AI-based Intelligent Monitoring System and Method for High and Low Voltage Switchgear Operation Faults

Abstract

This invention relates to the field of secondary circuit fault monitoring technology for high and low voltage switchgear, specifically to an AI-based intelligent monitoring system and method for operational faults in high and low voltage switchgear. The system comprises a distortion event detection unit, a multi-dimensional feature extraction unit, a state evolution trend analysis unit, and a loose connection degree determination unit. This invention performs sliding window analysis on the voltage waveform of the secondary circuit of the switchgear, identifies instantaneous voltage drop distortion events, extracts multi-dimensional features, correlates in-phase distortion across cycles and constructs a time-series event sequence, uses a long short-time memory network model to predict feature evolution trends, generates predicted trajectories, and calculates the comprehensive deviation of actual features. When both the comprehensive deviation and distortion frequency exceed preset thresholds, a loose connection fault is determined at the secondary circuit node, and an early warning is output. This method can accurately identify faults and reduce false alarms and missed alarms, achieving early warning of loose connection faults and ensuring the safe and stable operation of the switchgear.

Description

Technical Field

[0001] This invention relates to the field of secondary circuit fault monitoring technology for high and low voltage switchgear, specifically to an AI-based intelligent monitoring system and method for operational faults in high and low voltage switchgear. Background Technology

[0002] Secondary circuit fault monitoring in high and low voltage switchgear is an important technology, specifically applied to the intelligent identification and early warning of loose connection faults in the secondary circuits of switchgear. It achieves accurate fault identification through voltage distortion detection and trend analysis. Loose connection faults in the secondary circuits of switchgear can cause instantaneous voltage drop distortion. However, monitoring of voltage distortion only uses a single instantaneous amplitude threshold comparison method, without combining multi-dimensional distortion characteristics, in-phase cross-cycle correlation, and temporal evolution laws for comprehensive judgment. This makes it impossible to effectively distinguish between instantaneous interference signals and the continuous distortion caused by loose connection faults, leading to false alarms, missed alarms, and difficulty in achieving early warning of loose connection faults. To solve this problem, we provide an AI-based intelligent monitoring system and method for high and low voltage switchgear operation faults. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-based intelligent monitoring system and method for high and low voltage switchgear operation faults, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, an AI-based intelligent monitoring system for operational faults in high and low voltage switchgear is provided, including:

[0005] The distortion event detection unit is used to perform sliding window analysis on the voltage waveform signal of the secondary circuit node of the distribution cabinet. When the instantaneous voltage drop within the window exceeds the preset threshold, it is recorded as a distortion event.

[0006] The multidimensional feature extraction unit is used to extract multidimensional feature parameters for each distortion event. The multidimensional feature parameters include at least the distortion duration, voltage drop depth, power frequency phase angle at the time of distortion, slope of the leading and trailing edges of the distorted waveform, and time interval between two adjacent distortion events. It also performs cross-cycle correlation marking on distortion events with the same phase angle to construct the distortion event sequence of the secondary circuit node of the distribution cabinet.

[0007] The state evolution trend analysis unit is used to predict the changing trends of various characteristic parameters in the distorted event sequence using a time series prediction model based on a long short-term memory network, with a fixed statistical period as the unit, to generate the predicted trajectory of each characteristic parameter, and to compare the actual characteristic parameters of the distorted event in the current statistical period with the predicted trajectory to calculate the comprehensive deviation.

[0008] The loose connection determination unit is used to determine that there is a loose connection fault in the secondary circuit node of the power distribution cabinet and output a warning signal when the overall deviation exceeds the preset deviation threshold and the frequency of distortion events exceeds the preset frequency threshold.

[0009] The second objective of this invention is to provide a method for implementing the AI-based intelligent monitoring system for high and low voltage switchgear operation faults as described in any of the above-mentioned claims, comprising the following steps:

[0010] S1. Perform sliding window analysis on the voltage waveform signal of the secondary circuit node of the distribution cabinet. When the instantaneous voltage drop within the window exceeds the preset threshold, it is recorded as a distortion event.

[0011] S2. Extract multi-dimensional feature parameters for each recorded distortion event. The multi-dimensional feature parameters include at least the distortion duration, voltage drop depth, power frequency phase angle at the time of distortion, slope of the leading and trailing edges of the distortion waveform, and time interval between two adjacent distortion events. Cross-cycle correlation marking is performed on distortion events with the same phase angle to construct a distortion event sequence of the secondary circuit node of the distribution cabinet.

[0012] S3. Using a fixed statistical period as a unit, a time-series prediction model based on a long short-term memory network is used to predict the changing trends of various characteristic parameters in the distorted event sequence, generate the predicted trajectory of each characteristic parameter, and compare the actual characteristic parameters of the distorted event in the current statistical period with the predicted trajectory to calculate the comprehensive deviation.

