A multi-feature cooperative judgment method for power equipment monitoring and early warning

Abstract

The application discloses a kind of multi-feature coordination research and judgment's electric power equipment monitoring early warning method, belong to electric power equipment intelligent detection early warning technical field;Including 1, obtain standardization time series dataset;2, carry out multidimensional feature extraction, construct four-dimensional time series feature matrix;3, build load-environment double-factor threshold compensation model, and calculate working condition adaptive dynamic alarm threshold;4, carry out rule engine preliminary determination, determine whether to issue preliminary alarm start signal;5, after receiving preliminary alarm start signal, carry out depth time series research and judgment, obtain fault trend and fault probability result;6, receive the prediction result of 5, and the prediction result is bidirectional feedback closed-loop correction;7, read the prediction result of 5 and carry out graded early warning and disposal;The application uses the above method, by real-time inner loop fine tuning+long-term outer loop iteration double-layer nested structure, greatly reduce early warning false alarm, realize latent fault early prediction and system autonomous optimization.

Description

Technical Field

[0001] This invention belongs to the field of intelligent detection and early warning technology for power equipment, and in particular relates to a method for monitoring and early warning of power equipment based on multi-feature collaborative analysis. Background Technology

[0002] Power equipment such as power transformers, high-voltage switchgear, circuit breakers, power cables, and instrument transformers are the core carriers of power transmission and distribution systems. The operating status of these equipment directly determines the reliability of the power grid. Online monitoring and early warning are important means to promptly detect latent equipment faults and avoid unplanned power outages.

[0003] Current power equipment monitoring and early warning technologies have many shortcomings: 1. Limited monitoring dimensions: Most monitoring systems only collect a single electrical parameter and lack heterogeneous information from multiple physical fields such as vibration, temperature, and partial discharge, which cannot comprehensively reflect the health status of the equipment. 2. Static alarm threshold defects: The alarm threshold remains fixed for a long time, without taking into account the impact of equipment load changes, ambient temperature and humidity, seasonal factors, and power grid voltage fluctuations on operating conditions. Under high load and high temperature conditions, it is easy to generate a large number of false alarms, and it is easy to miss alarms when there are latent faults under low load. 3. The feature extraction level is simple, and most of them only use the raw monitoring values ​​to directly issue alarms. They lack time-domain and frequency-domain deep feature mining and cannot capture the characteristics of gradual equipment degradation and instantaneous sudden failure. 4. Insufficient predictive and judgment capabilities: Some existing technologies can only realize over-limit alarms and can only detect anomalies that have already occurred, lacking the ability to predict the long-term time series of equipment failure risks. 5. The TCN-LSTM algorithm is mostly used independently and is not linked in a two-way closed loop with the working condition adaptive threshold system. The algorithm output cannot optimize the alarm threshold and there is no real-time optimization mechanism. 6. The early warning-operation and maintenance-optimization link does not form a closed loop and lacks a double-layer nested optimization structure. The early warning results and on-site operation and maintenance information are independent of each other. The model cannot achieve self-iterative optimization based on historical operation and maintenance data, and the early warning accuracy is difficult to continuously improve over time. Summary of the Invention

[0004] The purpose of this invention is to provide a power equipment monitoring and early warning method based on multi-feature collaborative judgment. It integrates four-layer hierarchical multi-dimensional feature extraction, load-environment dual-factor threshold compensation, rule engine, TCN-LSTM time series prediction and judgment, real-time inner loop feedback, and long-term outer loop operation and maintenance iteration. Through a double-layer nested structure of real-time inner loop fine-tuning and long-term outer loop iteration, it significantly reduces false alarms and missed alarms, and realizes advanced prediction of latent faults and autonomous system optimization.

[0005] To achieve the above objectives, this invention provides a power equipment monitoring and early warning method based on multi-feature collaborative analysis, comprising the following steps: S1. Collect multi-source heterogeneous data from power equipment and preprocess it to obtain a standardized time-series dataset; S2. Perform multi-dimensional feature extraction on the standardized time series dataset and construct a four-dimensional time series feature matrix; S3. Construct a load-environment dual-factor threshold compensation model and calculate the adaptive dynamic alarm threshold for operating conditions; S4. Based on the adaptive dynamic alarm threshold, the rule engine makes a preliminary judgment to determine whether to issue a preliminary alarm start signal; S5. After receiving the primary alarm activation signal, perform in-depth time series analysis through the TCN-LSTM hybrid model to obtain fault trend and fault probability results. S6. Receive the prediction results from S5 and perform bidirectional feedback closed-loop correction on the prediction results; S7. Read the prediction results from S5 to conduct tiered early warning and response.

[0006] Preferably, the process of S1 is as follows: S11. Deploy edge acquisition terminals around the power equipment to be monitored; S12. Collect electrical quantities, non-electrical quantities, environmental parameters, and equipment operating condition labels of the power equipment to be monitored through the edge acquisition terminal; S13. Add a unified BeiDou / NTP timestamp to the data in S12 to align the data from the sensor and the gateway to the same timeline. S14. Perform data cleaning to remove dirty data from the data processed in S13, and sequentially perform missing value completion, gross outlier removal, sensor drift correction and impulse interference filtering. S15. Normalize the data after removing dirty data to achieve dimensional uniformity. S16. After processing, each monitoring sample in the data is labeled with its operating condition, including high load, low load, start-stop transition and steady state, and finally outputs a standardized time series dataset with operating condition labels.

