Emergency alarm system for power safety protection

Abstract

This invention discloses an emergency alarm system for power supply safety protection, specifically in the field of power supply protection alarm technology, including an intelligent detection module, a multi-level alarm module, and an automatic processing module. The intelligent detection module achieves accurate identification and low false alarm detection of power supply anomalies through an integrated learning model and online learning mechanism; the multi-level alarm module adaptively generates tiered alarm signals based on dynamic risk assessment and scene perception; the automatic processing module employs reinforcement learning decision-making and digital twin simulation verification to automatically execute optimal safety measures and form a closed-loop control. Through the coordinated operation of these three modules, this invention solves the problems of high false alarm rates, slow response times, and reliance on manual intervention in traditional systems, significantly improving the intelligence level, response speed, and overall reliability of power supply safety protection.

Description

Technical Field

[0001] This invention relates to the field of power supply protection alarm technology, and more specifically, to an emergency alarm system for power supply safety protection. Background Technology

[0002] As the energy source for various electrical equipment and electronic systems, the stability and safety of the power supply system are of paramount importance. Power supply anomalies such as overvoltage, undervoltage, overcurrent, and short circuits can not only cause permanent equipment damage and data loss, but may even trigger serious safety accidents such as fires. Therefore, emergency alarm systems for power supply safety protection have become an indispensable component in critical fields such as data centers, industrial control systems, and intelligent buildings.

[0003] Currently, most common power safety alarm systems use comparator circuits or simple microcontrollers based on fixed thresholds for monitoring. Their working principle involves collecting parameters such as voltage and current through sensors and comparing them with preset safety thresholds. Once these thresholds are exceeded, a general alarm signal is triggered. This type of solution has significant drawbacks: First, its detection mechanism is rigid and cannot effectively distinguish between genuine continuous anomalies and transient electromagnetic interference or normal fluctuations caused by load switching, resulting in a high false alarm rate and frequent unnecessary interference. Second, the alarm modes are limited, typically only offering "normal" and "abnormal" states, failing to provide tiered warnings based on the severity and development trend of the anomaly. This may lead to insufficient response to high-risk alarms or overreaction to low-risk events by operators. Finally, existing systems generally lack intelligent automatic processing capabilities, heavily relying on manual intervention for analysis and processing after an alarm is issued. The valuable time delay between anomaly detection and manual implementation of protective measures can delay effective fault isolation and control at critical moments, failing to meet the stringent requirements of modern high-reliability power systems for real-time active protection.

[0004] Furthermore, although some improvement solutions attempt to introduce simple delay judgments or multi-parameter verifications, their adaptability and accuracy remain insufficient in complex and ever-changing real-world operating environments. The system's decision-making logic is fixed, unable to learn from historical data and self-optimize, potentially leading to long-term performance degradation. Therefore, there is an urgent need in this field for an integrated power safety protection solution capable of accurate detection, intelligent grading, and rapid autonomous response to overcome the limitations of existing technologies in terms of reliability, efficiency, and intelligence. Summary of the Invention

[0005] To overcome the above-mentioned deficiencies of the prior art, embodiments of the present invention provide an emergency alarm system for power safety protection.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An emergency alarm system for power supply safety protection includes the following modules:

[0008] The intelligent detection module is used to collect power supply voltage, current and frequency data through sensors. After wavelet transform denoising and normalization preprocessing, mean features are extracted. Then, a heterogeneous model integrated stack is used for anomaly detection. Anomalies are judged by dynamic weight fusion and confidence assessment. The model is optimized through online learning. The analysis is performed with a 1-second sliding window. If the anomaly ratio exceeds 80%, it is confirmed to be a real anomaly.

[0009] The multi-level alarm module calculates a comprehensive severity index based on the results of the intelligent detection module, from three dimensions: basic deviation, trend acceleration, and correlation coupling. This index is then mapped to and executed as level one, two, and three alarms. If the anomaly is not resolved after an alarm is triggered, the alarm is automatically escalated. At the same time, alarm history is recorded for threshold optimization.

[0010] The automatic processing module selects actions from the measure library according to the alarm level and abnormality type after the alarm is triggered. The microcontroller controls the actuator in PWM mode. The module evaluates the effect and scores it in fuzzy mode. The strategy is adjusted according to the score. At the same time, the log is recorded to non-volatile memory and a processing report is sent to the monitoring center.

