Electric equipment fault prediction method based on adaptive deep reinforcement learning
By using an adaptive deep reinforcement learning method, the problem of fault prediction for electrical equipment in complex multi-source data and dynamic environments is solved, achieving high-precision and efficient fault prediction and management, and improving the efficiency of equipment health status monitoring and maintenance.
Patent Information
- Application Number
- CN202511651915.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-02-10
AI Technical Summary
Existing methods for predicting electrical equipment faults are inadequate in handling complex multi-source data and adapting to dynamic environments. They struggle to achieve high robustness and generalization ability, and their prediction accuracy is low under complex operating conditions, making them unable to adaptively adjust.
An adaptive deep reinforcement learning-based approach is adopted. Through multi-source data preprocessing, attention mechanism feature extraction network and deep reinforcement learning model, a state space, action space and immediate reward function are constructed. The agent is trained by combining adaptive learning rate algorithm, and the fault risk index is output and visualized.
It achieves high-precision fault prediction, improves equipment health management and maintenance efficiency, and enhances prediction accuracy and robustness in complex dynamic environments.
Smart Images
Figure CN121502295A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electrical equipment fault prediction technology, and in particular to a method for electrical equipment fault prediction based on adaptive deep reinforcement learning. Background Technology
[0002] The safe operation of electrical equipment is crucial to the stability of industrial production, energy management, and social life; its failure can lead to production interruptions, equipment damage, and even safety accidents. In recent years, with the increasing complexity and intelligence of equipment, real-time monitoring and fault prediction have become core requirements for ensuring equipment reliability. However, existing methods have significant shortcomings in processing complex, multi-source data and adapting to dynamic environments, making it difficult to meet the high demands of modern equipment operation scenarios.
[0003] Current fault prediction methods typically rely on single types of sensor data or static models, making it difficult to address the interactive effects of multiple physical parameters during equipment operation. For example, the coupling relationships between electrical, thermal, and mechanical vibration parameters are complex and highly susceptible to environmental changes, making it difficult for traditional methods to accurately capture these dynamic interaction characteristics. Furthermore, existing methods often lack adaptive adjustment capabilities when faced with data uncertainty, leading to decreased prediction accuracy in small samples or new scenarios. These limitations make it difficult for equipment fault prediction to achieve high robustness and generalization ability under complex operating conditions.
[0004] Furthermore, in current industrial scenarios, the operating environment of electrical equipment is complex and variable, and data acquisition often faces issues such as time asynchrony or missing values. At the same time, significant differences in equipment type and operating stage make it difficult for feature extraction and fault prediction models to adaptively adjust, directly affecting the accuracy of predictions. Therefore, how to achieve accurate and rapid fault prediction in complex and dynamic environments has become a key issue in improving equipment reliability and reducing maintenance costs. Summary of the Invention
[0005] The purpose of this invention is to provide a fault prediction method for electrical equipment based on adaptive deep reinforcement learning, which can achieve high-precision fault prediction and significantly improve equipment health management and maintenance efficiency.
[0006] To achieve the above objectives, the present invention provides the following solution:
[0007] A fault prediction method for electrical equipment based on adaptive deep reinforcement learning includes:
[0008] Collect multi-source operating data of electrical equipment, preprocess the multi-source operating data, and obtain a synchronous data stream;
[0009] The synchronous data stream is input into an attention-based adaptive feature extraction network, which outputs high-dimensional fault-sensitive features.
[0010] The high-dimensional fault-sensitive features are input into the trained fault prediction model to obtain the action probability distribution output by the agent and convert it into the fault risk index of the electrical equipment. The fault prediction model is a deep reinforcement learning model, which is constructed by establishing the equipment state space, action space and instant reward function and trained using an adaptive learning rate algorithm.
[0011] Based on the equipment failure risk index, the equipment health status curve and risk heat map are displayed in real time through a visual interface.
[0012] Optionally, preprocessing the multi-source operational data to obtain a synchronized data stream includes:
[0013] A pre-established timestamp alignment mechanism is used to perform time consistency calibration on the multi-source running data to obtain a preliminary time-aligned dataset;
[0014] Based on the initial dataset after time alignment, interpolation methods are used to compensate for missing values to obtain a complete intermediate dataset.
