A method for processing dynamic operation control data based on power operation recorder

Abstract

The present invention discloses a method for processing data of dynamic operation control based on an electric power operation recorder, including integrated adaptation and development of an electric power operation recorder, construction of an electric power behavior sample database, construction of an electric power operation behavior recognition algorithm, and optimization of the electric power operation behavior recognition algorithm based on the independently developed a-combined data constraint process and b-path cross data hierarchical decomposition process. On the one hand, the present invention integrates and integrates the existing technologies, especially carries out targeted research and development of multiple independent algorithms to optimize the directly integrated algorithm platform and artificial intelligence platform for electric power operation behavior recognition.

Description

Technical Field

[0001] The present invention relates to the field of intelligent and information-based power grid technology, and in particular to a method for processing dynamic operation control data based on an electric power operation recorder. Background Art

[0002] At present, the technology of power operation recorders and all-in-one machines has developed to a relatively mature stage. After the start-up period (2008-2010, the first generation of audio and video recording devices were first used on a large scale in the public security system) and the growth period (2011-2015, the audio and video information collection, transmission, storage and management gradually formed a system, and the industry standard GA / T947-2001 was implemented), the application field of law enforcement recorders has gradually expanded. It has now entered the standardized development period (2016 to date, the industry standard GA / T947-2015 has been implemented), product demand has become diversified, the application field has been further enriched, the software and hardware have developed in coordination, and the information technology has gradually matured. In the past two years, with the rapid development of mobile Internet and mobile media, law enforcement recorders and various types of operation recorders have entered a blowout development stage. However, the research on the charging and debugging interface of recorder equipment at home and abroad focuses on the charger power and battery capacity. According to the different interfaces of recorders from different manufacturers, it is necessary to develop special equipment to charge multiple recorders at the same time and collect the historical data of the recorders to realize automatic collection, transmission and archiving.

[0003] Regarding intelligent recognition technology, it has been applied in the power industry, where images are processed and analyzed by computer image processing and other technologies to extract the features of targets and objects in the image, thereby determining the different target and object areas in the image; image discrimination refers to the use of computer image processing and analysis technologies to extract the difference features of the same target and object in multiple images of the same monitoring scene, the same target and object at different times, and find the target and object areas when there are differences. Domestic and foreign research has been mainly focused on the image transmission recognition of recorders, but not much research and implementation has been conducted in conjunction with the on-site application of power operation recorders. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for processing dynamic operation control data based on an electric power operation recorder. On the basis of integrating and integrating the existing technologies, further targeted independent algorithm development is carried out to specifically optimize the efficiency of electric power operation behavior recognition.

[0005] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows.

[0006] A method for processing dynamic operation control data based on an electric power operation recorder, the method comprising at least the following sub-items: integrated adaptation and development of the electric power operation recorder, construction of an electric power behavior sample database, and construction and / or optimization of an electric power operation behavior recognition algorithm.

[0007] As a preferred technical solution of the present invention, the integrated adaptation development of the power operation recorder includes: building an all-in-one recorder for power operation, the screen of the all-in-one recorder adopts an LCD screen, supports multi-point capacitive touch, the chip adopts an Intel series processor, supports SSD+high-capacity HDD flexible expansion, the acquisition interface supports Mini USB / Micro / typeC, and the software management system adopts a B / S architecture, supports distributed download, supports remote download of recorder data, automatic collection, automatic deletion, automatic charging, unattended, and a full metal shell to prevent strong electric field and strong magnetic field interference. It also supports 7*24*365 all-weather operation, supports real-time access and parameter configuration through a Web browser, and realizes standard SDK development and support; the parameter configuration of the all-in-one recorder for power operation is as follows: Processor: 8-core high-speed processor; System: Android operating system; Communication: 4G full network access and 5G support; Call: support two-way interaction, real-time intercom, and cluster intercom functions; Video compression: It is suitable to adopt Use H.264 / H.265 high-definition encoding; storage capacity: 64G or 128G; battery: capacity greater than 3000mAh, battery life more than 10 hours, low-temperature battery; with uninterrupted battery replacement function, preferably with ultra-low power consumption recording mode; positioning: GPS+Beidou dual positioning, preferably with positioning accuracy of <1m; lens width: 120-degree wide-angle; night vision capability: 5 meters to see facial features, 10 meters to see human body outlines; weight: ≤200g; protection performance: IP6-8; AI function: support basic face recognition functions; anti-shake: intelligent anti-shake; screen: 2 inches or more, touch operation; others: optional WiFi, Bluetooth function, one-key SOS function.

[0008] As a preferred technical solution of the present invention, the construction of the power behavior sample database includes: establishing a power behavior sample database library based on collecting diversified power operation data. The database contains different types of operations and tasks, covering different industries and industry specifications, ensuring that the sample database is diverse and can be generalized to multiple power operation tasks; the construction sub-items include at least: 1) Sample annotation: the data in the sample database has correct annotations to ensure that the subsequent algorithm model learns meaningful features; 2) Data quality: Data quality is constructed as the underlying factor of an effective sample database. The data in the sample database is accurate, complete and reliable. In order to ensure data quality, Various measures include data cleaning, data deduplication, and data verification; 3) Collection strategy: achieve maximum coverage of possible scenarios when collecting data; 4) Data protection: data in the sample database is protected from leakage and abuse, and the data path at least includes data storage security, access control, and data encryption; The application-oriented construction of the power behavior sample database includes: in response to the safety requirements of power production site operations, the data sub-items of the power behavior sample database include at least 7 types of intelligent recognition algorithm sample data: workers wearing safety helmets, wearing safety belts, wearing insulating gloves, smoking, welding operations not in accordance with labor protection requirements, not wearing masks in dusty operations, throwing objects from high places, and illegal fire.

[0009] As a preferred technical solution of the present invention, the construction of the power operation behavior recognition algorithm includes:

[0010] A. Framework model for behavior analysis of power workers: The video stream obtained by the recorder camera is abstracted into quantifiable behavior data using a deep learning algorithm, including the following data steps: (1) obtaining an image of known behavior type; (2) performing convolution and maximum pooling calculations on the image, and extracting feature data used to represent image features; (3) inputting the feature data of the image and the corresponding behavior type into a pre-built neural network model; (4) classifying the feature data according to the corresponding behavior type; (5) repeating steps (1)-(4) until the training end condition is met;

[0011] B. Data flow of constructing the behavior analysis model of power workers: The behavior analysis model of power workers is constructed using a convolutional neural network. The steps of constructing the model are as follows: A. Forward propagation process: The forward propagation process of the constructed convolutional neural network specifically includes: (1) Padding: When performing a convolution operation on the input image, the pixels at the corners or edges are used relatively less frequently, which will weaken the edge information in image recognition. Therefore, a padding operation is used to fill p layers of data around the original image data, retaining valid information while ensuring that the height and width of the convolution calculation do not change before and after the convolution calculation; (2) Single-step convolution: In the convolution operation, the number of filter channels is Same as the input image, the height and width are odd numbers and are constructed as a multidimensional array; (3) Convolution layer: In the single-step convolution process, there is only one filter. In the CNN convolution layer, multiple filters are constructed. When outputting, the images after convolution of each filter are superimposed together; (4) Pooling layer: The role of the pooling layer is to reduce the size of the network, increase the calculation speed, and at the same time improve the robustness of the extracted features. It uses maximum pooling max-pooling and / or average pooling average-pooling; B. Back propagation process: In the case of a deep learning framework, as long as the forward propagation process is guaranteed, the framework will automatically perform the back propagation process.