[0011] Specifically, the execution process of the intelligent detection module is as follows:

[0012] Data acquisition and preprocessing: Voltage, current and frequency data of the power supply are continuously acquired by sensors. The sampling frequency is set to 1000 times per second. The acquired raw data is preprocessed, including denoising and normalization. Denoising is performed using wavelet transform algorithm to remove high-frequency noise. Normalization scales the data to the range of [0,1] to eliminate the influence of dimensions.

[0013] Feature extraction: Key features, including mean, variance, peak factor, and waveform entropy, are extracted from the preprocessed data to characterize the dynamic changes in power supply state; the feature extraction cycle is set to once every 100 milliseconds.

[0014] Anomaly detection: An integrated and adaptive approach is used for anomaly identification, and the model is given the ability to learn online and optimize itself to cope with new anomaly patterns that constantly emerge in the power system;

[0015] Results verification: To reduce instantaneous false alarms, a sliding window analysis was performed on the continuous detection results. The window size was 10 cycles, or 1 second. Only when the proportion of abnormalities in the window exceeded 80% was it confirmed as a real abnormality and output.

[0016] Specifically, the anomaly detection process in the intelligent detection module is as follows:

[0017] Ensemble Model Inference: Anomaly classification is performed using a pre-trained heterogeneous model ensemble stack. The stack consists of multiple heterogeneous base models for capturing local waveform features, a support vector machine for handling high-dimensional feature spaces, and a lightweight gradient boosting tree for efficiently processing tabular features in parallel. Each base model independently infers from the extracted feature vectors and outputs an initial anomaly probability score. , where i represents the i-th base model;

[0018] Dynamic weighted fusion: A dynamic weighted fusion mechanism based on real-time performance evaluation is introduced, maintaining a record of the performance of each base model over the past 100 detection cycles; the final anomaly probability for the current cycle is also included. The probabilities output by each model are calculated as a weighted average based on their recent performance weights: ,in, This represents the final anomaly probability for the current detection period. The dynamic weights of the i-th base model are... N represents the total number of base models in the heterogeneous model integration stack. Let be the recent AUC value of the i-th model;

[0019] Confidence assessment and uncertainty quantification: Calculate the confidence score C for this ensemble inference; the confidence score C is defined as 1 - entropy, where the entropy H is calculated using the following formula: ,in, This represents the normalized distribution of the output probabilities for each model, where K is the number of classes. It is a logarithmic function with base 2; when C is less than 0.7, the model has high uncertainty in judging this sample, so this sample is marked as pending and the subsequent processing is triggered.

[0020] Design an online learning pipeline based on confidence feedback.

[0021] Specifically, the online learning pipeline based on confidence feedback in the intelligent detection module is as follows:

[0022] Automatic sample labeling: For a decision result with a confidence score C ≥ 0.9, the feature vector and decision label are automatically used as new labeled samples and stored in a fixed-size first-in-first-out buffer.

[0023] Incremental training trigger: When the number of new samples in the buffer reaches the predetermined batch size, the incremental training process is automatically started;

[0024] Elastic Incremental Updates: Incremental training selectively updates the model in the ensemble stack; at the same time, all data in the buffer is periodically used to fine-tune the 1D-CNN part of the ensemble stack to capture deeper pattern evolution.

[0025] Output and triggering: Compare with an adaptively adjustable dynamic threshold; if If the value exceeds the dynamic threshold, an anomaly flag is triggered, along with its probability value and confidence level, and output to the result verification step.

[0026] Specifically, the execution process of the multi-level alarm module is as follows:

[0027] Anomaly Classification: Receives anomaly flags and their confidence levels output by the intelligent detection module, and calculates a comprehensive anomaly severity index; the anomaly severity index is synthesized through a multi-dimensional evaluation model, which not only considers the instantaneous deviation of parameters, but also covers dynamic change trends and the correlation and coupling effects with other parameters;

[0028] Level mapping: The severity index is mapped to three alarm levels: Level 1 is low risk, with a comprehensive severity index < 0.5; Level 2 is medium risk, with a comprehensive severity index of 0.5 ≤ comprehensive severity index < 1.0; Level 3 is high risk, with a comprehensive severity index ≥ 1.0; Each level corresponds to a different alarm method: Level 1 triggers a visual alarm, Level 2 triggers an audible alarm, and Level 3 triggers a comprehensive alarm.

