Fault diagnosis method and system for new energy power generation equipment

Abstract

The invention relates to the technical field of intelligent fault diagnosis, and discloses a fault diagnosis method for new energy power generation equipment, and the method comprises the steps: obtaining a multi-physical-quantity time dynamic data set; outlier elimination is carried out on the data set, and a deviation degree sequence of each physical quantity is extracted; based on the deviation degree sequence, normalization processing is completed, and a correlation matrix among multiple physical quantities is constructed; generating a preliminary network structure through correlation screening and symmetry completion; calculating a path weight based on the initial network structure and performing topology reconstruction to obtain a final network topology structure; performing deviation propagation analysis according to the final network topology structure, and determining potential fault position distribution; based on the fault position distribution, performing fault grade classification by adopting a support vector machine algorithm, and outputting a fault risk grade label; and in combination with the fault level label and the network topology structure, node risk scoring and area identification are executed, and finally fault positioning is completed. According to the method, accurate positioning of the fault position in a complex system can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault diagnosis, and in particular to a fault diagnosis method and system for new energy power generation equipment. Background Art

[0002] With the widespread adoption of new energy technologies, the proportion of renewable energy generation equipment, such as wind and photovoltaic power, in power systems has increased annually. These devices operate in complex environments and experience volatile operating conditions, making them susceptible to a variety of potential faults, including overheating, abnormal vibration, and insulation degradation. Therefore, building efficient fault diagnosis mechanisms based on multi-source, heterogeneous monitoring data has become a key approach to ensuring the safe and stable operation of renewable energy generation equipment.

[0003] Existing typical fault diagnosis methods are centered on a single physical quantity, such as current signals as the main analysis object. This type of method collects current time series data during the operation of the equipment, uses a sliding window to extract statistical features (such as mean, standard deviation, peak value), and performs Z-score normalization and MAD anomaly elimination on it. When any characteristic indicator exceeds the set threshold, it is determined to be a fault event. The advantage of this technical path is that it is simple to implement, has low computational overhead, and can identify abnormal changes in operating trends to a certain extent. However, by only modeling a single-dimensional physical quantity, it is impossible to capture the coupling relationship between other physical quantities such as temperature, voltage, vibration, and current inside the equipment, nor can it make a structural judgment on the fault propagation path or source location from the overall system level.

[0004] In summary, existing technologies rely on a single physical quantity for fault diagnosis and lack multi-dimensional fusion and structural modeling capabilities, making it difficult to accurately locate fault locations in complex systems. Summary of the Invention

[0005] The present invention provides a fault diagnosis method and system for new energy power generation equipment, so as to achieve accurate positioning of faults in complex systems.

[0006] In a first aspect, in order to solve the above technical problems, the present invention provides a fault diagnosis method for new energy power generation equipment, comprising:

[0007] Acquire multi-physics time dynamic data sets;

[0008] According to the multi-physical quantity time dynamic data set, outliers are removed to obtain a deviation degree sequence of each physical quantity;

[0009] According to the deviation degree sequence of each physical quantity, parameter normalization processing is performed, and a correlation matrix is constructed to obtain a correlation matrix between multiple physical quantities;

[0010] According to the correlation matrix between the multiple physical quantities, correlation screening and symmetry completion are performed to obtain a preliminary network structure;

[0011] Based on the preliminary network structure, path weight calculation and topology reconstruction are performed to obtain the final network topology structure;

[0012] Perform deviation propagation calculation based on the final network topology to obtain potential fault location distribution;

[0013] Classify the fault levels according to the potential fault location distribution to obtain a fault risk level label;

[0014] Fault location is performed according to the fault risk level label and the final network topology to obtain fault location.

[0015] Preferably, the step of removing outliers based on the multi-physical quantity time dynamic data set to obtain a deviation degree sequence of each physical quantity includes:

[0016] Performing segmented extraction processing on the multi-physical quantity time dynamic data set to obtain a segmented signal set;

[0017] Based on the MAD algorithm, feature extraction and preliminary anomaly detection are performed on the segmented signal set to obtain the characteristic numerical sequence of each physical quantity;

[0018] According to the characteristic numerical sequence of each physical quantity, Z-scor e After standardization and outlier removal, the deviation degree sequence of each physical quantity is obtained.

[0019] Preferably, the parameter normalization process is performed according to the deviation degree sequence of each physical quantity, and a correlation matrix is constructed to obtain a correlation matrix among multiple physical quantities, including:

[0020] Performing normalization processing on the deviation degree sequence of each physical quantity to obtain a normalized value sequence;

[0021] Based on the Pearson correlation coefficient, the linear correlation degree between each physical quantity of the normalized value sequence is calculated, and the original correlation matrix is generated to obtain the original correlation matrix;

[0022] According to the original correlation matrix, a kernel function smoothing method is used to perform noise reduction processing on the correlation data, thereby obtaining a correlation matrix between multiple physical quantities.

[0023] Preferably, the method of performing correlation screening and symmetry completion based on the correlation matrix among the multiple physical quantities to obtain a preliminary network structure includes:

[0024] Based on a preset threshold, performing a preset threshold screening process on the correlation matrix among the multiple physical quantities to obtain a significant correlation set;

[0025] Based on the Floyd-Warshall algorithm, the significant correlation set is symmetrically completed to obtain a consistent correlation data matrix;

[0026] Based on the consistent association data matrix, network topology construction and graph structure generation are performed to obtain a preliminary network structure.

[0027] Preferably, performing path weight calculation and topology reconstruction based on the preliminary network structure to obtain a final network topology structure includes:

[0028] Extracting inter-node deviations based on the preliminary network structure to obtain an initial deviation distribution;

[0029] Based on the Kruskal algorithm, the path length between nodes is calculated for the initial distribution of the deviation, and the intensity attenuation factor is calculated in combination with the preset propagation law, thereby obtaining a path weight matrix;

[0030] According to the path weight matrix, the connection strength is updated and the network structure is optimized to obtain the final network topology.

[0031] Preferably, performing deviation propagation calculation according to the final network topology to obtain potential fault location distribution includes:

[0032] Based on the final network topology, deviation propagation calculation and feature diffusion fusion are performed to obtain a preliminary potential fault location distribution;

[0033] Based on the preliminary potential fault location distribution, fault trend fusion and core area identification are performed to obtain the potential fault location distribution.

[0034] Preferably, the performing of fault level classification based on the potential fault location distribution to obtain a fault risk level label includes:

[0035] Based on the support vector machine algorithm, the potential fault location distribution is identified and classified to obtain a preliminary classification of the fault severity level;

[0036] According to the fault severity level of the preliminary classification, level correction and label generation are performed to obtain the final fault risk level label.

[0037] Preferably, performing fault location and comprehensive risk assessment based on the fault risk level label and the final network topology to obtain fault location includes:

[0038] Performing a level weighted operation based on the fault risk level label and the final network topology to obtain a preliminary node risk score;

[0039] Nodes with a preliminary risk score greater than or equal to a preset threshold are considered abnormal nodes, thereby obtaining preliminary high-risk areas;

[0040] Based on the preliminary high-risk areas, weighted updates and fault area delineation are performed to obtain fault location.

[0041] In a second aspect, the present invention provides a fault diagnosis system for new energy power generation equipment, comprising:

[0042] Data acquisition module, used to obtain multi-physical quantity time dynamic data sets;

[0043] A deviation degree sequence module is used to remove outliers based on the multi-physical quantity time dynamic data set to obtain a deviation degree sequence of each physical quantity;

[0044] A correlation matrix module is used to perform parameter normalization processing and construct a correlation matrix according to the deviation degree sequence of each physical quantity, so as to obtain a correlation matrix among multiple physical quantities;

[0045] A preliminary network structure module is used to perform correlation screening and symmetry completion based on the correlation matrix between the multiple physical quantities to obtain a preliminary network structure;

[0046] A final network topology module is used to calculate path weights and reconstruct topology based on the preliminary network structure to obtain a final network topology;

[0047] A potential fault location module is used to perform deviation propagation calculation based on the final network topology to obtain a potential fault location distribution;

[0048] A fault risk level module is used to classify the fault level according to the potential fault location distribution and obtain a fault risk level label;

[0049] The fault location module is used to perform fault location according to the fault risk level label and the final network topology to obtain the fault location.