[0012] A data optimization method optimizes an electric power operation behavior recognition algorithm based on a self-developed a-combination data constraint process and / or b-path cross-data hierarchical decomposition process.

[0013] As a preferred technical solution of the present invention, the a-combined data constraint process is specifically as follows: by constructing data constraint conditions, the video data collected by the power operation recorder is pre-classified to improve the data recognition algorithm's dynamic, accurate and continuous recognition of the same subject or object in the power operation site, wherein the data constraint conditions include independent and cross-limited event constraints and data constraints, wherein the constraint connotation of the event constraint conditions is based on the relevant subject and object factors implied by the power operation event itself, and these power operation behavior subject personnel and various object equipment associated with the current detected power event are limited to the defined monitoring scope of the constraint conditions. The data is read, recorded and multi-linearly processed within the scope; the constraint connotation of the data constraint condition is mainly based on the data format and data value limit conditions matched when the multi-dimensional components of spatio-temporal events are allocated to each power operation subject and object in the multi-linear dimension in the subsequent data processing process, and the power operation subject and object detection data that do not conform to the data format and exceed the upper and lower limits of the allowable limit will be eliminated or corrected in advance; the event constraint conditions and data constraint conditions of the constraint conditions are used together, and all related detection subjects and objects in the target power operation event that are compatible with subsequent multi-dimensional component allocation and multi-linear data processing are all defined as within the scope of the current data processing.

[0014] As a preferred technical solution of the present invention, event-based constraints and data-based constraints are independent and compatible with each other and can be cross-applied in no particular order; the optimized application process is to first perform event-based constraints, and define the subject and object associated with the current target power operation event within the data scope; secondly, apply the data-based constraints, and this data application process includes two optional paths: ① When the framework and data range of subsequent data processing can be defined in advance, the corresponding data-based constraints can be defined in advance through reverse deduction, and at this time, the constraints are applied before or at the same time as the event-based constraints; ② If the framework and data range of subsequent data processing cannot be accurately defined in advance, the data-based constraints are applied in the data process, that is, the data constraints are initially set to infinity, and when the data format does not match or the data value exceeds the limit when the multi-dimensional components of spatiotemporal events are allocated to each power operation subject and object in the multi-linear dimension, the data constraints are applied to the associated data to eliminate the error data; in the behavior recognition scenario at the real power operation site, the incidence of path ② is higher than that of path ①;

[0015] As a preferred technical solution of the present invention, the hierarchical decomposition process of the b-path cross data is as follows:

[0016] In general, an executable database group compatible with the four arithmetic operations is constructed. This database group contains multiple micro sub-data groups, and each sub-data group has independent attributes and interactive attributes based on the four arithmetic operations. The independent or mutual data operation results of each micro sub-data group in the executable database group are used as the decision basis for the dynamic trajectory tracking algorithm of the power operation site personnel subject or device object; the above database group adopts the bottom-level local self-organization to form a high-level overall data organization configuration. The specific data process includes:

[0017] ① For any power operation site personnel subject or device object, a discrete micro-database with a unique label is constructed respectively. The spatial coordinate data (two-dimensional topological coordinate data of three-dimensional physical space coordinate data) of the power operation site personnel subject or device object under the selected time interval Δτ is stored in this discrete micro-database. For example, under the unique label ak-Δτ, the data phenotype in this discrete micro-database is (ak-c0, ak-c 1Δτ ,ak-c 2Δτ ,ak-c 3Δτ , ...); for different discrete micro-databases, ak corresponds to a natural number value sequence (k = 1, 2, 3, ...); and (c0, c 1Δτ , c 2Δτ , c 3Δτ, …) sequences usually have global invariance. This is based on the convenience of data interaction of the micro-sub-databases within the database group that are compatible with the four arithmetic operations. On the other hand, the data sources of different discrete micro-databases are all the same hardware system and its data acquisition and storage algorithm, that is, the power operation recorder that has been integrated and adapted and developed before, so that the obtained data itself has natural coordination, which is consistent with the different types of c- Δτ The sequence is significantly improved.

[0018] ② For the executable database group constructed above, take its local data to construct an executable sub-database group under the premise of maintaining the original data configuration, where maintaining the original data configuration means that the sub-database group and its internal micro-sub-database adopt the same data labeling logic as the original executable database group, each micro-sub-database adopts a discrete data sequence model, and each micro-sub-database and different micro-sub-data have their own independence and data interactivity compatible with four arithmetic operations; where local data refers to the geometric intersection of two or more video image elements and the dynamic trajectory of the power operation site personnel subject or device object represented by them on the two-dimensional topological surface of the three-dimensional physical space, and the local data is the data sequence distributed near this intersection. Specifically, the number of local data sequences is limited to between (2-200), generally taking the number 5, then a corresponding executable database group containing k micro-sub-databases is obtained, where each micro-sub-database contains five data units represented as vector data pairs, and the data labeling model is represented as: (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ ), where k corresponds to a natural number value sequence k = 1, 2, 3, ...; and the data intersection point corresponds to the data point ak-c p+2Δτ or ak-c p+2Δτ With ak-c p+1Δτ ,ak-c p+3Δτ between;

[0019] ③Finally, a decision-making algorithm based on data verification is constructed. For any (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ) discrete data sequence, note that each data corresponds to a two-dimensional coordinate data, that is, vector data. Vector data has dual attributes of size and direction. The key is that in order to adapt to the subsequent data execution algorithm, its size attribute should be eliminated. To this end, the two-dimensional coordinate data pair and its corresponding vector data are first represented by complex numbers based on the Euler formula, that is, any discrete data unit is represented as ak-c p+mΔτ Characterized by e l+iθ form, where e l Corresponding vector length, e iθ Characterize the angle data of the data pair; in order to eliminate the vector size attribute data with interference attributes, only e is used in the subsequent data execution process iθ ; finally, an executable algorithm is constructed for the angle-characterized data item, and its construction standard is: based on the four arithmetic operations, the numerical value of the operation result is linearly related to (positively correlated with or negatively correlated with) the data mismatch probability of the executable database group or the micro-sub-database in the executable sub-database group, thereby completing the tracking decision of the dynamic data intersection; the executable algorithm that meets this data standard can have many options, including: ③-ω, for the data sequence (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ ) derived from the five groups of angle characterization data items, firstly, the first two and the last two are added, then the two data are subtracted and the absolute value is taken. The size of the obtained absolute value is positively correlated with the data mismatch probability of the micro-sub-data in the executable database group or the executable sub-database group; ③-ψ, for the data sequence (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ ) are derived from five groups of angle representation data items, with the first or last digit as the subtrahend, and the remaining four groups of data as the minuends in turn. Four numerical differences are constructed by subtraction operation and their absolute values ​​are taken respectively. Finally, the four positive numbers are added together. The size of the sum is positively correlated with the data mismatch probability of the micro-sub-data in the executable database group or the executable sub-database group.