[0029] Alarm execution: Based on the mapping results, activate the corresponding alarm devices; for level 3 alarms, simultaneously activate the local audible and visual alarms and send SMS or emails to the monitoring center via the wireless module; before alarm execution, perform device self-checks to ensure the alarm channels are available;

[0030] Feedback and Adjustment: Monitor alarm response status. If the anomaly persists after an alarm, the alarm level will be automatically escalated. If the anomaly is not eliminated within 10 seconds after a Level 1 alarm, it will be escalated to Level 2. At the same time, alarm history will be recorded for subsequent optimization of the grading thresholds.

[0031] Specifically, the anomaly classification process in the multi-level alarm module is as follows:

[0032] Calculate the basic deviation index : Instantaneous deviation percentage based on power supply voltage and current and its duration The basic severity is calculated using the following formula: ,in , These are the weighting factors calibrated through experiments;

[0033] Calculate the trend acceleration factor To capture the dynamic deterioration potential of anomalies, the acceleration of changes in anomaly parameters over the past 5 sampling periods was analyzed; the second derivative of the parameter values ​​was calculated and normalized to obtain the acceleration factor. ,in, Let P be the maximum value function, P be the outlier parameter, and t be time. This is the estimate of the second derivative of the outlier parameter P. This is a normalization coefficient used to scale the acceleration dimension to a reasonable range of magnification.

[0034] Calculate the correlation coupling coefficient Monitor a set of predefined correlation parameters: Where M is the total number of predefined associated parameters, This is an anomaly indicator for the m-th associated parameter, with 0 for normal and 1 for abnormal. The predefined weights represent their importance in relation to the main anomaly, and ;

[0035] Composite Severity Index The assessment results from the three dimensions are combined to form the final comprehensive severity index. .

[0036] Specifically, the execution process of the automatic processing module is as follows:

[0037] Measure selection: Select a matching action from the predefined measure library based on the level and anomaly type of the multi-level alarm module;

[0038] The measures library includes: disconnecting part of the load, enabling backup power, adjusting voltage regulator parameters, or completely disconnecting the main power supply; the selection logic is based on a rule engine.

[0039] Execution control: Control signals are sent to the actuator via the microcontroller; the control signals use pulse width modulation to ensure smooth operation; a safety check is performed before execution;

[0040] Effect evaluation: After the measures are implemented, the power parameters are monitored in real time; the evaluation period is set to once every 500 milliseconds. If the parameters do not return to normal within the set time, backup measures or upgrade processing will be triggered; if the voltage is still too high after disconnecting part of the load, it will automatically switch to completely disconnecting the main power supply.

[0041] Log recording: The automatic processing is recorded in non-volatile memory, including timestamps, action types and results, for subsequent analysis and system optimization; at the same time, processing reports are sent to the monitoring center through the communication module to achieve remote monitoring.

[0042] Specifically, the automatic processing module monitors changes in power parameters in real time, including:

[0043] Indicator recovery rate, system stability, recovery speed, and cost function;

[0044] Fuzzy comprehensive evaluation: Using a fuzzy logic system, the measured values ​​of the indicators are transformed into fuzzy evaluations of excellent, good, average, and poor, and then weighted and fused to obtain a comprehensive performance score. ;

[0045] Dynamic strategy adjustment: based on Decide on the next steps:

[0046] Excellent results. Record this status - the action pair is a positive sample, used for reinforcement learning model updates. Process ends.

[0047] The effect is mediocre; trigger backup measures or fine-tune the parameters.

[0048] If the effect is poor, the current measure is determined to have failed. Immediately provide feedback to the measure selection team, along with the current status information, to initiate other processing options.

[0049] The technical effects and advantages of this invention are as follows:

[0050] First, at the anomaly identification level, the system has achieved a leap from simple threshold judgment to intelligent pattern recognition through an intelligent detection module. This module employs integrated machine learning algorithms and real-time confidence assessment, effectively distinguishing between transient interference and genuine faults, and possesses online self-learning capabilities to adapt to performance drift during long-term system operation. This fundamentally reduces false alarm and false negative rates, ensuring the accuracy and reliability of alarm signals.

