New energy station video intelligent analysis method based on intelligent question and answer

Abstract

The invention discloses a new energy station video intelligent analysis method based on intelligent questions and answers, and relates to the field of logistics monitoring, and the method comprises the steps: firstly, obtaining data in a station, carrying out the analysis of a data package based on a deep learning algorithm, carrying out the interactive analysis of questions and answers through employing an NLP, and carrying out the detection of a monitoring video of a new energy station based on video analysis, data statistics and archiving are performed based on a video analysis detection result, early warning information is generated based on an overall data analysis result, a deep learning algorithm and a dynamic threshold adjustment mechanism are adopted, abnormal events in a complex scene can be adaptively recognized, and secondly, the method has self-learning and feedback optimization functions, and can be used for monitoring the abnormal events in a complex scene according to historical data and feedback of operation and maintenance personnel. According to the method, model parameters and detection thresholds are dynamically adjusted, so that the accuracy and efficiency of anomaly detection are continuously improved, in addition, multi-dimensional analysis and trend prediction of data are supported, and a generated report not only contains detailed historical data, but also can predict future high-risk time periods and potential equipment faults.

Description

Technical Field

[0001] This invention relates to the field of logistics monitoring, and in particular to a video intelligent analysis method for new energy power stations based on intelligent question answering. Background Technology

[0002] With the continuous development of new energy power stations, video surveillance and data analysis have become crucial for their safety and operation. Currently, video surveillance is commonly used in new energy power stations to monitor key areas in real time. At the same time, sensors collect status data of the station's equipment and environment, such as temperature and wind speed. After preprocessing, this data can provide basic information for the station's operation and maintenance, helping managers understand the equipment's operating status, personnel activity trajectories, and environmental changes within the station. Traditional monitoring methods typically employ anomaly detection based on preset rules, combined with historical data and sensor records, to identify and process abnormal events within the station. This can identify fires, equipment malfunctions, and unauthorized intrusions, and issue alarms when anomalies occur, allowing managers to handle them promptly.

[0003] However, traditional monitoring methods rely too heavily on preset rules and manually set thresholds, making it difficult to cope with complex and ever-changing scenarios and emerging anomaly patterns. In these scenarios, the false alarm rate is high, and it is impossible to accurately identify abnormal events in complex environments.

[0004] Secondly, traditional monitoring methods typically perform data analysis in a static manner, lacking real-time feedback mechanisms and self-learning capabilities. They cannot dynamically optimize the model based on historical data and feedback results, which leads to a decline in detection accuracy and efficiency after long-term operation, making it difficult to meet the increasingly complex monitoring needs of new energy power plants.

[0005] Furthermore, the generated reports are often limited to simple statistical data, lacking in-depth analysis and predictive capabilities, making it difficult to provide effective decision support for managers. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing an intelligent video analysis method for new energy power stations based on intelligent question answering. It aims to utilize deep learning algorithms and a dynamic threshold adjustment mechanism to adaptively identify abnormal events in complex scenarios, reducing false alarms and missed alarms. Secondly, it possesses self-learning and feedback optimization capabilities, dynamically adjusting model parameters and detection thresholds based on historical data and feedback from maintenance personnel, thereby continuously improving the accuracy and efficiency of anomaly detection. Furthermore, it supports multi-dimensional data analysis and trend prediction. The generated reports not only include detailed historical data but also predict future high-risk periods and potential equipment failures, providing managers with precise decision support and improving the overall safety and operational efficiency of the power station.

[0007] Therefore, this application provides a video intelligent analysis method for new energy power stations based on intelligent question answering, including the following steps:

[0008] Step 100: Obtain data from within the facility;

[0009] The data includes surveillance video data, historical video files, and sensor data;

[0010] Step 200: Analyze the data packets based on deep learning algorithms;

[0011] Step 300: Use NLP to perform interactive analysis on the question and answer process;

[0012] Step 400: Based on video analysis, detect the monitoring video of the new energy power station;

[0013] Step 500: Perform data statistics and archiving based on the video analysis and detection results;

[0014] Step 600: Generate early warning information based on the overall data analysis results.

[0015] In some specific implementations, acquiring data within the facility includes:

[0016] Step 100.1: Call data from external devices, including temperature and wind speed sensors, using the API interface;

[0017] Step 100.2: Based on the acquired monitoring video data, historical video files, temperature sensor data, and wind speed sensor data, preprocess the data and convert them into data packets.

