Fault diagnosis method and system for double-fed wind power converter based on multi-source data fusion

Abstract

The application discloses a double-fed wind power converter fault diagnosis method and system based on multi-source data fusion. When the converter detects that an electrical quantity exceeds a threshold value or receives a manual starting instruction of an operation and maintenance personnel, each acquisition module is triggered to work. A preprocessing unit performs interpolation, time axis alignment and denoising processing on the acquired multi-source data, and generates a data set in a unified format. A feature extraction unit extracts fault characteristic quantities by using Fourier algorithm, symmetrical component method, four sampling value method and other methods, and outputs a feature data set. Multiple subsystems are independently diagnosed, and respective preliminary diagnosis results are output. A fusion decision unit substitutes the preliminary results of each subsystem into a preset rule base, performs multi-source data fusion by using a weighted voting method, outputs a final diagnosis conclusion, and outputs the result to a closed loop feedback. The application significantly improves the accuracy and engineering practicability of double-fed wind power converter fault diagnosis by using multi-source data fusion and hierarchical diagnosis, and reduces operation and maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for wind power electronic equipment, and in particular to a fault diagnosis method and system for doubly fed wind power converters based on multi-source data fusion. Background Technology

[0002] Existing fault diagnosis technologies for doubly-fed wind power converters are mainly based on a single data source or a single algorithm, and can be categorized as follows: 1) Diagnostic technologies based on switching signals: These collect switching signals such as circuit breaker status, contactor status, and alarm signals from the converter, and judge faults based on preset logic rules (such as input / output signal consistency). However, without electrical waveform data support, they cannot identify gradual faults such as power module aging, and their recognition rate is less than 50% during long-term operation. 2) Diagnostic technologies based on event sequence recording: These record the occurrence time (usually at millisecond resolution) of various events in the converter (such as start-up commands, fault alarms, and state transitions), analyze the timing relationship of events to locate faults. For example, the temporal correlation between excitation contactor closing failure and subsequent turbine-side start-up failure can be used to determine excitation system faults. 3) Diagnostic technologies based on fault waveform recording: These collect voltage and current waveforms at the time of faults, and use Fourier algorithms, symmetrical component methods, etc., to extract feature quantities (such as fundamental amplitude and zero-sequence component) to identify faults such as overcurrent, short circuit, and phase sequence errors.

[0003] In existing technologies, the aforementioned diagnostic methods are mostly applied independently without forming a data fusion mechanism, resulting in a limited diagnostic scope. For example, some systems only analyze electrical quantity anomalies through fault waveform recording, ignoring the impact of switch status on the fault; or they rely solely on event timing, lacking waveform data support, making it difficult to determine the nature of the fault.

[0004] Problems and shortcomings of existing technologies: 1) Limited diagnostic dimensions and insufficient accuracy: A single data source is insufficient to cover complex fault scenarios. For example, power module overcurrent may be caused by module damage (current waveform verification requires waveform recording data) or cooling system failure (fan operation status signal in switch data required). Relying on only one type of data can easily lead to misjudgment. 2) Lack of fault correlation analysis: Converter faults often exhibit a "chain reaction," and existing technologies do not integrate the time-series correlation of multi-source data. For example, in low voltage ride-through faults, grid voltage drops (waveform recording data) can trigger converter pulse blocking. If only a single data point is analyzed, the root cause of the fault cannot be traced. 3) Poor engineering applicability: Existing methods do not incorporate the fault characteristics of the entire life cycle of the doubly-fed converter (e.g., communication failures are prone to occur during the commissioning phase, and power module aging failures are prone to occur during long-term operation), resulting in large differences in the identification rate of faults at different stages. 4) High operational complexity: Maintenance personnel need to interpret switch data, waveform recording data, etc., and rely on professional experience for comprehensive judgment. For inexperienced personnel, fault location can take several hours. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a fault diagnosis method and system for doubly-fed wind power converters based on multi-source data fusion. Through multi-source data fusion and hierarchical diagnosis, the accuracy and engineering practicality of fault diagnosis for doubly-fed wind power converters are significantly improved, and operation and maintenance costs are reduced.

[0006] The objective of this invention is achieved in one aspect as follows: a fault diagnosis system for doubly fed wind power converters based on multi-source data fusion, comprising a data acquisition layer, a data processing layer, a fusion diagnosis layer, a result display layer, and a closed-loop diagnosis layer;

[0007] The data acquisition layer is used to acquire switch quantities, event sequence records, fault waveform data, and main control pitch data in real time, and transmit the data to the data processing layer.

[0008] The data processing layer is used to align the received data along the time axis, calibrate the time precision of the data based on the timestamp of the waveform data, and generate a dataset with a unified time axis; and to process the fault waveform data and extract the fault feature quantities.

[0009] The fusion diagnostic layer is used to analyze the feature dataset separately through multiple subsystems, output their respective preliminary diagnostic results, substitute each preliminary result into a preset rule base, perform multi-source data fusion, and output the final diagnostic conclusion.

[0010] The results display layer is used to clarify the fault level classification criteria and display the integrated diagnostic conclusions in a unified format on the visual interactive interface.

[0011] The closed-loop diagnostic layer is used to dynamically adapt and optimize the parameters at each level of the system based on the final conclusion of the fusion diagnosis, following the logic of "fault feature extraction - optimization instruction generation - hierarchical parameter adjustment - effect verification - data accumulation - iterative optimization".

[0012] The data acquisition layer, data processing layer, fusion diagnosis layer, result display layer, and closed-loop diagnosis layer are connected via a data bus.

[0013] Furthermore, the data acquisition layer includes a switch quantity acquisition module, an event sequence recording module, a fault waveform recording module, and a wind power main control and pitch data transmission module;

[0014] The digital input and output acquisition module is used to acquire digital input and output quantities and is connected to the converter controller via hardwiring.

[0015] The event sequence recording module is used to record the occurrence time of converter events;

[0016] The fault recording module is used to collect analog quantities of grid voltage, stator current, rotor current, and DC bus voltage.

[0017] The wind power main control and pitch data transmission module is used to collect the generator speed, grid-side power, torque command of the wind power main control and the blade angle, pitch speed and pitch motor current data of the pitch system. It communicates with the main control / pitch system using Modbus TCP / IP protocol.

[0018] Furthermore, the data processing layer includes a preprocessing unit and a feature extraction unit;

[0019] The preprocessing unit is used to perform secondary Lagrange interpolation on the waveform data. Based on the sampling frequency of the waveform module, it interpolates and supplements the low-frequency channel data to unify the sampling frequency of all channels. The dynamic window alignment algorithm is used to align the time axis of switch quantity, event sequence recording data, waveform data, etc. Based on the timestamp of the waveform data, the time accuracy of the data is calibrated to ≤1ms to generate a dataset with a unified time axis.

[0020] The feature extraction unit is used to calculate the voltage and current fundamental amplitude and phase using the Fourier algorithm; extract positive sequence, negative sequence, and zero sequence components using the symmetrical component method; locate the fault time using the abrupt change detection algorithm; and determine the fault phase using the fault phase selection algorithm.

[0021] Furthermore, the fusion diagnostic layer includes a switch quantity diagnostic system, an event sequence recording diagnostic subsystem, a fault recording diagnostic subsystem, a wind power main control and pitch data processing diagnostic subsystem, and a fusion decision unit;

[0022] The switch quantity diagnostic subsystem is used to determine hardware faults based on the logical relationship between input and output signals.

[0023] The event sequence recording and diagnostic subsystem is used to compare normal event sequences and locate the first abnormal event.

[0024] The fault recording and diagnostic subsystem is used to identify electrical faults through waveform characteristics;

[0025] The wind power main control and pitch data processing and diagnostic subsystem is used to extract two types of fault features, static features and dynamic features, from the pre-processed main control and pitch data, and to quantitatively characterize the fault status.

[0026] The fusion decision unit is used to input the preliminary results of each subsystem into the preset rule base, perform multi-source data fusion through weighted voting, output the final diagnostic conclusion, and output the results with closed-loop feedback.

[0027] Furthermore, the result display layer includes a fault classification standard module and a visual interactive interface module;

[0028] The fault classification standard module is used to clarify the fault level classification standard, including Class I faults: power module short circuits and DC bus overvoltages that directly cause shutdowns; Class II faults: cooling system abnormalities and communication interruptions that require timely handling; Class III faults: parameters deviating from thresholds and minor alarms that allow continued operation; and Class IV faults: false alarms.

[0029] The visualization interactive interface module is designed using MATLAB GUI and includes a data import area, waveform display area, event timing diagram, switch status table, and diagnostic conclusion area.

[0030] Another aspect of the objective of this invention is achieved as follows: a fault diagnosis method for doubly-fed wind power converters based on multi-source data fusion, comprising the following steps:

[0031] S1) Data Acquisition: Real-time acquisition of switch quantities, event sequence records, fault waveform data, and main control pitch data;

[0032] S2) Data Preprocessing: Based on the sampling frequency of the waveform recording module, interpolation fitting algorithm is used to interpolate and supplement points for low-frequency channel data to unify the sampling frequency of all channels; a dynamic window alignment algorithm is used to align the time axis of switch quantity, event sequence recording data, and waveform data; based on the timestamp of the waveform data, the time precision of the data is calibrated to generate a dataset with a unified time axis; and fault waveform data is processed and fault feature quantities are extracted.

[0033] S3) Fusion Diagnosis: The feature dataset is analyzed by multiple subsystems, and each subsystem outputs its preliminary diagnostic results. The preliminary results of each subsystem are then substituted into a preset rule base, and multi-source data are fused using a weighted voting method to output the final diagnostic conclusion.

[0034] S4) Results Display: Clearly define the fault level classification criteria and display the integrated diagnostic conclusions in a unified format on the visual interactive interface;

[0035] S5) Closed-loop feedback: Based on the final conclusion of the fusion diagnosis, the system dynamically adapts and optimizes parameters at each level according to the logic of "fault feature extraction - optimization instruction generation - hierarchical parameter adjustment - effect verification - data accumulation - iterative optimization".

[0036] Furthermore, the dataset for generating a unified timeline described in S2) specifically includes:

[0037] Preprocessing of waveform data: The data value x stored in the data file is converted into the actual sample value using the conversion formula ax+b, where a is the channel multiplier and b is the channel offset addend; after processing by the formula, the actual sample value of each channel is obtained, which is the sample value of the corresponding time in each channel;

[0038] The low-sampling-frequency data is converted into high-sampling-frequency data using the quadratic Lagrange interpolation algorithm. First, the interpolation position is determined. Then, the values ​​of three adjacent sampling points are selected, and the interpolation polynomial is used to calculate the value of the required interpolation position. After obtaining the estimated value of the interpolation point, if the number of samples required for the high sampling frequency is not yet reached, the obtained value is used as the sampling value for further calculation until the sampling frequency is consistent. Finally, a set of analog channel waveform data with consistent sampling frequency is obtained.