[0051] Secondly, at the alarm response level, the multi-level alarm module introduces a dynamic risk assessment and scenario awareness mechanism. By comprehensively calculating the severity index of the anomaly and combining it with contextual information such as time and system mode, it achieves precise mapping of alarm levels and adaptive adjustment of execution strategies. This enables the system to provide alarm information that matches the risk level appropriately, avoiding insufficient response or alarm fatigue, and greatly improving the efficiency and experience of human-computer interaction.

[0052] Third, at the safety closed-loop level, the automatic processing module constitutes an autonomous response system with predictive and learning capabilities. It utilizes reinforcement learning for optimal decision-making, performs pre-execution safety verification through digital twins, and adjusts processing strategies in real time based on multi-indicator fuzzy evaluation. This not only achieves zero-delay linkage from alarm to response, minimizing potential damage, but also continuously optimizes the knowledge base and decision-making model through a closed-loop learning mechanism, enabling the entire system to possess constantly evolving intelligence, ultimately constructing a precise, rapid, and adaptive proactive power safety protection system. Attached Figure Description

[0053] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0054] 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.

[0055] like Figure 1 As shown, the emergency alarm system module for power safety protection is as follows:

[0056] The intelligent detection module first collects power supply voltage, current, and frequency data 1000 times per second using sensors. After wavelet transform denoising and normalization preprocessing, features such as the mean are extracted every 100 milliseconds. Then, a heterogeneous model integration stack is used for anomaly detection, combining dynamic weight fusion and confidence assessment to determine anomalies. The model is optimized through online learning, and finally, analysis is performed using a 1-second sliding window (10 cycles). If the anomaly rate exceeds 80%, a genuine anomaly is confirmed. The module includes the following steps:

[0057] Data Acquisition and Preprocessing: Voltage, current, and frequency data of the power supply are continuously acquired using sensors (such as voltage and current sensors). The sampling frequency is set to 1000 times per second to ensure high resolution. The acquired raw data is preprocessed, including denoising and normalization. Denoising is performed using wavelet transform algorithm to remove high-frequency noise, and normalization scales the data to the [0,1] range to eliminate the influence of dimensions.

[0058] Feature extraction: Key features are extracted from the preprocessed data, including mean, variance, peak factor and waveform entropy. These features are used to characterize the dynamic changes in power state. The feature extraction period is set to once every 100 milliseconds to achieve real-time updates.

[0059] Anomaly detection: An integrated, adaptive method is employed for anomaly identification, and the model is endowed with online learning and self-optimization capabilities to cope with emerging anomaly patterns in the power supply system. The specific execution steps are as follows:

[0060] Ensemble Model Inference: Anomaly classification is performed using a pre-trained heterogeneous model ensemble stack. This stack consists of multiple heterogeneous base models (e.g., a one-dimensional convolutional neural network (1D-CNN) for capturing local waveform features, a support vector machine (SVM) for handling high-dimensional feature spaces, and a lightweight gradient boosting tree (LightGBM) for efficiently handling tabular features) operating in parallel. Each base model independently infers from the extracted feature vectors and outputs an initial anomaly probability score. (where i represents the i-th base model).

[0061] Dynamic weighted fusion: A dynamic weighted fusion mechanism based on real-time performance evaluation is introduced, instead of simple voting or averaging. The system maintains a recent (e.g., past 100 detection periods) performance record for each base model (e.g., based on the area under the AUC-ROC curve). The final anomaly probability for the current period... The probabilities output by each model are calculated as a weighted average based on their recent performance weights:

[0062] ;

[0063] in, This represents the final anomaly probability for the current detection period. The dynamic weights of the i-th base model are... N represents the total number of base models in the heterogeneous model integration stack. Let be the recent AUC value of the i-th model. This mechanism automatically increases the influence of models that perform better in the current operating environment, making the integrated system more robust and adaptable to different environments.

[0064] Confidence assessment and uncertainty quantification: Calculate the confidence score C for this ensemble inference. The confidence score C is defined as 1 - entropy, where the entropy H is calculated using the following formula:

[0065] ;

[0066] in, This represents the normalized distribution of the output probabilities for each model, where K is the number of categories (normal / abnormal). It is a logarithmic function with base 2. When C is lower than a preset threshold (e.g., 0.7), it indicates that the model has high uncertainty in judging this sample, so the sample is marked as pending and subsequent processing is triggered.