[0020] As a preferred technical solution of the present invention, the following optional data processes are also included: ④ For two or more groups of local data sequences containing five data bits distributed according to intersection points, the median data, i.e., the third bit, is kept unchanged, and the data before and after the median data are cross-interchanged to construct a new local data sequence, and an executable algorithm that meets the data standard, such as ③-ω or ③-ψ, is applied to the obtained local data sequence, and the mismatch probability obtained by the data sequences before and after the cross-interchange under the executable algorithm is compared, so as to more accurately complete the tracking decision of dynamic data intersections.

[0021] As a preferred technical solution of the present invention, in step ③, the Euler form complex representation of the two-dimensional coordinate data and the elimination of interfering data sub-items are allowed to be completed before applying executable algorithms such as ③-ω, ③-ψ, etc., or to be integrated into the executable algorithm afterwards.

[0022] The beneficial effect of adopting the above technical solution is that this study automatically collects the situation of the operation site through the power operation recorder, aiming to identify and alarm the image data of personnel behavior during the operation based on image recognition and discrimination technology. Specifically, on the one hand, the present invention integrates and integrates the existing technology, such as the design of the all-in-one power-specific recorder, and in the video and image data recognition, the convolutional neural network and image recognition algorithm developed by the Hinton team are adopted, as well as the integration and access to the existing neural network platform such as TensorFlow, etc. On the other hand, a number of independent algorithm research and development are carried out in a targeted manner to optimize the directly integrated algorithm platform and artificial intelligence platform for the recognition of power operation behavior. DETAILED DESCRIPTION

[0023] The following examples illustrate the present invention in detail. Various raw materials and equipment used in the present invention are conventional commercial products, which can be directly obtained by market purchase. In the description of the following embodiments, specific details such as specific system structures, technologies, etc. are proposed for illustration rather than limitation, so as to thoroughly understand the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits and methods are omitted to prevent unnecessary details from hindering the description of the present application. It should be understood that when used in the present application specification and the attached claims, the term "including" indicates the presence of the described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their collections. It should also be understood that the term "and / or" used in the present application specification and the attached claims refers to any combination of one or more of the items listed in association and all possible combinations, and includes these combinations. As used in the specification of the present application and the appended claims, the term "if" can be interpreted as "when..." or "once" or "in response to determination" or "in response to detection" according to the context. Similarly, the phrase "if it is determined" or "if [the described condition or event] is detected" can be interpreted as meaning "once it is determined" or "in response to determination" or "once [the described condition or event] is detected" or "in response to detection [the described condition or event]" according to the context. In addition, in the description of the specification of the present application and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance. The reference "one embodiment" or "some embodiments" described in the specification of the present application means that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Thus, the sentences "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in the specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "include", "comprising", "having" and variations thereof all mean "including but not limited to", unless specifically emphasized otherwise.

[0024] Embodiment 1, recorder and integrated machine

[0025] The all-in-one screen uses an LCD screen and supports multi-point capacitive touch. The chip uses an Intel series processor and supports flexible expansion of SSD+high-capacity HDD. The acquisition interface supports Mini USB / Micro / typeC. In terms of software, the management system adopts a B / S architecture, supports distributed downloading, supports remote downloading of recorder data, automatic collection, automatic deletion, automatic charging, unattended operation, and a full metal shell to prevent interference from strong electric and magnetic fields. It also supports 7*24*365 all-weather operation, supports real-time access and parameter configuration through a Web browser, and realizes standard SDK development and support.

[0026] The recommended parameter configuration of the recorder dedicated to the power industry is as follows: Processor: 8-core high-speed processor; System: Android operating system; Communication: 4G full network access should support 5G; Call: support two-way interaction, real-time intercom, and cluster intercom functions; Video compression: H.264 / H.265 high-definition encoding is recommended; Storage capacity: 64G or 128G; Battery: Capacity greater than 3000mAh, battery life more than 10 hours, low-temperature battery; With the function of uninterrupted battery replacement, it is recommended to have an ultra-low power consumption recording mode; Positioning: GPS+Beidou dual positioning, it is recommended to have a positioning accuracy of <1m; Lens width: 120 degrees wide angle; Night vision capability: 5 meters to see facial features, 10 meters to see the outline of the human body; Weight: ≤200g; Protection performance: IP6-8; AI function: support basic face recognition functions; Anti-shake: intelligent anti-shake; Screen: 2 inches or more, touch operation; Others: Optional WiFi, Bluetooth function, one-button SOS function.

[0027] Example 2: Establishing a power behavior sample database

[0028] The basis for establishing a sample database of power behavior is to collect diverse data. The sample database should contain different types of jobs and tasks, covering different industries and industry specifications to ensure that the sample database is diverse. This helps ensure that the model can generalize to multiple industries and tasks. 1) Sample annotation: The data in the sample database should be correctly annotated to ensure that the model can learn meaningful features. For example, when identifying whether a worker is wearing a hard hat, it should be marked whether the head is wearing a hard hat, which helps the model learn features such as the appearance and position of the hard hat. 2) Data quality: Data quality is a key factor in establishing an effective sample database. The data in the sample database should be accurate, complete and reliable. In order to ensure data quality, a variety of measures can be taken, such as data cleaning, data deduplication, data verification, etc. 3) Collection strategy: When collecting data, it is necessary to consider how to maximize coverage of possible situations. For example, when identifying whether a worker is wearing a hard hat, you can consider taking photos in different scenes, such as indoors and outdoors, different lighting conditions and angles, etc. 4) Data protection: The data in the sample database should be fully protected to avoid leakage and abuse. This includes measures to ensure data storage security, access control and data encryption. In particular, the AI ​​recognition algorithm needs to establish an algorithm sample database according to different business needs. In view of the safety supervision needs of power production site operations, it is planned to establish 7 types of intelligent recognition algorithm sample databases, including personnel wearing safety helmets, wearing safety belts, wearing insulating gloves, smoking, welding operations not in accordance with labor protection requirements, not wearing masks for dusty operations, throwing objects from high places, and illegal fire. The details are as follows:

[0029] 1) Wearing of helmets. Identify all personnel who have worn helmets but cannot tell whether they are worn correctly. If the face can be seen, the face part needs to be identified. Identify all personnel who do not wear helmets on their heads. Personnel who are not wearing helmets need to be identified. If the face can be seen, the face part needs to be identified. Situations where helmets are not worn include but are not limited to situations where they are not worn on the head, held in the hand, placed on the ground / table / stone pier / cabinet, etc. Situations where helmets are not worn correctly at the work site include but are not limited to: not wearing the chin strap, wearing it backwards, not tightening the chin strap, the back hoop is too tight or too loose, and the chin strap is placed inside the helmet, on the side or on the top. Examples are shown in the figure below. 2) Wearing of safety belts. Nowadays, there are a wide range of types of aerial work, covering various medium and high altitude scenes, such as outdoor air conditioning installation, bridge construction, iron scaffolding on construction sites, and overhead wire pulling, and many other facilities within the scope of the project and buildings outside the site. Therefore, the environmental characteristics of the work site vary depending on the type of construction, and the following factors need to be considered when constructing the data set. 1) The posture of the workers, the degree of crowding, and the occlusion of people and buildings. 2) Objective external factors such as changes in lighting and weather. 3) The video images collected from the site vary greatly due to the different camera installation positions, resulting in occlusion. 4) The monitoring range of the camera at the work site generally needs to cover the entire construction site. At this time, the image target scale will be small due to the distance between the workers and the camera. The collected image data needs to be manually labeled before it can be used to train the model. For the safety belt detection task proposed in this paper, there are two categories for labeling the upper body area of ​​the worker, namely wearing a safety belt and not wearing a safety belt. In addition to labeling the category of the upper body area of ​​the worker, this task also requires labeling the location information of the upper body area of ​​the worker (upper left and lower right corner coordinates). 3) Wearing of insulating gloves. The detection of the wearing of insulating gloves requires the detection of the hands and insulating gloves of the person, which is a typical small target detection problem. The images used for training the model should come from images taken at the power grid maintenance site. The data set is divided into a training set and a validation set. The data set should contain 6 types of samples, including correct wearing of insulating gloves, abnormal wearing of insulating gloves (not worn or wrongly worn), and other related categories. In order to improve efficiency, the data set was manually annotated before training. 4) Smoking. The principles for constructing a smoking sample database include collecting online smoking images, capturing smoking scenes in online videos, and taking photos of smoking behaviors in real scenes. The smoking scenes are taken on the construction site to expand the sample database. 5) Electric welding operations are not dressed in accordance with labor protection requirements. Electric welding is a special but also common occupation. It is special because this work has a certain technical content and dangerous operation. Generally, it is required to hold a certificate to work. When establishing a work clothes detection sample database, a corresponding sample database should be established separately according to the type of work clothes. 6) Do not wear a mask for operations with large dust.The mask sample database should be divided into two labels, wearing masks and not wearing masks, according to the type of masks. The sample database can be divided into two labels by web crawlers. 7) Throwing objects from high places. The detection of throwing objects from high places currently has flying birds or drying clothes, which are similar to throwing objects from high places. It is necessary to divide them into two labels in the construction of the sample database: "Throwing objects from high places" and "No throwing objects from high places". Similar behaviors such as flying birds are included in the data set of throwing objects from high places. 8) Illegal fire. The target detection of flames, so in order to ensure that the accuracy of the trained model is sufficient to be used in practical applications, all flame data sets annotated with detection labels are selected to ensure the detection accuracy of the model. The sample database of flames can be selected from the public sample database on the Internet plus the collection of on-site flame samples, including car flames, forest flames, household appliance flames, electrical flames, residential flames, etc. The model trained based on this will be applicable to many scenarios and has good robustness.

[0030] Example 3: Integration of recognition algorithms based on power operation behavior

[0031] 3.1 Establishment of the personnel behavior analysis model. With the development of computer hardware technology, it is possible to combine video intelligent analysis with deep learning. Therefore, deep learning networks are of great significance for video intelligent analysis. Deep learning algorithms have strong performance in object classification and detection. Although the calculation of the previous conduction and back propagation of the deep network brings huge time complexity, the model has the characteristics of high accuracy, good generalization performance and strong robustness for intelligent object classification and detection. This project uses deep learning algorithms to abstract the video stream obtained by the recorder camera into quantifiable behavior data. It includes: (1) obtaining images of known behavior types; (2) performing convolution calculations and maximum pooling calculations on the images, and extracting feature data used to represent image features; (3) inputting the feature data of the image and the corresponding behavior type into the pre-built neural network model; (4) classifying the feature data according to the corresponding behavior type; (5) repeating steps (1)-(4) until the training end conditions are met.

[0032] 3.2 The process of constructing a personnel behavior analysis model. The behavior analysis model needs to be constructed using a convolutional neural network. The steps of constructing the model are as follows: ① Forward propagation process. The forward propagation process of the convolutional neural network includes: padding, convolution operation (conv), activation function (Relu), pooling, full connection (FC), and softmax classification. The activation, full connection, and softmax are consistent with the calculation methods in deep neural networks. Specifically, (1) Padding: When performing a convolution operation on the input image, we will find a problem: the number of pixels in the corners or edges is relatively small. This will weaken the edge information in image recognition. Therefore, we use the padding operation to fill p layers of data around the original image data, as shown in the figure below. When the padded data is 0, we call it Zero-padding. In addition to being able to retain more effective information, the padding operation can also ensure that the height and width of the volume do not change before and after the convolution calculation. (2) Single-step convolution: In the convolution operation, the first concept we need to clarify is the filter (kernel), which is a multidimensional array (f, f, n_c) with the same number of channels as the input image, but with a small odd number of height and width (usually 1, 3, 5, 7, etc., we can represent this hyperparameter with f). (3) Convolution layer: In the single-step convolution process, there is only one filter, but in the CNN convolution layer, there will be multiple filters. The operation will be slightly more complicated, but the principle is the same. It is just that the images after convolution of each filter are superimposed together for output. (4) Pooling layer: The role of the pooling layer is to reduce the size of the network, increase the calculation speed, and improve the robustness of the extracted features. There are two main types: max-pooling and average-pooling. ② Back-propagation process: When using a deep learning framework, we only need to ensure the forward propagation process, and the framework will automatically perform the back-propagation process. The back-propagation process of CNN is relatively complicated.

