Electric power inspection scheduling method and system based on artificial intelligence

Abstract

The invention provides a power inspection scheduling method and system based on artificial intelligence, and the method comprises the steps: obtaining a long-term change trend from historical operation data through employing a convolutional neural network for a key feature vector, and calculating a quantitative score of a current health state of a device based on the change trend; comparing the quantitative score with a preset health threshold value, and if the quantitative score exceeds the preset health threshold value, obtaining a fault occurrence probability prediction value from the time sequence data through a long short-term memory network; in combination with the fault occurrence probability prediction value and the real-time monitoring data, a database management system is adopted to update an equipment operation database, and a dynamic state file containing the health state and the fault trend is generated; extracting historical change trend data from the dynamic state file, adjusting an early warning threshold value by using a support vector machine according to the historical change trend data, and determining an update threshold value range adapted to the current operation state.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a power inspection and dispatching method and system based on artificial intelligence. Background Technology

[0002] As the lifeline of modern society, the stability and security of the power system are of paramount importance. With the rapid growth of electricity demand and the increasing complexity of equipment, traditional inspection and dispatch methods are no longer sufficient to meet the requirements of efficient and intelligent management. The introduction of artificial intelligence (AI) technology has provided new possibilities for power inspection and dispatch, becoming a key direction for promoting the intelligent transformation of the power system. However, how to deeply integrate AI with the actual needs of power equipment management remains one of the key areas of current research. Therefore, exploring the application methods and system design of AI in power inspection and dispatch is not only related to equipment operating efficiency but also directly affects the overall reliability of the power system.

[0003] Currently, power line inspections largely rely on manual checks or fixed-cycle equipment monitoring. This approach has significant shortcomings in terms of coverage, real-time performance, and data utilization efficiency. Manual inspections are limited by labor costs and subjective judgment, making it difficult to achieve comprehensive coverage and high-frequency monitoring. While traditional automatic monitoring systems can collect data, they lack the ability to deeply analyze complex operating conditions, leading to delays in fault prediction and early warning responses. These limitations make it difficult for the supervision and management of power equipment to shift from passive maintenance to proactive prevention, exposing the shortcomings of existing methods in adapting to the dynamic needs of modern power systems.

[0004] Against this backdrop, AI-driven inspection and scheduling methods face several core challenges. First, achieving both comprehensiveness and real-time data acquisition is difficult; existing technologies are mostly limited to localized monitoring, lacking a holistic view of equipment operating status. Second, the intelligence level of data analysis is insufficient, making it difficult to dynamically adjust warning thresholds or accurately predict faults based on long-term operational data. These unresolved technical factors prevent the system from forming a complete equipment lifecycle management file, thus limiting the accuracy and reliability of health status assessment and fault prevention. Furthermore, effectively combining wide-angle capture technology with AI analysis remains a critical technical challenge that needs to be overcome.

[0005] Therefore, the key issues of this research are how to deploy wide-angle capture devices in critical locations to achieve comprehensive real-time monitoring, and how to build a database of equipment operation through artificial intelligence analysis to dynamically assess equipment health and predict faults. Solving this problem will directly promote the transformation of power inspection and dispatching from the traditional model to intelligent, preventative management. Summary of the Invention

[0006] This invention provides an artificial intelligence-based power inspection and dispatch method, mainly comprising: Multi-mode sensors are deployed in key locations as wide-angle capture devices to acquire multi-angle data streams including images, temperature, and vibration from the equipment operating environment. Sensor fusion technology is used to integrate the multi-angle data streams to generate a comprehensive raw dataset of the equipment status. Wavelet transform denoising technology is used to process multi-source signals from the original dataset to obtain a denoised dataset. Then, principal component analysis is used to extract key feature vectors reflecting the operating status of the equipment from the denoised dataset. For key feature vectors, a convolutional neural network is used to obtain long-term trends from historical operating data, and a quantitative score of the current health status of the device is calculated based on the trends. The quantitative score is compared with a preset health threshold. If the quantitative score exceeds the preset health threshold, the probability prediction of failure is obtained from the time series data through a long short-term memory network. By combining the predicted failure probability with real-time monitoring data, a database management system is used to update the equipment operation database, generating a dynamic status profile that includes health status and failure trends. Historical trend data is extracted from the dynamic status archive, and the early warning threshold is adjusted based on the historical trend data using a support vector machine to determine the range of updated thresholds that are suitable for the current operating status. Within the update threshold range, abnormal fluctuation characteristics are extracted from real-time monitoring data. If the abnormal fluctuation characteristics exceed the update threshold range, an inspection and scheduling signal for the equipment is generated and transmitted to the management system. The system obtains inspection and scheduling signals from the management system, extracts the lifecycle data of the corresponding equipment from the equipment operation database, and obtains the potential time window for failure to occur from the lifecycle data through time series analysis to generate preventive maintenance time planning data. By fusing preventative maintenance time planning data with real-time status data, and using a distributed computing framework to analyze the health status of multiple devices in parallel from the fused data, health status assessment data for the entire power system is generated.