[0007] Preferably, the process of S2 is as follows: S21. Set the sliding window duration and sliding step size, and use the sliding window to segment and truncate the standardized time series dataset to obtain several continuous time series subsequences; use each time series subsequence as a calculation sample for a single feature extraction. S22. Read the original sampled values ​​within each monitoring parameter window from the time series subsequence to obtain the original physical characteristics. ; S23. Based on the original time series values ​​within the sliding window, calculate the time-domain statistical indices of the signal within the window to obtain the time-domain statistical characteristics. ; S24. Perform a Fast Fourier Transform on the time-series signal within the sliding window to convert the time-domain signal to the frequency domain and extract the frequency domain features. ; S25. Combining the load rate, season identifier, and ambient temperature of the power equipment to be monitored; and using the pre-queried operating condition-feature correction weight table, obtain the weighted correction coefficients for the corresponding operating conditions; use the weighted correction coefficients to eliminate feature offset interference caused by high load and high temperature environment, and obtain the operating condition fusion feature quantity after removing operating condition interference. ; S26. Combine all the features obtained from S22 to S25 to obtain a four-dimensional time series feature matrix. .

[0008] Preferably, the process of S3 is as follows: S31. Real-time acquisition of operating conditions and environmental parameters of the power equipment to be monitored, including: real-time load rate of the equipment. ambient temperature Ambient humidity Seasonal operating condition markings and grid voltage fluctuation coefficient Simultaneously, retrieve the fixed baseline alarm threshold from the equipment's factory technical specifications. ; S32. Retrieve the normal steady-state operation dataset of the power equipment to be monitored in recent years, and remove historical fault data, maintenance data and abnormal disturbance data to construct a clean model training sample set; S33. By using multiple linear regression fitting, the model is trained on the training sample set to obtain the six-dimensional dynamic compensation coefficients. , , , , and ; S34. Calculate the dynamic alarm threshold based on the relevant data from S31 and S33. The process is as follows: ; in , , , , and These represent the reference threshold coefficient, load compensation coefficient, temperature compensation coefficient, humidity compensation coefficient, seasonal compensation coefficient, and voltage fluctuation compensation coefficient, respectively. S35. Set the minimum allowable threshold for alarms. and maximum boundary And perform threshold validity verification, when the condition is met S4 is performed at that time.

[0009] Preferably, the process of S4 is as follows: S41. Read the key monitoring features in the four-dimensional time series feature matrix S2, and perform short-time mean filtering on the features to obtain real-time feature values. S42. Compare the real-time feature value with the dynamic alarm threshold and perform a two-level branch decision. S43. Record each judgment result, including dynamic threshold value, real-time feature value, working condition label, judgment result, trigger status and timestamp; S44. Set up a confirmation mechanism for consecutive N frames to verify the stability of the status signal. Only when the feature value exceeds the limit for 3 to 5 consecutive monitoring cycles will the TCN-LSTM hybrid model of S5 be triggered stably. Otherwise, the in-depth analysis will not be triggered to further filter out instantaneous noise interference.

[0010] Preferably, the two-level branch determination process of S42 is as follows: When real-time feature value When the process enters the suspected abnormal branch, it determines that the power equipment under monitoring is currently in a suspected abnormal operating condition. It immediately generates a primary alarm trigger signal, wakes up the TCN-LSTM deep prediction model that is in a dormant state, and marks the current time, over-limit amplitude and operating condition information. The process then enters the S5 deep time series analysis stage. When real-time feature value When the time comes, the system enters the normal operation branch, determines that the power equipment under monitoring is operating normally, does not trigger a primary alarm, maintains the TCN-LSTM model in a dormant state, and returns to S1 to continue the next round of data acquisition and monitoring cycle.

[0011] Preferably, the process of S5 is as follows: S51. Receive the primary alarm trigger signal from S4, switch the awakened TCN-LSTM deep prediction model to the working state, and input the multi-dimensional time series data in the four-dimensional time series feature matrix into the TCN-LSTM deep prediction model. S52. Perform sequence normalization on the input temporal feature matrix. The process involves interpolating and filling in missing points, normalizing and aligning the feature dimensions, and unifying the length of the model input sequence and the order of feature arrangement. S53. Utilize the dilated convolution and causal convolution characteristics of the TCN temporal convolutional network to perform multi-layer convolution operations on long temporal sequences to obtain advanced temporal features. S54. Input the advanced time series features into the LSTM memory unit to obtain long-term trend features; S55. Weighted fusion of advanced temporal features and long-term trend features to construct a spatiotemporal joint advanced feature representation; S56. Combining spatiotemporal joint high-level feature representation, the TCN-LSTM deep prediction model is used to predict the state of future multiple time periods and quantify the probability of failure. Two types of judgment results are output through the fully connected layer: multi-scale equipment state time-series prediction curve and failure probability value. S57. Based on the multi-scale equipment state time-series prediction curve, the fault trend sequence is obtained. The fault trend sequence and the fault occurrence probability value are synchronously divided into two parallel outputs, which are then processed in S6 and S7 respectively.