[0033] 3.3 Principles of constructing a human behavior analysis model. The following five principles should be followed when constructing a neural network model: (1) The principle of simplicity; First, build a relatively simple model: build a small model with a single hidden layer and verify it; then gradually add complexity to the model while verifying whether each level of the model structure (additional layers, parameters, etc.) is effective. Second, train the model on a single data node: one or two training data points can be used to confirm whether the model is overfitting. The neural network should overfit immediately with a training accuracy of 100%, which indicates that the model is in line with the requirements; if the model cannot overfit these data points, it proves that it is too small or has bugs. (2) Confirm the model loss principle; model loss is the main way to evaluate model performance and the basis for setting important parameters for evaluation. Therefore, it is necessary to ensure that: the model loss is suitable for the task (use classification cross entropy loss for multi-classification problems or use focal loss to solve imbalanced problems); correctly measure the importance of loss functions. If you use multiple types of loss functions, such as MSE, adversarial, L1, feature loss, then make sure that all losses are sorted in the right way. (3) Check the intermediate output and connection principles; if the gradient value is 0, it means that the learning rate in the optimizer may be too small and the expression of the gradient update is incorrect. In addition to paying attention to the absolute value of the gradient, make sure to monitor the size of the activation and weight matching of each layer. For example, the size of the parameter update (weight and bias) should be 1-e3. It should be pointed out that in a phenomenon called "Dying ReLU" or "gradient disappearance", the ReLU neuron will output 0 after learning the negative bias term of its weight. These neurons will not be activated at any data point. Then gradient checking can be used to approximate the gradient through numerical methods to check these errors. If it is close to the calculated gradient, the back propagation is implemented correctly. (4) Optimize the diagnostic parameter principle; the neural network has a large number of parameters that interact with each other, so it is necessary to achieve the principle of optimizing the model by constantly adjusting the parameters. Learning rate: too low will lead to slow convergence or the risk of falling into a local minimum, and too high will lead to optimization divergence. Gradient clipping: In back propagation, the maximum value or maximum norm used to clip the parameter gradient. Batch normalization: used to normalize the input of each layer to combat the internal covariate shift problem. Stochastic Gradient Descent (SGD): uses momentum, adaptive learning rate, Nesterov update. Regularization: essential for building generalizable models because it adds penalties for model complexity or extreme parameter values.At the same time, it significantly reduces the variance of the model without significantly increasing the bias. (5) Tracking work principles: By better tracking the work, previous experiments can be easily viewed and reproduced to reduce duplication of work. Compared with manually recording information, tools like comet.ml can help automatically track datasets, code changes, experimental history, and production models, including key information about the model, such as hyperparameters, model performance indicators, and environment details.

[0034] 3.4 Analysis of work quality based on behavioral trajectories. In the scenario of job safety risk identification, it is feasible to evaluate the quality of work by analyzing the behavioral trajectories of workers. The specific research paths include the following aspects. 1) Define behavioral indicators: Before analyzing the behavioral trajectories, it is necessary to define a set of behavioral indicators to measure the quality of workers' behavior. For example, on a construction site, the following indicators can be defined: the time it takes for workers to complete tasks, the quality of tasks completed, the correctness of using tools and equipment, etc. 2) Collect data: Collecting workers' behavioral data is the key to evaluating the quality of work. Workers' behavioral data can be collected using sensors, cameras, or other devices and stored in a database. 3) Process data: After the data is collected, the data needs to be preprocessed in preparation for analysis. This may include data cleaning, deduplication, standardization, etc. 4) Build a model: Build a deep learning model to analyze workers' behavioral trajectories to evaluate the quality of work. Models such as recurrent neural networks (RNN) or long short-term memory networks (LSTM) can be used to perform sequence modeling on workers' behavioral data to capture temporal correlations. 5) Train the model: Use labeled data sets to train the model to learn how to associate workers' behavior trajectories with job quality. To improve the accuracy of the model, techniques such as cross-validation can be used to evaluate the performance of the model. 6) Deploy the model: Deploy the model to the production environment to monitor the workers' behavior trajectories in real time and evaluate the workers' job quality based on pre-defined behavioral indicators. Real-time data stream processing technology can be used to process large amounts of data and perform real-time analysis of workers' behavior trajectories. It can be seen that it is feasible to evaluate job quality by analyzing workers' behavior trajectories, and deep learning technology can be used to build and train models to achieve automated job quality evaluation.

[0035] Example 4: Embedded solution for behavior recognition algorithm of power operation recorder

[0036] The embedded research on the behavior recognition algorithm of the power operation recorder is essentially the development of deep learning algorithms based on embedded artificial intelligence devices. Therefore, it is necessary to consider that the computing power of the embedded platform is relatively lacking compared to personal computers and the feasibility of transplanting the model to the platform. Therefore, how to improve the convolutional neural network algorithm in a lightweight way while maintaining a high recognition accuracy has become the focus of the research. In order to realize the behavior recognition function of the power operation recorder, the behavior recognition algorithm based on the convolutional neural network is improved. In addition, this algorithm needs to be applied to the embedded platform to realize the complete function. The realization of this function can rely on the neural network acceleration unit NNIE (Neural Network Inference Engine) in the Hi3516dv300 system-on-chip to load the deep learning neural network algorithm to process image data. Taking HiSilicon as an example, the MPI function interface provided by the intelligent vision platform SVP it provides interacts with the operating system, and then the operating system calls the NNIE related driver to control the work of the neural network unit hardware. NNIE currently supports a variety of network models, such as Alexnet, VGG16, Resnet18 and other type-distinguishing network models, YOLO, SSD, RFCN and other target recognition network models, and SegNet, FCN and other semantic segmentation network models. However, the NNIE development environment and tools provided by HiSilicon only support the Caffe1.0 version of the deep learning framework, and only support 8-bit integer or 16-bit integer precision mode. For other network models using deep learning architectures such as TensorFlow and Pytorch, it is necessary to convert the original model architecture into a network model under the Caffe architecture after training. Then the NNIE Mapper tool generates a .wk file, and the model accuracy is checked on the PC, and finally burned on the hardware platform for operation. The following paths are recorded in particular: (1) yolo file conversion to onnx file; ONNX, as an open neural network conversion platform, supports a variety of neural network frameworks. Currently, deep learning frameworks such as Tensorflow, MxNet, and Pytorch have their own unique ways to construct and store network models. When completing the full functionality of the project, different network models need to be matched, which makes it difficult to simulate and implement deep learning algorithms on devices. Onnx can convert models using different deep learning frameworks into onnx models, which greatly facilitates developers to flexibly choose the appropriate framework for development in the early stage and use onnx for framework conversion in the later stage. In addition, Onnx also defines an extensible network graph model, data type standards, and embedded operators.(2) Convert onnx files to caffe files; because the NNIE hardware acceleration unit can only run wk files converted from the Caffe1.0 framework, Caffe is developed based on the C++ language and is similar to Pytorch, which is also a widely used modular framework in the field of deep learning. Therefore, it is also necessary to convert the onnx network model file into the caffemodel file that describes the training weight parameters and the prototxt file that describes the network model and configuration parameters required by the Caffe framework. (3) Convert caffe files to wk files; the conversion process requires the use of the toolkit configured by HiSilicon for the hardware acceleration unit development, including the NNIE mapper tool, which converts the training model under the Caffe framework into a wk model file that can be loaded on the Hi3516dv300 platform. For both Windows and Linux operating systems, HiSilicon provides RuyiStudioIDE, which has functions such as compiling code, debugging, and running.