[0015] For the complete intermediate dataset, a frequency adjustment operation is performed. By using preset sampling rules, data of different frequencies are converted into a unified sampling frequency to obtain a dataset with consistent frequency.
[0016] Data stream integration technology is used to obtain the synchronous data stream from datasets with consistent frequencies.
[0017] Optionally, the synchronous data stream is input into an adaptive feature extraction network based on an attention mechanism, and the high-dimensional fault-sensitive features are output by automatically adjusting the convolutional kernel size, number of channels, and number of attention heads according to the device type and operating stage.
[0018] Optionally, automatically adjusting the kernel size, number of channels, and number of attention heads based on device type and operational phase includes:
[0019] By using a pre-established equipment type classification library, the operating parameters and historical data records of different equipment are obtained, the equipment types are initially classified, and the basic dataset after equipment classification is obtained.
[0020] Based on the basic dataset after equipment classification, and combined with the rules for dividing the operation phase, time series segmentation technology is used to group the data according to the operation phase to obtain the data units after phase division.
[0021] Based on the data units after stage division and the dynamic adjustment requirements of convolutional kernel size, an initial framework for an adaptive network is constructed to obtain convolutional kernel size configuration schemes suitable for different stages.
[0022] Based on the convolution kernel size configuration scheme, if the number of channels exceeds the preset threshold in the current running stage due to the data complexity, the number of channels is increased to obtain the adjusted channel configuration result.
[0023] Based on the adjusted channel configuration results, the number of attention heads is adjusted using the allocation rules of the attention mechanism to determine the final number of attention heads configuration;
[0024] Based on the final configuration of the number of attention heads, the synchronized data stream is mapped to a high-dimensional space.
[0025] Optionally, the state space is composed of the high-dimensional fault-sensitive features, equipment health indicators, and environmental disturbances.
[0026] The action space includes fault warning level output, data acquisition frequency adjustment command, and feature extraction network parameter update range.
[0027] The instant reward function includes fault prediction accuracy, false alarm / false alarm cost, and communication and computing resource consumption.
[0028] Optionally, the acquisition of the state space includes: acquiring multi-dimensional information input based on the high-dimensional fault sensitivity features, health indicators and environmental disturbances, and combining the multi-dimensional information into a unified device state representation through data integration to acquire the state space.
[0029] Optionally, training the fault prediction model using an adaptive learning rate algorithm includes:
[0030] By using a pre-established training environment, relevant data on changes in state distribution are obtained, and the initial learning rate adjustment range is determined.
[0031] Based on the initial learning rate adjustment range, the dynamic changes in state distribution are monitored in real time during the commentator network optimization. If the changes in state distribution exceed a preset threshold, the learning rate is dynamically adjusted to obtain an updated learning rate configuration.
[0032] Based on the updated learning rate configuration, a gradient penalty term is introduced into the actor network training to obtain stable training parameters;
[0033] Based on stable training parameters, the direction data for policy updates is obtained by adjusting the structure of the deep learning network, and the optimized policy output is obtained.
[0034] Based on the optimized strategy output, the synchronization update frequency of the commentator network and the actor network is obtained. If the synchronization frequency is lower than the preset threshold, the delayed update mechanism is triggered to determine whether the overall training has reached the preset convergence condition and obtain the trained fault prediction model.
[0035] Optionally, a real-time visualization interface can be used to display device health status curves and risk heatmaps, including:
[0036] Based on the equipment failure risk index, the dynamic change characteristics of the equipment health status are extracted, and a health status curve is generated using a time series processing method to obtain the equipment health status curve. Based on the equipment health status curve, a visualization tool is used to perform spatial distribution mapping of the risk index to obtain the risk heat map.
[0037] The beneficial effects of this invention are as follows: The method of this invention addresses the interconnected business problems in traditional monitoring, such as low accuracy of fault warning, high resource consumption, and model failure due to concept drift caused by asynchronous data, unsuitable feature extraction, and the inability of prediction models to be dynamically optimized. By collecting multi-source operational data in real time and performing timestamp alignment and missing value compensation to generate a synchronous data stream, an adaptive feature extraction network is constructed to automatically adjust parameters and output high-dimensional fault-sensitive features. A reinforcement learning controller is introduced to maximize the prediction reward by retaining the decision dimension. A state space, action space, and reward function are established. The agent is trained using a dual-delay deep deterministic policy gradient adaptive learning rate algorithm. Finally, the action probability is transformed into an interpretable risk index and visualized, thereby achieving high-precision fault prediction and robust model adaptation, significantly improving equipment health management and maintenance efficiency. Attached Figure Description
[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0039] Figure 1 This is a flowchart of a fault prediction method for electrical equipment based on adaptive deep reinforcement learning, according to an embodiment of the present invention. Detailed Implementation
[0040] 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.