[0012] Preferably, the process of S6 is as follows: S61. Perform slope analysis, trend fitting, and risk classification on the fault trend sequence, and classify the power equipment to be monitored into three states: continuous risk deterioration, instantaneous disturbance fluctuation, and stable state. S62. Based on the status of the power equipment to be monitored, perform corresponding dynamic threshold intelligent fine-tuning, as follows: If the risk continues to worsen and the equipment is in a deterioration phase, the adaptive threshold for the next cycle should be slightly lowered, and the system alarm sensitivity should be increased. If the disturbance is transient, it is determined that there is a brief disturbance, the threshold for the next cycle is slightly increased, and the transient over-limit interference is blocked. If the state is stable, the device is in a stable state and the threshold remains the original calculated value. S63. Set a fine-tuning amplitude constraint mechanism to ensure that the threshold fine-tuning amount in a single cycle does not exceed the benchmark threshold. ; S64. The fine-tuned threshold parameters are fed back to the load-environment dual-factor threshold compensation model to cover the initial calculated threshold for the next monitoring cycle; at the same time, the updated threshold is subjected to first-order smoothing filtering. S65. Record the parameters for each round of fine-tuning, including the original threshold, fine-tuning direction, fine-tuning magnitude, predicted failure probability, trend type, and timestamp.

[0013] Preferably, the criteria for judging the three states in S61 are as follows: Risk continues to worsen: the future forecast curve continues to rise, and the probability of failure increases cycle by cycle; Instantaneous disturbance fluctuation: Currently slightly exceeded, but the future trend is rapidly declining and the probability is decreasing; Stable and stable status: The prediction curve is flat, the probability is stable and without fluctuation, and the equipment operating status does not change significantly.

[0014] Preferably, the process of S7 is as follows: S71. Preset a three-segment probability threshold range, and standardize the classification of equipment operation risk based on the probability value of failure occurrence, as follows: Normal level, the probability of failure is within The equipment is running smoothly, showing no signs of deterioration, and no alarms or work orders are generated. Pay attention to the warning; the probability of a fault occurring is at [a certain level]. If the equipment has potential abnormalities, slight misalignment, or slow deterioration, a daily inspection work order will be generated. Emergency warning, the probability of failure is at The equipment exhibits a clear trend of failure evolution and poses a short-term risk of failure, thus generating an emergency repair work order. S72. Automatically generate standardized operation and maintenance work orders based on the warning level. The work order content includes: equipment number, current operating status, out-of-limit characteristic quantity, dynamic threshold, failure probability, future trend curve, risk level, suggested handling measures, and occurrence timestamp. S73. Implement tiered and differentiated push notifications through the operation and maintenance platform and mobile terminals, pushing ordinary warnings to inspection personnel, and pushing attention warnings and emergency warnings to operation and maintenance managers and emergency repair teams. S74. The relevant staff shall take appropriate measures, and after the measures are completed, the actual operation and maintenance results shall be backfilled through either manual or system methods. S75. After cleaning the actual operation and maintenance results, the data is fed back to the system's historical training dataset to complete the early warning process.

[0015] Therefore, the power equipment monitoring and early warning method based on multi-feature collaborative analysis described above has the following advantages: 1. Multi-source heterogeneous frequency conversion adaptive acquisition was implemented, covering multiple dimensions of parameters such as electrical, vibration, temperature and environment, overcoming the shortcomings of traditional monitoring with single dimensions and incomplete data, and fully restoring the true operating status of the equipment.

[0016] 2. A multi-factor dynamic threshold compensation system based on load, environment, season, and voltage has been implemented, which breaks away from the rigid judgment mode of traditional fixed thresholds, adapts to all operating conditions, and significantly reduces false alarms due to high temperature and high load, as well as missed faults due to low load.

[0017] 3. Four-layer hierarchical multi-dimensional feature extraction was carried out to deeply mine equipment status features from multiple dimensions such as physical, time domain, frequency domain, and operating conditions, effectively distinguishing between normal fluctuations and fault deterioration, and the judgment accuracy is far higher than that of traditional single feature monitoring schemes.

[0018] 4. The TCN-LSTM model is triggered on demand using a rule engine, avoiding the high computing power consumption of continuous long-term model operation and adapting to edge terminal deployment scenarios; at the same time, a two-way feedback link between the prediction model and the threshold system is built, breaking the technical limitations of traditional models that can only predict in one direction and have no optimization linkage.

[0019] 5. A double-layer nested closed-loop optimization system is set up. Real-time inner loop fine-tuning achieves accurate adaptation to instantaneous operating conditions, and long-term outer loop iteration achieves continuous model upgrades, forming a full-dimensional autonomous optimization mechanism of "real-time optimization + long-term iteration".

[0020] 6. A complete closed-loop link of monitoring-analysis-early warning-operation and maintenance-optimization has been constructed. The system's early warning accuracy can be continuously improved with the accumulation of operational data and operation and maintenance experience, realizing advanced and accurate early warning of power equipment failures, and effectively ensuring the reliability of power grid operation.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of a power equipment monitoring and early warning method based on multi-feature collaborative analysis according to the present invention; Figure 2 This is a flowchart illustrating the initial judgment of the rule engine in a power equipment monitoring and early warning method based on multi-feature collaborative analysis according to the present invention. Figure 3 This is a flowchart illustrating the deep time-series analysis in a power equipment monitoring and early warning method based on multi-feature collaborative analysis according to the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Specific model specifications need to be selected and determined according to the actual specifications of the device, etc. The specific selection calculation method adopts existing technology in the art, and therefore will not be described in detail.