[0039] Using fault waveform data as the baseline time axis, the data to be aligned includes event sequence records, switch inputs, and main control pitch data. A dynamic window alignment algorithm is employed.

[0040] First, construct a reference time axis, using the data acquired by the fault recording module as the reference T. base Based on sampling frequency and millisecond-level precise timestamps, a continuous reference time axis T is constructed. base ={t0,t1,t2,...,t n}, where the time interval is Δt, covering the recording period from 2 seconds before the fault is triggered to 5 seconds after the fault is triggered;

[0041] Then, the three types of data to be aligned are preprocessed to extract valid timestamps and corresponding data values, and to unify the timestamp format; event sequence record data: extract <event description, timestamp t soe Event status >, timestamp precision ≤ 1ms; Switch data: Extract <switch name, timestamp t> switch Switch status (0 / 1) >, fill in missing timestamps; Main control / pitch data: extract < data items (speed / pitch angle / power), timestamp t mc Data value > 50ms timestamp collected by Modbus TCP / IP, formatted uniformly in milliseconds;

[0042] Initialize the dynamic window and its core parameters, with the original timestamp t of the data to be aligned at the center of the window. x Initial window size: set according to the sampling frequency ratio of the data to be aligned to the reference data; Event sequence data recording: window size W soe =±1ms; Main control / pitch data: Window size W mc =±50ms; Switch data: Window size W switch =±1000ms; Window adjustment step size: ΔW=±1ms;

[0043] Set window dynamic adaptive rules and fault shrinkage rules: When the reference waveform data detects fault characteristics, the window size of all data to be aligned is automatically shrunk by 50%; Data fluctuation adjustment rules: If the adjacent timestamp interval of the data to be aligned deviates by more than ±10%, the algorithm automatically adjusts the window size according to the deviation ratio to ensure that the window contains a valid reference time point.

[0044] Perform in-window timestamp matching and data mapping, for each data point to be aligned, the timestamp t... x In its dynamic window [t x -W,t x Within +W], match the reference time axis T base The optimal reference time point The matching rules are as follows:

[0045]

[0046] That is, the reference time point within the window that is closest to the timestamp to be aligned is used to map the data value to be aligned one-to-one to the reference time point, thus completing the time alignment of a single data item;

[0047] Finally, full data fusion and continuous time axis generation are performed, fusing all the data values ​​to be aligned to the base time axis T according to the mapped base time point. base In the process, for points in the reference time axis without matching data, a nearest neighbor preservation strategy or a linear interpolation strategy is used to complete the data, and finally a multi-source fusion dataset with a unified time axis is generated. The time accuracy is consistent with the fault recording, and the dataset is directly input into the feature extraction unit for subsequent processing.

[0048] Furthermore, the processing of fault recording data and extraction of fault feature quantities described in S2) specifically includes: analyzing fault recording data using algorithms based on a simple non-sinusoidal wave model; including the symmetric component method, Fourier algorithm, and abrupt change phase selection algorithm;

[0049] In wind power converter systems, when an asymmetrical fault occurs, the symmetrical component method is used for analysis. The asymmetrical three-phase phasors are decomposed into three sets of symmetrical components: positive sequence, negative sequence, and zero sequence, thereby simplifying the problem of asymmetrical phasors into the analysis of symmetrical phasors.

[0050] A full-cycle Fourier algorithm is used to analyze the waveform data of the converter. The orthogonal properties of sine and cosine functions are used to extract a component of a specific frequency from the signal. First, the sampled signal is assumed to be a continuous periodic time function. Then, the amplitudes of the sine and cosine terms of the fundamental component are obtained according to the basic principles of Fourier series. The amplitudes of the sine and cosine terms of the fundamental component have eliminated the influence of the DC component and integer harmonic components. The effective value of the signal is obtained by combining the phasor form analysis of the signal. Using each sampling point as the starting position and the subsequent cycle as the data window, a full-cycle Fourier algorithm is performed. Through rolling recursive analysis, the overall characteristics of the wind power converter waveform data are reflected.

[0051] A mutation-based phase selection algorithm is used to determine fault initiation, using ΔI. AB ΔI BC ΔI BC To represent the sudden change in current between the three phases; depending on the fault, when there is a three-phase short circuit, ΔI AB =ΔI BC =ΔI BC When a phase is short-circuited to ground, the current difference between the other two unfaulted phases is zero. When a two-phase fault occurs, the current difference between the two faulty phases is the largest. The presence or absence of a zero-sequence current component is used to distinguish between a two-phase short circuit to ground and a short circuit between two phases.

[0052] Furthermore, the analysis of the feature dataset by multiple subsystems described in S3), and the output of their respective preliminary diagnostic results, specifically includes:

[0053] Switch quantity diagnostic subsystem: The basic logic of switch quantity fault analysis is as follows: First, all input and output quantities are imported, and then the current working status of the wind power converter system is determined. If it is in the running state, the fault judgment will be performed according to the normal operation state above; if the wind power converter system is in the shutdown state, the fault judgment will be performed according to the shutdown state.

[0054] Event Sequence Recording Diagnostic Subsystem: By collecting and analyzing the sequence records of fault events of a wind farm converter and comparing them with normal start-up and shutdown event records, the system can identify the first fault or the first abnormal action, thereby accurately locating the fault location.

[0055] The fault recording diagnostic subsystem first determines whether the relationship between the phases of the three-phase voltage and current is normal through phase analysis. By selecting sampling points on the waveform display interface using the mouse, the system calculates the amplitude and phase angle of each analog channel and displays them on the visualization interface. Each time a different sampling point is selected, the system performs a corresponding full-cycle Fourier algorithm calculation starting from the new sampling point and updates the values ​​on the interface, thus achieving dynamic display of phase analysis and diagnosing whether the converter has a missing phase or phase sequence error. Then, the symmetrical component method and fault phase selection algorithm are used to extract the fault component from the wind power converter fault recording. Finally, the fault time is determined, and the sudden change in phase current difference is used to determine and analyze the fault occurrence time.

[0056] Wind power main control and pitch data processing and diagnostic subsystem: For pre-processed main control and pitch data, extract two types of fault features, namely static features and dynamic features, and quantify the fault status; adopt a fusion mode of rule base diagnosis and machine learning diagnosis to achieve accurate judgment of main control / pitch system faults;

[0057] Static features are extracted based on threshold comparison to reflect the degree to which data deviates from the normal range; these include: blade angle deviation: actual blade angle - target blade angle, threshold > 5°, indicating pitch control execution deviation; pitch motor current over-limit: actual pitch motor current > 1.2 times rated current, indicating pitch motor overload; communication delay: time difference between master control command issuance and pitch feedback, threshold > 100ms, indicating communication link failure; and speed deviation: actual generator speed - master control target speed, threshold > 50 r / min, indicating abnormal speed control.

[0058] Dynamic features are extracted based on time-series analysis, reflecting the data's variation over time; these include pitch speed volatility: the standard deviation of pitch speed over 10 consecutive frames, with a threshold > Determine pitch mechanical jamming; Torque command mutation rate: the change in torque command between adjacent frames, threshold > If the main control command is abnormal, the blade angle response is delayed: if the time difference between the issuance of the pitch command and the change in the blade angle is greater than 200ms, the pitch reducer is determined to be faulty.

[0059] Furthermore, S4) includes: displaying the integrated diagnostic conclusions in a unified format on the visual interactive interface of the result display layer, specifically including diagnostic information: fault type, occurrence time, location component, overall confidence level, and fault level; operation and maintenance handling suggestions: outputting standardized implementation suggestions for different fault types; highlighting warning rules: Class I faults are highlighted with a red background, Class II faults with orange, Class III faults with yellow, and suspected faults with blue; data export function: supporting one-click export of diagnostic results, fault characteristic data, and waveform time series diagrams.

[0060] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) The diagnostic accuracy is significantly improved: multi-source data fusion solves the problem of misjudgment by single data; by integrating multi-source data such as switch quantity, event sequence record data, fault waveform, wind power main control and pitch data, the invention upgrades from "single-dimensional judgment" to "multi-dimensional cross-verification", improves the accuracy of complex fault diagnosis, and solves specific faults that traditional methods cannot detect by combining multi-source data fusion technology.

[0061] 2) Enhanced fault tracing capability: The fault propagation path is clarified through time-series correlation analysis; a time-series correlation analysis mechanism for multi-source data is established to trace the fault propagation path. Combined with multi-source data fusion technology, it solves complex and chain faults such as "power module overcurrent caused by cooling system failure" and "chain pulse blockage fault caused by grid voltage drop" that cannot be detected by traditional single methods.

[0062] 3) Engineering Applicability Optimization: The fault model was optimized for different stages of the doubly-fed converter. In a wind power plant, the identification rate for "communication faults" was improved by 30% during the commissioning phase, and the identification rate for "power module aging" was improved by 25% during long-term operation. Based on actual project fault data (918 records), the classification model was optimized to adapt to scenarios such as commissioning and long-term operation, achieving a 100% identification rate for one type of fault (such as power module short circuit).

[0063] 4) Improved maintenance efficiency: The visual interface integrates multi-source data, reducing fault location time from 2-4 hours using traditional methods to less than 30 minutes, thus minimizing wind turbine downtime losses. By integrating multi-source data diagnostic results through the visual interface, operational complexity is reduced, enabling on-site personnel to quickly locate faults.

[0064] 5) Full life cycle coverage: Adapt to the fault characteristics of different operating stages of doubly fed converters, optimize diagnostic models for on-site commissioning, long-term operation and other stages, and improve the fault identification rate throughout the entire life cycle.

[0065] 6) Optimization of algorithm robustness: The four-sampling value method is adopted to detect the sudden change in current, which improves the ability to resist frequency deviation and the fault time judgment error is ≤5ms; the symmetrical component method has an accuracy of 100% in identifying three-phase unbalanced faults. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0067] Figure 1This is a schematic diagram of the overall structure of the system of the present invention.

[0068] Figure 2 This is a schematic diagram of a switch quantity diagnostic system.

[0069] Figure 3 This is a schematic diagram of a waveform diagnostic system. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] like Figure 1 The system shown is a fault diagnosis system for a doubly fed wind power converter based on multi-source data fusion, comprising a data acquisition layer, a data processing layer, a fusion diagnosis layer, a result display layer, and a closed-loop diagnosis layer.