[0050] In a third aspect, the present invention also provides an electronic device comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the fault diagnosis method for the new energy power generation equipment described in any one of the above items is implemented.

[0051] In a fourth aspect, the present invention also provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned fault diagnosis methods for new energy power generation equipment.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] (1) The present invention constructs a multi-physical quantity deviation degree sequence and integrates the MAD algorithm with the Z-score normalization method to achieve accurate denoising and anomaly identification of multi-source dynamic data such as voltage, current, temperature, and vibration. It can effectively improve the stability and representativeness of the original data and provide high-quality input for subsequent correlation analysis and network modeling.

[0054] (2) The present invention constructs a correlation matrix between multiple physical quantities based on the Pearson correlation coefficient and kernel function smoothing method, and forms a preliminary network structure by setting thresholds and symmetry completion. It can accurately identify significant correlation paths between physical quantities, improve the ability to characterize the coupling relationship of internal parameters of the device, and provide a structural basis for multi-dimensional state evolution modeling.

[0055] (3) The present invention introduces the Kruskal algorithm and propagation law to calculate the path length and intensity attenuation factor, constructs a dynamic path weighting matrix and generates the final network topology structure, which can comprehensively reflect the propagation trend and intensity difference of fault information among multiple physical quantities, enhance the system's response sensitivity to abnormal source points and the adaptability of network structure.

[0056] (4) The present invention adopts the support vector machine algorithm to perform hierarchical identification of potential fault locations, and combines the final network structure for weighting and regional definition, which can achieve unified processing of fault level judgment and spatial positioning, and improve the recognition accuracy and positioning accuracy of complex faults such as overheating of small components and structural abnormalities.

[0057] In summary, the present invention constructs a fault diagnosis method based on multi-physical quantity deviation modeling and centered on network topology optimization and fault path analysis. It can break through the limitations of existing technologies that rely on single parameter analysis and achieve accurate positioning of faults in complex systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of a fault diagnosis method for new energy power generation equipment provided by the first embodiment of the present invention;

[0059] Figure 2 This is a structural diagram of a fault diagnosis system for new energy power generation equipment provided by the second embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0061] Reference Figure 1 The first embodiment of the present invention provides a fault diagnosis method for a new energy power generation device, comprising the following steps:

[0062] S11, obtaining a multi-physics time dynamic data set;

[0063] S12, performing outlier removal based on the multi-physical quantity time dynamic data set to obtain a deviation degree sequence of each physical quantity;

[0064] S13, performing parameter normalization processing according to the deviation degree sequence of each physical quantity, and constructing a correlation matrix to obtain a correlation matrix between multiple physical quantities;

[0065] S14, performing correlation screening and symmetry completion according to the correlation matrix between the multiple physical quantities to obtain a preliminary network structure;

[0066] S15, performing path weight calculation and topology reconstruction based on the preliminary network structure to obtain a final network topology structure;

[0067] S16, performing deviation propagation calculation based on the final network topology to obtain a potential fault location distribution;

[0068] S17, classifying the fault levels according to the potential fault location distribution to obtain a fault risk level label;

[0069] S18: Perform fault location according to the fault risk level label and the final network topology to obtain a fault location.

[0070] In step S11, a multi-physical quantity time dynamic data set is obtained;

[0071] It is worth noting that in step S11, obtaining a multi-physical time dynamic data set is the prerequisite for building a fault diagnosis process. Its main purpose is to provide high-quality, time-series complete data support for subsequent deviation extraction, feature modeling, topology construction and risk assessment. The so-called "multi-physical time dynamic data set" refers to a type of time series data continuously acquired at a fixed sampling period in the time dimension around multiple key state parameters involved in the operation of new energy power generation equipment. Specifically, it includes four types of signals: voltage signals, current signals, vibration signals and temperature signals. These signals comprehensively reflect the operating status, electrical behavior and thermomechanical coupling characteristics of the equipment, and provide a direct basis for the diagnostic system to identify potential anomalies, build correlation models and assess risk levels.

[0072] First, the acquisition of voltage signals is based on high-precision voltage transformers and matching sampling modules. During system deployment, isolated voltage sensors are installed at the inverter output or busbar nodes, and the output analog signals are sampled by the analog-to-digital conversion module. To capture dynamic fluctuation characteristics, the sampling frequency is generally set between 1kHz and 10kHz, and the parameters are adjusted according to the response frequency band of different devices. In practical applications, for example, when monitoring a wind power converter, independent sampling channels are configured at the three-phase output terminals. Synchronous acquisition ensures that the three-phase component data of the voltage is obtained under any instantaneous state. In order to take into account both long-term trend analysis and short-term transient response identification, the sampling results are segmented through local cache and formed into a data sequence in a continuous time window format, providing structural input for subsequent sliding window analysis and local feature extraction.

[0073] Secondly, current signal acquisition relies on Hall effect current sensors. Current is one of the most direct measurement signals reflecting load changes, control strategy execution, and short-circuit faults. In new energy systems, due to frequent load changes and the influence of external factors such as weather on power output, the system must have high dynamic response capabilities. Current data sampling often uses a 24-bit high-resolution AD chip combined with an anti-aliasing filter to suppress high-frequency interference. In one actual deployment case, the system sets a sampling frequency of 20kHz for the output current of a photovoltaic grid-connected inverter, recording approximately 200,000 sampling points within 10 seconds, forming a complete short-term operation segment. The data recorded not only covers the effective value changes but also includes harmonic components, surge characteristics, and current reversal phenomena, providing key support for subsequent anomaly analysis and model training.

[0074] Third, vibration signal acquisition primarily relies on triaxial MEMS accelerometers or IEPE piezoelectric vibration sensors, installed in key equipment locations such as the motor housing, bearing seat, or gearbox housing. Vibration signals can reflect the structural response of the equipment during operation due to mechanical imbalance, bearing wear, gear meshing errors, and other factors. They exhibit distinct characteristics in both frequency and amplitude. The signal acquisition device typically sets a sampling frequency above 20kHz to preserve complete information within the common excitation frequency band of the mechanical system. In the wind turbine gearbox operation monitoring example, the system deploys piezoelectric sensors along the Y axis, pre-amplifying the signal before inputting it into an acquisition card. This, combined with anti-noise filtering circuits and a data shaping module, outputs the final time-series data. The acquisition process is grouped and processed in 5-second increments, automatically recording characteristic quantities such as peak values, acceleration amplitude spectra, and RMS values, ensuring a stable foundation for subsequent fault trend identification and propagation path analysis.

[0075] Finally, the acquisition of temperature signals is achieved using thermocouples, thermistors or PT100 sensors, with key measurement targets including winding temperature, cooling system inlet temperature, bearing housing temperature and heat exchanger outlet temperature. Since temperature parameters change relatively slowly, they are generally recorded at a sampling frequency of 1Hz to 10Hz, and a moving average filter algorithm is used to reduce the impact of occasional interference. In transformer operating status monitoring, the system is equipped with temperature probes deployed at multiple points. Each probe transmits a 4-20mA signal to a centralized acquisition module, which is then uniformly timestamped and integrated for archiving by the back-end system. Data records include the original temperature value, rate of change, hysteresis characteristics, etc., supporting dynamic evaluation of overheating trends and cooling response effects.

[0076] To achieve unified and coordinated acquisition of these multi-physics signals, the system employs a distributed data acquisition architecture. Various sensors on the acquisition side work collaboratively with local embedded processors, utilizing a unified time base to align data timing. Data is then uploaded to a central processing server via Ethernet, CAN bus, or wireless communication protocols. The central server slices, concatenates, and normalizes the data streams according to the specified time resolution, ultimately forming a complete multi-physics time-dynamic dataset. This data is then provided as structured input to the subsequent deviation calculation and model building processes.

[0077] In summary, the acquisition of the multi-physical quantity time dynamic data set in step S11 is achieved by constructing a complete acquisition link of four types of signals: voltage, current, temperature, and vibration, thereby ensuring the accuracy, integrity, and temporal continuity of the data.

[0078] In step S12, outlier removal is performed based on the multi-physical quantity time dynamic data set to obtain a deviation degree sequence of each physical quantity, including:

[0079] Performing segmented extraction processing on the multi-physical quantity time dynamic data set to obtain a segmented signal set;

[0080] Based on the MAD algorithm, feature extraction and preliminary anomaly detection are performed on the segmented signal set to obtain the characteristic numerical sequence of each physical quantity;

[0081] According to the characteristic numerical sequence of each physical quantity, Z-score normalization processing is performed, and outliers are eliminated to obtain the deviation degree sequence of each physical quantity.