[0024] Example 1; like Figures 1-3 As shown, this invention provides a power equipment monitoring and early warning method based on multi-feature collaborative analysis, comprising the following steps: S1. Collect multi-source heterogeneous data from power equipment and preprocess it to obtain a standardized time-series dataset; S11. Deploy edge acquisition terminals around the power equipment to be monitored; S12. Collect electrical quantities, non-electrical quantities, environmental parameters, and equipment operating condition labels of the power equipment to be monitored through the edge acquisition terminal; S13. Add a unified BeiDou / NTP timestamp to the data in S12 to align the data from the sensor and the gateway to the same timeline. S14. Perform data cleaning to remove dirty data from the data processed in S13, and sequentially perform missing value completion, gross outlier removal, sensor drift correction and impulse interference filtering. S15. Normalize the data after removing dirty data to achieve dimensional uniformity. S16. After processing, each monitoring sample in the data is labeled with its operating condition, including high load, low load, start-stop transition and steady state, and finally outputs a standardized time series dataset with operating condition labels.

[0025] S2. Perform multi-dimensional feature extraction on the standardized time series dataset and construct a four-dimensional time series feature matrix; S21. Set the sliding window duration and sliding step size. The duration is 30-120 min and the sliding step size is 5-15 min. Use the sliding window to segment and truncate the standardized time series dataset to obtain several continuous time series subsequences. Use each time series subsequence as a calculation sample for a single feature extraction. S22. Read the original sampled values ​​within each monitoring parameter window from the time series subsequence to obtain the original physical characteristics. The original physical characteristics include: transformer top oil temperature, winding temperature, ambient temperature, vibration acceleration amplitude, partial discharge amplitude, three-phase voltage RMS value, three-phase current RMS value, active power, reactive power, and load rate; the original physical characteristics retain the most direct operating information of the equipment; S23. Based on the original time series values ​​within the sliding window, calculate the time-domain statistical indices of the signal within the window to obtain the time-domain statistical characteristics. The time-domain statistical features include: mean, maximum, minimum, variance, standard deviation, kurtosis, skewness, peak factor, impulse factor, waveform factor, temperature / vibration change rate, and duration of continuous over-threshold. These time-domain features are used to characterize the fluctuation level, distribution pattern, and rate of change of parameters over a period of time, and are used to capture fault signals such as slow equipment degradation and trend drift. S24. Perform a Fast Fourier Transform on the time-series signal within the sliding window to convert the time-domain signal to the frequency domain and extract the frequency domain features. The frequency domain features include: total signal spectral energy, main frequency, main frequency amplitude, harmonic amplitude, harmonic distortion rate, and energy proportion of specific fault characteristic frequency bands; the frequency domain features mainly identify fault signals with periodic abrupt changes such as vibration shock and partial discharge, and capture instantaneous abnormal disturbances; S25. Combining the load rate, season identifier, and ambient temperature of the power equipment to be monitored; and using the pre-queried operating condition-feature correction weight table, obtain the weighted correction coefficients for the corresponding operating conditions; use the weighted correction coefficients to eliminate feature offset interference caused by high load and high temperature environment, and obtain the operating condition fusion feature quantity after removing operating condition interference. ; S26. Combine all the features obtained from S22 to S25 to obtain a four-dimensional time series feature matrix. .

[0026] S3. Construct a load-environment dual-factor threshold compensation model and calculate the adaptive dynamic alarm threshold for operating conditions; S31. Real-time acquisition of operating conditions and environmental parameters of the power equipment to be monitored, including: real-time load rate of the equipment. ambient temperature Ambient humidity Seasonal operating condition markings and grid voltage fluctuation coefficient Simultaneously, retrieve the fixed baseline alarm threshold from the equipment's factory technical specifications. ; S32. Retrieve the normal steady-state operation dataset of the power equipment to be monitored in recent years, and remove historical fault data, maintenance data and abnormal disturbance data to construct a clean model training sample set; S33. By using multiple linear regression fitting, the model is trained on the training sample set to obtain the six-dimensional dynamic compensation coefficients. , , , , and ; S34. Calculate the dynamic alarm threshold based on the relevant data from S31 and S33. The process is as follows: ; in , , , , and These represent the reference threshold coefficient, load compensation coefficient, temperature compensation coefficient, humidity compensation coefficient, seasonal compensation coefficient, and voltage fluctuation compensation coefficient, respectively. S35. Set the minimum allowable threshold for alarms. and maximum boundary And perform threshold validity verification, when the condition is met S4 is performed at that time.

[0027] S4. Based on the adaptive dynamic alarm threshold, the rule engine makes a preliminary judgment to determine whether to issue a preliminary alarm start signal; S41. Read the key monitoring features in the four-dimensional time series feature matrix S2, and perform short-time mean filtering on the features to obtain real-time feature values. S42. Compare the real-time feature value with the dynamic alarm threshold and perform a two-level branch decision. The two-level branch determination process is as follows: When real-time feature value When the process enters the suspected abnormal branch, it determines that the power equipment under monitoring is currently in a suspected abnormal operating condition. It immediately generates a primary alarm trigger signal, wakes up the TCN-LSTM deep prediction model that is in a dormant state, and marks the current time, over-limit amplitude and operating condition information. The process then enters the S5 deep time series analysis stage. When real-time feature value When the time comes, the system enters the normal operation branch, determines that the power equipment under monitoring is operating normally, does not trigger a primary alarm, maintains the TCN-LSTM model in a dormant state, and returns to S1 to continue the next round of data acquisition and monitoring cycle.