[0067] To address model degradation and adapt to new anomalies, an online learning pipeline based on confidence feedback is designed, as follows:

[0068] Automatic sample labeling: For judgment results with high confidence (C≥0.9), the system will automatically store its feature vector and judgment label (normal / abnormal) as new labeled samples in a fixed-size first-in-first-out (FIFO) buffer.

[0069] Incremental training trigger: When the number of new samples in the buffer reaches the predetermined batch size (e.g., 200), or when the system detects that the proportion of "low confidence" samples has been rising recently, the incremental training process will be automatically started.

[0070] Elastic Incremental Updates: Incremental training does not update the entire complex ensemble stack, but rather selectively updates specific models within the ensemble that are suitable for online learning (such as LightGBM). Simultaneously, the system periodically (e.g., every 24 hours) uses all data in the buffer to fine-tune the 1D-CNN portion of the ensemble stack to capture deeper pattern evolution. The entire process is executed asynchronously in the background, without affecting the real-time performance of forward inference.

[0071] Output and Trigger: Ultimately, will Compare with an adaptively adjustable dynamic threshold (initial value 0.5). If If the value exceeds the threshold, an anomaly flag is triggered, along with its probability value and confidence level, and output to the result verification step.

[0072] Results verification: To reduce instantaneous false alarms, a sliding window analysis was performed on the continuous detection results (the window size is 10 cycles, i.e., 1 second). Only when the proportion of abnormalities in the window exceeds 80% is it confirmed as a real abnormality and output.

[0073] The multi-level alarm module calculates a comprehensive severity index based on the results of the intelligent detection module, considering three dimensions: basic deviation, trend acceleration, and correlation coupling. This index is mapped to three alarm levels: Level 1 (<0.5, LED flashing), Level 2 (0.5-1.0, buzzer), and Level 3 (≥1.0, audible and visual alarm + remote notification), and then executed. If the anomaly is not resolved after an alarm, the alarm is automatically escalated (e.g., if Level 1 alarm fails to clear within 10 seconds, it is escalated to Level 2). Simultaneously, alarm history is recorded for threshold optimization. The process includes the following steps:

[0074] Anomaly Classification: The system receives anomaly flags and their confidence levels from the intelligent detection module and calculates a comprehensive anomaly severity index. This index is synthesized through a multi-dimensional evaluation model, considering not only instantaneous parameter deviations but also dynamic trends and their correlation with other key parameters, thus achieving a more accurate and proactive assessment of anomaly risk. The specific calculation steps are as follows:

[0075] Calculate the basic deviation index : Instantaneous deviation percentage based on key power supply parameters (such as voltage and current) and its duration The underlying severity is calculated using the following formula: ,in Based on the basic deviation index, , For weighting factors calibrated experimentally (e.g., when voltage deviation exceeds 10%) Set to 0.7, the duration increases by 1 second. (0.3)

[0076] Calculate the trend acceleration factor To capture the dynamic deterioration potential of anomalies, the acceleration of change of the anomaly parameter within a recent time window (e.g., the past 5 sampling periods) is analyzed. The second derivative of the parameter value (approximately the difference between adjacent rates of change) is calculated and normalized to obtain the acceleration factor. ,in, Let P be the maximum value function, P be the outlier parameter, and t be time. This is the estimate of the second derivative of the outlier parameter P. This is a normalization factor (e.g., 0.2) used to scale the acceleration measure to a reasonable range. This factor ensures that progressively worsening anomalies receive higher severity scores;

[0077] Calculate the correlation coupling coefficient Assess whether the current anomalous parameter has triggered a cascading anomaly in other key parameters to measure systemic risk. Monitor a predefined set of associated parameters (e.g., when the primary anomaly is "voltage," the associated parameters are "current" and "frequency"). Where M is the total number of predefined associated parameters, This is the anomaly indicator for the m-th associated parameter (0 for normal, 1 for abnormal). The predefined weights represent their importance in relation to the main anomaly, and (For example, current weighting for voltage anomalies) It can be set to 0.6, frequency weight. (Set to 0.4).

[0078] Composite Severity Index The assessment results from the above three dimensions are combined to form the final comprehensive severity index. .