[0007] This invention also provides an artificial intelligence-based power inspection and dispatch system, mainly comprising: The multi-mode sensor deployment module is used to deploy multi-mode sensors at key locations as wide-angle capture devices to acquire multi-angle data streams including images, temperature, and vibration from the equipment operating environment; The data fusion module is used to integrate multi-angle data streams through sensor fusion technology to generate a comprehensive raw dataset of device status. The denoising and feature extraction module is used to process multi-source signals from the original dataset using wavelet transform denoising technology to obtain a denoised dataset, and then extract key feature vectors reflecting the operating status of the equipment from the denoised dataset through principal component analysis. The health status quantification module is used to extract long-term trends from historical operating data based on key feature vectors using a convolutional neural network, and calculate a quantitative score of the device's current health status based on the trends. The fault prediction module is used to compare the quantitative score with a preset health threshold. If the quantitative score exceeds the preset health threshold, the predicted probability of fault occurrence is obtained from the time series data through the long short-term memory network. The dynamic status profile generation module combines the predicted failure probability with real-time monitoring data, uses the database management system to update the equipment operation database, and generates a dynamic status profile that includes health status and failure trends. The early warning threshold adjustment module is used to extract historical trend data from the dynamic status archive, and use support vector machine to adjust the early warning threshold based on the historical trend data to determine the updated threshold range that adapts to the current operating status. The inspection and scheduling module is used to extract abnormal fluctuation characteristics from real-time monitoring data within the update threshold range. If the abnormal fluctuation characteristics exceed the update threshold range, an inspection and scheduling signal for the equipment is generated and transmitted to the management system. The preventive maintenance planning module is used to obtain inspection and scheduling signals from the management system, extract the life cycle data of the corresponding equipment from the equipment operation database, obtain the potential time window for failure occurrence from the life cycle data through time series analysis, and generate preventive maintenance time planning data.

[0008] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a method for monitoring the health status and preventing the maintenance of power equipment. The method deploys multi-mode sensors at key locations to collect multi-angle data streams of the equipment's operating environment and utilizes sensor fusion technology to generate a comprehensive raw dataset. After signal processing and feature extraction, the invention applies deep learning algorithms to analyze the equipment's health status, predict failure probabilities, and generate dynamic status profiles. Based on historical data, the invention dynamically adjusts early warning thresholds to achieve timely detection of abnormal fluctuations. Combined with lifecycle analysis, the invention formulates preventative maintenance plans and assesses the health status of the entire power system through distributed computing. This method achieves intelligent monitoring, precise early warning, and scientific maintenance of power equipment, effectively improving equipment reliability and system operating efficiency. Attached Figure Description

[0009] Figure 1 This is a flowchart of an artificial intelligence-based power inspection and scheduling method according to the present invention.

[0010] Figure 2 This is a schematic diagram of an artificial intelligence-based power inspection and dispatching method and system according to the present invention.

[0011] Figure 3 This is another schematic diagram of an artificial intelligence-based power inspection and dispatching method and system according to the present invention.

[0012] Figure 4 This is a schematic diagram of the composition of an artificial intelligence-based power inspection and dispatching method and system according to the present invention. Detailed Implementation

[0013] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0014] like Figure 1-4 This embodiment of a power inspection and dispatching method and system based on artificial intelligence may specifically include: S101. Deploy multi-mode sensors at key locations as wide-angle capture devices to acquire multi-angle data streams including images, temperature, and vibration from the equipment operating environment. Integrate the multi-angle data streams through sensor fusion technology to generate a comprehensive raw dataset of the equipment status.

[0015] Raw data streams, including temperature, vibration, and image data, are acquired from sensors. Sensor fusion technology is used to integrate these raw data streams, generating a fused dataset of device status. Features are extracted from the fused dataset to generate a feature set. Temperature data in the feature set is compared with a preset threshold. If the temperature exceeds the preset threshold, a convolutional neural network is used to perform pattern recognition on the image data to identify abnormal regions. Simultaneously, abnormal points are extracted from the vibration data. The location information of abnormal regions and abnormal points is fused to generate descriptive data for the device status. Key indicators are extracted from the descriptive data and classified using a support vector machine to obtain the status assessment result.

[0016] For example, raw data streams are acquired from sensors, typically including temperature, vibration, and image data. This data originates from multi-dimensional monitoring during device operation.

[0017] For example, temperature sensors record heat changes on the surface of a device in real time, vibration sensors capture the amplitude and frequency of mechanical parts, and image sensors capture the appearance or internal state of a device through a camera.

[0018] In one possible implementation, suppose an industrial motor is running, a temperature sensor collects data at 45°C, a vibration sensor records an amplitude of 5 mm / s, and an image sensor generates an image showing slight wear marks on the motor surface. This raw data provides the foundation for subsequent fusion. Sensor fusion technology is used to integrate the raw data streams to generate a fused dataset of the equipment status.

[0019] Specifically, methods such as weighted averaging or Kalman filtering can be used to unify multi-source data onto a single time axis and remove noise.

[0020] For example, temperature and vibration data may reflect the physical state of a motor's operation, while image data provides visual evidence.

[0021] In one embodiment, the fusion process may correlate a temperature of 45°C, an amplitude of 5 mm / s, and wear marks in the image to form a timestamp-integrated cube. The advantage of this fusion technique is that it improves data integrity and consistency, laying a reliable foundation for subsequent analysis. Features are then extracted from the fused dataset to generate a feature set.

[0022] For example, average and peak values ​​can be extracted from temperature data, frequency components and anomalous amplitudes can be extracted from vibration data, and edge features or color changes can be extracted from image data.

[0023] For example, temperature features might be an average of 40°C and a peak of 50°C over 24 hours; vibration features might include a dominant frequency of 60Hz and an abnormal amplitude of 7mm / s; and image features might include a pixel percentage of 15% for the worn area. These feature sets condense key information from the raw data, facilitating further analysis. The temperature data in the feature sets is compared with preset thresholds.