[0013] S4. When the overall deviation exceeds the preset deviation threshold and the frequency of distortion events exceeds the preset frequency threshold, it is determined that there is a loose connection fault in the secondary circuit node of the power distribution cabinet, and a corresponding warning signal is output.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0015] This invention effectively overcomes the limitations of traditional methods that rely solely on single-instantaneous amplitude threshold comparison by integrating distortion event detection, multi-dimensional feature extraction, long short-term memory network temporal prediction, and dual-threshold joint judgment. It can accurately distinguish between instantaneous interference signals and persistent distortions caused by loose connections, reducing the false alarm and false alarm rates of secondary circuit loose connection fault monitoring. This enables early and accurate warnings of loose connection faults, comprehensively capturing the temporal evolution of loose connection faults in the secondary circuit of the distribution cabinet, accurately locating the occurrence node and development trend of loose connection faults, and ensuring the safe and stable operation of high and low voltage distribution cabinets. Attached Figure Description

[0016] Figure 1 This is an overall block diagram of the present invention;

[0017] Figure 2This is the overall flowchart of the present invention.

[0018] The meanings of the labels in the diagram are as follows:

[0019] 1. Distortion event detection unit; 2. Multidimensional feature extraction unit; 3. State evolution trend analysis unit; 4. Loose connection degree determination unit. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides an AI-based intelligent monitoring system for operational faults in high and low voltage switchgear. Please refer to [link / reference]. Figure 1 As shown, it includes:

[0022] The distortion event detection unit 1 is used to perform sliding window analysis on the voltage waveform signal of the secondary circuit node of the distribution cabinet. When the instantaneous voltage drop within the window exceeds the preset threshold, it is recorded as a distortion event.

[0023] The multidimensional feature extraction unit 2 is used to extract multidimensional feature parameters for each distortion event. The multidimensional feature parameters include at least the distortion duration, voltage drop depth, power frequency phase angle at the time of distortion, slope of the leading and trailing edges of the distorted waveform, and time interval between two adjacent distortion events. It also performs cross-cycle correlation marking on distortion events under the same phase angle to construct the distortion event sequence of the secondary circuit node of the distribution cabinet.

[0024] The state evolution trend analysis unit 3 is used to predict the changing trends of various characteristic parameters in the distorted event sequence using a time series prediction model based on a long short-term memory network with a fixed statistical period as the unit, generate the predicted trajectory of each characteristic parameter, and compare the actual characteristic parameters of the distorted event in the current statistical period with the predicted trajectory to calculate the comprehensive deviation.

[0025] The loose connection determination unit 4 is used to determine that there is a loose connection fault in the secondary circuit node of the power distribution cabinet and output a warning signal when the overall deviation exceeds the preset deviation threshold and the frequency of distortion events exceeds the preset frequency threshold.

[0026] Further explanation is needed regarding the distortion event detection unit 1, which performs real-time acquisition and distortion identification of the voltage waveform. The distortion event detection unit 1 advances forward on the continuous voltage waveform signal with a fixed step size, while simultaneously conducting real-time monitoring using a sampling window with a fixed time width. The fixed step size refers to the fixed time interval between each movement of the sampling window. This parameter is preset by the system based on the characteristics of the power frequency signal, and a typical value is 1 millisecond. This ensures the continuity of voltage waveform monitoring without wasting computational resources due to an excessively small step size. The fixed-time-width sampling window is a pre-set fixed-length window used to extract segments of the voltage waveform for anomaly analysis. For a 50Hz power frequency voltage signal, the system sets the sampling window's time width to 20 milliseconds. This duration covers a complete power frequency cycle, enabling the complete capture of instantaneous changes and abrupt changes in the voltage waveform, avoiding the omission of distortion signals.

[0027] Within each moved sampling window, the distortion event detection unit 1 collects instantaneous voltage values ​​at a preset sampling frequency. The preset sampling frequency is the acquisition rate set by the system to ensure the accuracy of waveform restoration. This invention uses a sampling frequency of 10kHz, at which 10 instantaneous voltage values ​​can be collected every millisecond, which can accurately restore the subtle changes in voltage waveform and fully meet the accuracy requirements of secondary circuit voltage distortion detection.

[0028] The distortion event detection unit 1 continuously compares the instantaneous voltage values ​​collected within the sampling window in real time, and identifies the amplitude drop of the instantaneous voltage values ​​within a preset time. The preset time is a short duration set by the system to judge rapid voltage changes. This invention sets it to 1 millisecond. Only voltage amplitude drops occurring within 1 millisecond are judged as sudden distortions, thus excluding interference from normal voltage fluctuations.

[0029] When the distortion event detection unit 1 detects that the difference in voltage amplitude drops exceeds the preset drop amplitude threshold, the system immediately determines that the voltage fluctuation is an abnormal state and officially records the state as a distortion event. The preset drop amplitude threshold is a critical value calibrated by the system based on the rated normal voltage of the secondary circuit of the distribution cabinet. In this invention, the rated normal voltage is 100V, and the drop amplitude threshold is set to 10V, that is, when the instantaneous voltage drop exceeds 10V, the distortion event recording is triggered.

