Intelligent operation and maintenance fault early warning method for rail transit variable-frequency power supply

Abstract

The invention discloses an intelligent operation and maintenance fault early warning method for a rail transit variable-frequency power supply, and relates to the technical field of power electronics and intelligent operation and maintenance, and the method comprises the following steps: setting a multi-point synchronous collection device to carry out the high-precision real-time collection of voltage and current signals of a variable-frequency power supply system, and carrying out the digital coding processing of data; and performing frequency domain analysis on the acquired voltage and current signals by using a preset spectrum analysis algorithm, and extracting harmonic amplitudes and phase characteristic parameters of different orders. According to the invention, through high-precision synchronous acquisition and spectrum analysis, rapid and accurate identification of harmonic abnormity is realized; a machine learning model and dynamic threshold adjustment are fused, and the adaptability and intelligent recognition capability of the system to fault modes under multiple working conditions are enhanced; and a trend prediction and graphical early warning platform is introduced, so that the fault risk is pre-judged in advance and visually responded, and the operation and maintenance efficiency and the operation safety of the rail transit power supply system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power electronics and intelligent operation and maintenance technology, specifically to a method for early warning of faults in intelligent operation and maintenance of frequency converter power supplies for rail transit. Background Technology

[0002] Intelligent operation and maintenance fault early warning for frequency converter power supplies in rail transit refers to the use of intelligent means (such as sensor acquisition, edge computing, artificial intelligence algorithms, big data analysis, etc.) to monitor the operating status of frequency converter power supply equipment that undertakes key functions such as traction power supply and power conversion in rail transit systems, and to continuously analyze historical data. This allows for timely identification of potential fault signs, early warning and fault location, thereby guiding maintenance personnel to carry out targeted maintenance, reducing the rate of sudden failures, improving the reliability and safety of system operation, and ultimately achieving intelligent, precise and efficient operation and maintenance work.

[0003] Existing technologies have the following shortcomings: In the current intelligent operation and maintenance fault early warning process for frequency converter power supplies in rail transit, the difficulty in timely identification of harmonic anomalies is a far-reaching problem. Harmonics, caused by nonlinear loads, lead to voltage and current distortions. Although the initial impact may not be significant, their long-term presence can cause hidden dangers such as equipment overheating, insulation aging, and malfunction of protection devices. In severe cases, it may lead to power supply system instability, communication interference, or even train operation interruptions. Because the early characteristics of harmonic faults are weak and easily confused with normal fluctuations, they are often overlooked by traditional monitoring methods, thus forming a hidden risk in the operation and maintenance system and affecting the safety and reliability of the rail transit system.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent operation and maintenance fault early warning method for rail transit frequency converter power supplies. This method achieves rapid and accurate identification of harmonic anomalies through high-precision synchronous data acquisition and spectrum analysis; it integrates machine learning models and dynamic threshold adjustment to enhance the system's adaptability and intelligent identification capabilities under multiple operating conditions; and it introduces a trend prediction and graphical early warning platform to achieve advance prediction and visual response to fault risks, significantly improving operation and maintenance efficiency and the operational safety of rail transit power supply systems, thereby solving the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent operation and maintenance fault early warning of frequency converter power supply for rail transit, comprising the following steps: A multi-point synchronous acquisition device is set up to acquire the voltage and current signals of the frequency converter power supply system with high precision in real time, and the data is digitally encoded. The collected voltage and current signals are analyzed in the frequency domain using a preset spectrum analysis algorithm to extract the amplitude and phase characteristic parameters of different orders of harmonics. An anomaly detection module based on a machine learning model is constructed, which inputs current feature parameters and compares them with historical operating data to uncover potential harmonic anomaly patterns. A dynamic threshold mechanism is introduced, which combines factors such as equipment operating environment, load type and time period to set an adaptive identification threshold for each order harmonic, thereby improving the accuracy of anomaly identification. Using time series trend prediction algorithms, abnormal harmonic parameters are predicted over multiple periods to determine their development trend and potential risk level. The judgment results are displayed to the operation and maintenance platform in real time through a graphical interface, and an early warning mechanism is triggered to push the diagnosis results and suggested measures to the maintenance personnel's terminals.

[0007] Preferably, in the high-precision real-time acquisition step, the acquisition device is a synchronous multi-channel analog-to-digital converter unit, which is deployed at the input end, output end and key load nodes of the frequency converter power supply system, respectively, with a sampling accuracy of not less than 16 bits and a sampling frequency of not less than 10kHz. The data encoding process uses a discrete-time window segmented encoding method to ensure the time alignment of data in each channel. Each data segment is automatically tagged and timestamped after acquisition, sent to local buffer, and uploaded to the central processing unit via CAN bus or fiber optic channel to ensure high real-time performance and stable transmission of data.

[0008] Preferably, the spectrum analysis algorithm employs a joint processing mechanism of windowed Fourier transform and short-time Fourier transform to perform multi-scale frequency domain analysis on the sampled signal. The specific steps are as follows: Hamming windows are used to segment the original sampled signal to reduce spectral leakage. The amplitude and phase of the fundamental wave and the 2nd to 25th harmonics are extracted from the output data of the window function based on the Fast Fourier Transform (FFT) operation. The harmonics of each order are averaged over multiple time windows to construct a harmonic characteristic matrix.