[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0042] like Figure 1As shown, this embodiment proposes a fault prediction method for electrical equipment based on adaptive deep reinforcement learning, including:
[0043] Collect multi-source operating data of electrical equipment, preprocess the multi-source operating data, and obtain synchronous data streams;
[0044] The synchronous data stream is input into an attention-based adaptive feature extraction network, which outputs high-dimensional fault-sensitive features.
[0045] High-dimensional fault-sensitive features are input into a pre-trained fault prediction model to obtain the action probability distribution of the agent output and convert it into a fault risk index of electrical equipment. The fault prediction model is a deep reinforcement learning model, which is constructed by establishing the equipment state space, action space and instant reward function and trained using an adaptive learning rate algorithm.
[0046] Based on the equipment failure risk index, the system displays the equipment health status curve and risk heat map in real time through a visual interface.
[0047] Furthermore, preprocessing of multi-source operational data to obtain synchronized data streams includes:
[0048] A pre-established timestamp alignment mechanism is used to perform time consistency calibration on multi-source running data to obtain a preliminary time-aligned dataset.
[0049] Based on the initial dataset after time alignment, interpolation methods are used to compensate for missing values to obtain a complete intermediate dataset.
[0050] For a complete intermediate dataset, a frequency adjustment operation is performed. By using preset sampling rules, data of different frequencies are converted into a unified sampling frequency to obtain a dataset with consistent frequencies.
[0051] Data stream integration techniques are used to obtain synchronous data streams for datasets with consistent frequencies.
[0052] Specifically, data is collected from electrical devices using IoT technology. Pre-established interface technologies are used to acquire multi-source operational data from multiple devices. This data includes at least electrical measurement information, resulting in a preliminary dataset. Based on this preliminary dataset, the multi-source operational data is uploaded to a cloud processing platform in real time using a data transmission mechanism. When collecting data from electrical devices using IoT technology, a smart grid scenario can be envisioned, where multiple household electrical devices, such as air conditioners and electric water heaters, are monitored in real time. The IoT uses sensors and communication modules to collect electrical measurement information such as voltage, current, and power from the devices and transmit it to a local gateway, forming a preliminary dataset. For example, an air conditioner might have a voltage of 220V and a current of 5A, and this data is recorded in real time. In one possible implementation, the MQTT protocol can be used to upload the multi-source operational data to the cloud processing platform, ensuring low latency and high reliability of data transmission. Assuming a data upload frequency of once per minute, if some current data is found to be missing in a transmission, it is cleaned according to preset threshold rules. For example, the missing values are filled with the average value of 5A from the previous minute, resulting in a cleaned running dataset. This ensures data integrity and lays the foundation for subsequent analysis.