[0037] Example 5. Development of autonomous optimization algorithm: a-Combined data constraint process

[0038] a- The combined data constraint process is specifically as follows: by constructing data constraint conditions, the video data collected by the power operation recorder is pre-classified to improve the data recognition algorithm's dynamic, accurate and continuous recognition of the same subject or object at the power operation site.

[0039] Among them, the data constraints include independent and cross-limited event constraints and data constraints. The constraint connotation of the event constraints is based on the relevant subject and object factors implied by the power operation event itself. The power operation subject personnel and various object equipment related to the current detected power event are limited to the monitoring range of the constraints and data reading, recording and multi-linear processing are carried out; the constraint connotation of the data constraints is mainly based on the data format and data value limit conditions matched when the multi-dimensional components of spatio-temporal events are allocated to each power operation subject and object in the multi-linear dimension in the subsequent data processing process. The power operation subject and object detection data that does not conform to the data format and exceeds the allowable upper and lower limits will be eliminated or corrected in advance.

[0040] Obviously, event-based constraints and data-based constraints are independent of each other and can be cross-applied in a partial order. Generally, event-based constraints are first applied to define the subject and object associated with the current target power operation event within the data scope. There are two optional ways to apply the data-based constraints: when the framework and data range of subsequent data processing can be defined in advance, the corresponding data-based constraints can also be defined in advance through reverse reasoning. At this time, the constraints can be applied before or at the same time as the event-based constraints. However, under normal circumstances, the framework and data range of subsequent data processing cannot be accurately defined in advance, so the data-based constraints are applied in the data process, that is, the data constraints are initially set to infinity. When the data format does not match or the data value exceeds the limit when the multi-dimensional components of spatiotemporal events are allocated to each power operation subject and object in the multilinear dimension, the data constraints are applied to the associated data to eliminate the error data.

[0041] In general, event-based constraints and data-based constraints are used together to define all associated detection subjects and objects in the target power operation event that are compatible with subsequent multi-dimensional component allocation and multi-linear data processing within the scope of the current data processing; this is also the original intention of setting the data process of data constraints.

[0042] Example 6: Development of autonomous optimization algorithm: b-path cross-data hierarchical decomposition process

[0043] After the existing video image recognition neural network was systematically integrated and imported for application, after the initial large-scale trial, the data of collecting, identifying and tracking deviations was collected and compared and data traced back, and it was found that the main guiding traceability points of data deviations corresponded to the following data attribute sites: the dynamic trajectories of two or more video image elements and the power operation site personnel or device objects they represented, and the geometric intersection points on the two-dimensional topological surface of the three-dimensional physical space; further, the existing data algorithm used adopted a one-dimensional curve continuity model, and it was found that its curve tracking selection after the intersection of the dynamic trajectory was random, which constituted the data root cause of the video image recognition deviation; thus, the b-path intersection data was constructed. The hierarchical decomposition process improves the original random tracking algorithm into a decision-making tracking algorithm that is consistent with the physical dynamic trajectory of the personnel subject or device object at the power operation site. Generally speaking, an executable database group compatible with the four arithmetic operations of data is constructed. This database group contains multiple micro sub-data groups, and each sub-data group has independent attributes and interactive attributes based on the four arithmetic operations of data. The independent or mutual data operation results of each micro-sub-data group in the executable database group are used as the decision basis for the dynamic trajectory tracking algorithm of the personnel subject or device object at the power operation site; the above-mentioned database group adopts the bottom-level local self-organization to form a high-level overall data organization configuration, and the specific data process includes:

[0044] ① For any power operation site personnel subject or device object, a discrete micro-database with a unique label is constructed respectively. The spatial coordinate data (two-dimensional topological coordinate data of three-dimensional physical space coordinate data) of the power operation site personnel subject or device object under the selected time interval Δτ is stored in this discrete micro-database. For example, under the unique label ak-Δτ, the data phenotype in this discrete micro-database is (ak-c0, ak-c 1Δτ ,ak-c 2Δτ ,ak-c 3Δτ , ...); for different discrete micro-databases, ak corresponds to a natural number value sequence (k = 1, 2, 3, ...); and (c0, c 1Δτ , c 2Δτ , c 3Δτ , …) sequences usually have global invariance. This is based on the convenience of data interaction of the micro-sub-databases within the database group that are compatible with the four arithmetic operations. On the other hand, the data sources of different discrete micro-databases are all the same hardware system and its data acquisition and storage algorithm, that is, the power operation recorder that has been integrated and adapted and developed before, so that the obtained data itself has natural coordination, which is consistent with the different types of c- Δτ The sequence is significantly improved.

[0045] ② For the executable database group constructed above, take its local data to construct an executable sub-database group under the premise of maintaining the original data configuration, where maintaining the original data configuration means that the sub-database group and its internal micro-sub-database adopt the same data labeling logic as the original executable database group, each micro-sub-database adopts a discrete data sequence model, and each micro-sub-database and different micro-sub-data have their own independence and data interactivity compatible with four arithmetic operations; where local data refers to the geometric intersection of two or more video image elements and the dynamic trajectory of the power operation site personnel subject or device object represented by them on the two-dimensional topological surface of the three-dimensional physical space, and the local data is the data sequence distributed near this intersection. Specifically, the number of local data sequences is limited to between (2-200), generally taking the number 5, then a corresponding executable database group containing k micro-sub-databases is obtained, where each micro-sub-database contains five data units represented as vector data pairs, and the data labeling model is represented as: (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ ), where k corresponds to a natural number value sequence k = 1, 2, 3, ...; and the data intersection point corresponds to the data point ak-c p+2Δτor ak-c p+2Δτ With ak-c p+1Δτ ,ak-c p+3Δτ between;

[0046] ③Finally, a decision-making algorithm based on data verification is constructed. For any (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ ) discrete data sequence, note that each data corresponds to a two-dimensional coordinate data, that is, vector data. Vector data has dual attributes of size and direction. The key is that in order to adapt to the subsequent data execution algorithm, its size attribute should be eliminated. To this end, the two-dimensional coordinate data pair and its corresponding vector data are first represented by complex numbers based on the Euler formula, that is, any discrete data unit is represented as ak-c p+mΔτ Characterized by e l+iθ form, where e l Corresponding vector length, e iθ Characterize the angle data of the data pair; in order to eliminate the vector size attribute data with interference attributes, only e is used in the subsequent data execution process iθ ; finally, an executable algorithm is constructed for the angle-characterized data item, and its construction standard is: based on the four arithmetic operations, the numerical value of the operation result is linearly related to (positively correlated with or negatively correlated with) the data mismatch probability of the executable database group or the micro-sub-database in the executable sub-database group, thereby completing the tracking decision of the dynamic data intersection; the executable algorithm that meets this data standard can have many options, including: ③-ω, for the data sequence (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ ) derived from the five groups of angle characterization data items, firstly, the first two and the last two are added, then the two data are subtracted and the absolute value is taken. The size of the obtained absolute value is positively correlated with the data mismatch probability of the micro-sub-data in the executable database group or the executable sub-database group; ③-ψ, for the data sequence (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ) are derived from five groups of angle representation data items, with the first or last digit as the subtrahend, and the remaining four groups of data as the minuends in turn. Four numerical differences are constructed by subtraction operation and their absolute values ​​are taken respectively. Finally, the four positive numbers are added together. The size of the sum is positively correlated with the data mismatch probability of the micro-sub-data in the executable database group or the executable sub-database group.