[0009] Preferably, the machine learning model is based on a supervised learning structure, using gradient boosting decision trees and support vector machines for fusion training, and its training data comes from labeled harmonic feature samples of historical normal operation cycles and failure cycles. The model training process includes feature selection, sample balancing and cross-validation steps to ensure that the training process is not affected by data bias. Among them, the GBDT model is used to identify strong nonlinear patterns, and the SVM model is used to accurately delineate anomaly boundaries in the feature space. In actual deployment, the model uses a sliding window approach to input real-time feature sequences and outputs probability values ​​to score anomalies for each order harmonic.

[0010] Preferably, based on the extracted spectral feature parameters, the following process is executed sequentially to accurately identify weak harmonic anomalies in the operation of rail transit frequency converter power supplies. The process is as follows: For each order harmonic signal, continuous sampling is performed. The amplitude data within a certain period are used to construct a time series. By calculating the root mean square error of the time series with respect to its average amplitude, the stability of harmonic fluctuations within the current time period is evaluated. The calculation formula is as follows: In the formula, It is the first The first harmonic is at the 1st order. The amplitude during each cycle, It is the first Order harmonics in Average amplitude over a period of time It is the length of the time window used for calculation, in units of periods. It is the first The root mean square error of the order harmonics; To detect whether harmonics are on the rise, the current period is taken. The average of the differences between the data and the data from the previous 5 periods is calculated using the following formula: In the formula, It is the current number The first cycle The order harmonic amplitude represents the latest sampled value. It is the 5th cycle ago First harmonic amplitude, It is the first The average growth rate of the order harmonics over the past 5 periods, denominator This represents the interval between two sampling periods; The obtained mean square error Compared with average growth rate Weighted fusion is performed to generate anomaly scores, using the following formula: In the formula, It is the weighting coefficient of the mean square error index. It is the weighting coefficient of the growth rate indicator. It is the first Comprehensive anomaly score for order harmonics.

[0011] Preferably, the dynamic threshold mechanism adopts a multi-factor fusion judgment strategy, which adjusts the identification threshold of harmonic anomalies in real time according to the current operating status. The specific steps are as follows: Extract the average value and fluctuation amplitude of the target order harmonic over multiple past working cycles as an initial reference benchmark; By combining the current operating load level of the equipment, the ambient temperature, and the changes in interference factors in historical data during the same period, the degree of impact on harmonic behavior is evaluated, and adjustment factors are set accordingly. By integrating the state changes reflected by each adjustment factor through a built-in weighting strategy, a new identification reference range is generated, and the original threshold range is dynamically corrected.

[0012] Preferably, the trend prediction algorithm uses an improved long short-term memory network, combined with a gating mechanism to perform attention weighting on key features in the input sequence; The network consists of three LSTM units and one fully connected output layer. The input is the amplitude of each harmonic and its derivative sequence of the most recent M cycles, and the output is the trend classification label for the next K cycles. The training data incorporates three trends: “increased risk,” “stable volatility,” and “declining recovery,” through expanded labeling. The model outputs the risk level for each order during inference.

[0013] Preferably, after completing the classification output of the machine learning model, to improve the accuracy and interpretability of anomaly identification, higher-order statistics and spectral dynamics indicators are introduced to perform cluster analysis and fine-grained classification judgment on harmonic signals. The specific steps are as follows: The signal within each cycle is converted into a frequency distribution form using a Fast Fourier Transform to obtain the power spectrum. The information entropy of the power spectrum is then calculated to determine the uniformity characteristics of the current harmonic signal energy distribution. The calculation formula is as follows: In the formula, The harmonic signal is at the 1st Power percentage at each frequency point It is the number of frequency components analyzed in the spectrum. It is the first Spectral power entropy of order harmonic signals; The maximum, minimum, and median values ​​of harmonics within a selected time window are extracted, and a standardized nonlinear exponent is constructed. The expression for the constructed exponent is as follows: In the formula, It is the first The nonlinear fluctuation index of an order harmonic signal within a given period Within the current analysis period window, the first The maximum value of the order harmonic amplitude sequence reflects its peak energy. It is the minimum value within the same period, used to measure the base variation of the signal. It is the median of the sequence, serving as a normalized reference point for fluctuation intensity; Spectral power entropy Nonlinear volatility index and comprehensive abnormality score Weighted fusion is performed to form the final judgment score. The weighted fusion formula is as follows: In the formula, It is the first The fusion feature score of order harmonic signals It is the spectral power entropy Weights in the fused feature score It is a non-linear volatility index Weighting in the fusion score It is a comprehensive anomaly score The weight.

[0014] Preferably, the early warning mechanism includes three parts: local audio and video prompts, cloud message push, and automatic generation of maintenance suggestions; Audio and video prompts include voice broadcasts of alarm levels and fault types, and red highlighted flashing border prompts; cloud push is based on the MQTT protocol, pushing fault information to maintenance personnel's handheld terminal APP in real time, along with the location of the fault and a graph of harmonic index changes; The maintenance suggestion system uses a pre-built knowledge graph engine to call similar cases to automatically recommend fault levels, causes, and preferred troubleshooting paths.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention achieves high-precision, low-delay extraction of harmonic signals in rail transit frequency converter power supply systems through a multi-point synchronous acquisition and joint spectrum analysis technique. Traditional methods struggle to stably identify harmonic variations in complex electromagnetic environments and are easily affected by sampling drift and phase deviation, leading to warning failures. This invention employs a high-precision analog-to-digital converter and a time synchronization mechanism, combined with short-time Fourier transform for multi-scale extraction of frequency domain features, ensuring that weak harmonic anomalies are accurately captured within milliseconds. This method significantly improves the accuracy and timeliness of fault identification, providing a solid foundation of data for early warning systems.