[0053] When processing operational data of electrical appliances using IoT technology, we can envision a smart home scenario where data is collected and analyzed from multiple appliances such as refrigerators and washing machines. For multi-source operational data collection, assume there are time discrepancies in the data acquired from sensors of different devices; for example, the refrigerator's power data is recorded at 10:00:05, while the washing machine's current data is recorded at 10:00:08. Through a pre-established timestamp alignment mechanism, these data can be calibrated to a unified time point of 10:00:00, forming a time-aligned preliminary dataset. For handling missing values, assuming a refrigerator's voltage data is missing for a certain minute in the preliminary dataset, a linear interpolation method can be used. Combining the 220V of the previous minute and the 218V of the next minute, the missing value can be estimated as 219V, resulting in a complete intermediate dataset. This approach ensures data continuity. For instance, when performing frequency adjustment, assuming the refrigerator's data collection frequency is once per second, while the washing machine's is once every 5 seconds, by using preset sampling rules, both can be uniformly adjusted to a frequency of once every 3 seconds, resulting in a dataset with consistent frequency. This facilitates the consistency of subsequent data processing. To address data synchronization requirements, data stream integration technology is employed. For example, the voltage and current data streams of the refrigerator and washing machine are integrated into a single synchronized data stream, ensuring that the operating status of both can be acquired at the same time. For instance, at 10:01:00, the refrigerator voltage is 220V and the washing machine current is 3A, thus determining the synchronized data sequence. This integration facilitates overall analysis. In data segmentation processing, the synchronized data sequence for a day is divided into 24 time periods per hour, obtaining segmented data units. If the number of data points in a given hour is less than a preset threshold of 100, that time period is marked. For example, if only 80 data points are collected in the 1:00-2:00 AM segment, it is marked as incomplete, resulting in marked data units. For instance, for the marked data units, supplementary data collection is performed in the 1:00-2:00 AM segment by increasing the sensor sampling frequency or calling backup data sources to fill in the missing data points, ultimately forming a complete synchronized data stream. This processing ensures data comprehensiveness.
[0054] The steps of time alignment and frequency adjustment stem from the inconsistency of multi-source data. If left unaddressed, this inconsistency can lead to analytical biases. Segmentation and labeling, however, allow for precise identification of problematic data, ensuring the reliability of subsequent analyses. These steps collectively support the integrity and accuracy of data processing, providing a solid foundation for the monitoring of electrical equipment operation.
[0055] Furthermore, the synchronous data stream is input into an adaptive feature extraction network based on an attention mechanism, and the output high-dimensional fault-sensitive features include: automatically adjusting the convolutional kernel size, number of channels, and number of attention heads according to the device type and operating stage, and outputting high-dimensional fault-sensitive features.
[0056] Furthermore, the automatic adjustment of convolutional kernel size, number of channels, and number of attention heads based on device type and operational phase includes:
[0057] By using a pre-established equipment type classification library, the operating parameters and historical data records of different equipment are obtained, the equipment types are initially classified, and the basic dataset after equipment classification is obtained.
[0058] Based on the basic dataset after equipment classification, and combined with the rules for dividing the operation phase, time series segmentation technology is used to group the data according to the operation phase to obtain the data units after phase division.
[0059] Based on the data units after stage division and the dynamic adjustment requirements of convolutional kernel size, an initial framework for an adaptive network is constructed to obtain convolutional kernel size configuration schemes suitable for different stages.
[0060] Based on the convolution kernel size configuration scheme, if the number of channels exceeds the preset threshold in the current running stage due to the data complexity, the number of channels is increased to obtain the adjusted channel configuration result.
[0061] Based on the adjusted channel configuration results, the number of attention heads is adjusted using the allocation rules of the attention mechanism to determine the final number of attention heads configuration;
[0062] Based on the final configuration of the number of attention heads, the synchronous data stream is mapped to a high-dimensional space.
[0063] Specifically, in smart home scenarios, for processing operational data of electrical devices, a preliminary classification of devices such as refrigerators and air conditioners can be performed based on a pre-established device type classification library. Assuming refrigerators are classified as refrigeration devices and air conditioners as temperature control devices, historical power data for refrigerators and runtime records for air conditioners are obtained from the classification library to form the basic dataset after device classification. This classification method facilitates subsequent differentiated analysis based on the characteristics of different devices. Regarding the division of operational phases, refrigerator operational data can be grouped into startup, stable operation, and standby phases. Assuming that the refrigerator's power fluctuates significantly during startup and remains around 200W during stable operation, time series segmentation technology can be used to divide a day's data into multiple phase units. This grouping method facilitates the identification of operational patterns in different phases. Regarding the dynamic adjustment of convolutional kernel size, when constructing the adaptive network framework, the convolutional kernel size can be adjusted to 5x5 to capture a wider range of features, given the larger data fluctuations during the refrigerator startup phase, while it can be adjusted to 3x3 to focus on local details during the stable operation phase. This dynamic configuration scheme can adapt to the data characteristics of different phases. Regarding the channel number adjustment logic, if the data complexity during the refrigerator's startup phase is high, exceeding a preset threshold, the number of channels is increased from 16 to 32 to improve the processing capability for complex data. This adjusted channel configuration helps to extract equipment operating characteristics more comprehensively. In the dynamic allocation of attention heads, if the air conditioner's fault sensitivity data points exceed a preset threshold during peak operation, such as five instances of abnormal temperature fluctuations, the number of attention heads is increased from 4 to 8, focusing on key areas such as abnormal temperature points. This configuration method allows for more precise identification of potential problems.