[0047] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0048] In each embodiment, the hardware implementation of the technology can directly adopt existing intelligent devices, including but not limited to industrial computers, PCs, smart phones, handheld stand-alone machines, floor-standing stand-alone machines, etc. Its input device preferably adopts a screen keyboard, its data storage and calculation module adopts existing memory, calculator, controller, its internal communication module adopts existing communication ports and protocols, and its remote communication adopts existing GPRS networks, World Wide Internet, etc. It can be clearly understood by technicians in the relevant field that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of each functional unit and module are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working processes of the units and modules in the above system can refer to the corresponding processes in the above method embodiments, which will not be repeated here.

[0049] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A method for processing dynamic operation control data based on an electric power operation recorder, characterized in that: The method includes at least the following sub-items: integrated adaptation and development of power operation recorders, construction of a power behavior sample database, and construction and / or optimization of a power operation behavior recognition algorithm; The integrated adaptation development of the power operation recorder includes: building an all-in-one recorder for power operation, the screen of the all-in-one recorder adopts an LCD screen, supports multi-point capacitive touch, the chip adopts an Intel series processor, supports SSD+high-capacity HDD flexible expansion, and the acquisition interface supports Mini USB / Micro / typeC. In terms of software, the management system adopts a B / S architecture, supports distributed download, supports remote download of recorder data, automatic collection, automatic deletion, automatic charging, unattended, and a full metal shell to prevent strong electric field and strong magnetic field interference. It also supports 7*24*365 all-weather operation, supports real-time access and parameter configuration through a Web browser, and realizes standard SDK development and support; the parameter configuration of the all-in-one recorder for power operation is as follows: processor: 8-core high-speed processor; system: Android operating system; communication: 4G full network support 5G; call: support two-way interaction, real-time intercom, cluster intercom functions; video compression: Adopt H.264 / H.265 high-definition encoding; Storage capacity: 64G or 128G; Battery: Capacity greater than 3000mAh, battery life more than 10 hours, low-temperature battery; With uninterrupted battery replacement function, with ultra-low power consumption recording mode; Positioning: GPS+Beidou dual positioning, with positioning accuracy of <1m; Lens width: 120-degree wide angle; Night vision capability: 5 meters to see facial features, 10 meters to see human outline; Weight: ≤200g; Protection performance: IP6-8; AI function: Support basic face recognition function; Anti-shake: intelligent anti-shake; Screen: 2 inches or more, touch operation; Others: WiFi, Bluetooth function, one-key SOS function; The construction of the power behavior sample database includes: establishing a power behavior sample database based on collecting diversified power operation data. The database contains different types of operations and tasks, covering different industries and industry specifications, ensuring that the sample database is diverse and can generalize to multiple power operation tasks; the construction sub-items include at least: 1) sample labeling: the data in the sample database has correct labels to ensure that the subsequent algorithm model learns meaningful features; 2) data quality: data quality is constructed as the underlying factor of an effective sample database. The data in the sample database is accurate, complete and reliable. In order to ensure data quality, a variety of measures are taken including data cleaning, data deduplication and data verification; 3) acquisition strategy: to achieve maximum coverage of possible scenarios when collecting data; 4) data protection: data in the sample database is prevented from being leaked and abused, and the data path at least includes data storage security, access control and data encryption; the application-oriented construction of the power behavior sample database includes: in response to the safety requirements of power production site operations, the data sub-items of the power behavior sample database include at least 8 types of intelligent recognition algorithm sample data: workers wearing safety helmets, safety belts, insulating gloves, smoking, welding operations not in accordance with labor protection requirements, dusty operations without masks, high-altitude throwing, and illegal fire; The construction of the electric power operation behavior recognition algorithm includes: A. Framework model for behavior analysis of power workers: The video stream obtained by the recorder camera is abstracted into quantifiable behavior data using a deep learning algorithm, including the following data steps: (1) obtaining an image of known behavior type; (2) performing convolution and maximum pooling calculations on the image, and extracting feature data used to represent image features; (3) inputting the feature data of the image and the corresponding behavior type into a pre-built neural network model; (4) classifying the feature data according to the corresponding behavior type; (5) repeating steps (1)-(4) until the training end condition is met; B. Data flow of constructing the behavior analysis model of power workers: The behavior analysis model of power workers is constructed using a convolutional neural network. The steps of constructing the model are as follows: A. Forward propagation process: The forward propagation process of the constructed convolutional neural network specifically includes: (1) Padding: When performing a convolution operation on the input image, the pixels at the corners or edges are used relatively less frequently, which will weaken the edge information in image recognition. Therefore, a padding operation is used to fill p layers of data around the original image data, retaining valid information while ensuring that the height and width of the convolution calculation do not change before and after the convolution calculation; (2) Single-step convolution: In the convolution operation, the number of filter channels is proportional to the input. The input images are the same, the height and width are odd numbers and are constructed as a multidimensional array; (3) Convolution layer: In the case of a single-step convolution process, there is only one filter. In the CNN convolution layer, multiple filters are constructed. When outputting, the images after convolution of each filter are superimposed together; (4) Pooling layer: The function of the pooling layer is to reduce the size of the network, increase the calculation speed, and improve the robustness of the extracted features. It uses maximum pooling max-pooling and / or average pooling average-pooling; B. Back propagation process: In the case of a deep learning framework, as long as the forward propagation process is guaranteed, the framework will automatically perform the back propagation process; Furthermore, the data processing method is optimized based on the algorithm process, and the power operation behavior recognition algorithm is optimized based on a-combination data constraint process and / or b-path cross data hierarchical decomposition process.