[0028] S43. Record each judgment result, including dynamic threshold value, real-time feature value, working condition label, judgment result, trigger status and timestamp; S44. Set up a confirmation mechanism for consecutive N frames to verify the stability of the status signal. Only when the feature value exceeds the limit for 3 to 5 consecutive monitoring cycles will the TCN-LSTM hybrid model of S5 be triggered stably. Otherwise, the in-depth analysis will not be triggered to further filter out instantaneous noise interference.

[0029] S5. After receiving the primary alarm activation signal, perform in-depth time series analysis through the TCN-LSTM hybrid model to obtain fault trend and fault probability results. S51. Receive the primary alarm trigger signal from S4, switch the awakened TCN-LSTM deep prediction model to the working state, and input the multi-dimensional time series data in the four-dimensional time series feature matrix into the TCN-LSTM deep prediction model. S52. Perform sequence normalization on the input temporal feature matrix. The process involves interpolating and filling in missing points, normalizing and aligning the feature dimensions, and unifying the length of the model input sequence and the order of feature arrangement. S53. Utilizing the dilated and causal convolution characteristics of the TCN temporal convolutional network, multi-layer convolution operations are performed on long temporal sequences to obtain advanced temporal features. The process is as follows: 1. Extract millisecond-level, short-cycle, abrupt fault characteristics; 2. Capture instantaneous abnormal disturbances such as vibration jumps, partial discharge pulses, sudden increases in current harmonics, and sudden temperature rises; 3. Retain the temporal correlation information without destroying the causal relationship of time; The TCN module is specifically designed to identify sudden, transient, and impulsive anomalies, addressing the shortcoming of LSTM's insensitivity to short-term mutations. S54. The high-level time series features are fed into the LSTM memory cells to obtain long-term trend features. The process is as follows: 1. Automatically filter valid historical degradation information through an input gate, forget gate, and output gate structure; 2. The excavating equipment slowly drifts, gradually deteriorates, and shows a trend of increasing wear over several days or tens of hours; 3. Capture latent fault trends such as hidden aging, slight insulation degradation, temperature rise accumulation, and load accumulation damage; The LSTM module is specifically responsible for long-term trend prediction, hidden fault tracing, and learning state evolution patterns.

[0030] S55. Weighted fusion of advanced time-series features and long-term trend features is performed to construct a spatiotemporal joint advanced feature representation. The fused features simultaneously include: instantaneous anomaly details + long-term degradation trend + operating condition decoupling information + frequency domain distortion information, thereby achieving comprehensive and accurate coding of the equipment's operating status. S56. Combining spatiotemporal joint high-level feature representation, the TCN-LSTM deep prediction model is used to predict the state of future multiple time periods and quantify the probability of failure. Two types of judgment results are output through the fully connected layer: multi-scale equipment state time-series prediction curve and failure probability value. Two types of assessment results: 1. Output multi-scale equipment status time series prediction curves for the next 1 hour, 6 hours, 12 hours, 24 hours, and 72 hours; 2. Output a quantized fault occurrence probability value ranging from 0% to 100%; S57. Based on the multi-scale equipment state time-series prediction curve, the fault trend sequence is obtained. The fault trend sequence and the fault occurrence probability value are synchronously divided into two parallel outputs, which are then processed in S6 and S7 respectively.

[0031] S6. Receive the prediction results from S5 and perform bidirectional feedback closed-loop correction on the prediction results; S61. Perform slope analysis, trend fitting, and risk classification on the fault trend sequence, and classify the power equipment to be monitored into three states: continuous risk deterioration, instantaneous disturbance fluctuation, and stable state. The criteria for judging the three states are as follows: Risk continues to worsen: the future forecast curve continues to rise, and the probability of failure increases cycle by cycle; Instantaneous disturbance fluctuation: Currently slightly exceeded, but the future trend is rapidly declining and the probability is decreasing; Stable and stable status: The prediction curve is flat, the probability is stable and without fluctuation, and the equipment operating status does not change significantly.

[0032] S62. Based on the status of the power equipment to be monitored, perform corresponding dynamic threshold intelligent fine-tuning, as follows: If the risk continues to worsen and the equipment is in a deterioration phase, the adaptive threshold for the next cycle should be slightly lowered, and the system alarm sensitivity should be increased. If the disturbance is transient, it is determined that there is a brief disturbance, the threshold for the next cycle is slightly increased, and the transient over-limit interference is blocked. If the state is stable, the device is in a stable state and the threshold remains the original calculated value. S63. Set a fine-tuning amplitude constraint mechanism to ensure that the threshold fine-tuning amount in a single cycle does not exceed the benchmark threshold. ; S64. The fine-tuned threshold parameters are fed back to the load-environment dual-factor threshold compensation model to cover the initial calculated threshold for the next monitoring cycle; at the same time, the updated threshold is subjected to first-order smoothing filtering. S65. Record the parameters for each round of fine-tuning, including the original threshold, fine-tuning direction, fine-tuning magnitude, predicted failure probability, trend type, and timestamp.

[0033] S7. Read the prediction results from S5 to conduct tiered early warning and response.