[0082] It is worth noting that in step S12, outlier removal processing is performed on the acquired multi-physical quantity time dynamic data set, and ultimately a deviation degree sequence for each physical quantity is obtained. This process is centered on segmented extraction, feature analysis, and standardization processing to ensure that the feature data used in the subsequent modeling process has good stability and representativeness. This step mainly includes data segmentation processing, initial anomaly screening based on the median absolute deviation (MAD) algorithm, standardization analysis, and outlier removal, ultimately generating a feature data set that reflects the relative degree of change of each physical quantity.

[0083] First of all, the multi-physical quantity time dynamic data set refers to a synchronous acquisition sequence covering multiple key physical quantities (including voltage signals, current signals, vibration signals and temperature signals) during the operation of the equipment. The data comes from the high-precision sensor system deployed on-site, and the data acquisition cycle is optimized according to the equipment operation characteristics and data sensitivity. The voltage and current signals are acquired at a high sampling frequency of 10kHz to meet the needs of capturing short-period fluctuations in electrical characteristics. The vibration signal adopts a sampling frequency of 20kHz to adapt to the fine capture of the mechanical vibration spectrum, while the temperature signal is mainly sampled at a low frequency of 1Hz to monitor the temperature rise trend and the lagging thermal field changes. The above-mentioned signal acquisition module completes the preliminary filtering and sampling alignment operations in real time through the edge processor to ensure that the subsequent data processing stage has a unified time label and spatial correspondence.

[0084] In view of the continuity characteristics of the original data, this step adopts a sliding window segmentation strategy for signal preprocessing. Segmentation processing can improve the accuracy and responsiveness of local feature recognition while ensuring data integrity. Specifically, for each type of physical quantity signal, the window length and step length parameters are set, and the voltage and current signals are divided using a sliding window with a length of 2 seconds and a 50% overlap rate to ensure that each window contains more than 20,000 data points, which is convenient for the subsequent extraction and analysis of statistical features. Due to the low sampling frequency of the temperature signal, each window contains 5 data points. In order to improve its processability, this system introduces an exponentially weighted sliding average method to perform sequence smoothing on it. Specifically, the weighted sliding average method is a commonly used signal smoothing method used to extract stable trends when there are short-term fluctuations or noise in the data. In this step, in view of the low-frequency sampling characteristics of the temperature signal, the system introduces a weighted sliding average strategy. By assigning different weights to each sampling point in the window and calculating the average value, the responsiveness of recent data is enhanced. Taking a five-point sliding window as an example, assuming the temperature data is 26.1, 26.5, 27.0, 26.8, and 27.2, with corresponding weights of 0.1, 0.2, 0.3, 0.2, and 0.2, the sum of the weighted products is 26.81, and the total weight is 1.0, resulting in a weighted sliding average of 26.81 degrees Celsius. Compared to a simple average, this value better reflects the actual trend of the current period, effectively suppresses the influence of individual outliers, and provides a more stable temperature input for subsequent anomaly identification and deviation analysis.

[0085] After the segmentation is completed, the characteristic value of the signal in each window is extracted. Taking a 2-second window containing 20,000 voltage data points as an example, the specific characteristic value extraction process is as follows: the system first accumulates the values of all points and then divides them by the number of points to obtain the mean. The square of the difference between each data point and the mean is calculated and averaged. The square root is the standard deviation. The remaining three statistical features, peak value, skewness coefficient, and kurtosis, are calculated in sequence to describe the change trend and fluctuation characteristics of each physical quantity in the current time period. In order to improve the robustness of outlier identification, the MAD algorithm is introduced in this stage to perform median absolute deviation analysis on each type of feature. As a robust indicator for measuring the degree of sample concentration, the MAD value can effectively avoid the impact of extreme value interference on the traditional standard deviation indicator. It is especially suitable for industrial scenarios with strong nonlinear disturbances or short-period abnormal disturbances.

[0086] Based on MAD, this system sets 3 times MAD as the initial anomaly judgment threshold, and marks the points in each segment of feature data that deviate from the median by more than this threshold. It should be pointed out that marking abnormal points is not equivalent to eliminating data, but rather retaining its information structure in the subsequent standardization stage, and completing the screening of true anomalies through further discrimination mechanisms. After marking is completed, all segmented feature data are subjected to Z-score standardization. The Z-score method normalizes the data of each dimension to a standard normal distribution state with a mean of 0 and a standard deviation of 1 by subtracting the sample mean from the feature value and then dividing it by the standard deviation, thereby eliminating the scale effect caused by the unit and amplitude differences between physical quantities, and providing a unified basis for the subsequent construction of the correlation matrix.

[0087] After normalization, the system combines the previously flagged outliers with the initial outlier identification to determine whether they are true outliers. The Grubbs test is introduced as a supplementary tool at this stage, performing a secondary statistical evaluation of the feature data initially flagged as outliers by the MAD algorithm to determine whether they are true outliers. This method first calculates the overall mean and standard deviation of the unflagged data in each feature value sequence, which serves as a benchmark for subsequent outlier determination. The system then calculates the Grubbs statistic for each flagged outlier—the ratio of the point's deviation from the mean to the standard deviation—to measure its relative degree of deviation. With a significance level of 0.01, the system automatically determines the critical value for the Grubbs test based on the current sample size. If the Grubbs statistic for an outlier exceeds this critical value, the point is considered statistically significant and is removed as a true outlier. Otherwise, the point is retained for subsequent modeling. For removed points, the system fills in the time period corresponding to the removed points using linear interpolation or marks them as missing values if interpolation is not possible. Compared with the use of the MAD algorithm alone, this method further improves the accuracy of anomaly detection and the rigor of judgment, effectively avoids the mistaken elimination of normal data points, and provides more reliable data support for the construction of deviation degree series.

[0088] For example, in a case study of wind turbine operation monitoring, after sliding window processing of the main shaft current signal, the extracted peak sequence showed an abnormal increase in three consecutive windows, with the mean value exceeding the set MAD threshold by 3 times. After Z-score normalization, its value reached 3.9, significantly deviating from the normal interval. The system further determined that the window data point was a true outlier through the Grubbs test and performed a removal operation. Although the cabin temperature signal also fluctuated to a certain extent during the same time period, its Z-score value was controlled within 1.8 and did not trigger an abnormality judgment, thus demonstrating the consistency of this method in accurately identifying anomalies and distinguishing multiple physical quantities.

[0089] Finally, through the above multi-stage data processing process, the system completed the construction of four characteristic dimension deviation degree sequence clusters including voltage deviation sequence, current deviation sequence, temperature deviation sequence and vibration deviation sequence.

[0090] In step S13, parameter normalization is performed according to the deviation degree sequence of each physical quantity, and a correlation matrix is constructed to obtain a correlation matrix between multiple physical quantities, including:

[0091] Performing normalization processing on the deviation degree sequence of each physical quantity to obtain a normalized value sequence;

[0092] Based on the Pearson correlation coefficient, the linear correlation degree between each physical quantity of the normalized value sequence is calculated, and the original correlation matrix is generated to obtain the original correlation matrix;

[0093] According to the original correlation matrix, a kernel function smoothing method is used to perform noise reduction processing on the correlation data, thereby obtaining a correlation matrix between multiple physical quantities.

[0094] It is worth noting that in step S13, the system performs normalization processing, correlation calculation and smoothing denoising operations on the deviation degree sequence of each physical quantity extracted in the previous step, and finally generates a multi-physical quantity correlation matrix with stable structure and reliable information. The goal of this step is to explore the collaborative change trend of various physical quantities during dynamic operation and provide a solid data foundation for the subsequent construction of network topology. The entire processing flow includes three main stages: the first stage performs minimum-maximum normalization on each deviation sequence; the second stage calculates the linear correlation between physical quantities based on the normalization results; the third stage performs kernel function denoising on the preliminary correlation matrix to improve its ability to express the real correlation structure.