[0030] While recording distortion events, the distortion event detection unit 1 simultaneously collects and stores three core basic data items. The first item is the event timestamp of the distortion event, which records the time when the distortion event occurs and is used for subsequent time-series sorting and time interval calculation of distortion events. The second item is the instantaneous voltage value at the drop initiation point, which is the first sampling point where the voltage waveform begins to drop abnormally from its normal stable state. The voltage value at this point is the basis for calculating the drop amplitude and waveform slope. The third item is the instantaneous voltage value at the drop trough, which is the lowest point of the voltage waveform drop. The sampling point is also the turning point where the voltage begins to recover and rise. This value is the core parameter for calculating the voltage drop depth. After recording these three data points, the basic information collection work for a single distortion event is officially completed. After the distortion event detection unit 1 completes the identification and basic information recording of the single distortion event, the multi-dimensional feature extraction unit 2 immediately starts working to extract full-dimensional feature parameters for this distortion event, transforming the original voltage distortion signal into quantifiable, analyzable, and fault-determinable digital features. This step is a key bridge connecting the original monitoring data and the AI ​​trend analysis model. The multi-dimensional feature extraction unit 2 first calculates and determines the distortion duration. Distortion duration is a core parameter reflecting the length of voltage distortion. Specifically, it is determined by measuring the total time elapsed from the moment the voltage instantaneously drops to the moment the voltage recovers to the normal threshold range. The normal threshold range is the allowable fluctuation range of the secondary circuit voltage of the distribution cabinet during normal operation. In this invention, it is set to ±5% of the rated voltage, i.e., between 95V and 105V. The end of the distortion state is indicated when the instantaneous voltage value returns to this range. The unit of this duration is uniformly set to milliseconds, reflecting the duration of the distortion event. Subsequently, the multi-dimensional feature extraction unit 2 obtains the voltage drop depth through numerical calculation. Voltage drop depth is a key indicator for quantifying the severity of voltage distortion, and its specific calculation formula is as follows: ,in Represents the voltage drop depth, expressed as a percentage. This represents the normal voltage reference value before the voltage drop, which is the rated value for stable voltage operation before the distortion occurs. This represents the instantaneous voltage value at the bottom of the voltage drop. The larger the calculated percentage value, the greater the magnitude of the voltage drop and the more severe the distortion. This parameter is highly correlated with secondary circuit loose connection faults.

[0031] Multidimensional feature extraction unit 2 determines the power frequency phase angle at the moment of distortion occurrence using phase-locked loop (PLL) technology. PLL technology is an electronic circuit technology capable of accurately tracking and locking the phase of power frequency voltage. It can stably lock the 20-millisecond power frequency cycle of 50Hz mains power, unaffected by voltage fluctuations or distortion, and accurately locates the phase angle corresponding to the distortion initiation point within the voltage cycle. The power frequency phase angle ranges from 0° to 360°. This parameter reflects the specific location of the distortion event within the voltage cycle and is the core basis for identifying the periodic characteristics of loose connection faults. Simultaneously, multidimensional feature extraction unit 2 calculates the leading edge slope and trailing edge slope of the distorted waveform. The leading edge slope and trailing edge slope are characteristic parameters reflecting the voltage drop rate and recovery rate. The leading edge slope of the distorted waveform is obtained by calculating the tangent slope of the voltage waveform at the drop initiation point, and the trailing edge slope is obtained by calculating the tangent slope of the voltage waveform at the recovery point. The recovery point is the first sampling point where the voltage recovers from the trough to the normal threshold range. The specific formula for calculating the tangent slope is... ,in Represents the slope of the waveform tangent. This represents the instantaneous voltage difference between two adjacent sampling points. The time interval between two adjacent sampling points is represented by the leading edge slope. A larger leading edge slope indicates a faster voltage drop, while a larger trailing edge slope indicates a faster voltage recovery. These numerical characteristics can effectively distinguish between loose connection faults and other types of voltage anomalies. The multi-dimensional feature extraction unit 2 calculates the time interval between two adjacent distortion events. This parameter is used to analyze the frequency and periodicity of distortion events. Specifically, it is calculated by subtracting the event timestamp of the previous distortion event from the event timestamp of the current distortion event. The time difference obtained is the time interval between two adjacent distortion events. If this parameter exhibits a periodicity, it strongly points to a loose connection fault in the secondary circuit. In addition, the multi-dimensional feature extraction unit 2 also calculates two auxiliary feature parameters: harmonic distortion rate and three-phase imbalance, further enriching the feature dimensions of distortion events. The harmonic distortion rate is calculated by analyzing the spectral components of the voltage waveform during the distortion duration. Specifically, it is implemented by using a fast Fourier transform algorithm to convert the distorted voltage waveform from the time domain to the frequency domain, decomposing it to obtain the effective value of the fundamental voltage and the effective values ​​of each harmonic voltage, and then using the formula... Calculation, where Represents harmonic distortion rate. Represents the effective value of the fundamental voltage. to This parameter represents the effective value of the second and higher harmonic voltages, reflecting the degree of voltage waveform distortion. Distortion caused by a loose connection fault is accompanied by an increase in the harmonic distortion rate. The three-phase unbalance is calculated by comparing the voltage drop depths of different phases at the same time; the specific calculation formula is as follows. ,in Represents the three-phase imbalance. This represents the maximum voltage drop depth in the three-phase voltage range. This represents the minimum voltage drop depth in the three-phase voltage range. The average value of the three-phase voltage drop depth is represented by this parameter, which reflects the degree of imbalance in the three-phase voltage distortion. A single point of loose connection in the secondary circuit usually leads to a significant increase in the three-phase imbalance. At this point, the extraction of all multi-dimensional characteristic parameters of a single distortion event is complete.