[0018] In some specific implementations, data packets are analyzed based on deep learning algorithms, specifically including:

[0019] Step 200.1: Based on the preprocessed data packets, a convolutional neural network is used to extract image features from the surveillance video and identify key targets;

[0020] The extraction process involves multi-layer convolution and pooling operations to extract key visual features and generate feature vectors.

[0021] The key objectives are: personnel, equipment, and environmental anomalies;

[0022] The image features are determined through a specific classification model, which is trained based on historical data.

[0023] Step 200.2: Based on the identified targets, the scenes in the video are tagged and generated.

[0024] The rules for generating the tags are based on the type of target and the detected behavior;

[0025] Among them, the behavior is divided into personnel tags and equipment tags;

[0026] The personnel tags are: normal walking and abnormal activity;

[0027] The equipment labels are: Normal Operation and Fault Warning;

[0028] Step 200.3: After tag generation, perform tag classification processing;

[0029] The tag classification process categorizes tags based on the target category and event type in the scene, specifically including: scene tags, behavior tags, and status tags;

[0030] The scene labels are: outdoor photovoltaic area and wind turbine maintenance area;

[0031] The behavioral tags are: personnel approaching the device and device activation;

[0032] The status labels are: equipment is operating normally and fault alarm is in progress.

[0033] In some specific implementations, NLP is used to perform interactive analysis of question and answer, specifically including:

[0034] Step 300.1: Based on NLP technology, identify language, convert it into structured data, and extract keywords and core meanings;

[0035] The language recognition employs a frequency-cephalic coefficient feature extraction algorithm, specifically:

[0036]

[0037] In the formula, For the nth spectral coefficient, Let i be the energy of the i-th frequency band of the input signal. Number of frequency bands;

[0038] Step 300.2: Use a semantic matching algorithm to convert the user's natural language question into an understandable query statement;

[0039] The semantic matching algorithm is as follows:

[0040]

[0041] In the formula, To search for sentences and descriptions in the database Similarity score between them Cosine similarity between word vectors The importance coefficient of words. For balance coefficient, The distance between words in a sentence;

[0042] Step 300.3.1: Analyze the core entities and intents in the user's question;

[0043] Step 300.3.2: Use an entity-relationship mapping algorithm to identify key elements, specifically:

[0044]

[0045] In the formula, Representing entities and entity The strength of the relationship, For word vector representations of entities, The normalization coefficient is... Parameters used to control the decay rate of the relationship;

[0046] Step 300.4: Based on the parsing results and combined with the generated scene tags, retrieve relevant video clips;

[0047] Step 300.5: After locating the relevant video, the system calculates the device health index to generate the corresponding answer, and combines it with semantic information to provide feedback to the user through speech synthesis or text output.

[0048] The health index of the computing device is specifically as follows:

[0049]

[0050] In the formula, The number of historical data entries. For the equipment The ideal value of a record This is the actual value. Importance weight;

[0051] Step 300.6: Combining knowledge graph reasoning capabilities, construct a knowledge graph from device status, historical data, and sensor data, and use reasoning formulas to calculate potential risks;

[0052] The potential risks of the reasoning are as follows:

[0053]

[0054] In the formula, Indicates the condition The probability of fire occurring under these conditions. For the influence factor weights, The normalization coefficient is... For the relevant condition numbers.

[0055] In some specific implementations, video analytics is used to detect monitoring videos of new energy power plants, specifically including:

[0056] Step 400.1: Identify abnormal events based on intelligent video analysis;

[0057] The abnormal events include: fire, equipment failure, and personnel not operating in accordance with regulations;

[0058] The identification of the abnormal events adopts an adaptive threshold anomaly detection algorithm, specifically as follows:

[0059]

[0060] in, This is the outlier determination coefficient. This is the current detection value. The average of historical data. The standard deviation of historical data;

[0061] The threshold is set to 3.5;

[0062] The when When this occurs, it is considered an abnormal event;

[0063] Step 400.2: Use clustering to compare consecutive video frames in the scene and perform comprehensive analysis by combining multiple parameters;

[0064] The comprehensive analysis includes: temperature, personnel behavior, and equipment status;

[0065] The clustering process consists of three steps;

[0066] First, extract multi-dimensional features from surveillance videos and sensor data;