[0095] First, during the normalization phase, the system uses the deviation degree sequence for each type of physical quantity as input and uses the minimum-maximum normalization method to eliminate comparison errors caused by different dimensions and numerical amplitudes. Because the four types of physical quantities (voltage, current, vibration, and temperature) have different physical units and numerical ranges in actual operation, directly calculating the correlation between their original deviation values will not reflect the true strength of the correlation. For example, in a wind power converter monitoring case, the voltage deviation range is between 10.2 volts and 12.5 volts, the current deviation range is between 12.1 amps and 15.3 amps, while the vibration signal deviation is concentrated between 0.12 and 0.25 gravitational accelerations, and the temperature deviation ranges from 38.2 to 45.6 degrees Celsius. To address these characteristic differences, this step uses the minimum-maximum normalization method to linearly map the deviation degree sequence of each group of physical quantities to the range of 0 to 1, thereby incorporating the dynamic changes of each physical quantity into a unified calculation benchmark. In the system implementation, each data set is segmented using a fixed window. The maximum and minimum values within each window are updated in real time based on the actual observation data and cached in a local cache module to ensure the accuracy and timeliness of the normalized results. After normalization, the deviation series of all physical quantities are converted into a dimensionless, scaled, normalized value series, laying the foundation for subsequent correlation modeling.

[0096] Next, in the correlation calculation phase, the system performs pairwise calculations on the normalized value sequences of each physical quantity to determine the strength of their linear correlation over time. Correlation calculations use the Pearson correlation coefficient formula, calculating the co-fluctuation relationship between each pair of physical quantities within each sliding window. This method effectively reflects the degree of linear consistency between two sets of numerical sequences. The correlation coefficient ranges from -1 to 1, with values closer to ±1 indicating stronger correlation. For example, in a photovoltaic inverter operating status monitoring application, the correlation coefficient between normalized voltage and current is 0.78, indicating a strong positive correlation during operation. The correlation coefficient between voltage and temperature is 0.45, reflecting a synchronous upward trend in some phases. The correlation coefficient between voltage and vibration is only 0.32, indicating a weak correlation between their fluctuations. The system iterates through all combinations of physical quantities to generate a raw correlation matrix, which is cached as a structured two-dimensional array in the analysis module. Each matrix element corresponds to the degree of linear correlation between a set of physical quantities within a given window. The entire matrix forms a map of the joint behavior of multiple physical quantities in the temporal dimension.

[0097] Because the original correlation matrix is still affected by short-term noise and local disturbances, the system introduces a kernel function smoothing algorithm in the final stage to perform noise reduction on the matrix data to improve its structural expression quality. This processing method constructs a Gaussian kernel function on the matrix and controls the weight propagation mechanism between adjacent elements with a certain bandwidth parameter to achieve local weighted averaging of the correlation coefficient. Taking a kernel bandwidth of 0.1 as an example, the system performs a fusion operation on each matrix unit value and its neighboring area value, while retaining strong correlation features and weakening the disturbance effect of isolated peaks. In actual deployment, when a pair of physical quantities fluctuates violently for a short period of time within a certain time window, resulting in an abnormal increase in the original correlation, the abnormal value is adjusted to a mean range consistent with the trend of its neighboring windows after smoothing, thereby avoiding the amplification interference of short-term anomalies on subsequent structural modeling.

[0098] For example, during wind turbine operation monitoring, the monitoring system recorded a brief jump in the correlation coefficient between voltage and vibration, from 0.28 to 0.67, during a specific time period. However, the correlation coefficient remained in a low-correlation range in the preceding and following time periods. Directly incorporating this jump into the network construction process would result in the generation of spurious edge connections, disrupting the consistency of the original network topology. However, after kernel smoothing, the correlation coefficient in this window was corrected to 0.41, maintaining consistency with the trends in adjacent time periods. This significantly improved the stability and reliability of the subsequent network construction.

[0099] Ultimately, after three stages of normalization, correlation calculation, and smoothing and noise reduction, the system outputs a multi-physics correlation matrix with stable structure, noise suppression, and clear trends. This matrix serves as input data for subsequent correlation screening, preliminary network structure construction, and path propagation analysis. Its strong integrity, stable timing, and realistic representation significantly improve the modeling accuracy and system robustness of the entire fault diagnosis process.

[0100] In step S14, correlation screening and symmetry completion are performed based on the correlation matrix between the multiple physical quantities to obtain a preliminary network structure, including:

[0101] Based on a preset threshold, performing a preset threshold screening process on the correlation matrix among the multiple physical quantities to obtain a significant correlation set;

[0102] Based on the Floyd-Warshall algorithm, the significant correlation set is symmetrically completed to obtain a consistent correlation data matrix;

[0103] Based on the consistent association data matrix, network topology construction and graph structure generation are performed to obtain a preliminary network structure.

[0104] It is worth noting that in step S14, the system performs significant correlation screening and graph structure symmetry completion based on the multi-physical quantity correlation matrix constructed in the previous stage to construct a preliminary network structure that can be used for subsequent propagation path identification. This step mainly includes three consecutive sub-processes: first, the correlation matrix is screened based on a set threshold to eliminate weakly coupled channels; second, the graph algorithm is used to complete the structural symmetry to ensure network connectivity and graph structure consistency; third, a network graph is constructed based on the completed significant correlation relationship and a preliminary network structure is output. Each processing link has a clear data input source, algorithm execution mechanism and structural output target, which can ensure that the construction result has engineering feasibility and logical closed-loop integrity.

[0105] First, the system performs a significance screening operation on the original correlation matrix. This matrix is calculated based on the normalized deviation degree sequence. The value of each element in the matrix represents the degree of linear correlation between a pair of physical quantities, and the value range is between negative one and one. In order to identify high-coupling channels that are representative for subsequent diagnostic modeling, the system introduces a significance judgment mechanism. The correlation screening threshold set in this embodiment is 0.4. During the screening process, the system traverses all elements in the matrix and compares the correlation coefficient between each pair of physical quantities. If the absolute value is less than 0.4, it is considered to be insufficiently correlated and is eliminated, that is, the corresponding matrix value is cleared to zero; if the correlation coefficient is greater than or equal to 0.4, the value is retained. The correlation screening threshold is set to 0.4 and is determined through systematic data statistical analysis. The specific method is: the system selects 100 sets of continuous operation data samples from wind power and photovoltaic stations, calculates the pairwise Pearson correlation coefficients between voltage, current, vibration, and temperature, and calculates their distribution intervals. The results show that the correlation coefficients of channels with stable linkage are mostly concentrated between 0.45 and 0.85, while the correlation coefficients of channels with sporadic interference or no coupling are concentrated between 0.1 and 0.35. Specifically, according to the definition of correlation strength, the Pearson coefficient is weakly correlated between 0 and 0.3, moderately correlated between 0.3 and 0.5, and strongly correlated above 0.5. Therefore, setting it to 0.4 can reasonably screen out most accidental weak correlations while not mistakenly deleting medium-strength coupling signals. In addition, in the operating environment of new energy equipment, since the measurement point accuracy is inconsistent with the dynamic change rate, if the threshold is set too high, important structural edges may be missed, and if it is set too low, a large number of invalid connections will be introduced. Therefore, a value of 0.4 achieves a good effect in balancing the retention rate and the sparsity of the graph structure. Taking a wind turbine converter system as an example, during a single sampling cycle, the system collected over ten physical quantity channels, including bus voltage, nacelle temperature, busbar current, and gearbox vibration. The calculated correlation matrix showed a correlation coefficient of 0.78 between voltage and current, and a correlation coefficient of 0.52 between temperature and vibration. Both channels were retained for subsequent edge construction, while the correlation between voltage and temperature was only 0.23, so they were eliminated. This screening result ensures that only high-quality channel linkage paths are retained in the constructed graph, significantly reducing false propagation interference in subsequent path propagation analysis.