[0024] Preferably, a temperature threshold of 48°C is set. If this value is exceeded, for example, if the temperature reaches 50°C, further in-depth analysis is triggered. This comparison method is simple and efficient, and can quickly screen for potential problems.

[0025] It should be noted that the selection of thresholds is usually based on the equipment's historical operating data or industry standards, and has strong business justification. If the temperature data exceeds the preset threshold, a convolutional neural network is used to perform pattern recognition on the image data to identify abnormal areas.

[0026] In one embodiment, a convolutional neural network can identify the shape and location of wear marks in an image of the motor surface and output the coordinates of abnormal areas, such as "the upper left 20% area." This technique utilizes deep learning to extract spatial features from images, achieving high accuracy and effectively locating problem areas, thus improving diagnostic precision. Simultaneously, abnormal points are extracted based on vibration data.

[0027] For example, time-frequency analysis revealed an anomaly of 7 mm / s amplitude that occurred after 10 hours of operation, with the frequency deviating from the normal range. This anomaly extraction can reflect potential internal faults in the equipment, such as bearing wear or loosening, complementing the abnormal areas in the image. By fusing the location information of the abnormal areas and anomalies, descriptive data about the equipment status is generated.

[0028] Specifically, the wear locations in the image can be correlated with the abnormal vibration time points to form a description: "The upper left corner of the motor surface is worn, and abnormal vibration occurs after 10 hours of operation." This fusion enhances the comprehensiveness of the state description and provides multi-dimensional evidence for fault location. Key indicators are extracted from the descriptive data and classified using a support vector machine to obtain the state assessment results.

[0029] Understandably, key indicators may include the extent of temperature exceeding the limit, the duration of abnormal vibration, and the percentage of wear area.

[0030] For example, given indicators such as "temperature exceeding the standard by 2°C, abnormal vibration for 2 hours, and wear rate of 15%", a support vector machine can classify the equipment status as "minor fault". This classification method combines the advantages of statistical learning, efficiently distinguishing between normal and abnormal states, and providing scientific support for maintenance decisions.

[0031] By extracting key indicators, support vector machines are used to classify the fused dataset. Based on the classification results, the operating status of the equipment is determined, and it is determined whether there are abnormal areas, thus obtaining the final status assessment result.

[0032] Based on the classification results, it is determined whether there are abnormal areas in the equipment's operating status, and the location information of these abnormal areas is recorded as status description data. The location coordinates of the abnormal areas are extracted from the status description data and used as input data for location information. K-means clustering is used to cluster the location coordinates to obtain the distribution pattern of the abnormal areas. If the number of data points in a certain category in the clustering results exceeds a preset threshold, the area of ​​that category is marked as an abnormal area. A logistic regression model is used to verify the stability of the equipment's operating status in the marked abnormal areas, inputting the operating parameters of the abnormal areas and outputting a stability score. The final classification of the equipment's operating status is determined based on the stability score, yielding equipment status assessment data.

[0033] For example, determining whether there are abnormal areas in the operating status of a device based on the classification results usually requires identifying potential problems from multi-dimensional data.

[0034] For example, during the operation of an industrial motor, the classification results show that the temperature in a certain area is consistently high and the vibration is abnormal, indicating that there may be an abnormal area. When the location information of the abnormal area is recorded as state description data, it can be specified as location coordinates.

[0035] For example, a portion of the motor surface might be marked as abnormal, with coordinates such as "top center area, x:50, y:30". This location information is extracted from the condition description data and used as input data for subsequent analysis. Using the K-means clustering algorithm to cluster the location coordinates can reveal the distribution patterns of the abnormal areas.

[0036] In one possible implementation, assume that the motor runs for one revolution and collects multiple abnormal coordinate points, such as "x:50, y:30", "x:52, y:28", "x:48, y:32", etc. K-means clustering divides these points into two classes, one of which is concentrated in the top region and contains 10 data points.

[0037] Preferably, the preset threshold is 8 data points. If a category exceeds this value, such as 10 points in the top region, it is marked as an abnormal region. This method intuitively reflects the central tendency of anomalies through clustering.

[0038] It should be noted that after marking outlier regions, it is particularly important to use a logistic regression model to verify its stability.

[0039] Specifically, the system inputs operating parameters for the abnormal region, such as a temperature of 55°C and a vibration amplitude of 6 mm / s. The logistic regression model analyzes the historical trends and fluctuations of these parameters and outputs a stability score, for example, 0.75 out of 1. The higher the score, the more stable the operating state of the abnormal region.

[0040] In one embodiment, if the score is below 0.6, the area is considered unstable and may require immediate maintenance; if the score is between 0.6 and 0.8, such as 0.75, it indicates that there is some risk but it can still be monitored. This scoring method provides a quantitative basis for condition assessment. The final classification of the equipment's operating status is determined based on the stability score, forming equipment condition assessment data.

[0041] For example, the top abnormal area with a score of 0.75 is ultimately classified as "a state requiring attention".

[0042] Understandably, this classification combines the distribution patterns of cluster analysis with the stability verification of logistic regression, supporting the reliability of the assessment from multiple perspectives.

[0043] In one possible implementation, if another region scores only 0.5 points, with coordinates "x:20, y:10" and 12 data points, it is classified as "fault state". The two results corroborate each other, indicating that the top region has a lower risk, while the other region needs to be addressed first.

[0044] Specifically, K-means clustering has the advantage of extracting patterns from scattered coordinate points, while logistic regression enhances the depth of evaluation through parameter analysis.