[0072] The data acquisition layer is used to acquire switch signals, event sequence records, fault waveform data, and main control pitch data in real time; and transmit the data to the data processing layer.

[0073] The data acquisition layer includes a switch quantity acquisition module, an event sequence recording module, a fault waveform recording module, and a wind power main control and pitch data transmission module;

[0074] The digital input module is used to acquire digital input quantities (such as grid-connected contactor status and main circuit breaker fault trip signal) and digital output quantities (such as main circuit breaker closing command and excitation contactor control signal), and is connected to the converter controller through hard wiring.

[0075] The event sequence recording module is used to record the occurrence time of converter events (such as "pre-charge complete" and "fault alarm") (resolution ≤ 1ms), and the storage format is "event number-description-timestamp".

[0076] The fault recording module is used to collect analog quantities such as grid voltage, stator current, rotor current, and DC bus voltage, and stores them in COMTRADE99 format, including header file (.HDR), configuration file (.CFG), data file (.DAT), and information file (.INF).

[0077] The wind power main control and pitch data transmission module is used to collect the generator speed, grid-side power, and torque command of the wind power main control, as well as the blade angle, pitch speed, and pitch motor current data of the pitch system. It communicates with the main control / pitch system using the Modbus TCP / IP protocol.

[0078] The data processing layer is used to align the received data to the time axis, calibrate the time precision of the data based on the timestamp of the waveform data, and generate a dataset with a unified time axis; and to process the fault waveform data and extract the fault feature quantities.

[0079] The data processing layer includes a preprocessing unit and a feature extraction unit;

[0080] The preprocessing unit performs secondary Lagrange interpolation on the waveform data. Based on the sampling frequency of the waveform module, it interpolates and fills in the low-frequency channel data to unify the sampling frequency of all channels. The dynamic window alignment algorithm is used to align the time axis of switch quantity, event sequence recording data, waveform data, etc. Based on the timestamp of the waveform data, the time accuracy of the data is calibrated to ≤1ms to generate a dataset with a unified time axis.

[0081] The feature extraction unit is used to calculate the voltage and current fundamental amplitude and phase using the Fourier algorithm; extract positive sequence, negative sequence and zero sequence components using the symmetrical component method; locate the fault time using the mutation detection algorithm (four sampling value method); and determine the fault phase using the fault phase selection algorithm.

[0082] The fusion diagnostic layer is used to analyze the feature datasets separately through multiple subsystems, outputting their respective preliminary diagnostic results; the preliminary results are then substituted into a preset rule base to perform multi-source data fusion and output the final diagnostic conclusion.

[0083] The integrated diagnostic layer includes a switch quantity diagnostic system, an event sequence recording diagnostic subsystem, a fault recording diagnostic subsystem, a wind power main control and pitch data processing diagnostic subsystem, and an integrated decision unit;

[0084] The switch quantity diagnostic subsystem is used to determine hardware faults (such as relay failure, board power failure, switch disconnection, etc.) based on the logical relationship between input and output signals (such as "consistency between closing command and feedback status").

[0085] The event sequence recording and diagnostic subsystem is used to compare normal event sequences (such as startup process: pre-charging → excitation contactor closing → grid connection) to locate the first abnormal event (such as "excitation contactor closing failure").

[0086] The fault recording and diagnostic subsystem is used to identify electrical faults by waveform characteristics (such as a sudden increase in zero-sequence current → ground fault, phase voltage phase deviation → phase sequence error).

[0087] The wind power main control and pitch data processing and diagnostic subsystem is used to extract two types of fault features, namely static features and dynamic features, from the pre-processed main control and pitch data, and to quantitatively characterize the fault status.

[0088] The fusion decision unit is used to input the preliminary results of each subsystem into the preset rule base, perform multi-source data fusion through weighted voting, output the final diagnostic conclusion, and output the results with closed-loop feedback.

[0089] The construction logic of the rule base (such as a three-dimensional rule base based on fault type, data characteristics, and correlation), specific rule entries (such as rule 1: waveform A phase overcurrent (current > 2 times rated value) + event record A phase module temperature > 85℃ + switch quantity fan running status = 0 → diagnosis is cooling fan failure causing A phase module overcurrent), and fusion judgment logic (such as using the "weighted voting method", waveform data weight 0.5, event record data weight 0.3, switch quantity data weight 0.2, and the final diagnosis conclusion is output when the comprehensive score is ≥ 0.8).

[0090] The results display layer is used to clarify the fault level classification criteria and to display the integrated diagnostic conclusions in a unified format on the visual interactive interface. The results display layer includes a fault classification criteria module and a visual interactive interface module.

[0091] The fault classification standard module is used to clarify the fault level classification standard, including Class I faults: power module short circuits and DC bus overvoltages that directly cause shutdowns; Class II faults: cooling system abnormalities and communication interruptions that require timely handling; Class III faults: parameters deviating from thresholds and minor alarms that allow continued operation; and Class IV faults: false alarms.

[0092] The visual interactive interface module is designed using MATLAB GUI and includes a data import area, a waveform display area (power grid / stator / rotor waveforms), an event timing diagram, a switch status table, and a diagnostic conclusion area (fault type, cause, and handling suggestions). The data import area supports importing waveform recording files, the waveform display area allows zooming in / out of the waveform at the time of the fault, and the diagnostic conclusion area supports exporting fault reports.

[0093] The closed-loop diagnostic layer is used to dynamically adapt and optimize the parameters at each level of the system based on the final conclusion of the fusion diagnosis, following the logic of "fault feature extraction - optimization instruction generation - hierarchical parameter adjustment - effect verification - data accumulation - iterative optimization".

[0094] The data acquisition layer, data processing layer, fusion diagnosis layer, result display layer, and closed-loop diagnosis layer are connected through a data bus.

[0095] A fault diagnosis method for doubly-fed wind power converters based on multi-source data fusion includes the following steps:

[0096] S1) Data Acquisition: Real-time acquisition of switch quantities, event sequence records, fault waveform data, and main control pitch data; trigger condition is converter fault report or manual start.

[0097] S2) Data Preprocessing: Based on the sampling frequency of the waveform recording module, interpolation fitting algorithm is used to interpolate and supplement points for low-frequency channel data to unify the sampling frequency of all channels; a dynamic window alignment algorithm is used to align the time axis of switch quantity, event sequence recording data, and waveform data; based on the timestamp of the waveform data, the time precision of the data is calibrated to generate a dataset with a unified time axis; and fault waveform data is processed and fault feature quantities are extracted.

[0098] S2-1) Preprocessing of waveform data: The data values ​​of the analog channels in the waveform file are not the actual values. It is necessary to use the conversion formula ax+b to convert the data values ​​x stored in the data file into actual sampled values, where a is the channel multiplier and b is the channel offset addend. After processing by the formula, the actual sampled value of each channel is obtained, which is the sampled value of the corresponding time in each channel.

[0099] However, these waveform recordings cannot be used directly because analog channel data is typically recorded at different sampling frequencies over different time periods. To unify the processing of the waveform data, an interpolation fitting algorithm is needed to convert low-sampling-frequency data into high-sampling-frequency data. A commonly used algorithm is the quadratic Lagrange interpolation algorithm. To convert low-sampling-frequency data into high-sampling-frequency data using the quadratic Lagrange interpolation algorithm, the interpolation position is first determined. Then, the values ​​of three adjacent sampling points are selected, and an interpolation polynomial is used to calculate the value at the required interpolation position. After obtaining the estimated sampling value at the insertion point, if the required number of samples for the high sampling frequency is not yet reached, the obtained value is used as the sampling value for further calculations until the sampling frequencies are consistent. This results in a set of analog channel waveform data with consistent sampling frequencies, facilitating subsequent data processing and computation.

[0100] S2-2) Multi-source data dynamic window alignment algorithm: Using fault waveform data as the reference time axis, the data to be aligned includes event sequence records, switch quantities, and main control pitch data. A dynamic window alignment algorithm is adopted:

[0101] First, construct a reference time axis, using the data acquired by the fault recording module as the reference T. base Based on the sampling frequency and millisecond-level precise timestamps (such as 2026-01-01 10:00:00.000, 2026-01-01 10:00:00.001), a continuous reference time axis T is constructed. base ={t0,t1,t2,...,t n}, where the time interval is Δt, covering the recording period from 2 seconds before the fault is triggered to 5 seconds after the fault is triggered;

[0102] Then, the three types of data to be aligned are preprocessed to extract valid timestamps and corresponding data values, and the timestamp format is standardized to "year-month-day hour:minute:second:millisecond" (power system standard format); event sequence record data: extract <event description, timestamp t soe Event status >, timestamp precision ≤ 1ms; Switch data: Extract <switch name, timestamp t> switch Switch status (0 / 1) >, fill in missing timestamps (interpolated based on the most recent sampling point); Main control / pitch data: extract < data items (speed / pitch angle / power), timestamp t mc Data value > 50ms timestamp collected by Modbus TCP / IP, formatted uniformly in milliseconds;

[0103] Initialize the dynamic window and its core parameters, with the original timestamp t of the data to be aligned at the center of the window. x (x represents soe / switch / mc); Initial window size: set according to the sampling frequency ratio of the data to be aligned to the reference data; Event sequence data recording: window size W soe =±1ms (because the timing of the recorded event data is close to the accuracy of the reference waveform timestamp, a small window ensures alignment accuracy); Main control / pitch data: Window size W mc =±50ms (consistent with the acquisition period, matching the data transmission frequency); Switch data: window size W switch =±1000ms (1s, matching the sampling period of the switch quantity); Window adjustment step size: ΔW=±1ms (minimum adjustment unit, consistent with the accuracy of the reference time axis).

[0104] A dynamic adaptive window rule is set, and the algorithm adjusts the window size in real time according to data characteristics to avoid alignment distortion. Fault-time shrinkage rule: When fault characteristics are detected in the reference waveform data (such as sudden current increase or voltage drop, determined by the four-sample value method), the window size of all data to be aligned is automatically shrunk by 50% (e.g., the main control / pitch window shrinks from ±50ms to ±25ms), improving the time alignment accuracy during critical fault periods. Data fluctuation adjustment rule: If the deviation between adjacent timestamps of the data to be aligned exceeds ±10% (e.g., the main control data acquisition cycle changes from 50ms to 55ms), the algorithm automatically adjusts the window size according to the deviation ratio (e.g., the window expands from ±50ms to ±55ms), ensuring that the window contains valid reference time points.