[0016] This invention effectively improves the intelligence and adaptability of anomaly identification by introducing a multi-model fusion and dynamic threshold adjustment mechanism. Traditional fixed threshold methods struggle to cover fluctuation ranges under different operating conditions, easily leading to false alarms or missed alarms. This invention integrates GBDT and SVM models to classify harmonic anomaly patterns and dynamically generates identification thresholds based on equipment status parameters, achieving a dual enhancement approach of "model self-learning + threshold self-adjustment." This mechanism not only improves the system's ability to handle sudden and complex harmonic problems but also ensures robustness and reliability under multiple operating conditions and scenarios.

[0017] This invention achieves a system transformation from traditional "fault response" to "proactive prediction" by constructing a trend prediction model and a graphical early warning platform. It utilizes an improved LSTM neural network to analyze the evolution trend of harmonic characteristics over time and combines this with high-order statistical indicators to form a multi-dimensional risk scoring system, enabling the system to predict future risk trends. Once a potential abnormal trend appears, the platform can immediately issue a graphical early warning and provide corresponding maintenance suggestions, significantly shortening maintenance response time. This approach not only reduces the incidence of sudden failures but also significantly improves the foresight and accuracy of operation and maintenance work, enhancing the safety assurance capabilities of the rail transit power supply system. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of a method for intelligent operation and maintenance fault early warning of frequency converter power supply in rail transit according to the present invention. Detailed Implementation

[0020] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0021] This invention provides, for example Figure 1 The method for intelligent operation and maintenance fault early warning of frequency converter power supply in rail transit, as shown, includes the following steps: A multi-point synchronous acquisition device is set up to acquire the voltage and current signals of the frequency converter power supply system with high precision in real time, and the data is digitally encoded. In the high-precision real-time acquisition step, the acquisition device is a synchronous multi-channel analog-to-digital conversion unit, which is deployed at the input end, output end and key load nodes of the frequency conversion power supply system. The sampling accuracy is not less than 16 bits and the sampling frequency is not less than 10kHz. The data encoding process adopts a discrete time window segmented encoding method to ensure the time alignment of data in each channel and avoid phase drift problems during the acquisition process; Each data segment is automatically tagged and timestamped after acquisition, sent to local buffer, and uploaded to the central processing unit via CAN bus or fiber optic channel to ensure high real-time performance and stable transmission of data.

[0022] This synchronous acquisition mechanism enables subtle changes in the early stages of a fault to be accurately captured within a millisecond timeframe, enhancing the early warning system's sensitivity to sudden harmonic events.

[0023] The collected voltage and current signals are analyzed in the frequency domain using a preset spectrum analysis algorithm to extract the amplitude and phase characteristic parameters of different orders of harmonics. The spectrum analysis algorithm employs a joint processing mechanism of windowed Fourier transform and short-time Fourier transform to perform multi-scale frequency domain analysis on the sampled signal. The specific steps are as follows: First, the Hamming window is used to divide the original sampled signal into windows to reduce spectral leakage. Second, the amplitude and phase of the fundamental wave and the 2nd to 25th harmonics are extracted from the output data of the window function based on the Fast Fourier Transform (FFT) operation. Third, the harmonics of each order are averaged over multiple time windows, and a harmonic characteristic matrix is ​​constructed.

[0024] This processing method not only improves the anti-interference capability of spectrum identification, but also maintains accurate modeling of harmonic structures under non-stationary load conditions, thereby enhancing the robustness of subsequent anomaly identification steps.

[0025] An anomaly detection module based on a machine learning model is constructed, which inputs current feature parameters and compares them with historical operating data to uncover potential harmonic anomaly patterns. The machine learning model is based on a supervised learning structure and uses gradient boosting decision tree (GBDT) and support vector machine (SVM) for fusion training. Its training data comes from labeled harmonic feature samples of historical normal operation cycles and failure cycles. The model training process includes feature selection, sample balancing and cross-validation steps to ensure that the training process is not affected by data bias. Among them, the GBDT model is used to identify strong nonlinear patterns, and the SVM model is used to accurately delineate anomaly boundaries in the feature space. In actual deployment, the model uses a sliding window approach to input real-time feature sequences and outputs probability values ​​to achieve anomaly scoring for each order harmonic. The scoring values ​​play a key role in subsequent dynamic threshold adjustments.

[0026] The model output is used not only for classification but also for predicting whether the input signal is in a risk-rising trend range.