[0045] For example, the 10 points in the top region have high concentration and small fluctuations in operating parameters, resulting in a reasonably high stability score; while the 12 points in another region, although concentrated in distribution, show significant parameter anomalies, leading to a lower score. This multi-level analysis ensures the comprehensiveness of anomaly area identification and provides a clear direction for maintenance decisions. Multi-angle verification also reduces the risk of misjudgment and improves the practicality of condition assessment.

[0046] S102. Wavelet transform denoising technology is used to process multi-source signals from the original dataset to obtain a denoised dataset. Then, principal component analysis is used to extract key feature vectors reflecting the operating status of the equipment from the denoised dataset.

[0047] Multi-source signals from the original dataset are acquired, and wavelet transform is performed on the signals using the PyWavelets library to obtain a denoised dataset. From the denoised dataset, principal component analysis (PCA) is used to extract key feature vectors, and their changing trends are calculated using the difference method to obtain at least one operating status parameter. If the operating status parameter exceeds a preset threshold, a bandpass filter is used to adjust the multi-source signals to obtain a corrected signal. Based on the updated feature vectors obtained through PCA-optimized dataset processing, support vector machine (SVM) is used for classification to obtain the equipment operating status assessment result.

[0048] For example, multiple sources of signals are acquired from the raw dataset, typically including various data streams such as the device's temperature, vibration, and sound.

[0049] For example, in the operation monitoring of an industrial pump, a temperature sensor records the pump body surface temperature as 42°C, a vibration sensor captures an amplitude of 4 mm / s, and a sound sensor collects the sound wave signal during operation. These signals form the basis of a multi-source signal. By processing these signals using wavelet transform, noise and valid information can be effectively separated.

[0050] In one possible implementation, wavelet transform decomposes the vibration signal into high-frequency noise and a low-frequency trend component. After removing high-frequency interference, the 4mm / s amplitude signal reflecting the equipment's operating pattern is retained. Similarly, environmental noise in the sound signal is filtered out, preserving the core frequency characteristics of the pump's operation. This denoising process generates a denoised dataset, providing a clearer data foundation for subsequent analysis. Key feature vectors are extracted from the denoised dataset, such as the rate of change over 24 hours from temperature data, the dominant frequency component from vibration data, and the intensity peak from sound data.

[0051] Specifically, the temperature change rate of an industrial pump might be 0.5°C / hour, the dominant vibration frequency is 50Hz, and the peak sound intensity is 70 dB. Calculating the trends in these eigenvectors yields the operating parameters.

[0052] Preferably, the temperature change rate trend shows a gradual increase, the dominant vibration frequency remains stable, and the sound peak fluctuates slightly. If the preset temperature change rate threshold is 0.8°C / hour, and the calculated result is 0.9°C / hour, exceeding the threshold, it indicates a possible anomaly. If the operating status parameters exceed the preset threshold, signal processing methods are used to adjust the multi-source signals and generate a correction signal.

[0053] In one embodiment, for cases where the rate of temperature change exceeds the limit, the temperature signal can be adjusted using a smoothing filter to reduce the impact of sudden noise and obtain a more stable corrected signal.

[0054] It should be noted that this adjustment preserves the signal's trend characteristics while avoiding the risk of misjudgment. Similarly, vibration signals can be enhanced with bandpass filtering to generate a corrected signal that better reflects normal operation. This method improves data usability. The operating status of the equipment is determined based on the updated feature vector obtained by optimizing the dataset through principal component analysis.

[0055] For example, principal component analysis integrates the rate of temperature change, the dominant frequency of vibration, and the peak value of sound into a comprehensive feature vector, which highlights the core operating mode of the equipment.

[0056] Understandably, this optimization reduces redundant information and focuses on key changes.

[0057] In one possible implementation, the updated feature vector of the industrial pump shows anomalies dominated by temperature trends. Combined with stable vibration and sound performance, the condition assessment result may be an "early overheat warning." This judgment logic is clear, provides a basis for maintenance decisions, and improves the accuracy and practicality of monitoring.

[0058] A denoised dataset is obtained by processing multi-source signals through wavelet transform. Key feature vectors are extracted based on principal component analysis, and the changing trend is calculated using the difference method to obtain operating status parameters. If the operating status parameters exceed the preset threshold, a bandpass filter is used to adjust the multi-source signals to obtain a corrected signal. Based on the updated feature vectors, a support vector machine is used for classification to obtain the equipment operating status assessment result.

[0059] After acquiring multi-source signals, wavelet transform is used to decompose high-frequency noise, with a decomposition layer of 5, to obtain a preliminary processed dataset. For this preliminary dataset, time-frequency features are extracted using short-time Fourier transform, and principal component analysis is used to reduce the feature dimension to 10, obtaining a compressed feature vector. Based on the compressed feature vector, a first-order difference method is used to calculate the time series difference to determine the state change parameters. If the state change parameters exceed a preset threshold, a Butterworth bandpass filter is used to filter out abnormal frequency bands from 1Hz to 50Hz, obtaining an optimized signal dataset. Based on the optimized signal dataset, the feature vector is recalculated using short-time Fourier transform to obtain an adjusted feature description vector. A support vector machine is used to perform pattern classification on the adjusted feature description vector, setting the kernel function to radial basis functions, and the model is trained to determine the device state category.

[0060] For example, after acquiring multi-source signals, wavelet transform is used to decompose high-frequency noise, with the number of decomposition layers set to 5, to obtain a preliminary processed dataset.

[0061] For example, in monitoring an industrial pump, the raw signal includes temperature, vibration, and sound data streams. Wavelet transform separates high-frequency noise from the vibration signal through multi-scale decomposition.