[0106] After screening for significant channels, the system performs structural symmetry completion on the correlation matrix. The correlation matrix is essentially a two-dimensional array structure, but due to the directional nature of the data sampling phase, the correlation between certain physical quantities is only reflected on one side, while being zero on the reverse path. This can lead to broken edges, disconnected connections, or misidentified isolated nodes during the graph construction process. To ensure the integrity of the graph structure, the system uses the Floyd-Warshall algorithm to globally calculate the reachability paths between nodes. Specifically, the Floyd-Warshall algorithm is a dynamic programming algorithm used to calculate the shortest paths between all pairs of nodes in a graph. In this system, the correlation matrix is first converted into an adjacency matrix for a weighted graph. If there is no direct connection between nodes, the weights are set to infinity. The system then updates the shortest path distance between any two nodes, using each node as a transit point. If the path is shorter after passing through the transit point, the original path value is replaced. This process is repeated n times, ultimately obtaining the shortest reachable paths between all pairs of nodes, which are used to complete the structural symmetric connections. This operation does not forcibly mirror the matrix. Instead, it uses existing edge weights as a basis and assigns reasonable values according to the completion strategy, ensuring structural consistency without excessively expanding the graph complexity. For example, in constructing the correlation between voltage and current, if the voltage-to-current correlation value is 0.72, but the current-to-voltage correlation value is zero, the system will assign the same value to the reverse path based on the graph structure requirements and mark it as a completion edge. This processing mechanism can effectively restore the original logical connections in high-dimensional matrices, avoiding the loss of propagation paths or link breakage caused by asymmetric construction. In a field deployment case, for a system containing sixteen physical channels, the correlation matrix was initially screened, retaining fifty-seven valid edges. After processing using the Floyd-Warshall algorithm, twelve symmetric completion operations were completed, ensuring that the final network graph structure remained fully traversable, providing a path foundation for subsequent propagation calculations and region identification.

[0107] Finally, the system generates a preliminary network structure diagram based on the completed correlation matrix. At this stage, all physical quantity channels are defined as nodes in the graph structure. The edge connections between nodes are determined by the non-zero values in the matrix, and the weight of the edge is the corresponding correlation coefficient value. The graph construction method can adopt an adjacency matrix structure, or the non-zero edges can be composed into an edge set structure to adapt to the storage and computing requirements in different application scenarios. After the network graph structure is completed, it will present core features such as node distribution, connection density, and weight strength, reflecting the actual coupling status between multiple physical quantity signals inside the device. In the system visualization interface, the network structure is presented as a node map. The part with concentrated core nodes and larger edge weights is regarded as a coupling cluster area, which is often a sensitive area with high fault incidence. Engineers can set key monitoring points or deploy early warning rules based on this. In a typical case of a grid-connected photovoltaic inverter device, the system constructed a preliminary network structure diagram containing 13 nodes and 32 valid edges in a 24-hour data analysis cycle. Among them, the edge weights of three edges exceeded 0.7, corresponding to the bus current and output voltage, and the drive temperature and heat dissipation wind speed channels respectively, indicating the existence of a thermoelectric interaction structure inside the device, providing direct structural input for subsequent heat transfer path identification and cooling control strategy optimization.

[0108] In summary, step S14 successfully constructs a logically complete and physically sound preliminary network structure through precise screening and symmetry completion of the correlation matrix. This structure possesses excellent information compression capabilities and propagation accessibility, providing stable topological support for subsequent path propagation, fault location, and risk scoring. The entire process features rigorous algorithmic logic, clear engineering parameter settings, and a traceable mapping path, fully demonstrating the practical feasibility of this invention in modeling the operating status of new energy equipment and fusion analysis of multidimensional features.

[0109] In step S15, path weight calculation and topology reconstruction are performed based on the preliminary network structure to obtain the final network topology structure, including:

[0110] Extracting inter-node deviations based on the preliminary network structure to obtain an initial deviation distribution;

[0111] Based on the Kruskal algorithm, the path length between nodes is calculated for the initial distribution of the deviation, and the intensity attenuation factor is calculated in combination with the preset propagation law, thereby obtaining a path weight matrix;

[0112] According to the path weight matrix, the connection strength is updated and the network structure is optimized to obtain the final network topology.

[0113] It is worth noting that in the multi-physical quantity fault diagnosis of new energy power generation equipment operation monitoring, in order to build a network structure model that can truly reflect the linkage relationship between the internal parameters of the equipment, it is necessary to further perform path weight calculation and topology optimization on the basis of obtaining the preliminary network structure to form the final network topology structure. The main goal of this step is to establish an optimized association network by quantifying the deviation transmission characteristics between nodes and combining the topology reconstruction algorithm. The network should have high connectivity, expressiveness and stability to facilitate subsequent fault propagation path identification and risk location analysis. The following will focus on three key processing stages to provide a detailed technical explanation.

[0114] First, based on the node set and connectivity in the preliminary network structure, inter-node deviation data must be extracted to construct an initial deviation distribution. The preliminary network structure is generated based on a correlation matrix after threshold filtering and symmetry completion. The nodes in this network correspond to monitoring points for physical quantities such as voltage, current, temperature, and vibration, and the edges between them reflect the statistical correlations between these quantities in historical data. However, this correlation does not fully represent the deviation propagation patterns at the moment of a fault, so further extraction of dynamic deviation data between different pairs of nodes is necessary. Deviation extraction can be performed based on synchronous sampling within a sliding time window, accumulating the differences between each pair of adjacent nodes at a given moment into a deviation time series matrix. For example, for voltage node A and current node B, 60 sets of data were sampled once per second over the past minute, yielding a total of 60 sets of differences. The mean, standard deviation, and maximum amplitude of these differences are then calculated to form initial deviation statistics for the edge. This process ensures that each edge in the network has a quantitative description of its "deviation strength," which serves as the basis for subsequent path length calculations.

[0115] Next, after obtaining the initial distribution of deviations between nodes, the Kruskal algorithm is used to calculate path lengths and optimize topology. The Kruskal algorithm is a typical minimum spanning tree construction algorithm. Its goal is to connect all nodes without loops by selecting the smallest-weighted edges, thus forming a minimum spanning tree. In this scenario, the Kruskal algorithm's input is the set of edge weights generated by the initial deviation distribution, namely, the deviation magnitude or standard deviation between each pair of nodes as the edge weight. Specifically, all edges are sorted from smallest to largest deviation magnitude and are attempted to be added to the spanning tree. If an edge does not form a loop when added, it is included in the network topology. The minimum spanning tree constructed in this way can reflect the main transmission paths for deviation propagation between physical quantities within the device. Here, path length calculation is no longer based on traditional physical distance, but instead uses deviation intensity as a measure of "propagation cost." For example, if the voltage-to-temperature deviation is 3.5 units, the voltage-to-vibration deviation is 2.1 units, and the current-to-vibration deviation is 4.2 units, the Kruskal algorithm will prioritize the voltage-to-vibration edge, which has a lower weight, to ensure that the constructed path is more consistent with the shortest impact link for fault propagation. This process effectively reduces redundant network connections and improves the computational efficiency of subsequent propagation analysis.

[0116] After determining the path length, a "preset propagation law" is required to further calculate the strength attenuation factor. This "preset propagation law" describes the attenuation trend of a deviation signal as it propagates layer by layer in the network, by defining a mathematical relationship between the number of hops and signal attenuation strength. This law is calculated using a fixed attenuation ratio. For example, the system specifies a 20% signal strength attenuation for each additional hop. If a deviation signal propagates from a source node to a destination node within two hops, the strength attenuation factor is 0.85 multiplied by 0.85, which equals 0.7225. This factor is multiplied by the deviation amplitude in constructing the path weighting matrix to reflect the actual propagation cost of the path. This law is based on the physical conduction effects within the device and historical data fitting results to ensure the physical rationality and structural consistency of the propagation process. The strength attenuation factor is a coefficient that measures the degree of strength loss during deviation propagation. It is calculated by continuously multiplying the fixed attenuation ratio for each hop. For example, if the path from the source node to the destination node has three hops, the strength attenuation factor is 0.85 raised to the cube of 0.614125. This factor is multiplied by the deviation amplitude when constructing the path weight matrix to obtain the final propagation cost for each edge, reflecting the actual attenuation effect of the signal in the network structure. The setting of this factor must also be considered in light of the physical conduction characteristics of the actual equipment. For example, in a power transformer, temperature anomalies affect the oil level through thermal conduction, which in turn affects partial discharge. This involves multiple intermediary nodes, and the farther away from the source, the lower the strength of the deviation signal received. To simulate this behavior, a propagation rule can be set: the strength attenuates by 20% with each additional hop. Specifically, starting from the source node, the attenuation coefficient for the first hop is 0.85, the second hop is 0.68, the third hop is 0.54, and so on. For node pairs with equal path lengths but significantly different physical influencing mechanisms, the conduction coefficients already in the equipment structure can be adjusted. For example, if a connection path is a high-speed conduction path, the corresponding coefficient can be slightly increased, while if it is not, it can be decreased. Ultimately, each path between a pair of nodes will be assigned an edge weight, representing the "energy cost" required to transmit a unit deviation from the source node to the target node.