[0105] Perform in-window timestamp matching and data mapping, for each data point to be aligned, the timestamp t... x In its dynamic window [t x -W,tx Within +W], match the reference time axis T base The optimal reference time point The matching rules are as follows:

[0106]

[0107] That is, the reference time point within the window that is closest to the timestamp to be aligned is used to map the data value to be aligned one-to-one to the reference time point, thus completing the time alignment of a single data item;

[0108] Finally, full data fusion and continuous time axis generation are performed. All values ​​of the data to be aligned are fused to the base time axis T according to the mapped base time point. base In the process, for points in the reference time axis that have no matching data, a nearest neighbor value preservation strategy (for non-fault periods) or a linear interpolation strategy (for fault periods to ensure data continuity) is used to complete the data. Finally, a multi-source fusion dataset with a unified time axis is generated. The time accuracy is consistent with the fault recording (≤1ms), which can be directly input into the feature extraction unit for subsequent processing.

[0109] S2-3) Feature Extraction: To further analyze the fault recording data of the wind power converter, it is necessary to process the fault recording data and extract the fault feature quantities. Algorithms based on a simple non-sinusoidal wave model are used to analyze the fault recording data, including the symmetrical component method, Fourier algorithm, and abrupt change phase selection algorithm.

[0110] In wind power converter systems, when a fault occurs, such as a single-phase grounding or open circuit, a short circuit between two phases, or a two-phase open circuit, an asymmetrical situation arises. In this case, the impedance, voltage, current, and phase of the three phases will all deviate. For this type of three-phase asymmetrical fault, the symmetrical component method is generally used for analysis. This method decomposes the asymmetrical three-phase phasors into three sets of symmetrical components: positive sequence, negative sequence, and zero sequence, thus simplifying the asymmetrical phasor problem into a symmetrical phasor analysis.

[0111] The power quality of wind power converters is not always optimal during operation, often containing DC components and harmonic components. Generally, the Fourier algorithm, derived from Fourier series, is used to analyze and process the recorded waveform data. The Fourier algorithm is an algorithm based on a simple non-sinusoidal model, which can extract a component of a specific frequency from a signal using the orthogonal function property of sine and cosine functions. This paper uses the full-cycle Fourier algorithm to analyze the converter's recorded waveform data. The specific algorithm is as follows: First, assume the sampled signal is a continuous periodic time function, containing non-attenuated DC components and harmonics in addition to the fundamental frequency component. Then, based on the basic principles of Fourier series, the amplitudes of the sine and cosine terms of the fundamental component are obtained. From the integration process, it can be seen that the amplitudes of the sine and cosine terms of the fundamental component have eliminated the influence of the DC components and harmonic components. Combining this with the phasor form of the signal, the effective value of the signal can be obtained. Furthermore, harmonic analysis can be further implemented using the above principles. To achieve real-time calculation of RMS values ​​and phase, this paper uses each sampling point as the starting position and performs full-cycle Fourier algorithm calculations using the next cycle as the data window. Through rolling recursive analysis, the overall characteristics of the wind power converter's waveform data can be better reflected.

[0112] To accurately determine the time of fault occurrence from wind power converter fault recording data, a fault occurrence timing detection algorithm is needed. Generally, wind power converter systems employ steady-state fault initiation algorithms, which set certain ranges for current and voltage values. When the current or voltage exceeds or falls below the set value, a fault is triggered (e.g., overcurrent or undervoltage). However, this method has low sensitivity and sometimes fails to trigger the fault. Another fault initiation algorithm is the abrupt change detection algorithm, which detects changes in measured values ​​before and after the fault occurs. A sudden change in the change is used as the fault initiation condition. Compared to voltage abrupt changes, current abrupt changes have higher initiation sensitivity, so in practical applications, the current abrupt change method is generally used to determine fault initiation. This paper uses a four-sample value method to calculate the actual abrupt change.

[0113] In fault diagnosis of wind power converters, it is often necessary to determine which phase the fault occurs in order to further identify the specific damaged components and the fault location. Depending on the selected power parameters, fault phase selection algorithms can be divided into steady-state phase selection algorithms and abrupt change phase selection algorithms. Steady-state phase selection algorithms select steady-state power parameters, such as steady-state phase voltage and phase current, and can utilize voltage phase selection, current phase selection, impedance phase selection, etc. This method is simple to operate but has poor accuracy and a limited application range. Abrupt change phase selection algorithms, on the other hand, select parameters representing changes in voltage or current. This method is relatively complex to calculate but is unaffected by load current, and has higher accuracy and reliability. Therefore, this paper adopts the abrupt change phase selection algorithm. Since current abrupt changes are more significant than voltage abrupt changes, current abrupt change phase selection is more accurate and sensitive. Depending on the selected parameters, current abrupt changes are further divided into phase current abrupt changes and interphase current abrupt changes. Among them, interphase current abrupt changes do not reflect zero-sequence fault components and are less affected by common-mode faults; therefore, it is the main method for current abrupt change phase selection and can be used... , , This is used to represent the relationship between the sudden changes in current between the three phases, depending on the type of fault. When a three-phase short circuit occurs, When a phase is short-circuited to ground, the current difference between the other two unfaulted phases is zero. When a two-phase fault occurs, the current difference between the two faulty phases is the largest. Specifically, the presence or absence of a zero-sequence current component can be used to distinguish between a two-phase short circuit to ground and a short circuit between two phases.

[0114] S3) Fusion Diagnosis: The feature dataset is analyzed by multiple subsystems, and each subsystem outputs its preliminary diagnostic results. The preliminary results of each subsystem are then substituted into a preset rule base, and multi-source data are fused using a weighted voting method to output the final diagnostic conclusion.

[0115] Switch quantity diagnostic subsystem: The basic logic of switch quantity fault analysis is as follows: First, all input and output quantities are imported, and then the current working status of the wind power converter system is determined. If it is in the running state, the fault judgment will be performed according to the normal operation state above; if the wind power converter system is in the shutdown state, the fault judgment will be performed according to the shutdown state.

[0116] Event Sequence Recording Diagnostic Subsystem: By collecting and analyzing the sequence records of fault events in a wind farm converter and comparing them with normal start-up and shutdown event records, it can be found that the fault event sequence records contain a large number of abnormal actions, mostly anomalies, faults, and alarm information. The basic principle of event sequence recording fault analysis and diagnosis is to analyze the events at the time of the fault, find the first fault or the first abnormal action, and thus accurately locate the fault point. This is a simple and effective method for converter fault analysis. Based on the above principle, an event sequence recording fault diagnosis system was created using MATLAB interface design and combined with the normal start-up and shutdown sequence of the converter. Importing and displaying fault event sequence records: First, the uigetfile() function is used to select the file, then the fullfile() function is used to combine the path and filename, and the xlsread() function is used to read the information. Finally, it is displayed on the interface. Find the first abnormal action in the fault event sequence record: After importing the fault event sequence record, compare it with the preset normal event sequence record to filter out all abnormal actions. Sort all abnormal actions according to their occurrence time and display them on the interface as a basis for analysis. Integrate the above procedures to build a fault analysis interface.

[0117] The fault recording and diagnostic subsystem first uses phase analysis to determine if the relationship between the phases of the three-phase voltage and current is normal. By selecting sampling points on the waveform display interface with the mouse, the system calculates the amplitude and phase angle of each analog channel and displays them on the visualization interface. Each time a different sampling point is selected, the system performs a corresponding full-cycle Fourier algorithm calculation starting from the new sampling point and updates the values ​​on the interface, thus achieving dynamic display of phase analysis and diagnosing whether the converter has a phase loss or phase sequence error.

[0118] Then, the fault components in the fault recordings of the wind power converter are extracted using the symmetrical component method and the fault phase selection algorithm. The symmetrical component method can extract the positive-sequence, negative-sequence, and zero-sequence components in a three-phase power system. The fault phase selection algorithm can determine the specific phase in which the fault occurs. When a system fault occurs, a fault component is generated. The system state at this time can be considered as the superposition of the normal state and the fault state caused by the fault component. When analyzing the recording data using the symmetrical component method, it can be found that the positive-sequence component always exists, regardless of whether a fault occurs. When the system is operating normally, only the normal load current exists in the line, while the fault component can be regarded as the fault component current generated by the faulty power source acting alone in the line. The fault component current can be obtained by subtracting the current under normal conditions before the fault from the current actually measured after the fault. The three-phase fault components can then be calculated, and the positive-sequence fault component can be obtained using the symmetrical component method. By analyzing the sudden changes in the three-phase current at the fault point using the fault phase selection algorithm, the judgment of faults such as three-phase short circuits and grounding in phases A, B, and C can be obtained, thereby extracting the specific fault phase.

[0119] Then, the fault time is determined. In fault waveform data analysis, the sudden change in phase voltage or phase current is generally used to detect the fault time. This paper uses the sudden change in phase current difference to determine and analyze the fault occurrence time. When the system is operating normally, although the sudden change in phase current difference changes, it will not change significantly within a frequency cycle. By detecting the sudden change in the related channels of the sudden change current, the specific time of the fault occurrence can be obtained. The determined fault time is then exported to the final fault analysis results. In addition, since the fault waveform triggering method of wind power converter is divided into manual triggering and automatic triggering, if the sudden change in phase voltage or phase current does not change significantly, it is considered manual triggering under normal conditions.

[0120] Faults that can be identified in a doubly-fed induction generator (DFIG) wind turbine converter through fault waveform analysis include: DC voltage faults, grid voltage exceeding the range, single-phase grounding faults, phase sequence errors, two-phase short-circuit faults, excessively high or low frequency, excessively high current exceeding the protection value, and three-phase current imbalance. The characteristics of the more frequent fault waveforms are summarized to obtain specific fault waveform features. Based on these characteristics, various fault judgment conditions are pre-set in the program. If a fault waveform meets a certain fault characteristic, it is recorded in the fault analysis conclusion.

[0121] (1) DC voltage fault: The DC voltage of the wind power converter is out of range, that is, the maximum or minimum DC voltage exceeds the preset value.

[0122] (2) Grid voltage out of range: The grid voltage is out of range, that is, the maximum or minimum grid voltage exceeds the preset value.

[0123] (3) Single-phase grounding fault: The current of the faulted phase suddenly increases at the moment of the fault, and the voltage decreases accordingly. Zero-sequence voltage and zero-sequence current will appear. The voltage of the faulted phase is opposite to the zero-sequence voltage, and the current of the faulted phase is in the same direction as the zero-sequence current.

[0124] (4) Phase sequence error: The phase sequence of the grid voltage and stator voltage does not match the default phase sequence, and a certain phase is lagging or leading.