[0064] To address the need for high-dimensional feature extraction, an adaptive network is used to map refrigerator power, voltage, and other data into a high-dimensional space, obtaining fault-sensitive feature data units. For example, extracting power surge features during the startup phase can aid in subsequent analysis to determine if the equipment is in an abnormal state.
[0065] Furthermore, the state space consists of high-dimensional fault-sensitive features, equipment health indicators, and environmental disturbances.
[0066] The action space includes fault warning level output, data acquisition frequency adjustment instructions, and feature extraction network parameter update range;
[0067] The immediate reward function includes fault prediction accuracy, false alarm / false alarm cost, and communication and computing resource consumption.
[0068] Furthermore, the acquisition of the state space includes: obtaining multi-dimensional information input based on high-dimensional fault sensitivity characteristics, health indicators and environmental disturbances, and combining the multi-dimensional information into a unified device state representation through data integration to obtain the state space.
[0069] Specifically, in the smart home field, monitoring the operational status of refrigerators can involve acquiring information from multiple data sources to construct a complete representation of the device's status. For data integration addressing high-dimensional features, health indicators, and environmental interference, let's assume that power fluctuation data collected by sensors is used as a dimensionality-reducing feature, while simultaneously recording the compressor's health indicators and monitoring ambient temperature as a disturbance factor. Through data integration technology, this multi-dimensional information is combined into a unified state representation. For example, setting the power fluctuation value to 50 watts, the health indicator to 80%, and the ambient temperature to 25 degrees Celsius creates a complete state space description, providing a basis for subsequent decision-making. The definition of the action space and the determination of fault warning trigger conditions can be based on a preset rule base. Assuming the rule base stipulates that a warning should be triggered when the health indicator is below 70%, if the current refrigerator's health indicator is 65%, a secondary warning signal output instruction is generated. This method can promptly identify potential problems, providing a preliminary basis for subsequent processing. Regarding data acquisition frequency adjustment, if the real-time monitoring module detects a sudden increase in ambient temperature from 25 degrees Celsius to 30 degrees Celsius, exceeding the preset threshold of 2 degrees Celsius, the data acquisition frequency will be adjusted from once per hour to once every 30 minutes. This dynamic adjustment strategy helps to capture state changes more promptly and improves monitoring sensitivity. For parameter updates, if the feature extraction network configuration module detects an increase in the power fluctuation of the dimensionality reduction features from 50 watts to 80 watts, exceeding the preset limit of 20 watts, the parameter update mechanism will be triggered to reconfigure network parameters, such as adjusting the weight allocation for feature extraction. This mechanism adapts to data changes and ensures the accuracy of the analysis. In reward function calculation, a reinforcement learning method is used to comprehensively evaluate prediction accuracy and resource consumption. Assuming the current prediction accuracy is 85% and resource consumption is 60%, the immediate reward value is calculated according to a preset standard and is positive, indicating that the current action space execution effect is good. This evaluation method helps optimize the decision-making process.
[0070] For continuous monitoring of equipment status, a data feedback mechanism combines reward values with new status data. For example, if the reward value is positive and the power fluctuation data is updated to 75 watts, a new status representation is generated, providing a basis for the next round of decision-making. This feedback loop continuously improves monitoring effectiveness and enhances the system's adaptability.
[0071] Furthermore, training the fault prediction model using an adaptive learning rate algorithm includes:
[0072] By using a pre-established training environment, relevant data on changes in state distribution are obtained, and the initial learning rate adjustment range is determined.
[0073] Based on the initial learning rate adjustment range, the dynamic changes in state distribution are monitored in real time during the critic network optimization. If the changes in state distribution exceed the preset threshold, the learning rate is dynamically adjusted to obtain the updated learning rate configuration.
[0074] Based on the updated learning rate configuration, a gradient penalty term is introduced into the actor network training to obtain stable training parameters;
[0075] Based on stable training parameters, the direction data for policy updates is obtained by adjusting the structure of the deep learning network, and the optimized policy output is obtained.