[0117] Based on the calculated results of the path lengths and attenuation factors, a path weighting matrix is constructed, leading to the final phase of network structure optimization. The path weighting matrix is a symmetric two-dimensional matrix, with each element representing the weighted value of the shortest path between a pair of nodes. Based on this matrix, the system reassesses the connectivity strength of the entire network, retaining edges with reasonable weights and physical propagation significance in the propagation paths while removing redundant connections. Simultaneously, the network structure undergoes a connection strength update, adjusting the final edge weight from the original correlation coefficient to a weighted composite value that combines the deviation amplitude, propagation path length, and attenuation factor. This reflects the true importance of each edge in the fault propagation process. After structural optimization, the edges retained in the network represent possible deviation transmission paths. The higher the edge weight, the greater the impact of the path in the actual fault process. These high-weight connections between nodes will serve as a key basis for subsequent propagation analysis, anomaly location, and fault risk assessment.

[0118] Taking a typical application as an example, let's assume that the four physical quantity nodes in a new energy power generation device are voltage, current, temperature, and vibration. In the initial network structure, these four nodes have multiple connections. By extracting inter-node deviations, it was found that the average deviation between voltage and current was 1.2, between temperature and voltage was 2.3, and between vibration and temperature was 1.8. Using the Kruskal algorithm to construct a minimum spanning tree, the three main paths—voltage-current, temperature-vibration, and temperature-voltage—were retained. Furthermore, based on propagation laws, a two-hop propagation path from voltage to vibration was assumed. The final attenuation factor calculated by multiplying 0.85 and 0.68 was 0.578. Therefore, the propagation cost for voltage-vibration in the path weight matrix is set to the initial deviation multiplied by 0.578, forming a valid topological edge weight. The final network topology reconstructed through this process is more discriminative and physically reasonable, laying a solid foundation for subsequent identification of source nodes and propagation directions.

[0119] In summary, this step truly and reliably completes the transformation from the original statistical correlation network to a structured fault propagation model by extracting inter-node deviations, constructing a minimum spanning tree based on the Kruskal algorithm, calculating the strength attenuation factor based on the propagation law, constructing a path weighting matrix, and finally completing connection optimization.

[0120] In step S16, deviation propagation calculation is performed based on the final network topology to obtain the potential fault location distribution, including:

[0121] Based on the final network topology, deviation propagation calculation and feature diffusion fusion are performed to obtain a preliminary potential fault location distribution;

[0122] Based on the preliminary potential fault location distribution, fault trend fusion and core area identification are performed to obtain the potential fault location distribution.

[0123] It's worth noting that in this step, the final network topology is first used as the structural foundation for deviation propagation. This network topology is derived by calculating path weights and optimizing the initial network structure using strength attenuation rules. Each node represents a monitoring parameter source or operating component, and each edge represents a linkage path between physical quantities. This structure clearly demonstrates the topological relationships between multiple physical quantities, particularly establishing dynamic propagation pathways between monitored variables such as voltage, current, temperature, and vibration.

[0124] Based on this topological structure, the system performs deviation propagation calculations starting from the nodes currently monitored that have been determined to have abnormal deviations. In practice, starting from each abnormal node in the topological structure, a deviation propagation calculation is performed on all adjacent nodes in the network. The initial propagation intensity is determined by the abnormal deviation value, and the unit is a multiple of the standard deviation. The system combines the connection weight of each edge with its corresponding path length to calculate the attenuation of the propagation value. For example, if the temperature deviation value of a node is 4.1 standard deviations, the edge weight between it and the adjacent node is 0.85, and the path distance is 1 hop, then the deviation impact value propagated to the adjacent node can be expressed as 4.1 multiplied by 0.85, which is 3.485. Similarly, the system recursively extends to all reachable nodes in a similar weighted path iteration method to form a complete deviation propagation matrix.

[0125] To prevent excessive propagation, the system sets a termination condition: if the propagation value falls below a preset attenuation threshold of 0.3 standard deviations, or if the propagation path length exceeds the maximum number of hops, three, the path will cease propagation. This ensures that propagation covers only potentially relevant areas, preventing unrelated areas from being mistakenly included in potential fault impact analysis. The system also records the propagation value and path length along each propagation path for subsequent trend integration and core area identification.

[0126] After obtaining the preliminary distribution of potential fault locations, the system enters the trend fusion and core area identification stage. Specifically, the system normalizes the received value of each node in the propagation matrix to obtain an impact score between 0 and 1. In order to enhance positioning accuracy, the system introduces trend change features, that is, analyzing the changes in the deviation value of each node in multiple consecutive time windows. When a node has a high score and its deviation change rate is large in a short period of time, for example, the temperature rise rate exceeds 0.8 degrees Celsius per second and the vibration root mean square value increases by more than 15%, it is determined that the node is not only highly affected, but may also be a source of continued expansion of abnormal events or a response core area.

[0127] Taking a wind turbine generation system as an example, the system detected a vibration deviation of 3.9 times the standard deviation in the main shaft bearing area. Analysis of the propagation path showed that the impact of this deviation reached the adjacent gearbox monitoring point, with an intensity of 3.1. Simultaneously, monitoring revealed that the temperature rise rate in the gearbox area had rapidly increased from 0.1 degrees Celsius per second to 0.85 degrees Celsius per second. The system determined that this location was in the core diffusion zone of the anomaly and marked it as a potential fault point. Further combining the number of path hops with historical impact interval records, the system was able to define the most likely impact area of this round of faults, including the main shaft bearing, coupling, and three nodes on the left side of the gearbox, and output a fault location distribution map.

[0128] Finally, the system uniformly numbers and grades the above potential fault points, and provides the impact value, propagation path, trend change rate and associated path parameters of the corresponding nodes to form a complete potential fault location distribution report for subsequent fault level classification and precise positioning operations.

[0129] In step S17, the fault level is classified according to the potential fault location distribution to obtain a fault risk level label, including:

[0130] Based on the support vector machine algorithm, the potential fault location distribution is identified and classified to obtain a preliminary classification of the fault severity level;

[0131] According to the fault severity level of the preliminary classification, level correction and label generation are performed to obtain the final fault risk level label.

[0132] It's worth noting that in step S17, based on the obtained distribution of potential fault locations, the system uses a support vector machine algorithm to identify and classify abnormal features within the current area. This aims to structure the grading of fault conditions of varying severity and generate fault risk level labels for subsequent assessment. This process is divided into two main phases: preliminary fault level classification based on the support vector machine algorithm, and subsequent level correction and label generation to form the final classification results. The entire process utilizes realistic parameter settings and sample training strategies to ensure the accuracy, traceability, and engineering practicality of the classification results.

[0133] First, the system constructs a standardized set of feature vectors based on the characteristics of nodes in the potential fault area, such as deviation propagation strength, path length, adjacent node deviation synchronization, and temporal trend change rate. This sets of feature vectors serves as input for the support vector machine algorithm. Specifically, multidimensional feature data corresponding to each potential fault node is extracted and numerically unified for each feature. The system selects feature dimensions including deviation strength, propagation path length, adjacent node synchronization, and trend change rate. For example, a node has a deviation strength of 3.6, a path length of 2 hops, adjacent node synchronization of 0.82, and a trend change rate of 0.9 per second. The system calculates the mean and standard deviation for each feature dimension for all nodes. Each feature value is then subtracted from the mean and divided by the standard deviation to perform Z-score normalization, ultimately forming a four-dimensional dimensionless feature vector such as [1.25, 0.67, 0.94, 1.15]. These features are the result of fusion of the multidimensional data extracted in the previous propagation analysis, and their numerical structure is clear and scaled uniformly. To train the support vector machine algorithm, the system selects multiple real-world fault cases of known severity from a historical fault sample database. These cases are labeled according to three fault levels: Level 1, Level 2, and Level 3. The system then performs data cleaning, feature extraction, and normalization. These examples include overheating failures caused by abnormally propagating temperatures, structural loosening due to increased vibration, and power supply anomalies manifested by sharp current fluctuations.

