisition unit is used to synchronously acquire all channels of data from the target unit and set the acquisition cycle; The data processing module is used to perform synchronous timing calibration and abnormal timing processing on the full-channel data, as well as noise reduction preprocessing. It identifies and removes transient interference from the noise-reduced full-channel data, reconstructs the waveform of the removed interference segment, and determines the operating conditions of the current acquisition cycle. The model building module is used to build an electromagnetic thermal coupling model, and to build a fault life cycle evolution mapping table and a personalized four-field coupling prediction model for the target unit based on the simulation test data of the electromagnetic thermal coupling model. The fault early warning module is used to output the four-dimensional coupling coefficient and core mechanism state parameters under the current operating conditions based on the personalized four-field coupling prediction model. Based on the four-dimensional coupling coefficient, it performs pre-identification of coupling strength abrupt changes. Based on the pre-identification results, it determines whether to perform accurate identification of degradation rate based on the core mechanism state parameters. Based on the accurate identification results, it determines whether to perform critical inflection point risk assessment or trigger shutdown protection. Based on the critical inflection point risk assessment results, it generates an early warning level.

2. The fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration according to claim 1, characterized in that, The acquisition channels of the multi-parameter acquisition unit include: the total excitation current channel, the branch current channel of each carbon brush, the vibration acceleration channel of each carbon brush holder, the speed / phase pulse channel, the non-contact acquisition channel, and the unit operating condition channel. The full-channel data includes contact-based multidimensional heterogeneous data and contactless redundant data; The contact-type multidimensional heterogeneous data includes the total current of the excitation circuit, the full waveform data of the current of each carbon brush branch, the vibration acceleration signal of each carbon brush holder, the slip ring speed phase pulse signal, and the unit operating parameters. The non-contact redundant data includes the radial vibration of the carbon brush end face, the temperature field distribution of the slip ring surface and the carbon brush contact surface, and the partial discharge ultra-high frequency signal.

3. The fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration according to claim 2, characterized in that, The process of synchronizing timing calibration and handling abnormal timing for all channels of data includes: An absolute time base is established for all channel data, and full-channel delay calibration is performed on all acquisition channels. A full-link delay compensation table is established. Based on the absolute time base and the full-link delay compensation table, synchronous timing calibration is performed on the full channel data for each acquisition cycle. Sampling loss points are determined for the full channel data after synchronous timing calibration. Abnormal data segments are marked based on the sampling loss point determination results. Abnormal data segments include a first abnormal segment and a second abnormal segment. The first abnormal segment is interpolated and completed by adjacent normal data segments, and the second abnormal segment is removed.

4. The fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration according to claim 3, characterized in that, The process of noise reduction preprocessing and transient interference identification and removal for all channels of data includes: Establish a typical disturbance feature library for the excitation system, which includes multidimensional instantaneous features corresponding to different typical disturbances; Steady-state periodic interference and random noise reduction preprocessing are performed on the contact-type multidimensional heterogeneous data in the full channel data to obtain high signal-to-noise ratio basic signals, including excitation current signals and carbon brush vibration acceleration signals. A sliding detection window is set up, and multi-dimensional transient features are detected on the high signal-to-noise ratio basic signal within the sliding detection window to obtain multi-dimensional transient features, including amplitude change rate, rising edge slope, modulus maxima, and time-frequency domain energy entropy change. A transient threshold corresponding to each transient feature in the multi-dimensional transient features is preset. If any transient feature is greater than the corresponding transient threshold, the detection window is determined to be a transient event. When a transient event occurs, the instantaneous coherence coefficients of the excitation current signal and the carbon brush vibration acceleration signal within the detection window are obtained. The instantaneous coherence coefficients are compared with the preset instantaneous coherence coefficient threshold. If the instantaneous coherence coefficient is greater than the preset instantaneous coherence coefficient threshold, the transient event is determined to be a real fault transient and no rejection processing is performed. If the instantaneous coherence coefficient is less than or equal to the preset instantaneous coherence coefficient threshold, the multi-dimensional transient features are matched with the typical interference feature library of the excitation system using cosine similarity to obtain the matching degree between the multi-dimensional transient features and each typical interference. If there is a typical interference whose matching degree is greater than the preset matching degree threshold, the transient event is determined to be the interference interval corresponding to the typical interference. The high signal-to-noise ratio basic signal in the interference interval is removed and the waveform is reconstructed. If there is no typical interference and the matching degree is greater than the preset matching degree threshold, the corner position of the transient event is detected based on the slip ring speed phase pulse signal. It is determined whether the transient event repeats at the same corner position within 10 consecutive frequency cycles. If it repeats, it is determined to be a real fault transient and no rejection process is performed. If it does not repeat, the transient event is determined to be an interference range, the high signal-to-noise ratio basic signal in the interference range is removed, and the waveform is reconstructed.