[0032] After extracting the multidimensional feature parameters of a single distortion event, the multidimensional feature extraction unit 2 continues to perform cross-cycle association marking operations for distortion events with the same phase angle. This step can associate and integrate distortion events with different power frequency cycles but with highly consistent phases, and uncover the unique periodic distortion features of the loose connection fault, greatly improving the accuracy of subsequent fault determination. The multidimensional feature extraction unit 2 first presets a set of phase intervals indexed by the power frequency phase angle. The complete range of the power frequency phase angle is from 0° to 360°. The system evenly divides this range into 12 continuous and non-overlapping phase intervals, each covering a phase angle of 30°, namely 0°-30°, 30°-60°...330°-360°. Each phase interval corresponds to a unique numerical index identifier, numbered sequentially from 1 to 12, facilitating the subsequent classification and association of distortion events. When the multi-dimensional feature extraction unit 2 extracts the power frequency phase angle of the current distortion event, it assigns this phase angle to the phase interval with the closest value and binds the corresponding index identifier to this phase interval. For example, if the power frequency phase angle of the distortion event is 45°, it is assigned to the 30°-60° phase interval and bound to index identifier 2. Next, the multi-dimensional feature extraction unit 2 traverses all historical distortion event records stored locally in the system, filters out all distortion events with the same phase interval identifier, and uniformly associates these distortion events, even though they occur in different power frequency cycles. While the events occur at different times, their phases are highly consistent, suggesting a persistent secondary circuit connection fault. The multi-dimensional feature extraction unit 2 then adds a common cross-period association marker to these interconnected distortion events. This cross-period association marker is a unique identifier for groups of in-phase distortion events, comprising two core components: an index of the associated phase interval to clarify the unified occurrence phase of the group, and a cumulative occurrence value for the group, used to count the cumulative frequency of in-phase distortion events. This cumulative value automatically increments by 1 for each new in-phase distortion event. Through this phase classification, historical traversal, event association, and marker addition operation, multi-dimensional features... Extraction unit 2 can fully realize cross-cycle correlation marking of distortion events under the same phase angle, providing key periodic feature support for subsequent construction of distortion event sequences and fault trend analysis. Subsequently, after completing the multi-dimensional feature extraction and cross-cycle correlation marking of all distortion events, multi-dimensional feature extraction unit 2 begins to construct the distortion event sequence of the secondary circuit nodes of the distribution cabinet, providing standardized input data for subsequent state evolution trend analysis. Multi-dimensional feature extraction unit 2 treats each distortion event that has completed detection, basic information recording, multi-dimensional feature extraction, and cross-cycle correlation marking as an independent data entry. Each data entry contains complete core information, including at least the event timestamp, power frequency phase angle, and cross-cycle correlation marking.The system also includes a multi-dimensional feature vector composed of distortion duration, voltage drop depth, distortion waveform leading edge slope, distortion waveform trailing edge slope, time interval between two adjacent distortion events, harmonic distortion rate, and three-phase imbalance. This multi-dimensional feature vector is a one-dimensional vector composed of seven feature parameters combined in a fixed order, capable of fully capturing all the quantitative features of a single distortion event. The multi-dimensional feature extraction unit 2 arranges all generated data entries sequentially according to the event timestamps, forming an ordered sequence continuously sorted by time. This ordered sequence is the distortion event sequence of the secondary circuit nodes of the distribution cabinet, serving as the sole input data for the state evolution trend analysis unit 3 to perform AI time-series prediction, ensuring the continuity and accuracy of subsequent trend analysis. After the distortion event sequence is constructed and verified by the system, the state evolution trend analysis unit 3 first divides the data into fixed statistical periods, and then generates corresponding feature parameter time-series sub-sequences based on the distortion event sequence, preparing data for the subsequent training and prediction of the long short-term memory network model.