[0027] Based on the extracted spectral feature parameters, the following process is executed sequentially to accurately identify weak harmonic anomalies in the operation of rail transit frequency converter power supplies. The process is as follows: For each order harmonic signal, continuous sampling is performed. The amplitude data within a certain period are used to construct a time series. By calculating the root mean square error of the time series with respect to its average amplitude, the stability of harmonic fluctuations within the current time period is evaluated. The calculation formula is as follows: In the formula, It is the first The first harmonic is at the 1st order. The amplitude at each cycle is a harmonic characteristic extracted from the spectrum analysis, with units of voltage (V) or current (A), depending on the monitored object. It is the first Order harmonics in Average amplitude over a period of time It is the length of the time window used for calculation, and the unit is the number of periods. It is usually selected as multiple monitoring periods of continuous sampling (such as 10 or 20 periods). It is the first The root mean square error of a harmonic order represents the intensity of that harmonic fluctuation within the monitoring period. The larger the value, the more drastic the harmonic changes, which may indicate an anomaly. Larger The value usually indicates high harmonic activity or the presence of structurally abnormal fluctuations, which helps to screen for potential risk signals.

[0028] To detect whether harmonics are on the rise, the current period is taken. The average difference between the data and the data from the previous 5 periods is calculated. If the average growth rate is significantly positive, it indicates that the amplitude of that harmonic order is rapidly increasing. This trend may represent an early sign of a fault. The formula is as follows: In the formula, It is the current number The first cycle The order harmonic amplitude represents the latest sampled value. It is the 5th cycle ago The amplitude of the first harmonic is used as a past reference value; the difference between the two represents the change in amplitude. It is the first The average growth rate of the order harmonics over the past 5 cycles reflects whether the harmonics are in a continuous upward trend. (The denominator is...) This represents the interval between two sampling periods, and calculates the average rate of change per unit period, with the unit being the change in amplitude per period. The obtained mean square error Compared with average growth rate Weighted fusion is performed to generate an anomaly score, which reflects whether a certain harmonic is currently deviating from its typical fluctuation range and showing an increasing risk. It is an important basic indicator for triggering subsequent threshold judgments and alarm actions. The generation formula is as follows: In the formula, It is the weighting coefficient of the mean squared error index, used to adjust... The degree of influence on the final score. The value is selected based on historical experience with actual harmonic anomalies (e.g., 0.6). It is the weighting coefficient of the growth rate indicator, used for adjustment. The influence can be adjusted according to the need for sensitivity to trend changes (e.g., 0.4). It is the first The comprehensive anomaly score of order harmonics is the key indicator used to determine whether a harmonic has entered the warning range. The higher the score, the greater the harmonic fluctuation and the upward trend.

[0029] By constructing a triple-computation mechanism that combines harmonic fluctuation intensity (mean square deviation), trend growth rate, and a weighted fusion anomaly score, early, weak, but potentially risky harmonic anomalies in variable frequency power supply systems can be accurately identified. This mechanism not only considers the current severity of harmonic fluctuations but also comprehensively assesses whether they are in a continuous deterioration trend, thus possessing greater dynamic sensing and predictive capabilities than the traditional fixed threshold method. It significantly enhances the fault early warning system's ability to detect "precursor faults," effectively avoiding misjudgments or missed reports due to weak early signals, and improving the foresight and proactiveness of rail transit power supply system operation and maintenance.

[0030] A dynamic threshold mechanism is introduced, which combines factors such as equipment operating environment, load type and time period to set an adaptive identification threshold for each order harmonic, thereby improving the accuracy of anomaly identification. The dynamic threshold mechanism employs a multi-factor fusion judgment strategy, adjusting the harmonic anomaly identification threshold in real time based on the current operating status. The specific steps are as follows: First, the average value and fluctuation amplitude of the target order harmonic over the past multiple working cycles are extracted as the initial reference benchmark. Secondly, by combining the current operating load level of the equipment, the ambient temperature, and the changes in interference factors in historical data during the same period, the degree of impact on harmonic behavior is evaluated, and adjustment factors are set accordingly. Furthermore, by integrating the state changes reflected by each adjustment factor through a built-in weighting strategy, a new identification reference range is generated, and the original threshold range is dynamically corrected.

[0031] This method avoids the problem of misjudgment caused by static thresholds ignoring changes in equipment operating status, enhances the system's adaptability to high-frequency disturbances, seasonal load fluctuations, and special operating modes (such as morning and evening peak hours and nighttime maintenance), and effectively improves the accuracy and robustness of fault identification.

[0032] Using time series trend prediction algorithms, abnormal harmonic parameters are predicted over multiple periods to determine their development trend and potential risk level. The trend prediction algorithm uses an improved Long Short-Term Memory (LSTM) network, combined with a gating mechanism to perform attention weighting on key features in the input sequence; The network consists of three LSTM units and one fully connected output layer. The input is the amplitude of each harmonic and its derivative sequence of the most recent M cycles, and the output is the trend classification label for the next K cycles. The training data incorporates three trends: “increased risk,” “stable volatility,” and “declining recovery,” through expanded labeling. The model outputs the risk level for each order during inference.

[0033] The prediction results are dynamically marked with different colors when displayed graphically, making it easy for maintenance personnel to identify the severity and development trend of the problem at a glance, and realizing a fully closed-loop response system of "monitoring and warning while running".