[0079] Level mapping: The severity index is mapped to three alarm levels: Level 1 (low risk, comprehensive severity index < 0.5), Level 2 (medium risk, 0.5 ≤ comprehensive severity index < 1.0), and Level 3 (high risk, comprehensive severity index ≥ 1.0). Each level corresponds to a different alarm method: Level 1 triggers a visual alarm (such as LED flashing), Level 2 triggers an audible alarm (such as a buzzer), and Level 3 triggers a combined alarm (such as audible, visual, and remote notification).

[0080] Alarm execution: Based on the mapping results, activate the corresponding alarm devices; for level 3 alarms, simultaneously activate the local audible and visual alarms and send SMS or emails to the monitoring center via wireless modules (such as GSM or Wi-Fi); before alarm execution, perform device self-checks to ensure that the alarm channels are available.

[0081] Feedback and Adjustment: Monitor alarm response status. If the anomaly persists after an alarm, the alarm level will be automatically escalated. If the anomaly is not eliminated within 10 seconds after a Level 1 alarm, it will be escalated to Level 2. At the same time, alarm history will be recorded for subsequent optimization of the grading thresholds.

[0082] Upon alarm triggering, the automatic processing module selects an action (such as load disconnection or backup power activation) from the measure library based on the alarm level and anomaly type. The microcontroller controls the actuator via PWM. The module evaluates the effect and assigns a fuzzy score every 500 milliseconds. The strategy is adjusted based on the score (≥0.8: end; 0.5-0.8: fine-tune; <0.5: change solution). Simultaneously, a log is recorded to non-volatile memory, and a processing report is sent to the monitoring center. This includes the following steps:

[0083] Measure selection: Select an appropriate action from the predefined measure library based on the level and anomaly type of the multi-level alarm module (such as overvoltage, undervoltage, or overcurrent). The measure library includes: disconnecting part of the load, enabling the backup power supply, adjusting the voltage regulator parameters, or completely disconnecting the main power supply. The selection logic is based on the rule engine. For example, for a level 3 overvoltage anomaly, disconnecting part of the load is performed first. If this is ineffective, the system switches to the backup power supply.

[0084] Execution control: Control signals are sent to the actuator (such as a relay or solid-state switch) via a microcontroller (such as an ARM Cortex-M series); the control signal adopts PWM (pulse width modulation) to ensure smooth operation; before execution, a safety check is performed, such as checking whether the actuator is in normal condition to avoid conflicting operations.

[0085] Effectiveness evaluation: After implementing the measures, monitor changes in power parameters in real time, including:

[0086] Core indicators include recovery rate (e.g., the proportion of voltage deviation recovery), system stability (e.g., the variance of parameters within the evaluation period), recovery speed (e.g., the time required for parameters to recover to the normal range), and cost function (e.g., the load loss and number of switching caused by the measure).

[0087] Fuzzy comprehensive evaluation: Using a fuzzy logic system, the measured values ​​of the above indicators are transformed into fuzzy ratings such as "excellent," "good," "average," and "poor," and then weighted and fused to obtain a comprehensive performance score. (0-1 point).

[0088] Dynamic strategy adjustment: based on Decide on the next steps:

[0089] The results are excellent. Record the "state-measure" pair as a positive sample for use in updating the reinforcement learning model. The process ends here. The effect is generally poor, triggering backup measures or making minor parameter adjustments (such as slightly adjusting the voltage regulator output). If the effect is poor, the current measure is deemed a failure. Immediately report this to the "Measure Selection" module, along with the current status information, to initiate a higher-level processing plan. The evaluation cycle is set to once every 500 milliseconds. If the parameters do not return to normal within the set time (e.g., 2 seconds), backup measures or upgraded processing will be triggered. If the voltage is still too high after partially disconnecting the load, it will automatically switch to completely disconnecting the main power supply.

[0090] Log recording: The automatic processing is recorded in non-volatile memory, including timestamps, action types and results, for subsequent analysis and system optimization; at the same time, processing reports are sent to the monitoring center through the communication module to achieve remote monitoring.

[0091] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.