[0033] The State Evolution Trend Analysis Unit 3 first clarifies the specific definition of the fixed statistical period. The fixed statistical period is a standard time length pre-set by the system for segmented analysis of the evolution law of distortion events. Combining the evolution characteristics of the secondary circuit loose connection fault of high and low voltage distribution cabinets and the operation and maintenance monitoring requirements, the system sets the fixed statistical period to 1 hour. This period length can ensure that each period contains sufficient distortion event data, avoid data sparsity leading to prediction failure, and timely capture the early evolution trend of the fault to meet the requirements of real-time monitoring. Subsequently, the State Evolution Trend Analysis Unit 3 extracts all distortion event data entries contained in each fixed statistical period in chronological order from the constructed distortion event sequence. Then, the State Evolution Trend Analysis Unit 3 rearranges the multidimensional feature vectors of all data entries in each statistical period according to the time order of the distortion events, forming a feature parameter time sequence subsequence specific to that statistical period. This subsequence is a feature data set sorted by time, which can completely reflect the dynamic change law of distortion features within a single statistical period and is the standard input format adapted to the Long Short-Term Memory Network model.

[0034] Based on this, the State Evolution Trend Analysis Unit 3 initiates a time-series prediction model based on Long Short-Term Memory (LSTM) networks. It utilizes the generated time-series subsequences of feature parameters to train the model and predict future trends, ultimately generating the predicted trajectories of various feature parameters. The LSM network model is a deep learning model specifically designed for processing time-series data and effectively capturing long-term data dependencies. It perfectly adapts to the time-series evolution prediction needs of distorted event features, solving the gradient vanishing problem of traditional recurrent neural networks and accurately capturing the short-term fluctuations and long-term evolution patterns of distorted features. The State Evolution Trend Analysis Unit 3 first standardizes the time-series subsequences of feature parameters. Standardization maps feature data of different dimensions, numerical ranges, and magnitudes to a uniform range of 0 to 1, eliminating the impact of numerical differences between feature parameters on model training accuracy and ensuring the stability and convergence speed of model training.

[0035] After processing, the state evolution trend analysis unit 3 uses the standardized feature parameter time series subsequence as training data and inputs it into a pre-initialized Long Short-Term Memory (LSTM) network model. This LSM network model has three core layers: an input layer, at least one LSM layer, and a fully connected output layer. The input layer receives the standardized feature parameter time series subsequence data and smoothly transmits the data into the model. The LSM layer is the core computational layer of the model, containing forget gates, input gates, and output gates. It is specifically used to capture the long-term dependencies and short-term fluctuations of feature parameters over time and learn the historical change patterns of distortion features. The fully connected output layer is responsible for converting the time series features extracted by the LSM layer into the final prediction results and outputting the prediction data for multiple feature parameters in parallel. During the model training phase, the state evolution trend analysis unit continuously adjusts the weights and bias parameters within the model through iterative optimization. The core objective of iterative optimization is to minimize the difference between the predicted trajectory output by the model and the actual observed values ​​of the feature parameters in the next fixed statistical period. The system uses mean squared error as the loss function for model training, and the specific calculation formula is as follows: ,in Represents the model loss value. This represents the total number of sampling points within the statistical period. The predicted value of the feature parameters represented by the model output. The loss value represents the actual observed value of the feature parameter. The smaller the loss value, the higher the prediction accuracy of the model. When the loss value converges to the preset accuracy threshold, the model training is complete and it can be used for actual prediction. For a fixed statistical period to be predicted, the Long Short-Term Memory (LSTM) network model uses the time series subsequence of the feature parameter of the previous statistical period as input data. Through forward propagation, the model calculates and outputs multiple independent time series in parallel by the fully connected output layer. Each time series corresponds to the predicted value of a feature parameter at each sampling time in the future statistical period. Finally, the state evolution trend analysis unit 3 connects these discrete predicted values ​​in chronological order to form a complete prediction trajectory of each feature parameter in the future period. This prediction trajectory represents the future change trend of each feature parameter under normal operating conditions of the distribution cabinet, predicted by the model after learning historical data, and serves as the benchmark for subsequent deviation calculations.