[0034] S71. Preset a three-segment probability threshold range, and standardize the classification of equipment operation risk based on the probability value of failure occurrence, as follows: Normal level, the probability of failure is within The equipment is running smoothly, showing no signs of deterioration, and no alarms or work orders are generated. Pay attention to the warning; the probability of a fault occurring is at [a certain level]. If the equipment has potential abnormalities, slight misalignment, or slow deterioration, a daily inspection work order will be generated. Emergency warning, the probability of failure is at The equipment exhibits a clear trend of failure evolution and poses a short-term risk of failure, thus generating an emergency repair work order. S72. Automatically generate standardized operation and maintenance work orders based on the warning level. The work order content includes: equipment number, current operating status, out-of-limit characteristic quantity, dynamic threshold, failure probability, future trend curve, risk level, suggested handling measures, and occurrence timestamp. S73. Implement tiered and differentiated push notifications through the operation and maintenance platform and mobile terminals, pushing ordinary warnings to inspection personnel, and pushing attention warnings and emergency warnings to operation and maintenance managers and emergency repair teams. S74. The relevant staff shall take appropriate measures, and after the measures are completed, the actual operation and maintenance results shall be backfilled through either manual or system methods. S75. After cleaning the actual operation and maintenance results, the data is fed back to the system's historical training dataset to complete the early warning process.

[0035] Example 2; The 10kV power transformer in the substation was used as the monitoring object.

[0036] S1. Deploy edge acquisition terminals to collect transformer three-phase voltage, three-phase current, top oil temperature, tank vibration, ambient temperature and humidity, and load rate; enable frequency conversion adaptive sampling in sampling mode: sampling period of 5 minutes during stable load phase; when load fluctuation exceeds 10%, the sampling period is shortened to 30 seconds; all sensor data are tagged with BeiDou timestamps for alignment; sensor jump defects are removed, Z-score is standardized, high load condition labels are added, and a standardized time series dataset is output.

[0037] S2. Based on the preprocessed dataset, four layers of feature extraction are completed. The original physical features such as oil temperature and vibration are extracted in sequence, the time-domain statistical features such as the mean oil temperature fluctuation and kurtosis are extracted, the frequency-domain features such as the vibration main frequency and spectral energy are extracted, and the load-temperature condition fusion correction features are extracted and spliced ​​to obtain a four-dimensional time-series feature matrix.

[0038] S3, retrieve the transformer factory oil temperature alarm reference threshold. Substituting into the multivariate linear fitting formula, the compensation coefficients are obtained by fitting the historical operating data of the equipment. Combined with real-time load, summer high temperature environment, and grid voltage fluctuation parameters, the dynamic alarm threshold is calculated. Under the summer high load condition, the dynamic threshold is adaptively adjusted to 88℃ to avoid false alarms under normal high load conditions.

[0039] S4. Real-time comparison of oil temperature characteristic value and dynamic alarm threshold. When the monitored oil temperature reaches the dynamic threshold, a primary alarm is triggered, waking up the TCN-LSTM prediction model; if the equipment status is normal, continuous cyclic monitoring is performed.

[0040] S5. Input the four-dimensional time series feature matrix into the TCN-LSTM hybrid model. The TCN module captures the instantaneous fluctuation and sudden change features of oil temperature, and the LSTM module analyzes the long-term temperature rise trend of oil temperature, predicts the change law of transformer oil temperature in the next 24 hours, and quantifies the probability of fault occurrence.

[0041] S6. The model predicts that the oil temperature will continue to rise and the probability of failure will increase. The alarm threshold for the next cycle will be adjusted downward in real time to improve the warning sensitivity and capture potential faults in advance. If the oil temperature is predicted to fluctuate briefly and then drop rapidly, the threshold will be slightly raised to avoid invalid alarms and complete real-time inner loop optimization.

[0042] S7. Based on the probability of failure, a high-risk emergency warning is issued, and a repair work order is pushed to the operation and maintenance terminal. The operation and maintenance personnel conduct on-site investigation and handle abnormal temperature rise issues, and return the inspection records, fault causes, and handling results to the dataset. The threshold compensation coefficient is refitted and updated monthly using newly added operation and maintenance data to complete long-term iterative optimization of the model and continuously improve the system's early warning accuracy.

[0043] Therefore, the present invention employs a power equipment monitoring and early warning method based on multi-feature collaborative analysis, which has the following advantages: 1. Multi-source heterogeneous frequency conversion adaptive acquisition was implemented, covering multiple dimensions of parameters such as electrical, vibration, temperature and environment, overcoming the shortcomings of traditional monitoring with single dimensions and incomplete data, and fully restoring the true operating status of the equipment.

[0044] 2. A multi-factor dynamic threshold compensation system based on load, environment, season, and voltage has been implemented, which breaks away from the rigid judgment mode of traditional fixed thresholds, adapts to all operating conditions, and significantly reduces false alarms due to high temperature and high load, as well as missed faults due to low load.

[0045] 3. Four-layer hierarchical multi-dimensional feature extraction was carried out to deeply mine equipment status features from multiple dimensions such as physical, time domain, frequency domain, and operating conditions, effectively distinguishing between normal fluctuations and fault deterioration, and the judgment accuracy is far higher than that of traditional single feature monitoring schemes.

[0046] 4. The TCN-LSTM model is triggered on demand using a rule engine, avoiding the high computing power consumption of continuous long-term model operation and adapting to edge terminal deployment scenarios; at the same time, a two-way feedback link between the prediction model and the threshold system is built, breaking the technical limitations of traditional models that can only predict in one direction and have no optimization linkage.

[0047] 5. A double-layer nested closed-loop optimization system is set up. Real-time inner loop fine-tuning achieves accurate adaptation to instantaneous operating conditions, and long-term outer loop iteration achieves continuous model upgrades, forming a full-dimensional autonomous optimization mechanism of "real-time optimization + long-term iteration".