[0067] Secondly, cluster analysis is performed on the extracted feature data to identify data segments that deviate significantly from historical normal data;

[0068] Thirdly: By comparing historical records, abnormal change patterns are marked and abnormal event results are output;

[0069] Step 400.3.1: When an abnormal event is confirmed, the early warning mechanism triggers an alarm and sends an early warning message to the terminal device of the operation and maintenance personnel;

[0070] The early warning information includes: the time, type, and location of the anomaly, and suggested solutions;

[0071] Step 400.3.2: Early warning events are classified according to their severity and given priority, specifically as follows:

[0072]

[0073] In the formula, Rate the priority of alarms. This is the event risk coefficient, with a threshold range of 1-10; a higher value indicates a greater risk. The time for event detection, in seconds. This is the potential impact coefficient, ranging from 1 to 5, with higher values ​​indicating greater impact. and For custom weight parameters, the commonly used value is ;

[0074] Step 400.4: Based on complex abnormal events, the early warning mechanism combines historical data for reasoning and prediction, and uses the probability of failure occurrence to estimate potential future anomalies, specifically:

[0075]

[0076] In the formula, Indicates the device at time The probability of failure occurring at any time. The fault growth rate is defined as the value range of [value missing]. The time when the equipment temperature began to become abnormal;

[0077] Step 400.5: After handling the abnormal event, the processing result and early warning information are fed back. An adaptive feedback learning algorithm is used to dynamically optimize the anomaly detection threshold. Specifically:

[0078]

[0079] In the formula, The updated detection threshold is [value], the initial threshold is [value]. The learning rate is set to [default value]. For feedback rating, the range is: .

[0080] In some specific implementations, data statistics and archiving are performed based on the video analysis and detection results, specifically including:

[0081] Step 500.1.1: Statistically analyze the frequency of abnormal events, personnel activity trajectories, and equipment operation and maintenance status data;

[0082] The abnormal events mentioned in the statistics include fire alarms, equipment malfunctions, and abnormal personnel behavior;

[0083] Step 500.1.2: Personnel activity trajectory analysis. Based on the personnel position change data in the video, generate trajectory maps and perform cross-analysis in conjunction with abnormal events.

[0084] The trajectory analysis is: whether personnel enter the sensitive area of ​​the equipment;

[0085] Step 500.2: Generate an operation report based on the statistical results;

[0086] The report includes: operating status, equipment health status, and abnormal events;

[0087] The report is divided into daily, weekly, and monthly time periods;

[0088] Step 500.3: Combine big data analysis models to predict the overall operating trend of equipment and sites;

[0089] Step 500.4: Based on the analysis of historical data and current equipment status, generate a trend prediction chart;

[0090] The prediction graph includes: historical operating data of the equipment, and predictions of potential failure times and performance degradation trends of the equipment.

[0091] Step 500.5: Based on the frequency of abnormal events, predict future high-risk periods;

[0092] The high-risk period refers to the time when the equipment has a higher failure rate under specific seasonal and climatic conditions.

[0093] Step 500.6: After the report is generated, it will be sent to administrators and the operations and maintenance team;

[0094] The report is sent via email, SMS, or notification, and the user can choose the appropriate method to receive it.

[0095] In some specific implementations, early warning information is generated based on the overall data analysis results, including:

[0096] Step 600.1: After anomaly detection and early warning, provide feedback on the handling effect, response speed, and detection accuracy of the event;

[0097] The feedback will be recorded in the historical database;

[0098] Step 600.2.1: Dynamically adjust the detection model and threshold based on the evaluation;

[0099] The initial threshold is 3.5;

[0100] Step 600.2.2: When the feedback score is below 5, the threshold is automatically adjusted, specifically as follows:

[0101]

[0102] In the formula, The adjusted detection threshold, The current threshold is 3.5. To adjust the rate, the value range is 0.1. Rate it as feedback;

[0103] Step 600.3: When the feedback score is high, the algorithm will automatically maintain the current threshold;

[0104] Specifically, a weight adjustment method is used to optimize the weights of the abnormal event classification model based on feedback data. For high-frequency events, the model will increase their classification priority.

[0105]

[0106] In the formula, For the updated classification weights, For the original weights, This is the weighting adjustment factor, with a default value of 0.2;

[0107] Step 600.4.1: Analyze the feedback data and optimize the model based on the feedback information;

[0108] Step 600.4.2: Statistically analyze the feedback results and adjust the detection parameters for a certain type of abnormal event when the false alarm rate is high.