[0036] After the state evolution trend analysis unit 3 generates the predicted trajectories of various characteristic parameters, the unit immediately performs a comparison calculation between the actual characteristic parameters and the predicted trajectories. Through layer-by-layer statistics and weighted summation, it finally obtains the comprehensive deviation degree characterizing the degree of abnormality in the operating state of the distribution cabinet. The state evolution trend analysis unit 3 first obtains all the distortion events that actually occur within the current fixed statistical period, and then extracts the actual characteristic parameters of these distortion events according to the same time sequence and sampling time as the predicted trajectory. These parameters are then organized into actual characteristic parameter trajectories that correspond one-to-one with the predicted trajectories, ensuring the time synchronization and data consistency of the comparison calculation. Subsequently, for each independent characteristic parameter, the state evolution trend analysis unit 3 calculates the numerical difference between the actual characteristic parameter trajectory and the corresponding predicted trajectory at each same sampling time. After taking the absolute value of the numerical difference, the instantaneous deviation degree of that characteristic parameter at that time is obtained. The instantaneous deviation degree reflects the real-time abnormal deviation degree of the characteristic parameter at a single sampling time. Its specific calculation formula is as follows: ,in Represents instantaneous deviation. Represents the actual value of the characteristic parameter. The absolute value operation, representing the predicted characteristic parameter value, eliminates the influence of positive and negative deviations, retaining only the deviation magnitude. Next, the state evolution trend analysis unit 3 statistically processes all instantaneous deviations of this characteristic parameter within a statistical period, calculating their arithmetic mean. This mean is used as the periodic deviation of the characteristic parameter in this period. The periodic deviation reflects the overall abnormal deviation degree of a single type of characteristic parameter throughout the entire statistical period, and its calculation formula is... ,in Represents the degree of period deviation. This represents the total number of instantaneous deviations within the statistical period. The summation of all instantaneous deviations is then used. Next, the state evolution trend analysis unit 3 assigns different weighting coefficients to the periodic deviations of each characteristic parameter. These weighting coefficients are set based on the sensitivity and importance of the characteristic parameter to the loose connection fault. Voltage drop depth and distortion duration are most strongly correlated with loose connection faults, with weighting coefficients set to 0.25 and 0.25 respectively. The slope of the distorted waveform's leading and trailing edges has weighting coefficients of 0.15 and 0.15. Adjacent time intervals, harmonic distortion rate, and three-phase imbalance have weighting coefficients of 0.08, 0.06, and 0.06 respectively. The sum of all weighting coefficients is strictly equal to 1 to ensure the rationality and fairness of the weighted calculation. Finally, the state evolution trend analysis unit 3 sums the weighted periodic deviations of all characteristic parameters to obtain the final result, which is the comprehensive deviation. The comprehensive deviation characterizes the overall difference between the current periodic operating state of the distribution cabinet and the normal state predicted by the model. Its calculation formula is... ,in Represents the overall deviation. The weighting coefficients represent the weights of each characteristic parameter. The periodic deviation of each characteristic parameter represents the degree of deviation. The larger the overall deviation value, the more serious the abnormal operation and the higher the probability of the occurrence of loose connection faults.

[0037] Finally, after the state evolution trend analysis unit 3 completes the comprehensive deviation calculation and transmits it to the judgment module, the loose connection degree judgment unit 4 executes the final fault judgment logic and early warning signal output operation. This is the final execution link of the entire AI-based intelligent monitoring system for high and low voltage distribution cabinet operation faults. It can accurately identify secondary circuit loose connection faults and output early warnings in a timely manner to ensure the safe and stable operation of the distribution cabinet. The loose connection degree judgment unit 4 first performs a threshold judgment on the comprehensive deviation to determine whether the comprehensive deviation continuously exceeds the preset deviation threshold. The preset deviation threshold is a critical value calibrated by the system based on a large amount of historical data of normal operation of distribution cabinets and combined with on-site operation and maintenance experience. The deviation threshold is set to 0.3. A sustained deviation exceeding this value indicates a significant operational anomaly, not a random fluctuation. Simultaneously, the loose connection determination unit 4 monitors the frequency of distortion events, statistically analyzing the actual number of distortion events occurring in the current statistical period and several consecutive previous statistical periods. It then determines whether this frequency exceeds a preset frequency threshold. This preset frequency threshold is a critical value for judging the frequency of loose connection faults. In this invention, the preset frequency threshold is set to 5 times per period. Exceeding this frequency for three consecutive periods indicates high-frequency distortion events, persistent faults, and no self-healing trend. When the overall deviation consistently exceeds the preset deviation threshold, and distortion events occur in the current and subsequent periods... When both the actual occurrence frequency in the preceding consecutive statistical periods exceeds the preset frequency threshold, the loose connection determination unit 4 immediately triggers the preset fault determination logic, formally determining that a loose connection fault exists in the secondary circuit node of the distribution cabinet, eliminating misjudgments caused by transient interference and other types of faults. Subsequently, the loose connection determination unit 4 generates fault diagnosis information, which includes the fault node identifier, the overall deviation exceeding the standard, and the high frequency of event occurrence. The fault node identifier is used to clearly identify the specific secondary circuit node number and location where the loose connection fault occurs, facilitating maintenance personnel to quickly locate the fault point. The overall deviation exceeding the standard records the current overall deviation. The system records numerical values, exceedance multiples, and duration cycles. It also records the number of consecutive cycles of high-frequency events and the number of distortion events in each cycle, providing maintenance personnel with complete fault information. Finally, the loose connection severity determination unit 4 uses this diagnostic information as a formal early warning signal, outputting it to the local alarm device via the system's RS485 communication interface. Upon receiving the early warning signal, the local alarm device immediately activates the audible and visual alarm mode, using a high-decibel buzzer and a bright warning light to alert on-site maintenance personnel to promptly tighten and repair the loose connection fault, preventing the fault from escalating and causing a safety accident in the distribution cabinet. This achieves real-time monitoring and accurate determination of secondary circuit loose connection faults.