[0076] Based on the optimized strategy output, the synchronization update frequency of the commentator network and the actor network is obtained. If the synchronization frequency is lower than the preset threshold, the delayed update mechanism is triggered to determine whether the overall training has reached the preset convergence condition and obtain the trained fault prediction model.
[0077] Specifically, when optimizing the commentator network using an adaptive mechanism, changes in the refrigerator's operating state distribution can be monitored in real time. Assuming a preset threshold of state distribution change exceeding 10%, if a sudden increase in power from 40 watts to 55 watts is detected, exceeding the threshold, the learning rate is adjusted from 0.003 to 0.0025 to avoid training instability caused by overly rapid updates. This dynamic adjustment helps improve the network's adaptability to complex state changes. When introducing a gradient penalty term in actor network training, gradient fluctuation data can be analyzed to ensure training stability. Assuming a preset gradient fluctuation range of ±5%, if a fluctuation reaches 7%, training parameters are adjusted, such as decreasing the learning rate or increasing regularization weights, to ensure no overfitting or underfitting occurs during training. This method effectively smooths fluctuations during training. For the application of policy gradient methods, the update direction of the refrigerator monitoring policy can be obtained by adjusting the deep learning network structure. Assuming the current policy tends to frequently adjust the data collection frequency but consumes high resources, a more efficient policy output can be generated by reducing the number of network layers or optimizing the weight distribution. This adjustment can improve the efficiency of strategy execution when resources are limited.
[0078] When implementing the delayed update method, the synchronization frequency of the commentator network and the actor network can be monitored. Assuming a preset threshold of once every 10 iterations, if the actual frequency is once every 15 iterations, the delayed update mechanism is triggered, adjusting the update timing to once every 12 iterations. This mechanism can reduce unnecessary update operations and lower the computational burden.
[0079] Furthermore, the device's health status curves and risk heatmaps are displayed in real time through a visual interface, including:
[0080] Based on the equipment failure risk index, the dynamic change characteristics of equipment health status are extracted, and a health status curve is generated using time series processing methods. Based on the equipment health status curve, a visualization tool is used to spatially distribute the risk index and obtain a risk heat map.
[0081] Specifically, this embodiment generates a health status curve through time series analysis. Assuming the risk index gradually rises from 5.0 to 8.0 over the past 12 hours, with the curve showing a clear upward trend, it indicates that the equipment's condition may be deteriorating. This dynamic feature extraction helps identify patterns in equipment health status changes, providing data support for risk assessment. In constructing the risk distribution view, a risk heatmap is generated based on the status trend map. Assuming the risk indices for different areas inside the refrigerator are 6.0, 8.5, and 4.5 respectively, the heatmap will use varying shades of color to show high-risk areas concentrated in specific locations, such as near the freezer compartment. This intuitive view facilitates quick location of problem areas, improving monitoring accuracy. For generating maintenance guidance suggestions, the system performs rule matching based on the risk index and preset thresholds. Assuming the safety threshold is 6.0, and the current index for a certain area reaches 8.5, the system will automatically generate a prompt message suggesting checking the freezer temperature control device. This data-driven suggestion provides users with clear action guidelines, reducing the possibility of further escalation of the fault.
[0082] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for predicting faults in electrical equipment based on adaptive deep reinforcement learning, characterized in that, include: Collect multi-source operating data of electrical equipment, preprocess the multi-source operating data, and obtain a synchronous data stream; The synchronous data stream is input into an attention-based adaptive feature extraction network, which outputs high-dimensional fault-sensitive features. The high-dimensional fault-sensitive features are input into the trained fault prediction model to obtain the action probability distribution output by the agent and convert it into the fault risk index of the electrical equipment. The fault prediction model is a deep reinforcement learning model, which is constructed by establishing the equipment state space, action space and instant reward function and trained using an adaptive learning rate algorithm. Based on the equipment failure risk index, the equipment health status curve and risk heat map are displayed in real time through a visual interface.