[0062] For example, after setting a 5-level decomposition, the high-frequency component may contain environmental interference, while the low-frequency component retains the periodic characteristics of pump operation. This decomposition method can effectively separate different frequency components, providing a clear foundation for subsequent processing. For the initial processed dataset, time-frequency features are extracted using short-time Fourier transform.

[0063] Understandably, the short-time Fourier transform analyzes the signal in segments to generate a time-frequency distribution map.

[0064] Specifically, the vibration signal of an industrial pump may show a stable 50Hz dominant frequency over time, while the sound signal may exhibit brief 100Hz interference peaks. This time-frequency characteristic captures the dynamic changes of the signal. Then, principal component analysis is used to reduce the feature dimension to 10, obtaining a compressed feature vector.

[0065] In one possible implementation, the original features may have 30 dimensions, including temperature slope, vibration amplitude, etc. Principal component analysis retains 90% of the information, and the compressed features highlight key patterns, improving computational efficiency. Based on the compressed feature vector, the time series difference is calculated using the first-order differencing method to determine the state change parameters.

[0066] Preferably, the differential calculation of the dominant vibration frequency may show a change of 0.1 Hz per hour, and the temperature slope difference may be 0.2°C / hour. These parameters reflect subtle fluctuations in the equipment's condition. If the condition change parameters exceed preset thresholds, for example, if the temperature slope threshold is set to 0.5°C / hour but the actual measurement reaches 0.6°C / hour, further processing is required.

[0067] In one embodiment, a Butterworth bandpass filter is used to filter out abnormal frequency bands from 1 Hz to 50 Hz.

[0068] For example, low-frequency ambient noise in the audio signal is filtered out, retaining the core frequency of the pump's operation to generate an optimized signal dataset. This filtering method smooths out abnormal fluctuations and improves data quality. Based on the optimized signal dataset, the feature vector is recalculated using a short-time Fourier transform to obtain an adjusted feature description vector.

[0069] Specifically, the adjusted vibration signal may show a more stable 50Hz dominant frequency, and the interference peaks of the sound signal are reduced. This recalculation ensures the accuracy of the feature vector.

[0070] It should be noted that the adjusted vectors better reflect the true state of the device and avoid being misled by noise. A support vector machine is used to perform pattern classification on the adjusted feature description vectors, with the kernel function set to a radial basis function. After training the model, the device state category is determined.

[0071] In one embodiment, the training data includes feature vectors for both normal operation and overheating states.

[0072] For example, if an industrial pump has a high temperature slope and stable vibration, the model might classify it as "slight overheating." Examining the issue from multiple angles, if the sound peak also rises synchronously, it might strengthen the anomaly assessment; if the vibration deviates, it suggests a more complex problem.

[0073] Understandably, this classification logic is clear, and combined with the characteristics of multi-source signals, it can effectively distinguish state categories, providing a reliable basis for maintenance.

[0074] S103. For key feature vectors, use a convolutional neural network to obtain long-term change trends from historical operating data, and calculate a quantitative score of the current health status of the device based on the change trends.

[0075] Historical operational data of the equipment is acquired, and a convolutional neural network (CNN) is used to extract key feature vectors from the historical data, resulting in a feature vector sequence. The CNN is implemented using the TensorFlow framework and includes three convolutional layers and two fully connected layers. For the feature vector sequence, an ARIMA model is used for time series analysis to obtain long-term trends and determine trend feature values. If the trend feature value exceeds a preset threshold, a current health status score is calculated. The current health status score is calculated using a weighted average method, with weights determined based on the importance of the equipment's operating parameters. Key change points are extracted from the current health status score, and linear regression analysis is used to predict the future health status of the equipment, yielding a predicted trend result.

[0076] S104. The quantitative score is compared with the preset health threshold. If the quantitative score exceeds the preset health threshold, the predicted value of the failure probability is obtained from the time series data through the long short-term memory network.

[0077] Statistical and frequency domain features are extracted from time-series data. These extracted features are then input into a Long Short-Term Memory (LSTM) network. The LTM network processes these features to obtain the time-series trend. The probability distribution value for each time point in the time-series data is calculated and sorted from largest to smallest. The top 5% of probability distribution values ​​are selected as the high-risk probability interval. Based on the time range corresponding to the high-risk probability interval, corresponding time-series segments are extracted, and the Isolation Forest algorithm is used to detect abnormal patterns from these segments. The detected abnormal patterns are then matched against a pre-established fault feature database. If the matching degree exceeds 80%, a fault is identified. Based on the fault identification results, the time periods corresponding to the faults are marked in the time-series data. The labeled data is then used to retrain the LTM network, updating the network parameters.

[0078] S105. Combining the predicted failure probability with real-time monitoring data, the equipment operation database is updated using a database management system to generate a dynamic status file containing health status and failure trends.

[0079] Real-time monitoring data is acquired, and time-series analysis is performed using the ARIMA model to output the equipment operation trend. Based on the time-series analysis results, the equipment failure probability is calculated. If the failure probability exceeds a preset threshold, a predicted value is output. For temporary status files containing equipment health status, key features are extracted as input, and a random forest algorithm is used to output the equipment failure trend. The predicted value is compared with the failure trend; if both exceed the normal range, the real-time monitoring baseline data is updated.

[0080] S106. Extract historical trend data from the dynamic status archive, use support vector machine to adjust the early warning threshold based on the historical trend data, and determine the update threshold range that is suitable for the current operating status.