[0048] 6. A complete closed-loop link of monitoring-analysis-early warning-operation and maintenance-optimization has been constructed. The system's early warning accuracy can be continuously improved with the accumulation of operational data and operation and maintenance experience, realizing advanced and accurate early warning of power equipment failures, and effectively ensuring the reliability of power grid operation.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for monitoring and early warning of power equipment based on multi-feature collaborative analysis, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous data from power equipment and preprocess it to obtain a standardized time-series dataset; S2. Perform multi-dimensional feature extraction on the standardized time series dataset and construct a four-dimensional time series feature matrix; S3. Construct a load-environment dual-factor threshold compensation model and calculate the adaptive dynamic alarm threshold for operating conditions; S4. Based on the adaptive dynamic alarm threshold, the rule engine makes a preliminary judgment to determine whether to issue a preliminary alarm start signal; S5. After receiving the primary alarm activation signal, perform in-depth time series analysis through the TCN-LSTM hybrid model to obtain fault trend and fault probability results. S6. Receive the prediction results from S5 and perform bidirectional feedback closed-loop correction on the prediction results; S7. Read the prediction results from S5 to conduct tiered early warning and response.

2. The power equipment monitoring and early warning method based on multi-feature collaborative analysis according to claim 1, characterized in that, The process of S1 is as follows: S11. Deploy edge acquisition terminals around the power equipment to be monitored; S12. Collect electrical quantities, non-electrical quantities, environmental parameters, and equipment operating condition labels of the power equipment to be monitored through the edge acquisition terminal; S13. Add a unified BeiDou / NTP timestamp to the data in S12 to align the data from the sensor and the gateway to the same timeline. S14. Perform data cleaning to remove dirty data from the data processed in S13, and sequentially perform missing value completion, gross outlier removal, sensor drift correction and impulse interference filtering. S15. Normalize the data after removing dirty data to achieve dimensional uniformity. S16. After processing, each monitoring sample in the data is labeled with its operating condition, including high load, low load, start-stop transition and steady state, and finally outputs a standardized time series dataset with operating condition labels.

3. The power equipment monitoring and early warning method based on multi-feature collaborative analysis according to claim 2, characterized in that, The process of S2 is as follows: S21. Set the sliding window duration and sliding step size, and use the sliding window to segment and truncate the standardized time series dataset to obtain several continuous time series subsequences; use each time series subsequence as a calculation sample for a single feature extraction. S22. Read the original sampled values ​​within each monitoring parameter window from the time series subsequence to obtain the original physical characteristics. ; S23. Based on the original time series values ​​within the sliding window, calculate the time-domain statistical indices of the signal within the window to obtain the time-domain statistical characteristics. ; S24. Perform a Fast Fourier Transform on the time-series signal within the sliding window to convert the time-domain signal to the frequency domain and extract the frequency domain features. ; S25. Combining the load rate, season identifier, and ambient temperature of the power equipment to be monitored; and using the pre-queried operating condition-feature correction weight table, obtain the weighted correction coefficients for the corresponding operating conditions; use the weighted correction coefficients to eliminate feature offset interference caused by high load and high temperature environment, and obtain the operating condition fusion feature quantity after removing operating condition interference. ; S26. Combine all the features obtained from S22 to S25 to obtain a four-dimensional time series feature matrix. .

4. The power equipment monitoring and early warning method based on multi-feature collaborative analysis according to claim 3, characterized in that, The process of S3 is as follows: S31. Real-time acquisition of operating conditions and environmental parameters of the power equipment to be monitored, including: real-time load rate of the equipment. ambient temperature Ambient humidity Seasonal operating condition markings and grid voltage fluctuation coefficient Simultaneously, retrieve the fixed baseline alarm threshold from the equipment's factory technical specifications. ; S32. Retrieve the normal steady-state operation dataset of the power equipment to be monitored in recent years, and remove historical fault data, maintenance data and abnormal disturbance data to construct a clean model training sample set; S33. By using multiple linear regression fitting, the model is trained on the training sample set to obtain the six-dimensional dynamic compensation coefficients. , , , , and ; S34. Calculate the dynamic alarm threshold based on the relevant data from S31 and S33. The process is as follows: ； in , , , , and These represent the reference threshold coefficient, load compensation coefficient, temperature compensation coefficient, humidity compensation coefficient, seasonal compensation coefficient, and voltage fluctuation compensation coefficient, respectively. S35. Set the minimum allowable threshold for alarms. and maximum boundary And perform threshold validity verification, when the condition is met S4 is performed at that time.

5. The power equipment monitoring and early warning method based on multi-feature collaborative analysis according to claim 4, characterized in that, The process of S4 is as follows: S41. Read the key monitoring features in the four-dimensional time series feature matrix S2, and perform short-time mean filtering on the features to obtain real-time feature values. S42. Compare the real-time feature value with the dynamic alarm threshold and perform a two-level branch decision. S43. Record each judgment result, including dynamic threshold value, real-time feature value, working condition label, judgment result, trigger status and timestamp; S44. Set up a confirmation mechanism for consecutive N frames to verify the stability of the status signal. Only when the feature value exceeds the limit for 3 to 5 consecutive monitoring cycles will the TCN-LSTM hybrid model of S5 be triggered stably. Otherwise, the in-depth analysis will not be triggered to further filter out instantaneous noise interference.