[0034] After completing the classification output of the machine learning model, to improve the accuracy and interpretability of anomaly identification, higher-order statistics and spectral dynamics indicators are introduced to perform cluster analysis and fine-grained classification judgment on harmonic signals. The specific steps are as follows: The signal in each cycle is converted into a frequency distribution form using a Fast Fourier Transform (FFT) to obtain the power spectrum. The information entropy of the power spectrum is then calculated to determine the uniformity of the current harmonic signal energy distribution, reflecting the complexity of the system state. High entropy values ​​often correspond to equipment entering an abnormal high-frequency vibration or harmonic misalignment state. The formula is as follows: In the formula, The harmonic signal is at the 1st The power percentage at each frequency point is calculated by normalizing the squared amplitude obtained from the Fast Fourier Transform (FFT), and satisfies the following conditions: This forms a probability distribution. This refers to the number of frequency components analyzed in the spectrum, which typically depends on the FFT resolution and sampling rate. For example, when sampling a 10kHz signal with 512 points, It typically has 256 positive frequency components. It is the first The spectral power entropy of an order harmonic signal is used to measure the complexity and uncertainty of the energy distribution of the harmonic in the frequency domain. The larger the value, the more dispersed the frequency components of the signal, which may indicate that there are multi-frequency disturbances or harmonic energy splitting in the system. The maximum, minimum, and median values ​​of harmonics within a selected time window are extracted to construct a standardized nonlinear index. This index can effectively identify regions of strong fluctuations or drastic changes, and is particularly suitable for judging sporadic pulse-type harmonic interference. The formula is as follows: In the formula, It is the first The nonlinear fluctuation index of an order harmonic signal within a given period is used to assess the severity of harmonic variations. A larger value indicates more severe fluctuations and poorer stability. Within the current analysis period window, the first The maximum value of the order harmonic amplitude sequence reflects its peak energy. It is the minimum value within the same period, used to measure the base variation of the signal. It is the median of the sequence, serving as a normalization reference point for volatility intensity, which helps to eliminate the influence of outliers on the volatility index. Spectral power entropy Nonlinear volatility index and comprehensive abnormality score Weighted fusion is performed to form the final discrimination score. This score is not only used to determine the severity of harmonic anomalies, but also serves as input features for support vector machines (SVM) in multi-class classification tasks, enabling precise location of anomaly types and occurrence trends. The formula is as follows: In the formula, It is the first The fusion feature score of the order harmonic signal is used to comprehensively reflect whether it is in a high-risk abnormal state. The higher the score, the more the signal deviates from the normal range. It is a key input for the final classification judgment. It is the spectral power entropy The weights in the fusion feature score control the degree of influence of spectral entropy on the final score. Spectral entropy reflects the complexity of harmonic energy distribution. The larger the value, the more the system focuses on anomalies in the frequency distribution structure, making it suitable for identifying problems with significant changes in frequency domain characteristics, such as harmonic diffusion or multi-frequency disturbances. It is a non-linear volatility index The weights in the fusion scoring control the impact of signal strength fluctuation characteristics on anomaly detection; stronger nonlinear fluctuations generally indicate that the system has entered an unstable state. The larger the value, the more important the volatility is in fault early warning, making it more suitable for dealing with sudden spikes or impact load disturbances. It is a comprehensive anomaly score The weights reflect the system's emphasis on time series behavior (such as growth trends and historical deviations). If the magnitude is large, it emphasizes that trend evolution and volatility risk are the core of judging anomalies, making it suitable for early identification of gradual degradation or chronic failures.

[0035] By introducing three advanced feature parameters—spectral power entropy, nonlinear fluctuation index, and previous anomaly score—a multi-dimensional fusion feature score is constructed to achieve refined identification and high-precision classification of harmonic anomaly signals. This method overcomes the limitations of traditional methods that rely on single thresholds or single-point indicators for anomaly detection. It comprehensively reflects the signal's changing characteristics in terms of frequency distribution, fluctuation amplitude, and time trend, thus more comprehensively characterizing the intrinsic mechanism of harmonic instability. The scoring results not only serve as a key input to the final classification model but also provide a basis for anomaly level judgment and risk assessment for subsequent operation and maintenance systems, significantly improving the intelligent judgment and adaptive decision-making capabilities of the early warning system under complex operating conditions.

[0036] The judgment results are displayed to the operation and maintenance platform in real time through a graphical interface, and an early warning mechanism is triggered to push the diagnosis results and suggested measures to the maintenance personnel's terminals. The early warning mechanism consists of three parts: local audio and video alerts, cloud message push notifications, and automatically generated maintenance suggestions. Audio and video prompts include voice broadcasts of alarm levels and fault types, and red highlighted flashing border prompts; cloud push is based on the MQTT protocol, pushing fault information to maintenance personnel's handheld terminal APP in real time, along with the location of the fault and a graph of harmonic index changes; The maintenance suggestion system uses a pre-built knowledge graph engine to call similar cases to automatically recommend fault levels, causes, and preferred troubleshooting paths.

[0037] This mechanism ensures timely information coverage and clear instruction delivery, which helps shorten decision-making response time and assists maintenance personnel in quickly locating and handling harmonic-related hazards.

[0038] Implementation Method 1: This method primarily involves deploying high-precision synchronous acquisition devices at key nodes of the rail transit frequency converter power supply system. These devices acquire voltage and current signals in real time and perform spectrum analysis to accurately capture harmonic anomalies. In rail transit power supply networks, frequency converters typically play a crucial role in regulating power frequency and improving energy efficiency. However, their operation is susceptible to interference from complex power grid environments and nonlinear loads, easily introducing power quality issues such as harmonics and waveform distortion. To improve the sensitivity and processing efficiency of the operation and maintenance system to these faults, a multi-node, multi-dimensional, real-time synchronous data acquisition system must be established.