[0081] Support Vector Machines (SVMs) are used to analyze trend characteristics and obtain classification boundary parameters. If the current data deviates from the classification boundary parameters, the threshold range is updated based on the current data to obtain an adaptive threshold range. The training data for the SVM is adjusted based on the judgment results to obtain optimized classification boundary parameters, which are used for subsequent trend judgment. The SVM analyzes trend characteristics to obtain classification boundary parameters. After the classification boundary parameters are determined, deviation judgment is performed on the current data. If the current data deviates from the classification boundary parameters, the threshold range is recalculated based on the current data to generate an adaptive threshold range. Based on the deviation judgment results, the training data for the SVM is adjusted, and the SVM is retrained to obtain optimized classification boundary parameters. The optimized classification boundary parameters are used in the subsequent trend judgment process.

[0082] S107. Within the update threshold range, extract abnormal fluctuation features from real-time monitoring data. If the abnormal fluctuation features exceed the update threshold range, generate an inspection scheduling signal for the equipment and transmit it to the management system.

[0083] Acquire equipment operating data and extract abnormal fluctuation characteristics from the data. Calculate fluctuation characteristic values ​​using the mean-standard deviation method. If the fluctuation characteristic value exceeds a pre-set update threshold, the equipment is determined to be in an abnormal state. If an abnormal state is determined, an inspection scheduling signal is generated, and the inspection personnel list and inspection time are determined based on the signal. The inspection scheduling signal is sent to the management system, which receives the signal and updates the equipment status record to abnormal.

[0084] S108. Obtain inspection and scheduling signals from the management system, extract the lifecycle data of the corresponding equipment from the equipment operation database, obtain the potential time window for the occurrence of faults from the lifecycle data through time series analysis, and generate preventive maintenance time planning data.

[0085] Based on the device identifier, the device's lifecycle data is retrieved from a pre-established device operation database. At least one operational record is queried using SQL statements to obtain the device's historical state sequence. The historical state sequence undergoes data preprocessing, including converting timestamps to datetime format and handling missing and outlier values. The ARIMA model from Python's statsmodels library is used to perform time series analysis on the preprocessed historical state sequence, with model parameters set to p=1, d=1, and q=1, to calculate the state change trend. A linear regression method is used to calculate the slope of the state sequence to determine the time window for potential faults. If the slope of the state change trend exceeds a preset threshold, a potential fault is identified. Key time nodes are determined based on maintenance planning data, and the time nodes and device state information are encapsulated into a structured time planning dataset in JSON format.

[0086] S109. Integrate preventive maintenance time planning data with real-time status data, and use a distributed computing framework to analyze the health status of multiple devices in parallel from the integrated data to generate health status assessment data for the entire power system.

[0087] Fusion data is acquired from sensor networks and equipment monitoring systems, and the multi-source data is time-aligned and formatted. Real-time status features of at least one device are extracted from the fusion data, including parameters such as voltage, current, temperature, and vibration. A Spark-based parallel analysis method is used to compare the real-time status features with preset normal ranges to determine the device's health status. For the device health status, a Hadoop-based distributed processing method is used to calculate the device's health trend over a past period. If the health trend is below a preset threshold, an abnormal pattern is identified using the Isolation Forest algorithm to obtain a list of abnormal devices. Based on the list of abnormal devices, preventative maintenance strategies are adjusted, generating optimized maintenance plan data, including maintenance time and resource allocation. The fusion data is used to update a TensorFlow-based system evaluation model to calculate the overall health status of the power system, obtaining system evaluation data.

[0088] This invention provides an artificial intelligence-based power inspection and dispatch system, mainly comprising: The multi-mode sensor deployment module is used to deploy multi-mode sensors at key locations as wide-angle capture devices to acquire multi-angle data streams including images, temperature, and vibration from the equipment operating environment; The data fusion module is used to integrate multi-angle data streams through sensor fusion technology to generate a comprehensive raw dataset of device status. The denoising and feature extraction module is used to process multi-source signals from the original dataset using wavelet transform denoising technology to obtain a denoised dataset, and then extract key feature vectors reflecting the operating status of the equipment from the denoised dataset through principal component analysis. The health status quantification module is used to extract long-term trends from historical operating data based on key feature vectors using a convolutional neural network, and calculate a quantitative score of the device's current health status based on the trends. The fault prediction module is used to compare the quantitative score with a preset health threshold. If the quantitative score exceeds the preset health threshold, the predicted probability of fault occurrence is obtained from the time series data through the long short-term memory network. The dynamic status profile generation module combines the predicted failure probability with real-time monitoring data, uses the database management system to update the equipment operation database, and generates a dynamic status profile that includes health status and failure trends. The early warning threshold adjustment module is used to extract historical trend data from the dynamic status archive, and use support vector machine to adjust the early warning threshold based on the historical trend data to determine the updated threshold range that adapts to the current operating status. The inspection and scheduling module is used to extract abnormal fluctuation characteristics from real-time monitoring data within the update threshold range. If the abnormal fluctuation characteristics exceed the update threshold range, an inspection and scheduling signal for the equipment is generated and transmitted to the management system. The preventive maintenance planning module is used to obtain inspection and scheduling signals from the management system, extract the life cycle data of the corresponding equipment from the equipment operation database, obtain the potential time window for failure occurrence from the life cycle data through time series analysis, and generate preventive maintenance time planning data.

[0089] The above description is merely a specific implementation of this specification. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the scope of protection of this specification is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this specification, and these modifications or substitutions should all be covered within the scope of protection of this specification.