[0125] (5) Two-phase short circuit fault: The current in the two faulty phases increases and the voltage decreases accordingly. There is no zero-sequence voltage or zero-sequence current. The currents in the two faulty phases are basically opposite.

[0126] (6) Voltage frequency too high or too low: The voltage frequency is calculated based on the waveform period using the formula f=1 / T. Taking my country's 50Hz frequency as an example, a voltage frequency exceeding the "upper limit of stable access frequency" (54Hz) is considered too high. A voltage frequency below the "lower limit of stable access frequency" (46Hz) is considered too low.

[0127] Wind power main control and pitch data processing and diagnostic subsystem: For the pre-processed main control and pitch data, extract two types of core fault features: "static features + dynamic features" to quantitatively characterize the fault state, including static features (steady-state fault features) and dynamic features (transient fault features).

[0128] Static features are extracted based on threshold comparisons to reflect the degree to which data deviates from the normal range. These include: blade angle deviation: actual blade angle - target blade angle, threshold > 5° (indicating pitch control execution deviation); pitch motor current over-limit: actual pitch motor current > 1.2 times rated current (indicating pitch motor overload); communication delay: the time difference between the master control command and the pitch response, threshold > 100ms (indicating communication link failure); and speed deviation: actual generator speed - master control target speed, threshold > 50 r / min (indicating abnormal speed control).

[0129] Dynamic features are extracted based on time-series analysis, reflecting the data's variation over time. This includes pitch speed volatility: the standard deviation of pitch speed over 10 consecutive frames, with a threshold > (Determine pitch mechanical jamming); Torque command mutation rate: the change in torque command between adjacent frames, threshold > (Judging that the main control command is abnormal); Blade angle response lag: the time difference between the issuance of the pitch command and the change of the blade angle, the threshold is >200ms (judging that the pitch reducer is faulty).

[0130] A fusion approach combining "rule-based diagnosis (deterministic faults)" and "machine learning diagnosis (fuzzy faults)" is employed to achieve accurate fault identification in the main control / pitch system. Rule-based diagnosis leverages expert experience in wind power main control / pitch faults to construct a deterministic diagnostic rule base, matching feature quantities to determine fault types. Machine learning-assisted diagnosis addresses fuzzy faults not covered by the rule base (such as early wear of the pitch reducer and latent faults in the main control board) using a lightweight machine learning model. The training dataset includes historical main control / pitch operation data from wind farms (including fault labels); input features consist of 8-dimensional quantized features (static and dynamic features); output is a fault probability (0-1), with a threshold > 0.8 indicating a suspected fault, which is then pushed to the fusion decision unit for further verification.

[0131] The fusion decision unit inputs the preliminary results from each subsystem into a preset rule base, performs multi-source data fusion using a weighted voting method, and outputs the final diagnostic conclusion (including fault type, occurrence time, location of the faulty component, and confidence level). The results are then fed back into the closed-loop system.

[0132] S4) Results Display: Clearly define the fault level classification criteria and display the integrated diagnostic conclusions in a unified format on the visual interactive interface;

[0133] The integrated diagnostic conclusions are displayed in a unified format on the visual interactive interface of the results presentation layer, adapting to the practical needs of wind power field operation and maintenance. This includes core diagnostic information: clearly displaying the fault type, occurrence time, located component, overall confidence level, and fault severity; maintenance handling suggestions: outputting standardized implementation suggestions for different fault types, such as "cooling fan failure causing overcurrent in phase A module," with the corresponding suggestion being "immediately shut down, replace the cooling fan of phase A power module, and retest the motor current and module temperature"; highlighting warning rules: Class I faults are highlighted with a red background, Class II faults with orange, Class III faults with yellow, and suspected faults with blue, facilitating rapid identification of high-priority faults by maintenance personnel and shortening fault location time; and data export function: supporting one-click export of diagnostic results, fault characteristic data, and waveform time series diagrams, providing data support for fault review and subsequent operation and maintenance.

[0134] S5) Closed-loop feedback: Based on the final conclusion of the fusion diagnosis, the system dynamically adapts and optimizes parameters at each level according to the logic of "fault feature extraction - optimization instruction generation - hierarchical parameter adjustment - effect verification - data accumulation - iterative optimization".

[0135] Based on the final conclusion of the fusion diagnosis, and following the logic of "fault feature extraction - optimization instruction generation - hierarchical parameter adjustment - effect verification - data accumulation - iterative optimization," the system achieves dynamic adaptation and optimization of parameters at each level. All operations are completed automatically by the system without manual intervention, requiring only periodic review by maintenance personnel. The specific implementation steps are as follows:

[0136] Fault conclusions and feature data extraction: The system automatically triggers the fault feature analysis module to extract full-volume related data such as fault type, equipment operation stage, collected parameters, feature extraction threshold, and historical diagnosis patterns from the feature dataset and historical database, and stores them in a standardized label according to "fault type-feature dimension-parameter type";

[0137] Optimize requirements analysis and command generation: Combining wind power industry standards and system preset optimization rules, generate hierarchical optimization commands according to fault characteristics, clarifying the optimization objects, adjustment parameters, target values, and execution timing; for example, power module aging faults correspond to "increasing the waveform sampling frequency from 2900Hz to 3500Hz and reducing the negative sequence current judgment threshold from 0.1 times the rated current to 0.08 times", while communication faults correspond to "shortening the main control pitch acquisition cycle from 50ms to 10ms and improving the switch quantity acquisition refresh rate";

[0138] Layered optimization instruction issuance and execution: Optimization instructions are issued to the data acquisition layer and data processing layer respectively through the data bus. Each unit executes them in real time without interference, without affecting the normal operation of the wind turbine and diagnostic system. The data acquisition layer adjusts parameters such as sampling frequency, acquisition period, and refresh rate, while the data processing layer adjusts parameters such as dynamic window size, algorithm threshold, and feature extraction rules. After adjustment, parameter optimization records are generated to achieve process traceability.

[0139] Real-time verification of optimized diagnostic effect: The diagnostic effect verification module is launched to verify the effect through the reproduction of similar faults and real-time operation data monitoring. The optimization effect is verified from five dimensions: fault recognition rate, false judgment rate, feature extraction accuracy, positioning timeliness, and system stability. If the preset target is achieved (e.g., recognition rate improvement ≥5%, false judgment rate reduction ≥3%), the optimization is deemed effective. If the target is not met, a secondary optimization instruction is generated to fine-tune the parameters until the requirements are met.

[0140] Optimize full data accumulation and incremental updates: Store all data, including fault conclusions, optimization instructions, parameter adjustments, and verification results, into the historical database according to "timestamp-fault type-optimization dimension" to form a complete data link; at the same time, based on the fault characteristic patterns of this optimization, complete the incremental iteration of the fusion rule base (supplement matching conditions and adjust weights) and the retraining of the machine learning model (add incremental data and optimize feature weights).

[0141] Regular review and closed-loop iteration of optimization results: The system automatically generates a summary report of the optimization results of closed-loop feedback on a weekly or monthly basis and pushes it to the terminal of the operation and maintenance personnel. The operation and maintenance personnel can manually fine-tune the parameters according to the on-site working conditions. The system is driven by historical fault data and real-time diagnostic data, continuously triggering the closed-loop feedback process to achieve infinite iterative self-optimization, so that the diagnostic accuracy can be continuously improved with the equipment running time and the accumulation of fault data.

[0142] Example 1: Diagnosis of overcurrent fault in phase A power module caused by cooling fan failure

[0143] This embodiment takes the overcurrent of the A-phase power module caused by the converter cooling fan failure of a 1.5MW doubly fed wind turbine in a wind farm as an example to explain in detail the entire fault diagnosis process of the present invention and verify the diagnostic accuracy and tracing capability of the system under complex faults.

[0144] The wind turbine's doubly-fed induction generator has a rated current of 800A and a rated DC bus voltage of 1000V. The fault recording module has a sampling frequency of 2900Hz. Main control / pitch data is acquired based on the Modbus TCP / IP protocol at 50ms intervals. Event sequence recording timestamp accuracy is ≤1ms. All system layers are connected via an industrial Ethernet data bus. The overall architecture conforms to the appendix. Figure 1 The four-layer structure design.

[0145] S1: Data Acquisition Triggered

[0146] During the operation of the wind turbine, the fault recording module detected a sudden increase in the stator current of phase A to 1700A (exceeding 1.2 times the rated current), and the DC bus voltage instantaneously dropped to 920V, meeting the automatic triggering condition for electrical quantity exceeding the threshold. The system immediately started all modules of the data acquisition layer to work synchronously.

[0147] Switch quantity acquisition module ( Figure 1 Data Acquisition Layer: Data acquisition layer collects switch quantities such as cooling fan operating status, A-phase power module temperature alarm, and main circuit breaker status through hard wiring and stores them in the format of "switch name-timestamp-status (0 / 1)".

[0148] Event sequence recording module ( Figure 1 Data Acquisition Layer: Records timestamps for events such as "cooling fan power failure", "A-phase power module high temperature alarm", and "A-phase stator overcurrent protection activation", with a resolution of 1ms;

[0149] Fault recording module ( Figure 1 Data Acquisition Layer): Acquire analog quantities such as A / B / C phase voltage, stator / rotor A / B / C phase current, and DC bus voltage from the grid side in COMTRADE99 format, and generate .HDR / .CFG / .DAT / .INF files;

[0150] Main control / pitch data transmission module (with) Figure 1 Data Acquisition Layer: Data such as generator speed of 1500 r / min (rated speed), blade angle of 20° (normal operating angle), and normal pitch motor current are acquired via Modbus TCP / IP protocol, with no main control / pitch-related fault characteristics. All acquired data is transmitted in real time to the data processing layer via the data bus to complete the aggregation of multi-source data.

[0151] S2: Data Processing

[0152] The data processing layer performs interpolation completion, time axis alignment, noise reduction, and fault feature extraction on the aggregated data. The specific process is as follows:

[0153] Waveform recording data interpolation: The low-frequency channel data of the fault waveform recording is supplemented with a second-order Lagrange interpolation algorithm to unify the sampling frequency of all channels to 2900Hz, thereby eliminating the analysis error caused by inconsistent sampling frequencies;

[0154] Multi-source data time alignment: Based on the millisecond-level time axis of the fault waveform data, a dynamic window alignment algorithm is adopted. Figure 1 The data processing layer completes the time synchronization of all data, with the event sequence recording data window at ±1ms, the switch quantity data window at ±1ms, and the main control / pitch data window automatically shrinking to ±25ms due to fault triggering. Finally, the time accuracy of all data is calibrated to ≤1ms, generating a multi-source fusion dataset with a unified time axis.