[0039] Specifically, the system sets up three types of data acquisition terminals at the input, output, and typical load branches of the frequency converter. These terminals have built-in high-precision analog-to-digital conversion modules, enabling millisecond-level data sampling and time synchronization. The acquired raw voltage and current signals are processed in batches using window segmentation technology, accurately time-stamped, and cached in the local edge computing unit. These terminals transmit the data to the central controller via industrial Ethernet or high-speed fiber optic communication for unified data processing and spectrum conversion.

[0040] At the spectral analysis level, the acquired time-domain signal is first divided into several fixed time windows. Data within each window undergoes windowing processing to reduce spectral leakage. Subsequently, short-time spectral analysis techniques are used to extract the frequency domain amplitude and phase characteristics of multiple harmonics, constructing harmonic spectra and feature matrices. These feature matrices not only possess strong anti-interference capabilities but can also be used for subsequent intelligent identification and trend prediction analysis. The entire process achieves a closed-loop path of "data acquisition—real-time processing—harmonic feature generation," enabling the detection of abnormal signal fluctuations and the issuance of primary alarms during the acquisition phase, significantly improving the response speed to sudden power quality problems.

[0041] In addition, the system is equipped with a data fault tolerance mechanism and backup acquisition paths to ensure that acquisition tasks are not interrupted in the event of network failure or terminal malfunction. Through this approach, the rail transit frequency converter power supply system achieves high time resolution and high frequency coverage of full-domain power monitoring, providing a solid data foundation for subsequent fault early warning models and forming the first line of technical defense for proactive protection.

[0042] Implementation Method 2: This implementation method focuses on improving the intelligence level of harmonic anomaly identification. It utilizes various machine learning models to identify potential abnormal patterns in the operating data, while constructing a dynamically adjusted early warning threshold system to enhance the system's ability to cope with the complexity and suddenness of actual operating conditions.

[0043] During the model training phase, a historical sample library was constructed using long-term operational data collected from the rail transit system. The samples cover normal equipment conditions, minor disturbance conditions, and typical fault conditions. Each sample record contains statistical characteristics of multiple harmonic orders, such as mean amplitude, variation trend, and periodic changes. These features are extracted and formatted into standard vectors, serving as input to the machine learning algorithm. In terms of model structure, two types of models are used in combination: one is the gradient boosting tree model with strong feature fitting capabilities, used to capture the complex relationships between nonlinear features; the other is the support vector machine model, focused on high-dimensional feature boundary identification, used to accurately identify marginal and anomalous samples.

[0044] During actual operation, the data acquisition terminal inputs the latest feature vector into the model and outputs anomaly scores for each harmonic. The score can be interpreted as the model's confidence level in whether a data point is abnormal; a higher score indicates a more severe deviation from the normal state. To adapt to external factors such as equipment load changes, seasonal fluctuations, and power grid interference, the system introduces a dynamic threshold adjustment strategy. This strategy not only references historical statistical indicators but also incorporates current operating environment parameters, such as load level, temperature, and voltage fluctuation frequency, to dynamically calculate the warning threshold.

[0045] When the model score exceeds the current threshold, a system alarm is triggered, and the alarm result is uploaded to the central server for synchronous archiving. This dynamic threshold mechanism effectively prevents misjudgments or omissions that occur under dynamic operating conditions with traditional fixed-value judgments, making the entire fault identification process more accurate and reliable. Simultaneously, by continuously optimizing model training samples and strategy parameters, the system can sustainably improve identification accuracy and stability, providing a reliable basis for daily operation and maintenance management.

[0046] Implementation Method 3: This implementation method builds an intelligent operation and maintenance platform based on the concept of "early warning and proactive prevention". It uses deep learning technology to predict the development trend of harmonics and realizes real-time display of fault information and operation guidance through a graphical interface, which greatly improves the efficiency and accuracy of operation and maintenance response.

[0047] The platform consists of four main functional modules: a data access module, an intelligent prediction module, a risk assessment module, and a user interface. The data access module connects to the front-end acquisition system to receive and process all real-time harmonic parameters. Once the data is input, it is parsed into a standard format and sent to the intelligent prediction module for trend analysis.

[0048] The trend prediction module employs a time-series learning model based on a neural network architecture, capable of mining nonlinear trend change patterns in real-time data streams to predict harmonic variation trends in the near future. This module can identify subtle fluctuations in harmonic levels under different operating conditions, distinguishing between stable fluctuations and those indicating potential risk increases. The prediction model incorporates a key feature selection mechanism, assigning higher weight to representative harmonic changes to focus on the root causes of problems and improve prediction accuracy.

[0049] The system also includes a risk assessment module that categorizes trend results into different levels. Combining the platform's historical experience model with an expert knowledge base, it determines the potential hazard level of each type of abnormal trend and pushes different levels of alarm information based on severity. Alarm information is not only displayed on the platform's main interface but also sent to maintenance personnel's handheld terminals via a cloud push mechanism, ensuring that critical decision-making information is delivered in the shortest possible time.