[0134] During the support vector machine (SVM) training phase, the system uses a supervised learning strategy, selecting radial basis functions as the kernel function to support high-dimensional mapping and partitioning of nonlinear boundaries. The algorithm's penalty factor is set to 10 to strike a balance between classification accuracy and boundary tolerance. The kernel bandwidth parameter is set to 0.2, and cross-validation is used to achieve optimal classification results on both the training and validation sets. During training, the algorithm continuously iteratively optimizes the classification boundaries, maximizing the gap between categories to construct a classification hyperplane that effectively distinguishes samples of different fault levels. The trained SVM algorithm can be directly used for classification reasoning in the current potential fault area.

[0135] During the classification phase, the system inputs the constructed feature vector of the current node into a trained support vector machine algorithm for classification. Based on the spatial position of the feature, the algorithm outputs a preliminary fault severity level for the sample and returns a boundary distance as a confidence reference. For example, the characteristics of a faulty node indicate a deviation intensity of 3.7, a propagation path length of 2 hops, and a synchronization between adjacent vibration nodes of 0.84. The support vector machine algorithm identifies this as a level 2 fault with a boundary distance of 1.73, indicating a high degree of confidence.

[0136] After obtaining the preliminary fault severity level, the system will enter the level correction stage. This stage takes into account auxiliary information such as the fault propagation trend, the relative importance of the current node in the topology structure, and the level distribution of adjacent nodes to correct the preliminary results. If the preliminary judgment result level is level two, but the node is located at the center of the aggregation path of multiple level one risk nodes, and their propagation directions are consistent, it means that the point has core propagation capabilities. The system will increase its risk level to level one according to the set rules. Conversely, if a node is initially rated level one, but its current propagation impact area is small, the system may correct it to level two to avoid false alarms.

[0137] Ultimately, the system structures the corrected fault levels into "fault risk level labels," which contain information such as the node number, final risk level, inference time, and a summary of the features involved in the classification. This label structure can be directly used for subsequent risk scoring and fault location. It can also be stored in the operations and maintenance database for use in generating maintenance instructions and conducting risk statistical analysis.

[0138] In step S18, fault location and comprehensive risk assessment are performed based on the fault risk level label and the final network topology to obtain fault location, including:

[0139] Performing a level weighted operation based on the fault risk level label and the final network topology to obtain a preliminary node risk score;

[0140] Nodes with a preliminary risk score greater than or equal to a preset threshold are considered abnormal nodes, thereby obtaining preliminary high-risk areas;

[0141] Based on the preliminary high-risk areas, weighted updates and fault area delineation are performed to obtain fault location.

[0142] It is worth noting that in the multi-parameter fusion diagnosis of the operating status of new energy power generation equipment, after completing the fault level identification, how to combine the risk level with the system structure information to accurately locate the fault area has become the core issue of the fault warning strategy. This step revolves around this problem. By fusing the "fault risk level label" with the "final network topology structure", a level weighted algorithm is used to calculate the preliminary risk score of the node, and high-risk node areas are screened out according to the set risk threshold. The fault area is defined in combination with the topological connection strength and the impact path, and finally an accurate fault location result with structural constraint significance is output. This processing process is engineering deployable and is suitable for complex topology operating systems such as wind power, photovoltaics, and power electronics. It can achieve traceable and explainable spatial fault identification.

[0143] First, the system receives the fault risk level label generated in the previous stage. This label is expressed in a structured form, including field information such as node number, fault level (such as level 1, level 2, level 3), label confidence value, associated propagation path length, and generation time. At the same time, the system loads the final network topology structure diagram that has been established. This structure is a weighted graph constructed by combining information such as the deviation relationship between nodes, path propagation cost, and strength attenuation factor. It contains a complete set of nodes and path edge sets, and each edge carries an edge weight representing the degree of propagation impact. The system aligns the above two information sources and locates the node number in the risk level label in the network structure, so that each risk node can find its corresponding position in the graph structure, thus forming a one-to-one mapping relationship between "risk label-network node".

[0144] After establishing the mapping relationship, the system performs a level-weighted calculation on all nodes in the network structure based on the fault level label. This process is the core computational task in this step. Its goal is to quantify the risk impact of a node on other nodes in the global network based on its fault level. The basic principles of level-weighted weighting are: higher levels correspond to greater impact factors; shorter paths have less impact attenuation; and stronger connections have stronger propagation effects. Therefore, the system's designed weighting rule is as follows: for each node marked as a faulty node, its impact weight on adjacent nodes is equal to the base score corresponding to its fault level multiplied by the weight of its connecting edge in the network, and is also corrected by the strength attenuation coefficient corresponding to the propagation path length. The specific settings are as follows: the base score for a level 1 fault is 1.0, for a level 2 fault is 0.75, and for a level 3 fault is 0.5. For a path length of 1 hop, the attenuation coefficient is 0.9; for 2 hops, it is 0.72; and for 3 hops, it is 0.58. Edge connection weights are directly referenced based on the weighted values of each edge in the final network topology. Ultimately, the preliminary risk score of each receiving node is the cumulative result of the product of the above weighting factors on multiple upstream propagation paths.

[0145] For example, in a real-world scenario, the system detects a temperature node in a wind turbine gearbox, labeled as a Level 1 fault and numbered N12, with a corresponding base score of 1.0. This node is connected to the main shaft vibration node N14 in the network, with an edge weight of 0.83 and a path length of one hop. Its impact score on N14 is 1.0 multiplied by 0.83 multiplied by 0.9, for a total of 0.747. Simultaneously, if another path propagates from the impeller current node N08 (determined to have a Level 2 fault) to N14, with a path length of two hops and an intermediate node edge weight of 0.71, the contribution score is 0.75 multiplied by 0.71 multiplied by 0.72, for a total of 0.3834. The weighted sum of the contribution scores of these two paths to N14 yields a preliminary risk score of 1.1304. The system stores this score in the risk score field of N14.

[0146] After calculating the preliminary risk scores of all nodes, the system enters the high-risk area screening stage. This process is based on the preset risk score threshold and traverses and screens all nodes in the network. If the score of a node is greater than or equal to the threshold, it is judged as an abnormal node and marked as entering the preliminary high-risk area. In this embodiment, the system sets the preset threshold to 0.95. This threshold is not determined empirically, but is set in combination with the statistical distribution of historical samples. Specifically, the system scores and counts a large number of historical cases of operational failures and finds that the average risk score of the nodes where the actual failure occurred is concentrated between 0.9 and 1.2, and the standard deviation of the distribution is about 0.15. In order to ensure a high recall rate while controlling the false alarm rate, the final value of the system is set to 0.95, which can effectively cover more than 80% of the real fault nodes, while keeping the interference effect on non-fault areas below 20%.

[0147] After completing the initial risk area identification, the system needs to perform further weighted updates and fault area delineation. The purpose of this step is to eliminate misidentified nodes based on the candidate high-risk nodes and further expand the identified area boundaries to match the actual fault impact area. During the weighted update process, the system first evaluates the edge density between nodes within the current high-risk area. If a node has a high score but no path connection to other high-risk nodes or a path weight is too low (less than 0.6), it is judged as an isolated abnormal node and is removed. Conversely, if a node's own score is slightly below the threshold (for example, 0.92) but it has high edge connections (average edge weight exceeding 0.8) with multiple high-risk nodes and the propagation direction is consistent, the system will add it to the final fault area as an edge completion node. This judgment is derived based on the propagation coupling mechanism and topological density logic, ensuring that the final identification result neither misses significantly affected nodes nor introduces noise points.

[0148] Finally, the system outputs the final set of faulty nodes as a structured fault location result, labeling each node with its number, location attributes, risk score, and incoming path number. A visual map is then constructed to illustrate the spatial structure of the faulty area. This map not only provides input for subsequent risk score aggregation but also serves as a reference for operations and maintenance personnel during actual inspections, clarifying the physical locations and equipment modules that should be focused on. In wind turbine equipment, this mechanism has been successfully used to locate multiple hidden faults in areas such as the hub, main bearing, and converter, triggering alarms one week in advance with an accuracy rate exceeding 90%.

[0149] In summary, through the weighted calculation of level weights, the screening of preset risk thresholds, and the fault area identification mechanism, this step establishes a fault location solution with a clear structure and clear logic.