[0155] Denoising: Wavelet thresholding is used to denoise the recorded current / voltage waveforms to eliminate waveform glitches caused by electromagnetic interference from wind farms, and invalid redundant frames are removed from the switch quantity and event sequence recording data.

[0156] Feature extraction operations ( Figure 1 Data processing layer, core algorithm matching attachment Figure 3 (Algorithm module of the waveform diagnostic system)

[0157] Fourier algorithm ( Figure 3 ): Perform full-cycle calculation on 1024 sampling points of the recorded data, extract the fundamental amplitude of the stator current of phase A as 1700A, the fundamental phase shift is 5°, and the fundamental amplitude of the currents of the other phases is normal.

[0158] Improved symmetric component method ( Figure 3 The zero-sequence current was calculated to be 0.6A (>0.5A fault threshold), the negative-sequence current was 0.12 times the rated current, and the positive-sequence current showed no significant deviation, which is consistent with the component characteristics of a single-phase grounding / overcurrent fault.

[0159] Four-sample method ( Figure 3 (Sudden change detection algorithm): Locates the fault trigger moment and detects that a small fluctuation in the A-phase current has occurred 50ms before the fault.

[0160] Data feature extraction of sequential recording of switch events: extract logical anomaly features such as "cooling fan running status = 0 (power failure)", "A phase power module high temperature alarm = 1", and "overcurrent protection action = 1", as well as the timing chain features of "fan power failure - high temperature - overcurrent protection" in the event sequence record;

[0161] Main control / pitch feature extraction: The extracted static / dynamic features such as rotational speed, blade angle, and pitch motor current are all within the normal threshold range, with no abnormal features.

[0162] S3: Independent Diagnosis of Multiple Systems

[0163] Subsystem of fusion diagnostic layer ( Figure 1 Based on the feature dataset, independent diagnosis is performed, and preliminary diagnostic results are output. The switch quantity diagnostic system process is matched. Figure 2 Waveform diagnostic system process matching Figure 3 :

[0164] Switch quantity diagnostic system ( Figure 2 After importing all digital input / output data, the system determines that the converter is in operation and executes the operation status logic judgment. It detects "cooling fan start command = 1 (main controller issues normally) but fan operation status = 0" and "A-phase power module high temperature alarm = 1". The preliminary diagnosis is cooling fan failure and abnormal temperature of A-phase power module, with a confidence level of 100%. Figure 2 (Exporting fault diagnosis conclusions)

[0165] Event Sequence Recording Diagnostic System ( Figure 1 The fault events are compared with the preset normal operation event sequence, abnormal events are filtered out and sorted by timestamp to obtain the timing chain: cooling fan power failure, phase power module high temperature alarm, A phase stator overcurrent protection action, the timing interval is <30s, the preliminary diagnosis is that the cooling fan failure caused the subsequent chain of failures, the first failure is cooling fan power failure, with a confidence level of 100%;

[0166] Waveform Diagnostic System ( Figure 3 Follow the process of "Open fault recording file → Data preprocessing → Algorithm analysis → Export fault conclusions" ( Figure 3 Waveform analysis was completed. Phase analysis: No phase shift in the three-phase voltage / current, ruling out phase sequence errors and phase loss faults; Fault component extraction: Zero-sequence current of 0.6A and a sudden increase in phase A current indicate an overcurrent fault in the phase A power module; Fault time determination: The fault time was located by the sudden change in phase current difference and matched with the event record timestamp, initially diagnosing an overcurrent fault in the phase A stator with 100% confidence.

[0167] Main control / pitch data processing diagnostic system (with) Figure 1 The extracted static / dynamic features are all within the normal threshold range, the rule base diagnosis is fault-free, the machine learning model outputs a fault probability of 0.05 (<0.8), and the preliminary diagnosis is that the main control / pitch system is fault-free with a confidence level of 100%.

[0168] The four systems will simultaneously transmit the preliminary diagnostic results (fault type + suspected cause + confidence level) to the fusion decision unit. Figure 1 This prepares for the fusion of multi-source data.

[0169] S4: Fusion Decision Diagnosis

[0170] Integrated decision-making unit ( Figure 1 The preliminary diagnostic results of the four systems are substituted into a preset three-dimensional rule base, and multi-source data fusion is performed using a weighted voting method. The specific process is as follows:

[0171] Rule base matching: Matched rule entry: Waveform diagnosis A-phase overcurrent - Event log diagnosis fan power failure - High temperature - Overcurrent - Switch quantity diagnosis fan fault - Main control / pitch no abnormality, which is consistent with the fault characteristics of "cooling fan failure causing A-phase power module overcurrent";

[0172] Weighted voting calculation: The overall confidence level is calculated according to the weighting rules (waveform data 0.5, event log data 0.3, switch data 0.2, main control / pitch data 0.1). ;

[0173] Fault level determination: If the overall confidence level is 1.0 ≥ 0.9 and the fault directly causes the converter overcurrent protection to activate and the fan to shut down, it is determined to be a Class I fault (highest priority).

[0174] Final diagnostic conclusion: The output conclusion is that at time xx, the cooling fan failure of the doubly fed converter (the first failure) caused poor heat dissipation and a sudden temperature rise in the A-phase power module, which in turn triggered an overcurrent fault in the A-phase power module. The identified component is the cooling fan of the A-phase power module. There are no abnormalities in the main control / pitch system.

[0175] S5: Results Presentation and Closed-Loop Feedback

[0176] Results Display ( Figure 1 Results Display Layer): The fused diagnostic conclusions are displayed in a unified format on the visual interactive interface. Core content:

[0177] Key information: Class I fault, cooling fan failure causing overcurrent in phase A power module. At the time of the fault, the faulty component was the cooling fan of phase A power module. Overall confidence level: 100%.

[0178] Maintenance recommendations: Immediately shut down the power supply, replace the cooling fan of the A-phase power module, retest the fan motor current and module temperature, and restart the converter after confirming that there are no abnormalities.

[0179] Highlight warning: Highlighted against a red background (Class I fault rule), along with the fault waveform (see attached image). Figure 3 Features include waveform display, event sequence recording timing diagram, and switch status table; one-click export of diagnostic results and waveform data is supported.

[0180] Multi-source data linkage: The interface supports clicking on the fault time to simultaneously retrieve the waveform recording, sequential event records, and switch status at that time point, enabling linkage analysis of fault data.

[0181] Closed-loop feedback self-optimization ( Figure 1 (Full-process closed loop): Based on the conclusions of this integrated diagnostic analysis, the system automatically initiates the closed-loop feedback process. Specific operations are as follows:

[0182] Fault feature extraction: The core features of "cooling fan failure causing overcurrent" are extracted as follows: the fan switch state loses power 35ms earlier than the overcurrent fault, the zero-sequence current is 0.6A, and the module high temperature alarm is strongly correlated with the overcurrent.

[0183] Optimization instruction generation: For "power module overcurrent caused by cooling system failure", layered optimization instructions are generated: The data acquisition layer increases the refresh rate of the switch acquisition of cooling fan status and power module temperature from 50ms to 10ms, and the fault recording module automatically increases the sampling frequency of temperature-related overcurrent fault triggering to 3500Hz; The data processing layer fine-tunes the zero-sequence current fault judgment threshold from 0.5A to 0.45A, identifies slight grounding / overcurrent trends in advance, and increases the correlation weight between cooling fan status and power module temperature to 0.3;

[0184] Command execution and verification: The optimization command is sent to the data acquisition layer and the data processing layer through the data bus. Each module completes the parameter adjustment without interference. The system is verified by reproducing historical fault data of the same type. After optimization, the feature extraction accuracy of this type of fault is improved by 8%, and the fault identification lead time reaches 50ms. The verification effect meets the standard.

[0185] Data accumulation and iteration: The diagnostic data, optimization instructions, parameter adjustments, and verification results of this fault are classified and stored in the historical database according to "timestamp-type fault-cooling system". At the same time, the fusion rule base is incrementally iterated, and strong correlation matching conditions such as "cooling fan power failure-module high temperature-overcurrent fault" are added. The machine learning model is supplemented with fault feature data of this type, and the model is retrained to improve the diagnostic efficiency and accuracy of subsequent similar faults.

[0186] During this fault diagnosis process, the system took only 19 seconds from fault triggering to outputting the final diagnostic conclusion. The fault location was accurate to the specific component (cooling fan of phase A power module). Compared with the traditional single waveform recording diagnostic method (which is easily misjudged as damage to phase A power module itself), this system achieved accurate tracing of the root cause of the fault. Based on the system's diagnostic conclusion and handling suggestions, the on-site maintenance personnel completed the fault handling and fan restart in just 25 minutes. Compared with the 2-4 hours fault location time of the traditional method, the maintenance efficiency was improved by more than 90%, and the downtime losses of the fan were greatly reduced.

[0187] Meanwhile, through closed-loop feedback parameter optimization and rule base iteration, the subsequent diagnosis and identification rate of similar cooling system faults in this wind farm has been improved to 100%, and the fault trend can be identified 50ms in advance, realizing the upgrade from "post-fault diagnosis" to "pre-fault warning", which verifies the engineering practicality and self-optimization capability of the present invention.

[0188] Example 2: Diagnosis of overcurrent faults on the grid side of an air-cooled doubly-fed converter (Class I fault)

[0189] Case Background: In a 1.5MW wind-cooled doubly-fed induction generator (DFIG) project at a wind farm, the grid-side overcurrent protection tripped during converter operation, causing the wind turbine to shut down. The converter has a rated grid-side current of 800A, a fault recording module sampling frequency of 2900Hz, and uses air cooling. The grid-side power cabinet is equipped with four sets of cooling fans, and the fan operating status is collected through a switch module.

[0190] Diagnostic process:

[0191] S1: Data Acquisition Triggered

[0192] The converter grid-side current suddenly surged to 1800A (more than twice the rated current). The fault recording module detected that the electrical quantity exceeded the threshold and automatically triggered data acquisition.

[0193] Switch signal acquisition module: Acquires switch signals such as "Grid-side power cabinet cooling fan 1-4 running status = 0", "Grid-side module high temperature alarm = 1", and "Grid-side overcurrent protection action = 1";

[0194] Event Sequence Recording Module: Records the event sequence chain "Cooling Fan 3 Power Failure" (xx:05:20.120) - "Cooling Fan 1-2 Power Failure" (xx:05:20.150) - "Grid-Side Module High Temperature" (xx:05:22.300) - "Grid-Side Overcurrent Protection Activation" (xx:05:22.500), with a timestamp accuracy of 1ms;

[0195] Fault recording module: Collects phase A / B / C voltage, current and DC bus voltage on the grid side in COMTRADE99 format and generates complete recording files;

[0196] Main control / pitch data transmission module: Collects data such as generator speed and blade angle, with no abnormal characteristics.