[0109] Step 600.5: In special and unexpected abnormal situations, perform manual intervention and adjustment operations based on feedback, and record the manual intervention process;

[0110] Step 600.6: Automatically generate a standardized exception handling process based on historical feedback;

[0111] Step 600.7: Record the feedback and use it for analysis, summarize the standardized exception handling pattern, and form a processing rule base.

[0112] In summary, the intelligent question-and-answer-based video intelligent analysis method for new energy power stations provided in this application firstly acquires data within the power station, analyzes data packets using deep learning algorithms, performs interactive analysis of questions and answers using NLP, detects monitoring videos of the new energy power station based on video analysis, performs data statistics and archiving based on the video analysis detection results, and generates early warning information based on the overall data analysis results. Employing deep learning algorithms and a dynamic threshold adjustment mechanism, it can adaptively identify abnormal events in complex scenarios, reducing false alarms and missed alarms. Secondly, it possesses self-learning and feedback optimization functions, dynamically adjusting model parameters and detection thresholds based on historical data and feedback from maintenance personnel, thereby continuously improving the accuracy and efficiency of anomaly detection. Furthermore, it supports multi-dimensional data analysis and trend prediction; the generated reports not only include detailed historical data but also predict future high-risk periods and potential equipment failures, providing managers with precise decision support and improving the overall safety and operational efficiency of the power station. Attached Figure Description

[0113] Figure 1 This is an overall flowchart of the intelligent video analysis method for new energy power stations based on intelligent question answering provided in the embodiments of this application. Detailed Implementation

[0114] Please refer to Figure 1 It illustrates the flowchart of an embodiment of the intelligent video analysis method for new energy power stations based on intelligent question answering according to the present disclosure.

[0115] like Figure 1 As shown, the intelligent video analysis method for new energy power stations based on intelligent question answering includes the following steps:

[0116] Step 100: Obtain data from within the facility;

[0117] The data includes surveillance video data, historical video files, and sensor data;

[0118] Step 200: Analyze the data packets based on deep learning algorithms;

[0119] Step 300: Use NLP to perform interactive analysis on the question and answer process;

[0120] Step 400: Based on video analysis, detect the monitoring video of the new energy power station;

[0121] Step 500: Perform data statistics and archiving based on the video analysis and detection results;

[0122] Step 600: Generate early warning information based on the overall data analysis results.

[0123] In some specific implementations, acquiring data within the facility includes:

[0124] Step 100.1: Call data from external devices, including temperature and wind speed sensors, using the API interface;

[0125] Step 100.2: Based on the acquired monitoring video data, historical video files, temperature sensor data, and wind speed sensor data, preprocess the data and convert them into data packets.

[0126] In some specific implementations, data packets are analyzed based on deep learning algorithms, specifically including:

[0127] Step 200.1: Based on the preprocessed data packets, a convolutional neural network is used to extract image features from the surveillance video and identify key targets;

[0128] The extraction process involves multi-layer convolution and pooling operations to extract key visual features and generate feature vectors.

[0129] The key objectives are: personnel, equipment, and environmental anomalies;

[0130] The image features are determined through a specific classification model, which is trained based on historical data.

[0131] Step 200.2: Based on the identified targets, the scenes in the video are tagged and generated.

[0132] The rules for generating the tags are based on the type of target and the detected behavior;

[0133] Among them, the behavior is divided into personnel tags and equipment tags;

[0134] The personnel tags are: normal walking and abnormal activity;

[0135] The equipment labels are: Normal Operation and Fault Warning;

[0136] Step 200.3: After tag generation, perform tag classification processing;

[0137] The tag classification process categorizes tags based on the target category and event type in the scene, specifically including: scene tags, behavior tags, and status tags;

[0138] The scene labels are: outdoor photovoltaic area and wind turbine maintenance area;

[0139] The behavioral tags are: personnel approaching the device and device activation;

[0140] The status labels are: equipment is operating normally and fault alarm is in progress.