[0092] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0093] It should be understood that, in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0094] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0095] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0096] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0097] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0098] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An emergency alarm system for power safety protection, characterized in that, Includes the following modules: The intelligent detection module is used to collect power supply voltage, current and frequency data through sensors, and after wavelet transform denoising and normalization preprocessing, extract mean, variance, peak factor and waveform entropy features. Then, anomaly detection is performed using a heterogeneous model integration stack. Anomalies are judged by combining dynamic weight fusion and confidence assessment. The model is optimized through online learning. The analysis is performed with a 1-second sliding window. The anomaly rate exceeds 80%, confirming real anomalies. The multi-level alarm module calculates a comprehensive severity index based on the results of the intelligent detection module, from three dimensions: basic deviation, trend acceleration, and correlation coupling. This index is then mapped to and executed as level one, two, and three alarms. If the anomaly is not resolved after an alarm, the alarm is automatically escalated. At the same time, alarm history is recorded for threshold optimization. The calculation process for the comprehensive severity index is as follows: Calculate the basic deviation index : Instantaneous deviation percentage based on power supply voltage and current and its duration The basic severity is calculated using the following formula: ,in , These are the weighting factors calibrated through experiments; Calculate the trend acceleration factor Analyze the acceleration of the abnormal parameters over the past 5 sampling periods; calculate the second derivative of the parameter values ​​and normalize them to obtain the acceleration factor. ,in, Let P be the maximum value function, P be the outlier parameter, and t be time. Let be the estimate of the second derivative of the outlier parameter P with respect to time t. This is a normalization coefficient used to scale the acceleration dimension to a reasonable range of magnification. Calculate the correlation coupling coefficient Monitor a set of predefined correlation parameters: Where M is the total number of predefined associated parameters, This is an anomaly indicator for the m-th associated parameter, with 0 for normal and 1 for abnormal. The pre-defined weights represent their importance in relation to the main anomaly, and ; Synthetic composite severity index : The results of the assessment of the three dimensions are fused to form the final composite severity index: ; The automatic processing module selects actions from the measure library according to the alarm level and abnormality type after the alarm is triggered. The microcontroller controls the actuator in PWM mode. The module evaluates the effect and scores it in fuzzy mode. The strategy is adjusted according to the score. At the same time, the log is recorded to non-volatile memory and a processing report is sent to the monitoring center.

2. The emergency alarm system for power safety protection according to claim 1, characterized in that, The execution process of the intelligent detection module is as follows: Data acquisition and preprocessing: Voltage, current and frequency data of the power supply are continuously acquired by sensors. The sampling frequency is set to 1000 times per second. The acquired raw data is preprocessed, including denoising and normalization. Denoising is performed using wavelet transform algorithm to remove high-frequency noise, and normalization is performed to scale the data to the range of [0,1]. Feature extraction: Key features, including mean, variance, peak factor, and waveform entropy, are extracted from the preprocessed data to characterize the dynamic changes in power supply state; the feature extraction cycle is set to once every 100 milliseconds. Anomaly detection: An integrated and adaptive approach is used for anomaly identification, and the model is endowed with the ability to learn online and self-optimize. Result verification: Sliding window analysis was performed on the continuous detection results. The window size was 10 cycles, or 1 second. Only when the proportion of abnormalities in the window exceeded 80% was it confirmed as a real abnormality and output.

3. The emergency alarm system for power safety protection according to claim 2, characterized in that, The specific process of anomaly detection in the intelligent detection module is as follows: Ensemble Model Inference: Anomaly classification is performed using a pre-trained heterogeneous model ensemble stack, which consists of multiple heterogeneous base models in parallel. The base models include: a one-dimensional convolutional neural network for capturing local waveform features, a support vector machine for processing high-dimensional feature space, and a lightweight gradient boosting tree for efficiently processing tabular features. Each base model independently reasons over the extracted feature vector and outputs an initial anomaly probability score where i represents the i-th base model; Dynamic weighted fusion: A dynamic weighted fusion mechanism based on real-time performance evaluation is introduced, maintaining a record of the performance of each base model over the past 100 detection cycles; the final anomaly probability for the current cycle is also included. The probabilities output by each model are calculated as a weighted average based on their recent performance weights: ,in, This represents the final anomaly probability for the current detection period. The dynamic weights of the i-th base model are... N represents the total number of base models in the heterogeneous model integration stack. The area under the recent AUC-ROC curve of the i-th base model is used to quantitatively evaluate the performance of the base model. Confidence assessment and uncertainty quantification: Calculate the confidence score C for this ensemble inference; the confidence score C is defined as 1 - entropy, where the entropy H is calculated using the following formula: ,in, This represents the normalized distribution of the output probabilities for each model, where K is the number of classes. It is a logarithmic function with base 2; when C is less than 0.7, the model has high uncertainty in judging this sample, so this sample is marked as pending and the subsequent processing is triggered. Design an online learning pipeline based on confidence feedback.