[0197] S2: Multi-source data processing

[0198] Data preprocessing: The sampling frequency is unified by second-order Lagrange interpolation for the waveform data. The switching quantity and event sequence are synchronized with the waveform time axis by a dynamic window alignment algorithm (the window shrinks to ±25ms at the time of the fault).

[0199] Feature extraction: Fourier algorithm extracts the fundamental amplitude of grid-side current of 1800A; symmetrical component method calculates zero-sequence current of 0.7A (exceeding fault threshold); four-sampling value method locates the fault time as xx:05:22.500; switch quantity feature extraction reveals the "time-sequence correlation between fan power failure and high temperature and overcurrent".

[0200] S3: Independent Diagnosis of Multiple Systems

[0201] Switching quantity diagnostic system ( Figure 2 After importing the switch data, it was determined that the converter was in operation. The "cooling fan start command = 1 but operation status = 0" was detected, and a preliminary diagnosis of "grid-side power cabinet cooling fan failure" was made with a confidence level of 100%.

[0202] Event sequence recording diagnostic system: By comparing the normal operation event sequence, the initial fault was located as "cooling fan 3 power failure". Subsequent fans successively lost power, triggering a chain reaction. The preliminary diagnosis is "cooling system fault causing a sequential chain alarm", with a confidence level of 100%.

[0203] Waveform Diagnostic System ( Figure 3 Waveform analysis detected a sudden increase in grid-side current and an excessive zero-sequence current, leading to a preliminary diagnosis of "grid-side overcurrent fault" with 100% confidence.

[0204] Main control / pitch diagnostic system: No abnormal features, diagnosis "Main control / pitch system normal", confidence level 100%.

[0205] S4: Fusion Decision Diagnosis

[0206] The integrated decision-making unit calculates the overall confidence level using the weighted voting method. The fault was classified as a Class I fault. The final diagnosis was: the cooling fan of the grid-side power cabinet failed and shut down, resulting in poor heat dissipation and excessive temperature of the grid-side module, which triggered an overcurrent fault on the grid side. The initial fault was the loss of power to cooling fan 3.

[0207] S5: Results Presentation and Closed-Loop Feedback

[0208] Results Display: The visualization interface highlights a type of fault with a red background and outputs maintenance suggestions such as "immediately stop the machine and check the power supply circuit and fan body of the grid-side power cabinet cooling fan. After replacing the faulty fan, retest the module temperature and current." Simultaneously, it displays the recorded waveform and event sequence diagram.

[0209] Closed-loop feedback: The system optimization command increases the refresh rate of cooling fan status acquisition from 50ms to 10ms, the waveform recording module automatically increases the sampling frequency of "fan power failure + high temperature" associated faults to 3500Hz, and the fusion rule base adds strong correlation matching conditions between fan power failure and overcurrent.

[0210] On-site inspection revealed that the circuit breaker supplying the cooling fans to the grid-side power cabinet had tripped, causing all fans to shut down, which was completely consistent with the diagnostic conclusion. After the maintenance personnel handled the issue as recommended, the converter returned to normal operation. The fault location time was only 28 minutes, which is 85% shorter than the traditional method.

[0211] Example 3: Fault Diagnosis of Pre-charge Failure During Factory Debugging (Class II Fault)

[0212] Case Background: During the commissioning phase of a 2MW doubly fed induction generator (DFIG) of a certain brand, in order to verify the pre-charge fault diagnosis logic, the control line of the pre-charge contactor was manually disconnected and insulated. After starting the converter, the system reported a fault and shut down. It is necessary to locate the root cause of the fault through the diagnostic system.

[0213] Diagnostic process

[0214] S1: Data Acquisition Triggered

[0215] The inverter failed to complete the pre-charging process after startup, triggering data acquisition:

[0216] Switch quantity acquisition module: acquired "Pre-charge contactor closing command = 1", "Pre-charge contactor status = 0", and "Pre-charge failure alarm = 1";

[0217] Event Sequence Recording Module: Records events "Pre-charge Start Command Issued" (xx:30:10.000) - "Pre-charge Contactor Not Closed" (xx:30:20.000) - "Pre-charge Failure Alarm" (xx:30:20.100);

[0218] Fault recording module: The voltage across the pre-charging resistor was 0V, and there was no charging current.

[0219] Main control / pitch data: No abnormalities.

[0220] S2: Multi-source data processing

[0221] Preprocessing: The dynamic window alignment algorithm synchronizes the switch quantity and event sequence recording data with the waveform recording time axis (pre-charge stage window ±50ms).

[0222] Feature extraction: Extract core features such as "pre-charge command and contactor status logic abnormality" and "pre-charge timeout (10s)".

[0223] S3: Independent Diagnosis of Multiple Systems

[0224] The switch quantity diagnostic system detected an inconsistency between the pre-charging contactor closing command and the actual state, and preliminarily diagnosed "the pre-charging contactor did not operate" with a confidence level of 100%.

[0225] Event sequence recording diagnostic system: By comparing with the normal pre-charge sequence (start command - contactor closing - pre-charge completion), the initial fault was located as "pre-charge start-up timeout", and the preliminary diagnosis was "pre-charge circuit abnormality" with 100% confidence.

[0226] Waveform diagnostic system: No electrical fault characteristics such as overcurrent or short circuit are found; preliminary diagnosis is "no electrical faults" with 100% confidence.

[0227] Main control / pitch diagnostic system: No abnormalities found.

[0228] S4: Fusion Decision Diagnosis

[0229] Overall confidence level The fault was classified as a Class II fault. The diagnosis was that the pre-charging contactor failed to close as instructed. The initial fault was a pre-charging activation timeout, and the root cause was an abnormality in the pre-charging contactor control circuit.

[0230] S5: Results Presentation and Closed-Loop Feedback

[0231] Results Display: The visualization interface is displayed with an orange background, and the maintenance suggestion is to "check the pre-charge contactor control line, coil power supply and contactor body status";

[0232] Closed-loop feedback: The pre-charge fault detection threshold has been optimized, and the pre-charge timeout judgment time has been adjusted from 10s to 8s to trigger alarms in advance.

[0233] The commissioning personnel checked the pre-charge contactor control line according to the diagnostic recommendations. After restoration, the pre-charge process was normal, and the diagnostic results were completely consistent with the preset fault.

[0234] The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A fault diagnosis system for doubly-fed wind power converters based on multi-source data fusion, characterized in that, It includes a data acquisition layer, a data processing layer, a fusion diagnostic layer, a results display layer, and a closed-loop diagnostic layer; The data acquisition layer is used to acquire switch quantities, event sequence records, fault waveform data, and main control pitch data in real time. And transmit the data to the data processing layer; The data processing layer is used to align the received data to the time axis, calibrate the time precision of the data based on the timestamp of the recorded data, and generate a dataset with a unified time axis. The fault waveform data is processed and fault characteristic quantities are extracted. The fusion diagnostic layer is used to analyze the feature dataset separately through multiple subsystems, output their respective preliminary diagnostic results, substitute each preliminary result into a preset rule base, perform multi-source data fusion, and output the final diagnostic conclusion. The results display layer is used to clarify the fault level classification criteria and display the integrated diagnostic conclusions in a unified format on the visual interactive interface. The closed-loop diagnostic layer is used to dynamically adapt and optimize the parameters at each level of the system based on the final conclusion of the fusion diagnosis, following the logic of "fault feature extraction - optimization instruction generation - hierarchical parameter adjustment - effect verification - data accumulation - iterative optimization". The data acquisition layer, data processing layer, fusion diagnosis layer, result display layer, and closed-loop diagnosis layer are connected via a data bus.

2. The fault diagnosis system for doubly-fed wind power converter based on multi-source data fusion according to claim 1, characterized in that, The data acquisition layer includes a switch quantity acquisition module, an event sequence recording module, a fault waveform recording module, and a wind power main control and pitch data transmission module. The digital input and output acquisition module is used to acquire digital input and output quantities and is connected to the converter controller via hardwiring. The event sequence recording module is used to record the occurrence time of converter events; The fault recording module is used to collect analog quantities of grid voltage, stator current, rotor current, and DC bus voltage. The wind power main control and pitch data transmission module is used to collect the generator speed, grid-side power, torque command of the wind power main control and the blade angle, pitch speed and pitch motor current data of the pitch system. It communicates with the main control / pitch system using Modbus TCP / IP protocol.

3. The fault diagnosis system for doubly-fed wind power converters based on multi-source data fusion according to claim 1, characterized in that, The data processing layer includes a preprocessing unit and a feature extraction unit; The preprocessing unit is used to perform secondary Lagrange interpolation on the waveform data. Based on the sampling frequency of the waveform module, it interpolates and supplements the low-frequency channel data to unify the sampling frequency of all channels. The dynamic window alignment algorithm is used to align the time axis of switch quantity, event sequence recording data, waveform data, etc. Based on the timestamp of the waveform data, the time accuracy of the data is calibrated to ≤1ms to generate a dataset with a unified time axis. The feature extraction unit is used to calculate the voltage and current fundamental amplitude and phase using the Fourier algorithm; extract positive sequence, negative sequence, and zero sequence components using the symmetrical component method; locate the fault time using the abrupt change detection algorithm; and determine the fault phase using the fault phase selection algorithm.

4. The fault diagnosis system for doubly-fed wind power converter based on multi-source data fusion according to claim 1, characterized in that, The fusion diagnostic layer includes a switch quantity diagnostic system, an event sequence recording diagnostic subsystem, a fault waveform diagnostic subsystem, a wind power main control and pitch data processing diagnostic subsystem, and a fusion decision unit. The switch quantity diagnostic subsystem is used to determine hardware faults based on the logical relationship between input and output signals. The event sequence recording and diagnostic subsystem is used to compare normal event sequences and locate the first abnormal event. The fault recording and diagnostic subsystem is used to identify electrical faults through waveform characteristics; The wind power main control and pitch data processing and diagnostic subsystem is used to extract two types of fault features, static features and dynamic features, from the pre-processed main control and pitch data, and to quantitatively characterize the fault status. The fusion decision unit is used to input the preliminary results of each subsystem into the preset rule base, perform multi-source data fusion through weighted voting, output the final diagnostic conclusion, and output the results with closed-loop feedback.