[0141] In some specific implementations, NLP is used to perform interactive analysis of question and answer, specifically including:

[0142] Step 300.1: Based on NLP technology, identify language, convert it into structured data, and extract keywords and core meanings;

[0143] The language recognition employs a frequency-cephalic coefficient feature extraction algorithm, specifically:

[0144]

[0145] In the formula, For the nth spectral coefficient, Let i be the energy of the i-th frequency band of the input signal. Number of frequency bands;

[0146] Step 300.2: Use a semantic matching algorithm to convert the user's natural language question into an understandable query statement;

[0147] The semantic matching algorithm is as follows:

[0148]

[0149] In the formula, To search for sentences and descriptions in the database Similarity score between them Cosine similarity between word vectors The importance coefficient of words. For balance coefficient, The distance between words in a sentence;

[0150] Step 300.3.1: Analyze the core entities and intents in the user's question;

[0151] Step 300.3.2: Use an entity-relationship mapping algorithm to identify key elements, specifically:

[0152]

[0153] In the formula, Representing entities and entity The strength of the relationship, For word vector representations of entities, The normalization coefficient is... Parameters used to control the decay rate of the relationship;

[0154] Step 300.4: Based on the parsing results and combined with the generated scene tags, retrieve relevant video clips;

[0155] Step 300.5: After locating the relevant video, the system calculates the device health index to generate the corresponding answer, and combines it with semantic information to provide feedback to the user through speech synthesis or text output.

[0156] The health index of the computing device is specifically as follows:

[0157]

[0158] In the formula, The number of historical data entries. For the equipment The ideal value of a record This is the actual value. Importance weight;

[0159] Step 300.6: Combining knowledge graph reasoning capabilities, construct a knowledge graph from device status, historical data, and sensor data, and use reasoning formulas to calculate potential risks;

[0160] The potential risks of the reasoning are as follows:

[0161]

[0162] In the formula, Indicates the condition The probability of fire occurring under these conditions. For the influence factor weights, The normalization coefficient is... For the relevant condition numbers.

[0163] In some specific implementations, video analytics is used to detect monitoring videos of new energy power plants, specifically including:

[0164] Step 400.1: Identify abnormal events based on intelligent video analysis;

[0165] The abnormal events include: fire, equipment failure, and personnel not operating in accordance with regulations;

[0166] The identification of the abnormal events adopts an adaptive threshold anomaly detection algorithm, specifically as follows:

[0167]

[0168] in, This is the outlier determination coefficient. This is the current detection value. The average of historical data. The standard deviation of historical data;

[0169] The threshold is set to 3.5;

[0170] The when When this occurs, it is considered an abnormal event;

[0171] Step 400.2: Use clustering to compare consecutive video frames in the scene and perform comprehensive analysis by combining multiple parameters;

[0172] The comprehensive analysis includes: temperature, personnel behavior, and equipment status;

[0173] The clustering process consists of three steps;

[0174] First, extract multi-dimensional features from surveillance videos and sensor data;

[0175] Secondly, cluster analysis is performed on the extracted feature data to identify data segments that deviate significantly from historical normal data;

[0176] Thirdly: By comparing historical records, abnormal change patterns are marked and abnormal event results are output;

[0177] Step 400.3.1: When an abnormal event is confirmed, the early warning mechanism triggers an alarm and sends an early warning message to the terminal device of the operation and maintenance personnel;

[0178] The early warning information includes: the time, type, and location of the anomaly, and suggested solutions;

[0179] Step 400.3.2: Early warning events are classified according to their severity and given priority, specifically as follows:

[0180]

[0181] In the formula, Rate the priority of alarms. This is the event risk coefficient, with a threshold range of 1-10; a higher value indicates a greater risk. The time for event detection, in seconds. This is the potential impact coefficient, ranging from 1 to 5, with higher values ​​indicating greater impact. and For custom weight parameters, the commonly used value is ;

[0182] Step 400.4: Based on complex abnormal events, the early warning mechanism combines historical data for reasoning and prediction, and uses the probability of failure occurrence to estimate potential future anomalies, specifically:

[0183]

[0184] In the formula, Indicates the device at time The probability of failure occurring at any time. The fault growth rate is defined as the value range of [value missing]. The time when the equipment temperature began to become abnormal;

[0185] Step 400.5: After handling the abnormal event, the processing result and early warning information are fed back. An adaptive feedback learning algorithm is used to dynamically optimize the anomaly detection threshold. Specifically:

[0186]

[0187] In the formula, The updated detection threshold is [value], the initial threshold is [value]. The learning rate is set to [default value]. For feedback rating, the range is: .