5. A fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration according to claim 4, characterized in that, The waveform reconstruction process includes: The start and end timestamps of the interference interval are selected as the reference, and the normal waveform reference segment is extracted. The normal waveform reference segment is divided into a forward reference segment and a backward reference segment. Differential fitting is performed on the excitation current signal and carbon brush vibration acceleration signal in the normal waveform reference segment. The forward reference segment trend term and the backward reference segment trend term are extracted respectively. The forward reference segment trend term is subtracted from the high signal-to-noise ratio base signal of the forward reference segment to obtain the forward reference segment detail term. The backward reference segment detail term is obtained by subtracting the backward reference segment trend term from the high signal-to-noise ratio base signal of the backward reference segment. Based on the trend terms of the forward and backward reference segments, a continuous trend baseline is constructed within the interference interval. The cubic spline interpolation curve is superimposed on the continuous trend baseline to generate a reconstructed waveform. The detail component features of the detail terms of the forward and backward reference segments are extracted to obtain the detail components. Based on the slip ring rotation speed phase pulse signal, the normal detail component of the corresponding phase is matched for each sampling point within the interference interval to construct a detail compensation sequence for the interference interval. The detail compensation sequence is superimposed on the reconstructed waveform to generate the final reconstructed waveform.

6. A fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration according to claim 5, characterized in that, The process of determining the operating conditions of the current data acquisition cycle includes: The effective value of excitation current and the active power of the unit in the contact-type multidimensional heterogeneous data are selected as the operating condition characteristic indicators. An operating condition interval mapping table is preset. The operating condition interval mapping table includes the range of operating condition characteristic indicators and the transition zone corresponding to different operating conditions. The operating condition of the current collection cycle is determined according to the operating condition characteristic indicators and the operating condition interval mapping table.

7. A fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration according to claim 6, characterized in that, The process of constructing an electromagnetic motor thermal coupling model, and then building a fault life-cycle evolution mapping table and a personalized four-field coupling prediction model for the target unit based on the simulation test data of the electromagnetic motor thermal coupling model, includes: The electromagnetic-mechanical thermal coupling model includes a circuit field model, an electromagnetic field model, a mechanical field model, and a temperature field model. Simulation tests were conducted on the electromagnetic motor thermal coupling model under different fault types and operating conditions to obtain simulation test data of the electromagnetic motor thermal coupling model under different fault types and operating conditions. A fault lifecycle evolution mapping table is constructed based on simulation test data. The fault lifecycle evolution mapping table includes the mechanism core state parameter ranges of different evolution stages. The evolution stages include nonlinear degradation region, critical inflection point region and failure deterioration region. A four-field coupling prediction model is constructed. Simulation test data is used as training data to train the four-field coupling prediction model. The trained four-field coupling prediction model is obtained. Based on the equipment characteristics of the target unit, the individual parameters of the four-field coupling prediction model are calibrated to generate a personalized four-field coupling prediction model for the target unit. For each operating condition range, full-channel data of the target unit that has not generated a warning level within 72 hours is collected as healthy operation data. The healthy operation data is input into the personalized four-field coupling prediction model of the target unit, and a health benchmark library is constructed based on the output of the personalized four-field coupling prediction model.

8. A fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration according to claim 7, characterized in that, The process of pre-identifying abrupt changes in coupling strength based on four-dimensional coupling coefficients includes: The full-channel data of the target unit in the current acquisition cycle is input into the personalized four-field coupling prediction model. Based on the personalized four-field coupling prediction model, the four-dimensional coupling coefficient and core mechanism state parameters in the next few acquisition cycles under the current operating conditions are generated. The comprehensive coupling intensity increase and the duration of the increase of the four-dimensional coupling coefficient in the next few acquisition cycles are obtained. The coupling intensity increase threshold and time threshold are preset. If the comprehensive coupling intensity increase is greater than the coupling intensity increase threshold and the increase duration is greater than the time threshold, the degradation rate is accurately identified.

9. A fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration according to claim 8, characterized in that, The process of accurately identifying degradation rate based on core mechanistic state parameters includes: Extract the mean health benchmark of the core mechanism state parameters under the current working condition from the health benchmark library, compare the core mechanism state parameters with the corresponding mean health benchmark in several future acquisition cycles, obtain the comprehensive degradation degree in each future acquisition cycle, perform least squares fitting on the comprehensive degradation degree in each future acquisition cycle, generate degradation degree time series curve, and obtain the first and second derivatives of the degradation degree time series curve. Determine whether the second derivative changes from negative to positive and remains positive. If it does, mark the target unit as a nonlinear deterioration zone and perform a critical inflection point risk assessment. If the second derivative peaks and then falls back, but the first derivative continues to grow exponentially, the target unit is marked as a failure deterioration zone, triggering shutdown protection.

10. A fault early warning system based on the coupling characteristics of excitation DC high current and carbon brush vibration according to claim 9, characterized in that, The process of conducting a critical inflection point risk assessment and generating an early warning level based on the assessment results includes: Based on the fault lifecycle evolution mapping table, the critical benchmarks of the core mechanism state parameters in the critical inflection point region are obtained. Based on the average values ​​of the core mechanism state parameters, critical benchmarks, and health benchmarks in each future acquisition cycle, the safety margins in each future acquisition cycle are obtained. The warning levels corresponding to different safety margin intervals are preset. Based on the safety margins in each future acquisition cycle and the warning levels corresponding to different safety margin intervals, the warning levels in each future acquisition cycle are determined, and the preset operation and maintenance strategies corresponding to the warning levels are executed.

Images