Claims

1. A power grid inspection and dispatching method based on artificial intelligence, characterized in that, The method includes: Multi-mode sensors are deployed in key locations as wide-angle capture devices to acquire multi-angle data streams including images, temperature, and vibration from the equipment operating environment. Sensor fusion technology is used to integrate the multi-angle data streams to generate a comprehensive raw dataset of the equipment status. Wavelet transform denoising technology is used to process multi-source signals from the original dataset to obtain a denoised dataset. Then, principal component analysis is used to extract key feature vectors reflecting the operating status of the equipment from the denoised dataset. For key feature vectors, a convolutional neural network is used to obtain long-term trends from historical operating data, and a quantitative score of the current health status of the device is calculated based on the trends. The quantitative score is compared with a preset health threshold. If the quantitative score exceeds the preset health threshold, the probability prediction of failure is obtained from the time series data through a long short-term memory network. By combining the predicted failure probability with real-time monitoring data, a database management system is used to update the equipment operation database, generating a dynamic status profile that includes health status and failure trends. Historical trend data is extracted from the dynamic status archive, and the early warning threshold is adjusted based on the historical trend data using a support vector machine to determine the range of updated thresholds that are suitable for the current operating status. Within the update threshold range, abnormal fluctuation characteristics are extracted from real-time monitoring data. If the abnormal fluctuation characteristics exceed the update threshold range, an inspection and scheduling signal for the equipment is generated and transmitted to the management system. The system obtains inspection and scheduling signals from the management system, extracts the lifecycle data of the corresponding equipment from the equipment operation database, and obtains the potential time window for failure to occur from the lifecycle data through time series analysis to generate preventive maintenance time planning data. By fusing preventative maintenance time planning data with real-time status data, and using a distributed computing framework to analyze the health status of multiple devices in parallel from the fused data, health status assessment data for the entire power system is generated.

2. The method according to claim 1, characterized in that, The deployment of multi-mode sensors at key locations as wide-angle capture devices acquires multi-angle data streams including images, temperature, and vibration from the equipment's operating environment. Sensor fusion technology is then used to integrate these multi-angle data streams, generating a comprehensive raw dataset of the equipment's status, including: Obtain raw acquisition streams from sensors, including temperature, vibration, and image data; Sensor fusion technology is used to integrate the original data acquisition streams and generate a fused dataset of device status. Extract features from the fused dataset to generate a feature set; The temperature data in the feature set is compared with a preset threshold. If the temperature data exceeds the preset threshold, a convolutional neural network is used to perform pattern recognition on the image data to identify abnormal areas. At the same time, abnormal points are extracted based on vibration data; The location information of abnormal areas and abnormal points is fused to generate descriptive data of the device status; Key metrics are extracted from the descriptive data, and a support vector machine is used for classification to obtain the state evaluation results.

3. The method according to claim 2, characterized in that, Also includes: By extracting key indicators, a support vector machine is used to classify the fused dataset. Based on the classification results, the operating status of the equipment is determined, and the existence of abnormal areas is identified, resulting in a final status assessment. This assessment includes: Based on the classification results, determine whether there are abnormal areas in the equipment's operating status, and record the location information of the abnormal areas as status description data; Extract the location coordinates of the abnormal area from the status description data and use them as input data for location information; K-means clustering algorithm was used to cluster the location coordinates to obtain the distribution pattern of the abnormal area; If the number of data points in a certain category in the clustering results exceeds a preset threshold, then the region of that category is marked as an abnormal region. The stability of the equipment operating status marked as abnormal areas is verified by using a logistic regression model. The operating parameters of the abnormal areas are input, and the stability score is output. The final classification of the equipment's operating status is determined based on the stability score, thus obtaining equipment status assessment data.

4. The method according to claim 1, characterized in that, The process involves processing multi-source signals from the original dataset using wavelet transform denoising technology to obtain a denoised dataset, and then extracting key feature vectors reflecting the equipment's operating status from the denoised dataset using principal component analysis, including: The multi-source signals in the original dataset are obtained, and the PyWavelets library is used to perform wavelet transform on the multi-source signals to obtain a denoised dataset. From the denoised dataset, key feature vectors are extracted using principal component analysis, and their changing trends are calculated using the difference method to obtain at least one running state parameter; If the operating status parameters exceed the preset threshold, a bandpass filter is used to adjust the multi-source signal to obtain a corrected signal; Based on the updated feature vectors obtained by optimizing the dataset through principal component analysis, support vector machines are used for classification to obtain the equipment operating status assessment results.

5. The method according to claim 4, characterized in that, Also includes: A denoised dataset is obtained by processing multi-source signals using wavelet transform. Key feature vectors are extracted based on principal component analysis, and the changing trend is calculated using the difference method to obtain operating status parameters. If the operating status parameters exceed a preset threshold, a bandpass filter is used to adjust the multi-source signals to obtain a corrected signal. Based on the updated feature vectors, a support vector machine is used for classification to obtain the equipment operating status assessment result, which specifically includes: After acquiring the multi-source signals, wavelet transform was used to decompose the high-frequency noise. The number of decomposition layers was set to 5 to obtain the preliminary processed dataset. For the aforementioned preliminary processed dataset, time-frequency features are extracted by short-time Fourier transform, and the feature dimension is reduced to 10 dimensions by principal component analysis to obtain compressed feature vectors. Based on the compressed feature vector, the time series difference is calculated using the first-order difference method to determine the state change parameters; If the state change parameters exceed the preset threshold, a Butterworth bandpass filter is used to filter out the abnormal frequency band from 1Hz to 50Hz to obtain an optimized signal dataset. Based on the optimized signal dataset, the feature vector is recalculated using short-time Fourier transform to obtain the adjusted feature description vector; The adjusted feature description vectors are classified using a support vector machine, with the kernel function set as a radial basis function. After training the model, the device state category is determined.