[0050] The user interface adopts a modular and visual design, allowing maintenance personnel to view the trend lines, historical fluctuation charts, and future risk predictions for each harmonic order. When the system identifies a risk trend at a certain node, it automatically generates maintenance suggestions including the fault type, potential impact, and priority troubleshooting paths, significantly reducing the judgment burden on maintenance personnel and improving handling efficiency.

[0051] Through this implementation method, the system not only has the ability to respond passively, but also has the comprehensive ability to actively perceive, make forward-looking judgments and make intelligent recommendations, realizing a qualitative leap in the maintenance of rail transit power supply systems from "post-problem handling" to "pre-problem early warning".

[0052] This invention achieves high-precision, low-delay extraction of harmonic signals in rail transit frequency converter power supply systems through a multi-point synchronous acquisition and joint spectrum analysis technique. Traditional methods struggle to stably identify harmonic variations in complex electromagnetic environments and are easily affected by sampling drift and phase deviation, leading to warning failures. This invention employs a high-precision analog-to-digital converter and a time synchronization mechanism, combined with short-time Fourier transform for multi-scale extraction of frequency domain features, ensuring that weak harmonic anomalies are accurately captured within milliseconds. This method significantly improves the accuracy and timeliness of fault identification, providing a solid foundation of data for early warning systems.

[0053] This invention effectively improves the intelligence and adaptability of anomaly identification by introducing a multi-model fusion and dynamic threshold adjustment mechanism. Traditional fixed threshold methods struggle to cover fluctuation ranges under different operating conditions, easily leading to false alarms or missed alarms. This invention integrates GBDT and SVM models to classify harmonic anomaly patterns and dynamically generates identification thresholds based on equipment status parameters, achieving a dual enhancement approach of "model self-learning + threshold self-adjustment." This mechanism not only improves the system's ability to handle sudden and complex harmonic problems but also ensures robustness and reliability under multiple operating conditions and scenarios.

[0054] This invention achieves a system transformation from traditional "fault response" to "proactive prediction" by constructing a trend prediction model and a graphical early warning platform. It utilizes an improved LSTM neural network to analyze the evolution trend of harmonic characteristics over time and combines this with high-order statistical indicators to form a multi-dimensional risk scoring system, enabling the system to predict future risk trends. Once a potential abnormal trend appears, the platform can immediately issue a graphical early warning and provide corresponding maintenance suggestions, significantly shortening maintenance response time. This approach not only reduces the incidence of sudden failures but also significantly improves the foresight and accuracy of operation and maintenance work, enhancing the safety assurance capabilities of the rail transit power supply system.

[0055] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0056] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0057] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0058] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0059] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0060] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0061] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0062] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0063] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0064] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent operation and maintenance fault early warning of frequency converter power supply in rail transit, characterized in that, Includes the following steps: A multi-point synchronous acquisition device is set up to acquire the voltage and current signals of the frequency converter power supply system with high precision in real time, and the data is digitally encoded. The collected voltage and current signals are analyzed in the frequency domain using a preset spectrum analysis algorithm to extract the amplitude and phase characteristic parameters of different orders of harmonics. An anomaly detection module based on a machine learning model is constructed, which inputs current feature parameters and compares them with historical operating data to uncover potential harmonic anomaly patterns. A dynamic threshold mechanism is introduced, which combines factors such as equipment operating environment, load type and time period to set an adaptive identification threshold for each order harmonic; Using time series trend prediction algorithms, abnormal harmonic parameters are predicted over multiple periods to determine their development trend and potential risk level. The judgment results are displayed to the operation and maintenance platform in real time through a graphical interface, and an early warning mechanism is triggered to push the diagnosis results and suggested measures to the maintenance personnel's terminals.

2. The intelligent operation and maintenance fault early warning method for rail transit frequency converter power supply according to claim 1, characterized in that, In the high-precision real-time acquisition step, the acquisition device is a synchronous multi-channel analog-to-digital converter unit, which is deployed at the input end, output end and key load nodes of the frequency converter power supply system. The sampling accuracy is not less than 16 bits and the sampling frequency is not less than 10kHz. The data encoding and processing adopts a discrete time window segmented encoding method. Each data segment is automatically tagged with a timestamp after acquisition, sent to the local buffer and uploaded to the central processing unit through CAN bus or fiber optic channel.

3. The intelligent operation and maintenance fault early warning method for rail transit frequency converter power supplies according to claim 1, characterized in that, The spectrum analysis algorithm employs a joint processing mechanism of windowed Fourier transform and short-time Fourier transform to perform multi-scale frequency domain analysis on the sampled signal. The specific steps are as follows: The original sampled signal is segmented using a Hamming window; The amplitude and phase of the fundamental wave and the 2nd to 25th harmonics are extracted from the output data of the window function based on the Fast Fourier Transform (FFT) operation. The harmonics of each order are averaged over multiple time windows to construct a harmonic characteristic matrix.

4. The intelligent operation and maintenance fault early warning method for rail transit frequency converter power supply according to claim 1, characterized in that, The machine learning model is based on a supervised learning structure and uses gradient boosting decision trees and support vector machines for integrated training. Its training data comes from labeled harmonic feature samples from historical normal operation cycles and failure cycles. The model training process includes feature selection, sample balancing, and cross-validation steps to ensure that the training process is not affected by data bias. Among them, the GBDT model is used to identify strong nonlinear patterns, while the SVM model is used to accurately delineate anomaly boundaries in the feature space. In actual deployment, the model uses a sliding window approach to input real-time feature sequences and uses the model output probability values ​​to achieve anomaly scoring for each order harmonic.