[0038] The second objective of this invention is to provide a method for implementing the aforementioned AI-based intelligent monitoring system for high and low voltage switchgear operation faults, comprising the following steps:

[0039] S1. Perform sliding window analysis on the voltage waveform signal of the secondary circuit node of the distribution cabinet. When the instantaneous voltage drop within the window exceeds the preset threshold, it is recorded as a distortion event.

[0040] S2. Extract multi-dimensional feature parameters for each recorded distortion event. The multi-dimensional feature parameters include at least the distortion duration, voltage drop depth, power frequency phase angle at the time of distortion, slope of the leading and trailing edges of the distortion waveform, and time interval between two adjacent distortion events. Cross-cycle correlation marking is performed on distortion events with the same phase angle to construct the distortion event sequence of the secondary circuit node of the distribution cabinet.

[0041] S3. Using a fixed statistical period as the unit, a time series prediction model based on a long short-term memory network is used to predict the changing trends of various characteristic parameters in the distorted event sequence, generate the predicted trajectory of each characteristic parameter, and compare the actual characteristic parameters of the distorted event in the current statistical period with the predicted trajectory to calculate the comprehensive deviation.

[0042] S4. When the overall deviation exceeds the preset deviation threshold and the frequency of distortion events exceeds the preset frequency threshold, it is determined that there is a loose connection fault in the secondary circuit node of the distribution cabinet, and a corresponding warning signal is output.

[0043] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made without departing from the spirit and scope of the invention, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based intelligent monitoring system for operational faults in high and low voltage switchgear, characterized in that, include: The distortion event detection unit (1) is used to perform sliding window analysis on the voltage waveform signal of the secondary circuit node of the distribution cabinet. When the instantaneous voltage drop within the window exceeds the preset threshold, it is recorded as a distortion event. The multidimensional feature extraction unit (2) is used to extract multidimensional feature parameters for each distortion event. The multidimensional feature parameters include at least the distortion duration, voltage drop depth, power frequency phase angle at the time of distortion, slope of the leading and trailing edges of the distortion waveform, and time interval between two adjacent distortion events. The distortion events under the same phase angle are cross-cycle associated and marked to construct the distortion event sequence of the secondary circuit node of the distribution cabinet. Multidimensional feature parameters are extracted for each distortion event, specifically including: The distortion duration is determined by measuring the time elapsed from the onset of a voltage dip to the voltage recovering to the normal threshold range. The voltage dip depth is obtained by calculating the percentage difference between the instantaneous voltage value at the bottom of the voltage dip and the normal voltage reference value before the dip relative to that normal voltage reference value. The power frequency cycle is locked using phase-locked loop technology, and the phase angle corresponding to the distortion initiation point is determined as the power frequency phase angle at the moment of distortion. The tangent slope of the voltage waveform at the dip initiation point is calculated as the distorted waveform leading edge slope, and the tangent slope of the voltage waveform at the recovery point is calculated as the distorted waveform trailing edge slope. The time interval between two adjacent distortion events is obtained by calculating the difference between the event timestamp of the current distortion event and the event timestamp of the previous distortion event. The harmonic distortion rate is calculated by analyzing the spectral components of the voltage waveform during the distortion duration. The three-phase imbalance is calculated by comparing the voltage dip depths of different phases at the same time. Cross-cycle correlation labeling for distortion events at in-phase angles includes: A set of phase intervals indexed by the power frequency phase angle is pre-defined. After the power frequency phase angle of the distortion event is extracted, the power frequency phase angle is assigned to the closest phase interval, and an identifier is assigned to the phase interval. Historical distortion events are traversed, and distortion events with the same phase interval identifier are associated. A common cross-period association mark is attached to distortion events with the same phase interval identifier. The cross-period association mark contains the identifier of the associated phase interval and the cumulative value of the occurrence of the associated events, so as to realize cross-period association marking of distortion events under the same phase angle. Each distortion event that is recorded and whose multidimensional feature parameters are extracted is treated as a data entry. Each data entry contains at least the event timestamp, power frequency phase angle, cross-cycle correlation marker, and a multidimensional feature vector consisting of distortion duration, voltage drop depth, slope of the leading and trailing edges of the distortion waveform, time interval, harmonic distortion rate, and three-phase imbalance. Arrange all data entries in chronological order according to the event timestamps to form a time-sorted sequence, which is the aberration event sequence. The state evolution trend analysis unit (3) is used to predict the changing trends of various characteristic parameters in the distorted event sequence using a time series prediction model based on a long short-term memory network, with a fixed statistical period as the unit, and to generate the prediction trajectory of each characteristic parameter, specifically including: The fixed statistical period is a pre-set time length. Data entries of all distorted events within each fixed statistical period are extracted from the distorted event sequence in chronological order. The multidimensional feature vectors of the data entries within each statistical period are arranged in chronological order to form the time-series subsequence of the feature parameters of the statistical period. The time-series subsequences of the feature parameters are used as training data and input into a pre-trained long short-term memory network model. The long short-term memory network model learns the change patterns of the feature parameters in historical periods and outputs a prediction of the change trends of each feature parameter in the next fixed statistical period. This prediction result is represented as a set of trajectories that progress over time and reflect the expected values ​​of each feature parameter, i.e., the predicted trajectory of each feature parameter. The Long Short-Term Memory (LSTM) network model receives the standardized temporal subsequence of the feature parameters. The LSM network model includes an input layer, at least one LSM layer, and a fully connected output layer. The LSM layer is used to capture the long-term and short-term dependencies of the feature parameters in the time dimension. During the training phase, the LSM network model is iteratively optimized to minimize the difference between the predicted trajectory of the output and the actual observed values ​​of the feature parameters in the next fixed statistical period. For a fixed statistical period to be predicted, the Long Short-Term Memory Network model takes the time series of feature parameters from the previous period as input, and outputs multiple time series in parallel by a fully connected output layer. Each time series corresponds to the predicted value of a feature parameter at each sampling time in the future statistical period. The predicted values ​​are connected to form a complete prediction trajectory of the feature parameter in the future period. The actual feature parameters of the distortion event in the current statistical period are compared with the predicted trajectory to calculate the comprehensive deviation. The loose connection determination unit (4) is used to determine that there is a loose connection fault in the secondary circuit node of the power distribution cabinet and output a warning signal when the overall deviation exceeds the preset deviation threshold and the frequency of distortion events exceeds the preset frequency threshold.