4. The emergency alarm system for power safety protection according to claim 3, characterized in that, The process of the online learning pipeline based on confidence feedback in the intelligent detection module is as follows: Automatic sample labeling: For a decision result with a confidence score C ≥ 0.9, the feature vector and decision label are automatically used as new labeled samples and stored in a fixed-size first-in-first-out buffer. Incremental training trigger: When the number of new samples in the buffer reaches the predetermined batch size, the incremental training process is automatically started; Elastic Incremental Updates: Incremental training selectively updates the model in the ensemble stack; at the same time, all data in the buffer is periodically used to fine-tune the one-dimensional convolutional neural network part in the ensemble stack to capture deeper pattern evolution. Output and triggering: Compare with an adaptively adjustable dynamic threshold; if If the value exceeds the dynamic threshold, an anomaly flag is triggered, along with its probability value and confidence level, and output to the result verification step.

5. The emergency alarm system for power safety protection according to claim 1, wherein, The execution process of the multi-level alarm module is as follows: Anomaly Classification: Receives anomaly flags and their confidence levels output by the intelligent detection module, and calculates a comprehensive severity index; the comprehensive severity index is synthesized through a three-dimensional evaluation model of parameter base deviation, trend acceleration, and correlation coupling; Level Mapping: The comprehensive severity index is mapped to three alarm levels: Level 1 is low risk, with a comprehensive severity index < 0.5; Level 2 is medium risk, with a comprehensive severity index of 0.5 ≤ comprehensive severity index < 1.0; Level 3 is high risk, with a comprehensive severity index ≥ 1.0; each level corresponds to a different alarm method, with Level 1 triggering a visual alarm, Level 2 triggering an audible alarm, and Level 3 triggering a comprehensive alarm. Alarm execution: Based on the mapping results, activate the corresponding alarm devices; for level 3 alarms, simultaneously activate the local audible and visual alarms and send SMS or emails to the monitoring center via the wireless module; before alarm execution, perform device self-checks to ensure the alarm channels are available; Feedback and Adjustment: Monitor alarm response status. If the abnormality persists after an alarm is triggered, the alarm level will be automatically escalated. At the same time, alarm history will be recorded for subsequent optimization of the grading threshold.

6. The emergency alarm system for power supply safety protection according to claim 1, characterized in that, The execution process of the automatic processing module is as follows: Measure selection: Select a matching action from the predefined measure library based on the level and anomaly type of the multi-level alarm module; The measures library includes: disconnecting part of the load, enabling backup power, adjusting voltage regulator parameters, or completely disconnecting the main power supply; the selection logic is based on a rule engine. Execution control: Control signals are sent to the actuator via the microcontroller; the control signals use pulse width modulation to ensure smooth operation; a safety check is performed before execution; Effect evaluation: After the measures are implemented, the power parameters are monitored in real time; the evaluation period is set to once every 500 milliseconds. If the parameters do not return to normal within the set time, backup measures or upgrade processing will be triggered; if the voltage is still too high after disconnecting part of the load, it will automatically switch to completely disconnecting the main power supply. Log recording: The automatic processing is recorded in non-volatile memory, including timestamps, action types and results, for subsequent analysis and system optimization; at the same time, processing reports are sent to the monitoring center through the communication module to achieve remote monitoring.

7. The emergency alarm system for power safety protection according to claim 6, characterized in that, The automatic processing module monitors changes in power parameters in real time, including: Power supply parameters include index recoverability, system stability, recovery speed, and cost function; Fuzzy comprehensive evaluation: using fuzzy logic system, the measured value of the power parameter index is converted into excellent, good, medium, and poor fuzzy comments, and a comprehensive effect score is obtained by weighted fusion ; Dynamic policy adjustment: based on Decide on follow-up actions: : Effect is good, record this state - action pair as positive sample, use it to update reinforcement learning model, end of process; : Effect is generally minor, triggering a backup measure or a parameter tweak; : poor effect, determine that the current measure has failed, immediately feedback to the measure selection and carry the current state information to start other processing schemes.

Images