5. The fault diagnosis system for doubly-fed wind power converter based on multi-source data fusion according to claim 1, characterized in that, The results display layer includes a fault classification standard module and a visual interactive interface module; The fault classification standard module is used to clarify the fault level classification standard, including one type of fault: power module short circuit and DC bus overvoltage directly causing shutdown. Category II faults: Cooling system malfunctions and communication interruptions requiring immediate attention; Category III faults: Parameter deviations from thresholds and minor alarms that allow continued operation; Category IV faults: False alarms. The visualization interactive interface module is designed using MATLAB GUI and includes a data import area, waveform display area, event timing diagram, switch status table, and diagnostic conclusion area.

6. A fault diagnosis method for doubly-fed wind power converters based on multi-source data fusion, characterized in that, Includes the following steps: S1) Data Acquisition: Real-time acquisition of switch quantities, event sequence records, fault waveform data, and main control pitch data; S2) Data preprocessing: Based on the sampling frequency of the waveform recording module, interpolation fitting algorithm is used to interpolate and supplement the low-frequency channel data to unify the sampling frequency of all channels; dynamic window alignment algorithm is used to align the time axis of switch quantity, event sequence recording data and waveform data, and calibrate the time accuracy of the data based on the timestamp of the waveform data to generate a dataset with a unified time axis. The fault waveform data is processed and fault characteristic quantities are extracted. S3) Fusion Diagnosis: The feature dataset is analyzed by multiple subsystems, and each subsystem outputs its preliminary diagnostic results. The preliminary results of each subsystem are then substituted into a preset rule base, and multi-source data are fused using a weighted voting method to output the final diagnostic conclusion. S4) Results Display: Clearly define the fault level classification criteria and display the integrated diagnostic conclusions in a unified format on the visual interactive interface; S5) Closed-loop feedback: Based on the final conclusion of the fusion diagnosis, the system dynamically adapts and optimizes parameters at each level according to the logic of "fault feature extraction - optimization instruction generation - hierarchical parameter adjustment - effect verification - data accumulation - iterative optimization".

7. The fault diagnosis method for doubly-fed wind power converter based on multi-source data fusion according to claim 6, characterized in that, The dataset for generating a unified timeline as described in S2) specifically includes: Preprocessing of waveform data: The data value x stored in the data file is converted into the actual sample value using the conversion formula ax+b, where a is the channel multiplier and b is the channel offset addend; after processing by the formula, the actual sample value of each channel is obtained, which is the sample value of the corresponding time in each channel; The low-sampling-frequency data is converted into high-sampling-frequency data using the quadratic Lagrange interpolation algorithm. First, the interpolation position is determined. Then, the values ​​of three adjacent sampling points are selected, and the interpolation polynomial is used to calculate the value of the required interpolation position. After obtaining the estimated value of the interpolation point, if the number of samples required for the high sampling frequency is not yet reached, the obtained value is used as the sampling value for further calculation until the sampling frequency is consistent. Finally, a set of analog channel waveform data with consistent sampling frequency is obtained. Using fault waveform data as the baseline time axis, the data to be aligned includes event sequence records, switch inputs, and main control pitch data. A dynamic window alignment algorithm is employed. First, construct a reference time axis, using the data acquired by the fault recording module as the reference T. base Based on the sampling frequency and millisecond-level timestamps, a continuous reference time axis T is constructed. base ={t0,t1,t2,...,t n }, where the time interval is Δt, covering the recording period from 2 seconds before the fault is triggered to 5 seconds after the fault is triggered; Then, the three types of data to be aligned are preprocessed to extract valid timestamps and corresponding data values, and to unify the timestamp format; event sequence record data: extract <event description, timestamp t soe Event status >, timestamp precision ≤ 1ms; Switch data: Extract <switch name, timestamp t> switch Switch status (0 / 1) >, fill in missing timestamps; Main control / pitch data: extract < data items (speed / pitch angle / power), timestamp t mc Data value > 50ms timestamp collected by Modbus TCP / IP, formatted uniformly in milliseconds; Initialize the dynamic window and its core parameters, with the original timestamp t of the data to be aligned at the center of the window. x Initial window size: set according to the sampling frequency ratio of the data to be aligned to the reference data; Event sequence data recording: window size W soe =±1ms; Main control / pitch data: Window size W mc =±50ms; Switch data: Window size W switch =±1000ms; Window adjustment step size: ΔW=±1ms; Set window dynamic adaptive rules and fault shrinkage rules: When the reference waveform data detects fault characteristics, the window size of all data to be aligned is automatically shrunk by 50%; Data fluctuation adjustment rules: If the adjacent timestamp interval of the data to be aligned deviates by more than ±10%, the algorithm automatically adjusts the window size according to the deviation ratio to ensure that the window contains a valid reference time point. Perform in-window timestamp matching and data mapping, for each data point to be aligned, the timestamp t... x In its dynamic window [t x -W,t x Within +W], match the reference time axis T base The optimal reference time point The matching rules are as follows: That is, the reference time point within the window that is closest to the timestamp to be aligned is used to map the data value to be aligned one-to-one to the reference time point, thus completing the time alignment of a single data item; Finally, full data fusion and continuous time axis generation are performed, fusing all the data values ​​to be aligned to the base time axis T according to the mapped base time point. base In the process, for points in the reference time axis without matching data, a nearest neighbor preservation strategy or a linear interpolation strategy is used to complete the data, and finally a multi-source fusion dataset with a unified time axis is generated. The time accuracy is consistent with the fault recording, and the dataset is directly input into the feature extraction unit for subsequent processing.

8. The fault diagnosis method for doubly-fed wind power converter based on multi-source data fusion according to claim 6, characterized in that, The processing of fault recording data and extraction of fault feature quantities described in S2) specifically includes: analyzing fault recording data using algorithms based on a simple non-sinusoidal wave model; including the symmetrical component method, Fourier algorithm, and abrupt change phase selection algorithm; In wind power converter systems, when an asymmetrical fault occurs, the symmetrical component method is used for analysis. The asymmetrical three-phase phasors are decomposed into three sets of symmetrical components: positive sequence, negative sequence, and zero sequence, thereby simplifying the problem of asymmetrical phasors into the analysis of symmetrical phasors. A full-cycle Fourier algorithm is used to analyze the waveform data of the converter. The orthogonal properties of sine and cosine functions are used to extract a component of a specific frequency from the signal. First, the sampled signal is assumed to be a continuous periodic time function. Then, the amplitudes of the sine and cosine terms of the fundamental component are obtained according to the basic principles of Fourier series. The amplitudes of the sine and cosine terms of the fundamental component have eliminated the influence of the DC component and integer harmonic components. The effective value of the signal is obtained by combining the phasor form analysis of the signal. Using each sampling point as the starting position and the subsequent cycle as the data window, a full-cycle Fourier algorithm is performed. Through rolling recursive analysis, the overall characteristics of the wind power converter waveform data are reflected. A mutation-based phase selection algorithm is used to determine fault initiation, using ΔI. AB ΔI BC ΔI BC To represent the sudden change in current between the three phases; depending on the fault, when there is a three-phase short circuit, ΔI AB =ΔI BC =ΔI BC When a phase is short-circuited to ground, the current difference between the other two unfaulted phases is zero. When a two-phase fault occurs, the current difference between the two faulty phases is the largest. The presence or absence of a zero-sequence current component is used to distinguish between a two-phase short circuit to ground and a short circuit between two phases.

9. The fault diagnosis method for doubly-fed wind power converter based on multi-source data fusion according to claim 6, characterized in that, The analysis of the feature dataset by multiple subsystems, as described in S3), and the output of their respective preliminary diagnostic results, specifically include: Switch quantity diagnostic subsystem: The basic logic of switch quantity fault analysis is as follows: First, all input and output quantities are imported, and then the current working status of the wind power converter system is determined. If it is in the running state, the fault judgment will be performed according to the normal operation state above; if the wind power converter system is in the shutdown state, the fault judgment will be performed according to the shutdown state. Event Sequence Recording Diagnostic Subsystem: By collecting and analyzing the sequence records of fault events of a wind farm converter and comparing them with normal start-up and shutdown event records, the system can identify the first fault or the first abnormal action, thereby accurately locating the fault location. The fault recording and diagnostic subsystem first determines whether the relationship between the phases of the three-phase voltage and current is normal through phase analysis. By selecting sampling points on the waveform display interface with the mouse, the system calculates the amplitude and phase angle of each analog channel and displays them on the visualization interface. Each time a different sampling point is selected, the system performs a corresponding full-cycle Fourier algorithm calculation starting from the new sampling point and updates the values ​​on the interface, thus achieving dynamic display of phase analysis and diagnosing whether the converter has a missing phase or phase sequence error. Then, the symmetrical component method and fault phase selection algorithm are used to extract the fault component from the wind power converter fault recording. Finally, the fault time is determined, and the sudden change in phase current difference is used to determine and analyze the fault occurrence time. Wind power main control and pitch data processing and diagnostic subsystem: For pre-processed main control and pitch data, extract two types of fault features, namely static features and dynamic features, and quantify the fault status; adopt a fusion mode of rule base diagnosis and machine learning diagnosis to achieve accurate judgment of main control / pitch system faults; Static features are extracted based on threshold comparison to reflect the degree to which data deviates from the normal range; these include: blade angle deviation: actual blade angle - target blade angle, threshold > 5°, indicating pitch control execution deviation; pitch motor current over-limit: actual pitch motor current > 1.2 times rated current, indicating pitch motor overload; communication delay: time difference between master control command issuance and pitch feedback, threshold > 100ms, indicating communication link failure; and speed deviation: actual generator speed - master control target speed, threshold > 50 r / min, indicating abnormal speed control. Dynamic features are extracted based on time-series analysis, reflecting the data's variation over time; these include pitch speed volatility: the standard deviation of pitch speed over 10 consecutive frames, with a threshold > Determine pitch mechanical jamming; Torque command mutation rate: the change in torque command between adjacent frames, threshold > If the main control command is abnormal, the blade angle response is delayed: if the time difference between the issuance of the pitch command and the change in the blade angle is greater than 200ms, the pitch reducer is determined to be faulty.

10. The fault diagnosis method for doubly-fed wind power converter based on multi-source data fusion according to claim 6, characterized in that, The S4 includes: displaying the integrated diagnostic conclusions in a unified format on the visual interactive interface of the result display layer, specifically including diagnostic information: fault type, occurrence time, location component, overall confidence level, and fault level; operation and maintenance handling suggestions: outputting standardized implementation suggestions for different fault types; highlighting warning rules: Class I faults are highlighted with a red background, Class II faults with orange, Class III faults with yellow, and suspected faults with blue; data export function: supporting one-click export of diagnostic results, fault characteristic data, and waveform time series diagrams.

Images