6. The method according to claim 1, characterized in that, The process involves using a convolutional neural network to extract long-term trends from historical operational data for key feature vectors, and then calculating a quantitative score for the current health status of the device based on these trends. This includes: The system acquires historical operating data of the equipment and uses a convolutional neural network to extract key feature vectors from the historical operating data, resulting in a sequence of feature vectors. Convolutional neural networks are implemented using the TensorFlow framework and consist of three convolutional layers and two fully connected layers. For the feature vector sequence, the ARIMA model is used for time series analysis to obtain the long-term trend and determine the trend feature value; If the trend characteristic value exceeds the preset threshold, the current health status score of the device is calculated. The current health status score of the equipment is calculated using a weighted average method, with the weights determined based on the importance of the equipment's operating parameters. Key change points are extracted from the current health status score of the equipment, and the future health status of the equipment is predicted through linear regression analysis to obtain the prediction trend results.

7. The method according to claim 1, characterized in that, The step involves comparing the quantified score with a preset health threshold. If the quantified score exceeds the preset health threshold, a predicted probability of failure is obtained from the time-series data using a long short-term memory network. This includes: Statistical and frequency domain features are extracted from time series data. The extracted features are then input into a long short-term memory network. After processing by the long short-term memory network, the changing trend of the time series is obtained. Calculate the probability distribution value for each time point in the time series data, sort the probability distribution values ​​from largest to smallest, and select the top 5% of the probability distribution values ​​as the high-risk probability interval; Based on the time range corresponding to the high-risk probability interval, extract the corresponding time series segments, and use the Isolation Forest algorithm to detect abnormal patterns from the time series segments; The detected abnormal patterns are matched with a pre-established fault feature library. If the matching degree exceeds 80%, it is determined to be a fault. Based on the fault determination results, the time period corresponding to the fault is marked in the time series data, and the Long Short-Term Memory Network is retrained using the marked data to update the network parameters.

8. The method according to claim 1, characterized in that, The method combines the predicted failure probability with real-time monitoring data, and uses a database management system to update the equipment operation database, generating a dynamic status file containing health status and failure trends, including: Acquire real-time monitoring data, perform time series analysis on the monitoring data using the ARIMA model, and output the trend of equipment operation changes; Based on the time series analysis results, the probability of equipment failure is calculated. If the probability of failure exceeds the preset threshold, the predicted value is output. For temporary status files containing device health status, key features are extracted as input, and the random forest algorithm is used to output device failure trends. The predicted value is compared with the fault trend. If both are outside the normal range, the baseline data of real-time monitoring is updated.

9. The method according to claim 1, characterized in that, The process of extracting historical trend data from dynamic status archives, using a support vector machine to adjust early warning thresholds based on this historical trend data, and determining an update threshold range adapted to the current operating status includes: Support vector machines are used to analyze the changing trend features to obtain classification boundary parameters; If the current running data deviates from the classification boundary parameter, the threshold range is updated based on the current running data to obtain the adaptive threshold range; The training data of the support vector machine is adjusted based on the judgment results to obtain optimized classification boundary parameters, which are used for subsequent trend judgment. Support vector machines are used to analyze changing trend features and obtain classification boundary parameters. After the classification boundary parameters are determined, deviation judgment is performed on the current running data; If the current running data deviates from the classification boundary parameters, the threshold range is recalculated based on the current running data to generate an adaptive threshold range; Based on the deviation judgment results, the training data of the support vector machine is adjusted, the support vector machine is retrained, and the optimized classification boundary parameters are obtained. The optimized classification boundary parameters are used in the subsequent trend judgment process.

10. A power inspection and dispatch system based on artificial intelligence, characterized in that, The system includes: The multi-mode sensor deployment module is used to deploy multi-mode sensors at key locations as wide-angle capture devices to acquire multi-angle data streams including images, temperature, and vibration from the equipment operating environment; The data fusion module is used to integrate multi-angle data streams through sensor fusion technology to generate a comprehensive raw dataset of device status. The denoising and feature extraction module is used to process multi-source signals from the original dataset using wavelet transform denoising technology to obtain a denoised dataset, and then extract key feature vectors reflecting the operating status of the equipment from the denoised dataset through principal component analysis. The health status quantification module is used to extract long-term trends from historical operating data based on key feature vectors using a convolutional neural network, and calculate a quantitative score of the device's current health status based on the trends. The fault prediction module is used to compare the quantitative score with a preset health threshold. If the quantitative score exceeds the preset health threshold, the predicted probability of fault occurrence is obtained from the time series data through the long short-term memory network. The dynamic status profile generation module combines the predicted failure probability with real-time monitoring data, uses the database management system to update the equipment operation database, and generates a dynamic status profile that includes health status and failure trends. The early warning threshold adjustment module is used to extract historical trend data from the dynamic status archive, and use support vector machine to adjust the early warning threshold based on the historical trend data to determine the updated threshold range that adapts to the current operating status. The inspection and scheduling module is used to extract abnormal fluctuation characteristics from real-time monitoring data within the update threshold range. If the abnormal fluctuation characteristics exceed the update threshold range, an inspection and scheduling signal for the equipment is generated and transmitted to the management system. The preventive maintenance planning module is used to obtain inspection and scheduling signals from the management system, extract the life cycle data of the corresponding equipment from the equipment operation database, obtain the potential time window for failure occurrence from the life cycle data through time series analysis, and generate preventive maintenance time planning data.

Images