2. The AI-based intelligent monitoring system for high and low voltage switchgear operation faults according to claim 1, characterized in that: The distortion event detection unit (1) performs sliding window analysis on the voltage waveform signal of the secondary circuit node of the distribution cabinet, specifically including: A sampling window with a fixed time width is advanced on a continuous voltage waveform signal with a fixed step size. Within the sampling window, the instantaneous voltage value is acquired at a preset sampling frequency. The voltage instantaneous value is identified as decreasing within a preset time. When the difference in the decrease exceeds a preset drop threshold, it is recorded as a distortion event. The event timestamp, the instantaneous voltage value at the start of the drop, and the instantaneous voltage value at the bottom of the drop are also recorded.

3. The AI-based intelligent monitoring system for high and low voltage switchgear operation faults according to claim 1, characterized in that: The state evolution trend analysis unit (3) compares the actual characteristic parameters of the distortion events in the current statistical period with the predicted trajectory and calculates the comprehensive deviation, specifically including: Obtain all distortion events that actually occur within the current fixed statistical period, and extract the actual feature parameters according to the preset time sequence and sampling time to form the actual feature parameter trajectory; For each feature parameter, calculate the numerical difference between the actual feature parameter trajectory and the corresponding predicted trajectory at each same sampling time, and take the absolute value to obtain the instantaneous deviation of that feature parameter at that time. Statistical processing is performed on all instantaneous deviations of the feature parameter within a statistical period, and the average value is calculated as the periodic deviation of the feature parameter in this period. Different weights are assigned to the periodic deviations of each feature parameter, and the weighted periodic deviations of all feature parameters are summed. The result is the comprehensive deviation that characterizes the difference between the overall operating state and the predicted state in the current period.

4. The AI-based intelligent monitoring system for high and low voltage switchgear operation faults according to claim 3, characterized in that: When the overall deviation continuously exceeds the preset deviation threshold, and it is detected that the actual frequency of the distortion event also exceeds the preset frequency threshold in the current statistical period and several consecutive statistical periods before, the judgment logic is triggered to determine that there is a loose connection fault in the secondary circuit node of the power distribution cabinet. Then, diagnostic information including the fault node identifier, the overall deviation exceeding the standard, and the high frequency of the event is generated, and the diagnostic information is output to the local alarm device as a warning signal through the communication interface.

5. A method for implementing the AI-based intelligent monitoring system for high and low voltage switchgear operation faults as described in any one of claims 1-4, characterized in that: Includes the following steps: S1. Perform sliding window analysis on the voltage waveform signal of the secondary circuit node of the distribution cabinet. When the instantaneous voltage drop within the window exceeds the preset threshold, it is recorded as a distortion event. S2. Extract multi-dimensional feature parameters for each recorded distortion event. The multi-dimensional feature parameters include at least the distortion duration, voltage drop depth, power frequency phase angle at the time of distortion, slope of the leading and trailing edges of the distortion waveform, and time interval between two adjacent distortion events. Cross-cycle correlation marking is performed on distortion events with the same phase angle to construct a distortion event sequence of the secondary circuit node of the distribution cabinet. S3. Using a fixed statistical period as a unit, a time-series prediction model based on a long short-term memory network is used to predict the changing trends of various characteristic parameters in the distorted event sequence, generate the predicted trajectory of each characteristic parameter, and compare the actual characteristic parameters of the distorted event in the current statistical period with the predicted trajectory to calculate the comprehensive deviation. S4. When the overall deviation exceeds the preset deviation threshold and the frequency of distortion events exceeds the preset frequency threshold, it is determined that there is a loose connection fault in the secondary circuit node of the power distribution cabinet, and a corresponding warning signal is output.

Images