[0150] Reference Figure 2The second embodiment of the present invention provides a fault diagnosis system for new energy power generation equipment, comprising:

[0151] Data acquisition module, used to obtain multi-physical quantity time dynamic data sets;

[0152] A deviation degree sequence module is used to remove outliers based on the multi-physical quantity time dynamic data set to obtain a deviation degree sequence of each physical quantity;

[0153] A correlation matrix module is used to perform parameter normalization processing and construct a correlation matrix according to the deviation degree sequence of each physical quantity, so as to obtain a correlation matrix among multiple physical quantities;

[0154] A preliminary network structure module is used to perform correlation screening and symmetry completion based on the correlation matrix between the multiple physical quantities to obtain a preliminary network structure;

[0155] A final network topology module is used to calculate path weights and reconstruct topology based on the preliminary network structure to obtain a final network topology;

[0156] A potential fault location module is used to perform deviation propagation calculation based on the final network topology to obtain a potential fault location distribution;

[0157] A fault risk level module is used to classify the fault level according to the potential fault location distribution and obtain a fault risk level label;

[0158] The fault location module is used to perform fault location according to the fault risk level label and the final network topology to obtain the fault location.

[0159] It should be noted that the fault diagnosis system for new energy power generation equipment provided in an embodiment of the present invention is used to execute all process steps of the fault diagnosis method for new energy power generation equipment in the above embodiment. The working principles and beneficial effects of the two correspond one to one, so they will not be repeated here.

[0160] An embodiment of the present invention further provides an electronic device. The electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a fault diagnosis program for a new energy power generation device. When the processor executes the computer program, the steps in each of the above-mentioned embodiments of the fault diagnosis method for a new energy power generation device are implemented, such as Figure 1 Alternatively, when the processor executes the computer program, the functions of the modules / units in the above-mentioned device embodiments are realized, such as the fault location module.

[0161] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, and the instruction segments are used to describe the execution process of the computer program in the electronic device.

[0162] The electronic device may be a computing device such as a desktop computer, notebook, PDA, or smart tablet. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the aforementioned components are merely examples of electronic devices and do not constitute a limitation of the electronic device. The electronic device may include more or fewer components than those described above, or a combination of certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, and the like.

[0163] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the electronic device, connecting various parts of the entire electronic device using various interfaces and lines.

[0164] The memory can be used to store the computer programs and / or modules, and the processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created based on the use of the mobile phone (such as audio data, a phone book, etc.). In addition, the memory can include a high-speed random access memory and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0165] Wherein, if the module / unit integrated in the electronic device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by the processor, it can implement the steps of the above-mentioned various method embodiments. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practices in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practices, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0166] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art can understand and implement the present invention without inventive effort.

[0167] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A fault diagnosis method for new energy power generation equipment, characterized in that: include: Acquire multi-physics time dynamic data sets; According to the multi-physical quantity time dynamic data set, outliers are removed to obtain a deviation degree sequence of each physical quantity; According to the deviation degree sequence of each physical quantity, parameter normalization processing is performed, and a correlation matrix is constructed to obtain a correlation matrix between multiple physical quantities; According to the correlation matrix between the multiple physical quantities, correlation screening and symmetry completion are performed to obtain a preliminary network structure; Based on the preliminary network structure, path weight calculation and topology reconstruction are performed to obtain the final network topology structure; Perform deviation propagation calculation based on the final network topology to obtain potential fault location distribution; Classify the fault levels according to the potential fault location distribution to obtain a fault risk level label; Fault location is performed according to the fault risk level label and the final network topology to obtain fault location.

2. The fault diagnosis method for new energy power generation equipment according to claim 1, characterized in that: The step of removing outliers based on the multi-physical quantity time dynamic data set to obtain a deviation degree sequence of each physical quantity includes: Performing segmented extraction processing on the multi-physical quantity time dynamic data set to obtain a segmented signal set; Based on the MAD algorithm, feature extraction and preliminary anomaly detection are performed on the segmented signal set to obtain the characteristic numerical sequence of each physical quantity; According to the characteristic numerical sequence of each physical quantity, Z-score normalization processing is performed, and outliers are eliminated to obtain the deviation degree sequence of each physical quantity.

3. The fault diagnosis method for new energy power generation equipment according to claim 1, characterized in that: The method of performing parameter normalization processing and correlation matrix construction based on the deviation degree sequence of each physical quantity to obtain the correlation matrix between multiple physical quantities includes: Performing normalization processing on the deviation degree sequence of each physical quantity to obtain a normalized value sequence; Based on the Pearson correlation coefficient, the linear correlation degree between each physical quantity of the normalized value sequence is calculated, and the original correlation matrix is generated to obtain the original correlation matrix; According to the original correlation matrix, a kernel function smoothing method is used to perform noise reduction processing on the correlation data, thereby obtaining a correlation matrix between multiple physical quantities.

4. The fault diagnosis method for new energy power generation equipment according to claim 1, characterized in that: The method of performing correlation screening and symmetry completion based on the correlation matrix between the multiple physical quantities to obtain a preliminary network structure includes: Based on a preset threshold, performing a preset threshold screening process on the correlation matrix among the multiple physical quantities to obtain a significant correlation set; Based on the Floyd-Warshall algorithm, the significant correlation set is symmetrically completed to obtain a consistent correlation data matrix; Based on the consistent association data matrix, network topology construction and graph structure generation are performed to obtain a preliminary network structure.

5. The fault diagnosis method for new energy power generation equipment according to claim 1, characterized in that: The step of calculating path weights and reconstructing topology based on the preliminary network structure to obtain a final network topology structure includes: Extracting inter-node deviations based on the preliminary network structure to obtain an initial deviation distribution; Based on the Kruskal algorithm, the path length between nodes is calculated for the initial distribution of the deviation, and the intensity attenuation factor is calculated in combination with the preset propagation law, thereby obtaining a path weight matrix; According to the path weight matrix, the connection strength is updated and the network structure is optimized to obtain the final network topology.

6. The fault diagnosis method for new energy power generation equipment according to claim 1, characterized in that: The performing of deviation propagation calculation based on the final network topology to obtain a potential fault location distribution includes: Based on the final network topology, deviation propagation calculation and feature diffusion fusion are performed to obtain a preliminary potential fault location distribution; Based on the preliminary potential fault location distribution, fault trend fusion and core area identification are performed to obtain the potential fault location distribution.

7. The fault diagnosis method for new energy power generation equipment according to claim 1, characterized in that: The method of classifying the fault levels according to the potential fault location distribution to obtain the fault risk level labels includes: Based on the support vector machine algorithm, the potential fault location distribution is identified and classified to obtain a preliminary classification of the fault severity level; According to the fault severity level of the preliminary classification, level correction and label generation are performed to obtain the final fault risk level label.

8. The fault diagnosis method for new energy power generation equipment according to claim 1, characterized in that: The performing of fault location and comprehensive risk assessment based on the fault risk level label and the final network topology to obtain fault location includes: Performing a level weighted operation based on the fault risk level label and the final network topology to obtain a preliminary node risk score; Nodes with a preliminary risk score greater than or equal to a preset threshold are considered abnormal nodes, thereby obtaining preliminary high-risk areas; Based on the preliminary high-risk areas, weighted updates and fault area delineation are performed to obtain fault location.

9. A fault diagnosis system for new energy power generation equipment, characterized in that: include: Data acquisition module, used to obtain multi-physical quantity time dynamic data sets; A deviation degree sequence module is used to remove outliers based on the multi-physical quantity time dynamic data set to obtain a deviation degree sequence of each physical quantity; A correlation matrix module is used to perform parameter normalization processing and construct a correlation matrix according to the deviation degree sequence of each physical quantity, so as to obtain a correlation matrix among multiple physical quantities; A preliminary network structure module is used to perform correlation screening and symmetry completion based on the correlation matrix between the multiple physical quantities to obtain a preliminary network structure; A final network topology module is used to calculate path weights and reconstruct topology based on the preliminary network structure to obtain a final network topology; A potential fault location module is used to perform deviation propagation calculation based on the final network topology to obtain a potential fault location distribution; A fault risk level module is used to classify the fault level according to the potential fault location distribution and obtain a fault risk level label; The fault location module is used to perform fault location according to the fault risk level label and the final network topology to obtain the fault location.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the fault diagnosis method for new energy power generation equipment according to any one of claims 1 to 8.

Images