5. The intelligent operation and maintenance fault early warning method for rail transit frequency converter power supply according to claim 1, characterized in that, Based on the extracted spectral feature parameters, the following process is executed sequentially to accurately identify weak harmonic anomalies in the operation of rail transit frequency converter power supplies. The process is as follows: For each order harmonic signal, continuous sampling is performed. The amplitude data within a certain period are used to construct a time series. By calculating the root mean square error of the time series with respect to its average amplitude, the stability of harmonic fluctuations within the current time period is evaluated. The calculation formula is as follows: In the formula, It is the first The first harmonic is at the 1st order. The amplitude during each cycle, It is the first Order harmonics in Average amplitude over a period of time It is the length of the time window used for calculation. It is the first The root mean square error of the order harmonics; Take the current period The average of the differences between the data and the data from the previous 5 periods is calculated using the following formula: In the formula, It is the current number The first cycle The order harmonic amplitude represents the latest sampled value. It is the 5th cycle ago First harmonic amplitude, It is the first The average growth rate of the order harmonics over the past 5 periods, denominator This represents the interval between two sampling periods; The obtained mean square error Compared with average growth rate Weighted fusion is performed to generate anomaly scores, using the following formula: In the formula, It is the weighting coefficient of the mean square error index. It is the weighting coefficient of the growth rate indicator. It is the first Comprehensive anomaly score for order harmonics.

6. The intelligent operation and maintenance fault early warning method for rail transit frequency converter power supply according to claim 1, characterized in that, The dynamic threshold mechanism employs a multi-factor fusion judgment strategy, adjusting the harmonic anomaly identification threshold in real time based on the current operating status. The specific steps are as follows: Extract the average value and fluctuation amplitude of the target order harmonic over multiple past working cycles as an initial reference benchmark; By combining the current operating load level of the equipment, the ambient temperature, and the changes in interference factors in historical data during the same period, the degree of impact on harmonic behavior is evaluated, and adjustment factors are set accordingly. By integrating the state changes reflected by each adjustment factor through a built-in weighting strategy, a new identification reference range is generated, and the original threshold range is dynamically corrected.

7. The intelligent operation and maintenance fault early warning method for rail transit frequency converter power supply according to claim 1, characterized in that, The trend prediction algorithm uses an improved long short-term memory network and combines a gating mechanism to perform attention weighting on key features in the input sequence; The network consists of three LSTM units and one fully connected output layer. The input is the amplitude of each harmonic and its derivative sequence of the most recent M cycles, and the output is the trend classification label for the next K cycles. The training data incorporates three trends: "increased risk," "stable volatility," and "declining recovery," through extended labeling. The model outputs the risk level for each order during inference.

8. A method for intelligent operation and maintenance fault early warning of frequency converter power supply for rail transit according to claim 5, characterized in that, After completing the classification output of the machine learning model, to improve the accuracy and interpretability of anomaly identification, higher-order statistics and spectral dynamics indicators are introduced to perform cluster analysis and fine-grained classification judgment on harmonic signals. The specific steps are as follows: The signal within each cycle is converted into a frequency distribution form using a Fast Fourier Transform to obtain the power spectrum. The information entropy of the power spectrum is then calculated to determine the uniformity characteristics of the current harmonic signal energy distribution. The calculation formula is as follows: In the formula, The harmonic signal is at the 1st Power percentage at each frequency point It is the number of frequency components analyzed in the spectrum. It is the first Spectral power entropy of order harmonic signals; The maximum, minimum, and median values ​​of harmonics within a selected time window are extracted, and a standardized nonlinear exponent is constructed. The expression for the constructed exponent is as follows: In the formula, It is the first The nonlinear fluctuation index of an order harmonic signal within a given period Within the current analysis period window, the first The maximum value of the order harmonic amplitude sequence reflects its peak energy. It is the minimum value within the same period, used to measure the base variation of the signal. It is the median of the sequence, serving as a normalized reference point for fluctuation intensity; Spectral power entropy Nonlinear volatility index and comprehensive abnormality score Weighted fusion is performed to form the final judgment score. The weighted fusion formula is as follows: In the formula, It is the first The fusion feature score of order harmonic signals It is the spectral power entropy Weights in the fused feature score It is a non-linear volatility index Weighting in the fusion score It is a comprehensive anomaly score The weight.

9. A method for intelligent operation and maintenance fault early warning of frequency converter power supply in rail transit according to claim 1, characterized in that, The early warning mechanism consists of three parts: local audio and video prompts, cloud message push, and automatically generated maintenance suggestions; the audio and video prompts include voice broadcasts of the alarm level and fault type, and red highlighted flashing border prompts; The cloud-based push system, based on the MQTT protocol, pushes fault information to the maintenance personnel's handheld terminal APP in real time, along with the location of the fault and a graph showing changes in harmonic indicators. The maintenance suggestion system uses a pre-built knowledge graph engine to call similar cases and automatically recommend fault levels, causes, and preferred troubleshooting paths.

Images