2. The method according to claim 1, characterized in that: The a-combined data constraint process is specifically as follows: by constructing data constraint conditions, the video data collected by the power operation recorder is pre-classified to improve the data recognition algorithm's dynamic, accurate and continuous recognition of the same subject or object at the power operation site, wherein the data constraint conditions include independent and cross-limited event constraints and data constraints, wherein the constraint connotation of the event constraint conditions is based on the relevant subject and object factors implied by the power operation event itself, and the power operation subject personnel and various object equipment associated with the current detected power event are limited to the defined monitoring range of the constraint conditions and data is carried out. Reading, recording and multi-linear processing; the constraint connotation of the digitized constraint condition is based on the data format and data value limit conditions matched when allocating multi-dimensional components of spatio-temporal events to each power operation subject and object in the multi-linear dimension in the subsequent data processing process, and the power operation subject and object detection data that does not conform to the data format and exceeds the upper and lower limits of the allowable limit will be eliminated or corrected in advance; the event-based constraint conditions and digitized constraint conditions of the constraint conditions are used together, and all related detection subjects and objects in the target power operation event that are compatible with subsequent multi-dimensional component allocation and multi-linear data processing are all defined as within the scope of the current data processing.

3. The method according to claim 2, characterized in that: The event-based constraints and data-based constraints are independent and compatible with each other and can be cross-applied in no particular order; the application process is to first perform event-based constraints, and define the subject and object associated with the current target power operation event within the data scope; secondly, apply the data-based constraints, and this data application process includes two approaches: ① When the framework and data range of subsequent data processing can be defined in advance, the corresponding data-based constraints can be defined in advance through reverse deduction, and at this time, the constraints are applied before or at the same time as the event-based constraints; ② If the framework and data range of subsequent data processing cannot be accurately defined in advance, the data-based constraints are applied in the data process, that is, the data constraints are initially set to infinity. When the data format does not match or the data value exceeds the limit when the multi-dimensional components of spatiotemporal events are allocated to each power operation subject and object in multi-linear dimensions, the data constraints are applied to the associated data to eliminate the error data; In the behavior recognition scenario at a real power operation site, the occurrence rate of path ② is higher than that of path ①.

4. The method according to claim 1, characterized in that: The b-path cross data hierarchical decomposition process is: In general, an executable database group compatible with the four arithmetic operations is constructed. This database group contains multiple micro sub-database groups, and each sub-database group has independent attributes and interactive attributes based on the four arithmetic operations. The independent or mutual data operation results of each micro-sub-database group in the executable database group are used as the decision basis for the dynamic trajectory tracking algorithm of the power operation site personnel subject or device object; the above database group adopts the bottom-level local self-organization to form a high-level overall data organization configuration. The specific data process includes: ① For any power operation site personnel subject or device object, a discrete micro-database with a unique label is constructed respectively. The spatial coordinate data of the power operation site personnel subject or device object under the selected time interval Δτ is stored in this discrete micro-database. Under the unique label ak-Δτ, the data phenotype in this discrete micro-database is (ak-c0, ak-c 1Δτ ,ak-c 2Δτ ,ak-c 3Δτ ,…); for different discrete micro-databases, ak corresponds to a natural number value sequence, k = 1, 2, 3,…; and (c0, c 1Δτ , c 2Δτ , c 3Δτ , …) sequences usually have global invariance. This is based on the convenience of data interaction between the micro-sub-database groups within the database group, which are compatible with the four arithmetic operations. On the other hand, the data sources of different discrete micro-databases are all the same hardware system and its data acquisition and storage algorithm, that is, the power operation recorder that has been integrated and adapted and developed before, so that the obtained data itself has natural coordination. ② For the executable database group constructed above, its local data is taken to construct an executable sub-database group under the premise of maintaining the original data configuration, wherein maintaining the original data configuration means that the sub-database group and its internal micro-sub-database group adopt the same data labeling logic as the original executable database group, each micro-sub-database group adopts a discrete data sequence model, and each micro-sub-database group and different micro-sub-database groups have their own independence and data interactivity compatible with four arithmetic operations; wherein, local data refers to the geometric intersection of two or more video image elements and the dynamic trajectories of the power operation site personnel or device objects represented by them on the two-dimensional topological surface of the three-dimensional physical space, and the local data is the data sequence distributed near this intersection. Specifically, the number of local data sequences is limited to between (2-200), and the number is 5, then a corresponding executable database group containing k micro-sub-database groups is obtained, wherein each micro-sub-database group contains five data units represented as vector data pairs, and the data labeling model is represented as: (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ ), where k corresponds to a natural number value sequence k = 1, 2, 3, ...; and the data site is located at ak-c p+2Δτ or ak-c p+2Δτ With ak-c p+1Δτ ,ak-c p+3Δτ between; ③Finally, a decision-making algorithm based on data verification is constructed. For any (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ ) discrete data sequence, note that each data corresponds to a two-dimensional coordinate data, that is, vector data. Vector data has dual attributes of size and direction. The key is that in order to adapt to the subsequent data execution algorithm, its size attribute should be eliminated. To this end, the two-dimensional coordinate data pair and its corresponding vector data are first represented by complex numbers based on the Euler formula, that is, any discrete data unit is represented as ak-c p+mΔτ Characterized by e l+iθ form, where e l Corresponding vector length, e iθ Characterize the angle data of the data pair; in order to eliminate the vector size attribute data with interference attributes, only e is used in the subsequent data execution process iθ ; finally, an executable algorithm is constructed for the angle-characterized data item, and its construction standard is: based on the four arithmetic operations, the numerical value of the operation result is linearly related to the data mismatch probability of the executable database group or the micro-sub-database group within the executable sub-database group, so as to complete the tracking decision of the dynamic data intersection; the executable algorithms that meet this data standard include: ③-ω, for the data sequence (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ ) derived from the five groups of angle characterization data items, firstly, the first two and the last two items are added, then the two data are subtracted and the absolute value is taken. The size of the obtained absolute value is positively correlated with the data mismatch probability of the executable database group or the micro-sub-database group within the executable sub-database group; or ③-ψ, for the data sequence (ak-c pΔτ ,ak-c p+1Δτ ,ak-c p+2Δτ ,ak-c p+3Δτ ,ak-c p+4Δτ ) are derived from five groups of angle representation data items, with the first or last digit as the subtrahend, and the remaining four groups of data as the minuends in turn. Four numerical differences are constructed by subtraction operation and their absolute values ​​are taken respectively. Finally, the four positive numbers are added together. The size of the sum is positively correlated with the data mismatch probability of the executable database group or the micro-sub-database group within the executable sub-database group.

5. The method according to claim 4, characterized in that: It also includes the following data processes: ④ For two or more groups of local data sequences containing five data bits distributed according to intersection points, the median data, i.e. the third bit, is kept unchanged, and the data before and after the median data are cross-interchanged to construct a new local data sequence, and an executable algorithm that meets the data standard is applied to the obtained local data sequence: ③-ω or ③-ψ, and the mismatch probability obtained by the data sequences before and after the cross-interchange under the executable algorithm is compared, so as to more accurately complete the tracking decision of dynamic data intersections.

6. The method according to claim 4, characterized in that: In step ③, the Euler form complex representation of the two-dimensional coordinate data and the elimination of interfering data sub-items are allowed to be completed before applying the executable algorithms ③-ω and ③-ψ, or to be integrated into the executable algorithms afterwards.