[0188] In some specific implementations, data statistics and archiving are performed based on the video analysis and detection results, specifically including:

[0189] Step 500.1.1: Statistically analyze the frequency of abnormal events, personnel activity trajectories, and equipment operation and maintenance status data;

[0190] The abnormal events mentioned in the statistics include fire alarms, equipment malfunctions, and abnormal personnel behavior;

[0191] Step 500.1.2: Personnel activity trajectory analysis. Based on the personnel position change data in the video, generate trajectory maps and perform cross-analysis in conjunction with abnormal events.

[0192] The trajectory analysis is: whether personnel enter the sensitive area of ​​the equipment;

[0193] Step 500.2: Generate an operation report based on the statistical results;

[0194] The report includes: operating status, equipment health status, and abnormal events;

[0195] The report is divided into daily, weekly, and monthly time periods;

[0196] Step 500.3: Combine big data analysis models to predict the overall operating trend of equipment and sites;

[0197] Step 500.4: Based on the analysis of historical data and current equipment status, generate a trend prediction chart;

[0198] The prediction graph includes: historical operating data of the equipment, and predictions of potential failure times and performance degradation trends of the equipment.

[0199] Step 500.5: Based on the frequency of abnormal events, predict future high-risk periods;

[0200] The high-risk period refers to the time when the equipment has a higher failure rate under specific seasonal and climatic conditions.

[0201] Step 500.6: After the report is generated, it will be sent to administrators and the operations and maintenance team;

[0202] The report is sent via email, SMS, or notification, and the user can choose the appropriate method to receive it.

[0203] In some specific implementations, early warning information is generated based on the overall data analysis results, including:

[0204] Step 600.1: After anomaly detection and early warning, provide feedback on the handling effect, response speed, and detection accuracy of the event;

[0205] The feedback will be recorded in the historical database;

[0206] Step 600.2.1: Dynamically adjust the detection model and threshold based on the evaluation;

[0207] The initial threshold is 3.5;

[0208] Step 600.2.2: When the feedback score is below 5, the threshold is automatically adjusted, specifically as follows:

[0209]

[0210] In the formula, The adjusted detection threshold, The current threshold is 3.5. To adjust the rate, the value range is 0.1. Rate it as feedback;

[0211] Step 600.3: When the feedback score is high, the algorithm will automatically maintain the current threshold;

[0212] Specifically, a weight adjustment method is used to optimize the weights of the abnormal event classification model based on feedback data. For high-frequency events, the model will increase their classification priority.

[0213]

[0214] In the formula, For the updated classification weights, For the original weights, This is the weighting adjustment factor, with a default value of 0.2;

[0215] Step 600.4.1: Analyze the feedback data and optimize the model based on the feedback information;

[0216] Step 600.4.2: Statistically analyze the feedback results and adjust the detection parameters for a certain type of abnormal event when the false alarm rate is high.

[0217] Step 600.5: In special and unexpected abnormal situations, perform manual intervention and adjustment operations based on feedback, and record the manual intervention process;

[0218] Step 600.6: Automatically generate a standardized exception handling process based on historical feedback;

[0219] Step 600.7: Record the feedback and use it for analysis, summarize the standardized exception handling pattern, and form a processing rule base.

[0220] In practical applications, the monitoring video data, historical video files, and sensor data within the site are first obtained from external devices via API interfaces. The sensors include temperature sensors and wind speed sensors. All data undergoes preprocessing steps such as video compression, noise reduction, and image enhancement, and is then uniformly converted into data packets. Preprocessing ensures data quality, eliminates duplicate and invalid video clips, and dynamically adjusts image quality based on the current network bandwidth to guarantee the validity and integrity of the data.

[0221] Next, based on deep learning algorithms, the preprocessed data packets are intelligently analyzed, and convolutional neural networks are used to extract image features from the surveillance video. Through multi-layer convolution and pooling operations, corresponding feature vectors are generated.

[0222] Subsequently, key targets in the video were identified, including personnel, equipment, and environmental anomalies. Based on different target types, the identified objects were tagged and corresponding scene tags were generated. Scene tags include personnel behavior tags and equipment status tags.

[0223] After the video intelligent analysis is completed, the intelligent question answering module supports voice or text input queries by operation and maintenance personnel through natural language processing technology. Operation and maintenance personnel can ask questions through this module, which will parse the user's questions and automatically retrieve relevant video clips by combining scene tags and video content, and return the answer.