2. The method for predicting electrical equipment faults based on adaptive deep reinforcement learning according to claim 1, characterized in that, Preprocessing the multi-source operational data to obtain a synchronized data stream includes: A pre-established timestamp alignment mechanism is used to perform time consistency calibration on the multi-source running data to obtain a preliminary time-aligned dataset; Based on the initial dataset after time alignment, interpolation methods are used to compensate for missing values to obtain a complete intermediate dataset. For the complete intermediate dataset, a frequency adjustment operation is performed. By using preset sampling rules, data of different frequencies are converted into a unified sampling frequency to obtain a dataset with consistent frequency. Data stream integration technology is used to obtain the synchronous data stream from datasets with consistent frequencies.
3. The method for predicting electrical equipment faults based on adaptive deep reinforcement learning according to claim 1, characterized in that, The synchronous data stream is input into an adaptive feature extraction network based on an attention mechanism, and the high-dimensional fault-sensitive features are output by automatically adjusting the convolutional kernel size, number of channels, and number of attention heads according to the device type and operating stage.
4. The method for predicting electrical equipment faults based on adaptive deep reinforcement learning according to claim 3, characterized in that, Automatically adjusting the kernel size, number of channels, and number of attention heads based on device type and operating phase includes: By using a pre-established equipment type classification library, the operating parameters and historical data records of different equipment are obtained, the equipment types are initially classified, and the basic dataset after equipment classification is obtained. Based on the basic dataset after equipment classification, and combined with the rules for dividing the operation phase, time series segmentation technology is used to group the data according to the operation phase to obtain the data units after phase division. Based on the data units after stage division and the dynamic adjustment requirements of convolutional kernel size, an initial framework for an adaptive network is constructed to obtain convolutional kernel size configuration schemes suitable for different stages. Based on the convolution kernel size configuration scheme, if the number of channels exceeds the preset threshold in the current running stage due to the data complexity, the number of channels is increased to obtain the adjusted channel configuration result. Based on the adjusted channel configuration results, the number of attention heads is adjusted using the allocation rules of the attention mechanism to determine the final number of attention heads configuration; Based on the final configuration of the number of attention heads, the synchronized data stream is mapped to a high-dimensional space.
5. The method for predicting equipment faults based on adaptive deep reinforcement learning according to claim 1, characterized in that, The state space is composed of the high-dimensional fault sensitivity features, equipment health indicators, and environmental disturbances. The action space includes fault warning level output, data acquisition frequency adjustment command, and feature extraction network parameter update range. The instant reward function includes fault prediction accuracy, false alarm / false alarm cost, and communication and computing resource consumption.
6. The method for predicting electrical equipment faults based on adaptive deep reinforcement learning according to claim 5, characterized in that, The acquisition of the state space includes: obtaining multi-dimensional information input based on the high-dimensional fault sensitivity features, health indicators and environmental disturbances, and combining the multi-dimensional information into a unified device state representation through data integration to obtain the state space.
7. The method for predicting electrical equipment faults based on adaptive deep reinforcement learning according to claim 1, characterized in that, Training the fault prediction model using an adaptive learning rate algorithm includes: By using a pre-established training environment, relevant data on changes in state distribution are obtained, and the initial learning rate adjustment range is determined. Based on the initial learning rate adjustment range, the dynamic changes in state distribution are monitored in real time during the commentator network optimization. If the changes in state distribution exceed a preset threshold, the learning rate is dynamically adjusted to obtain an updated learning rate configuration. Based on the updated learning rate configuration, a gradient penalty term is introduced into the actor network training to obtain stable training parameters; Based on stable training parameters, the direction data for policy updates is obtained by adjusting the structure of the deep learning network, and the optimized policy output is obtained. Based on the optimized strategy output, the synchronization update frequency of the commentator network and the actor network is obtained. If the synchronization frequency is lower than the preset threshold, the delayed update mechanism is triggered to determine whether the overall training has reached the preset convergence condition and obtain the trained fault prediction model.
8. The method for predicting equipment faults based on adaptive deep reinforcement learning according to claim 1, characterized in that, The device's health status curve and risk heatmap are displayed in real time through a visual interface, including: Based on the equipment failure risk index, the dynamic change characteristics of the equipment health status are extracted, and a health status curve is generated using a time series processing method to obtain the equipment health status curve. Based on the equipment health status curve, a visualization tool is used to perform spatial distribution mapping of the risk index to obtain the risk heat map.