6. The power equipment monitoring and early warning method based on multi-feature collaborative analysis according to claim 5, characterized in that, The two-level branch determination process of S42 is as follows: When real-time feature value When the process enters the suspected abnormal branch, it determines that the power equipment under monitoring is currently in a suspected abnormal operating condition. It immediately generates a primary alarm trigger signal, wakes up the TCN-LSTM deep prediction model that is in a dormant state, and marks the current time, over-limit amplitude and operating condition information. The process then enters the S5 deep time series analysis stage. When real-time feature value When the time comes, the system enters the normal operation branch, determines that the power equipment under monitoring is operating normally, does not trigger a primary alarm, maintains the TCN-LSTM model in a dormant state, and returns to S1 to continue the next round of data acquisition and monitoring cycle.

7. The power equipment monitoring and early warning method based on multi-feature collaborative analysis according to claim 6, characterized in that, The process of S5 is as follows: S51. Receive the primary alarm trigger signal from S4, switch the awakened TCN-LSTM deep prediction model to the working state, and input the multi-dimensional time series data in the four-dimensional time series feature matrix into the TCN-LSTM deep prediction model. S52. Perform sequence normalization on the input temporal feature matrix. The process involves interpolating and filling in missing points, normalizing and aligning the feature dimensions, and unifying the length of the model input sequence and the order of feature arrangement. S53. Utilize the dilated convolution and causal convolution characteristics of the TCN temporal convolutional network to perform multi-layer convolution operations on long temporal sequences to obtain advanced temporal features. S54. Input the advanced time series features into the LSTM memory unit to obtain long-term trend features; S55. Weighted fusion of advanced temporal features and long-term trend features to construct a spatiotemporal joint advanced feature representation; S56. Combining spatiotemporal joint high-level feature representation, the TCN-LSTM deep prediction model is used to predict the state of future multiple time periods and quantify the probability of failure. Two types of judgment results are output through the fully connected layer: multi-scale equipment state time-series prediction curve and failure probability value. S57. Based on the multi-scale equipment state time-series prediction curve, the fault trend sequence is obtained. The fault trend sequence and the fault occurrence probability value are synchronously divided into two parallel outputs, which are then processed in S6 and S7 respectively.

8. The power equipment monitoring and early warning method based on multi-feature collaborative analysis according to claim 7, characterized in that, The process of S6 is as follows: S61. Perform slope analysis, trend fitting, and risk classification on the fault trend sequence, and classify the power equipment to be monitored into three states: continuous risk deterioration, instantaneous disturbance fluctuation, and stable state. S62. Based on the status of the power equipment to be monitored, perform corresponding dynamic threshold intelligent fine-tuning, as follows: If the risk continues to worsen and the equipment is in a deterioration phase, the adaptive threshold for the next cycle should be slightly lowered, and the system alarm sensitivity should be increased. If the disturbance is transient, it is determined that there is a brief disturbance, the threshold for the next cycle is slightly increased, and the transient over-limit interference is blocked. If the state is stable, the device is in a stable state and the threshold remains the original calculated value. S63. Set a fine-tuning amplitude constraint mechanism to ensure that the threshold fine-tuning amount in a single cycle does not exceed the benchmark threshold. ; S64. The fine-tuned threshold parameters are fed back to the load-environment dual-factor threshold compensation model to cover the initial calculated threshold for the next monitoring cycle; at the same time, the updated threshold is subjected to first-order smoothing filtering. S65. Record the parameters for each round of fine-tuning, including the original threshold, fine-tuning direction, fine-tuning magnitude, predicted failure probability, trend type, and timestamp.

9. The power equipment monitoring and early warning method based on multi-feature collaborative analysis according to claim 8, characterized in that, The criteria for determining the three states in S61 are as follows: Risk continues to worsen: the future forecast curve continues to rise, and the probability of failure increases cycle by cycle; Instantaneous disturbance fluctuation: Currently slightly exceeded, but the future trend is rapidly declining and the probability is decreasing; Stable and stable status: The prediction curve is flat, the probability is stable and without fluctuation, and the equipment operating status does not change significantly.

10. The power equipment monitoring and early warning method based on multi-feature collaborative analysis according to claim 9, characterized in that, The process for S7 is as follows: S71. Preset a three-segment probability threshold range, and standardize the classification of equipment operation risk based on the probability value of failure occurrence, as follows: Normal level, the probability of failure is within The equipment is running smoothly, showing no signs of deterioration, and no alarms or work orders are generated. Pay attention to the warning; the probability of a fault occurring is at [a certain level]. If the equipment has potential abnormalities, slight misalignment, or slow deterioration, a daily inspection work order will be generated. Emergency warning, the probability of failure is at The equipment exhibits a clear trend of failure evolution and poses a short-term risk of failure, thus generating an emergency repair work order. S72. Automatically generate standardized operation and maintenance work orders based on the warning level. The work order content includes: equipment number, current operating status, out-of-limit characteristic quantity, dynamic threshold, failure probability, future trend curve, risk level, suggested handling measures, and occurrence timestamp. S73. Implement tiered and differentiated push notifications through the operation and maintenance platform and mobile terminals, pushing ordinary warnings to inspection personnel, and pushing attention warnings and emergency warnings to operation and maintenance managers and emergency repair teams. S74. The relevant staff shall take appropriate measures, and after the measures are completed, the actual operation and maintenance results shall be backfilled through either manual or system methods. S75. After cleaning the actual operation and maintenance results, the data is fed back to the system's historical training dataset to complete the early warning process.

Images