[0224] For complex abnormal events, reasoning and prediction can be made based on historical data. When the equipment temperature rises abnormally, the historical operating data of the equipment is analyzed to predict the possibility of future failures and issue early warning information to help maintenance personnel take necessary maintenance measures. This process quantifies the status of the equipment by calculating the equipment health index and generates accurate analysis reports.

[0225] The data statistics and report generation module plays a key role in the entire process. It performs real-time statistics on the frequency of abnormal events, personnel activity trajectories, and equipment operation and maintenance status, and generates detailed operation reports. The report content covers the overall operation status of the site, equipment health status, and abnormal event handling records. The reports are automatically generated on a daily, weekly, and monthly basis, and can be sent to relevant management personnel via email or SMS according to user needs.

[0226] Finally, it has self-learning and feedback optimization functions. After each abnormal event is handled, the detection parameters of the model are dynamically adjusted based on the feedback from the operation and maintenance personnel. Through feedback optimization, the abnormal detection threshold will be gradually adjusted according to the actual usage, so as to make more accurate responses to future abnormal events. In addition, the manual handling process of the operation and maintenance personnel will also be recorded and used to optimize the automated handling process to further reduce the need for manual intervention.

Claims

1. A video intelligent analysis method for new energy power stations based on intelligent question answering, characterized in that, Includes the following steps: S100, Obtain data within the station; S200: Analyze data packets based on deep learning algorithms; S300: Uses NLP to perform interactive analysis of question and answer; S400, based on video analysis, detects monitoring videos of new energy power stations; S500 performs data statistics and archiving based on video analysis and detection results; S600 generates early warning information based on overall data analysis results.

2. The intelligent video analysis method for new energy power stations based on intelligent question answering as described in claim 1, characterized in that, Acquiring data within the site, specifically including: S100.1, Based on the API interface, call the data of external device temperature sensor and wind speed sensor; S100.

2. Based on the acquired monitoring video data, historical video files, temperature sensor data, and wind speed sensor data, preprocess the data and convert it into data packets.

3. The intelligent video analysis method for new energy power stations based on intelligent question answering as described in claim 1, characterized in that, Data packet analysis based on deep learning algorithms specifically includes: S200.1 Based on the preprocessed data packets, a convolutional neural network is used to extract image features from the surveillance video and identify key targets; S200.

2. Based on the identified targets, the scenes in the video are tagged and generated. S200.

3. After tag generation, perform tag classification processing; The tag classification process categorizes tags based on the target category and event type in the scene, specifically including: scene tags, behavior tags, and status tags.

4. The intelligent video analysis method for new energy power stations based on intelligent question answering as described in claim 1, characterized in that, Using NLP for interactive analysis of question and answer, specifically including: S300.1: Based on NLP technology, recognize language, convert it into structured data, and extract keywords and core meanings; The language recognition employs a frequency-cephalic coefficient feature extraction algorithm, specifically: ； In the formula, For the nth spectral coefficient, Let i be the energy of the i-th frequency band of the input signal. Number of frequency bands; S300.

2. Use a semantic matching algorithm to convert the user's natural language questions into understandable query statements; The semantic matching algorithm is as follows: ； In the formula, To search for sentences and descriptions in the database Similarity score between them Cosine similarity between word vectors The importance coefficient of words. For balance coefficient, The distance between words in a sentence; S300.3, Analyze the core entities and intents in the user's question; S300.3.

1. Employ an entity-relationship mapping algorithm to identify key elements, specifically: ； In the formula, Representing entities and entity The strength of the relationship, For word vector representations of entities, The normalization coefficient is... Parameters used to control the decay rate of the relationship; S300.

4. Based on the parsing results and combined with the generated scene tags, retrieve relevant video clips; S300.5 After locating the relevant video, the system uses the device health index to generate the corresponding answer and combines it with semantic information to provide feedback to the user through speech synthesis or text output. The health index of the computing device is specifically as follows: ； In the formula, The number of historical data entries. For the equipment The ideal value of a record This is the actual value. Importance weight; S300.

6. Combining knowledge graph reasoning capabilities, construct a knowledge graph from equipment status, historical data, and sensor data, and use reasoning formulas to calculate potential risks; The potential risks of the reasoning are as follows: ； In the formula, Indicates the condition The probability of fire occurring under these conditions. For the influence factor weights, The normalization coefficient is... For the relevant condition numbers.

5. The intelligent video analysis method for new energy power stations based on intelligent question answering according to claim 1, characterized in that, Based on video analytics, the monitoring videos of new energy power plants are inspected, specifically including: S400.1, based on intelligent video analysis, identifies abnormal events; The identification of the abnormal events adopts an adaptive threshold anomaly detection algorithm, specifically as follows: ； in, This is the outlier determination coefficient. This is the current detection value. The average of historical data. The standard deviation of historical data; The threshold is set to 3.5; The when When this occurs, it is considered an abnormal event; S400.

2. Clustering is used to compare consecutive video frames in the scene, and a comprehensive analysis is performed by combining multiple parameters. S400.3 When an abnormal event is confirmed, the early warning mechanism triggers an alarm and sends an early warning message to the terminal device of the operation and maintenance personnel; S400.3.

1. Warning events are classified according to their severity and prioritized accordingly, specifically as follows: ； In the formula, Rate the priority of alarms. This is the event risk coefficient, with a threshold range of 1-10; a higher value indicates a greater risk. The time for event detection, in seconds. This is the potential impact coefficient, ranging from 1 to 5, with higher values ​​indicating greater impact. and For custom weight parameters, the commonly used value is ; S400.

4. Based on complex abnormal events, the early warning mechanism combines historical data for reasoning and prediction, and uses the probability of failure occurrence to estimate potential future anomalies, specifically: ； In the formula, Indicates the device at time The probability of failure occurring at any time. The fault growth rate is defined as the value range of [value missing]. The time when the equipment temperature began to become abnormal; S400.5 After handling abnormal events, the processing results and early warning information are fed back. An adaptive feedback learning algorithm is used to dynamically optimize the anomaly detection threshold. Specifically: ； In the formula, The updated detection threshold is [value], the initial threshold is [value]. The learning rate is set to [default value]. For feedback rating, the range is: .

6. The intelligent video analysis method for new energy power stations based on intelligent question answering according to claim 1, characterized in that, Data statistics and archiving are performed based on video analysis and detection results, specifically including: S500.1 Statistical analysis of the frequency of abnormal events, personnel activity trajectories, and equipment operation and maintenance status data; S500.1.1 Personnel activity trajectory analysis: Based on the personnel position change data in the video, generate trajectory maps and perform cross-analysis in conjunction with abnormal events; S500.

2. An operational report generated based on statistical results; S500.

3. Combine big data analysis models to predict the overall operating trend of equipment and stations; S500.

4. Based on the analysis of historical data and current equipment status, generate a trend prediction chart; S500.5 Predict future high-risk periods based on the frequency of abnormal events; S500.6 After the report is generated, it will be sent to the management personnel and operations team.

7. The intelligent video analysis method for new energy power stations based on intelligent question answering according to claim 1, characterized in that, Early warning information is generated based on the overall data analysis results, specifically including: S600.1 After anomaly detection and early warning, feedback is provided on the handling effect, response speed and detection accuracy of the event; S600.2, Dynamically adjust the detection model and threshold based on evaluation; S600.2.1 When the feedback score is lower than 5, the threshold will be automatically adjusted, specifically as follows: ； In the formula, The adjusted detection threshold, The current threshold is 3.

5. To adjust the rate, the value range is 0.

1. Rate it as feedback; S600.3 When the feedback score is high, the algorithm will automatically maintain the current threshold; Specifically, a weight adjustment method is used to optimize the weights of the abnormal event classification model based on feedback data. For high-frequency events, the model will increase their classification priority. ； In the formula, For the updated classification weights, For the original weights, This is the weighting adjustment factor, with a default value of 0.2; S600.4 Analyze feedback data and optimize the model based on feedback information; S600.4.1 Statistical feedback results, and by analyzing the high false alarm rate of a certain type of abnormal event, the detection parameters of that type of event are adjusted; S600.5 In special and unexpected abnormal situations, manual intervention and adjustment operations shall be carried out based on feedback, and the manual intervention process shall be recorded. S600.6 Automatically generate standardized exception handling procedures based on historical feedback; S600.7 Record feedback and use it for analysis, summarize standardized exception handling patterns, and form a processing rule base.

Images