Research on Intelligent Recognition Method of Risk Levels of Electricity Marketing Field Operation Types Based on Data Mining

Data mining is mainly a process of obtaining hidden information from a large amount of incomplete, fuzzy, noisy, and random data through algorithms. Typically, data mining tasks are mainly categorized into two main types: prediction tasks and description tasks. The prediction task is to predict the value of the target variable or dependent variable attribute based on the value of the descriptive variable or independent variable attribute. Descriptive tasks, which are the discovery of underlying patterns in the given data, are usually exploratory and require interpretation of the results [21]. The four main tasks of data mining are predictive modeling, association analysis, cluster analysis, and anomaly detection.

The process of data mining generally refers to the process of automatically searching for information hidden in a large amount of data that has a special relational nature. Data mining is a cyclic process. When one of the steps does not achieve the desired goal, it is necessary to go back to the previous step and repeat the execution until the result meets the user’s needs. The whole process of data mining is shown in Figure 1.

Figure 1.

Data mining process

Association analysis is used to discover meaningful connections hidden in large datasets, which are usually represented in the form of association rules or frequent item sets.

Definition 1 (item): In general, let L = {x₁, x₂, ⋯, x_n} be the set of all items in the dataset and T = {t₁, t₂, …, t_i} be the set of all transactions. Each transaction t_i contains an itemset that is a subset of L. In correlation analysis, a set containing 0 or more items is called an item. If an item contains k items, it is said to be a k – item set.

Definition 2 (Association Rule): An association rule is an implication expression of the form X → Y, where X is called the rule antecedent, Y is called the rule consequent, and X and Y are disjoint sets of terms, i.e., X ∩ Y = ∅. The strength of an association rule can be measured in terms of support and confidence.

Definition 3 (Support): Support determines how frequently a rule can be used for a given data set. The degree of support has an aspirational nature; rules with low support may appear only by chance and are mostly meaningless. The formula calculates the support degree: 1S(X)=n(X)n(T)

where n(X) is the number of transactions that contain itemset X in all transactions, also known as the support count for itemset X, and n(T) is the number of all transactions.

Definition 4 (Confidence Level): the confidence level determines how often Y occurs in the transactions containing X. For a given rule X → Y, the higher the confidence level, the more likely it is that Y will occur in transactions containing X. The formula calculates the confidence level: 2C(X→Y)=S(X∪Y)S(X)

Definition 5 (Lift): Since the confidence measure ignores the support of the item that appears in the posterior of the rule, and the high-confidence rule may also be meaningless, this paper adds an objective measure called the interest factor or called the lift, which can make up for the shortcomings of the “support-confidence” framework to a certain extent, and its calculation formula is: 3I( X→Y)=n( T)*n( X∪Y)n( X)*n( Y)

Property 1: All non-empty subsets of a frequent itemset are frequent.

Property 2: If any (k − 1) − term set X, 1 < k, X of I is not frequent, then any k – term set Y is also not frequent, this property is used to terminate the search for frequent term sets operation.

The most commonly used association rule mining algorithms are the Apriori algorithm and the FP-growth algorithm. 1)

Apriori algorithm

The Apriori algorithm is the first association rule mining algorithm. Apriori algorithm adopts a layer-by-layer search strategy, first through the support to filter out all the frequent 1-item set, will leave the frequent item set of the two splicings, and then support “pruning”, will repeat the above steps until it can not be found (n + 1)-itemsets can no longer be found. As the number of items in the itemset increases, the number of times the algorithm scans the dataset gradually increases, the storage space required gradually increases, and the running time appears to increase exponentially with the number of transactions.

The main body of the FP-growth algorithm is to use a compact data structure called FP-tree to organize and process the data. FP-tree has only one root node, NullSet in the initial period of its construction. After that, the FP-tree connects the transactions in each path in the FP-tree one by one by reading the transactions in the original database. The construction of the FP-tree is shown in Fig.2. The prefix paths of the elements D in the FP tree in the figure are {A, B}, {B}, so these Ds are called similar nodes, and the red dotted line is used to connect the similar nodes in the FP tree.

Figure 2.

Construction of FP tree

After preprocessing the image, the image is further analyzed to extract the features of the field operation violation. The image is first filtered, where the feature extraction layer utilizes a 3×3 convolution kernel for depth processing of the image to effectively extract key features in the operational image. The depth processing process is as follows: 4D=Γ(a)

where a is the initial value of the nonlinear mapping used to extract the features of the power marketing metering field operation image.

Then, the adaptive filter is selected as the a priori expectation of no signal and no noise to reject the extracted noise features and reconstruct the image using the effective features [23]. The power marketing metering field operation violation parameter f(x) is: 5f(x)=D(x)+k(x)

where D(x) is the modulation parameter of the input signal in the adaptive filter. k(x) is the original error signal.

The sampled data is used as the core of the weighted average to reduce the error. The phase frequency weighting and averaging parameter d(x) of the low-pass filter is: 6d(x)=∑i=1nsu(x−p)∑i=1nhp

where s is the output error of the filter. p is the optimal convergence signal. h is the least squares parameter of the smooth signal.

The 3 information connection layers play a key role in the convolutional neural network, which are responsible for fusing the shallow features, deep features, and overall features of the power marketing metering field operation images, respectively. Each information connection layer contains an enhancement module and a compression module, and through the cooperative work of these 2 modules, the output of the 3 information connection layers is: 7Dn=g(x)d(x)n−1,n=1,2,3

where g is the nonlinear mapping of the 3 information connection layers. D_n is the output of the 3 information connection layers, and the output result is the fusion result of power marketing metering field operation violation features.

After completing the power marketing metering field operation violation feature extraction, the violation recognition model is constructed to realize the automatic recognition of violations. Assuming that there are multiple network nodes in the machine vision network architecture, the power marketing metering field operation violation recognition model is: 8M=φ1+σe

where σ is the regression parameter. e is the power marketing metering field operation violation data vector.

Violation training and identification can be realized with the help of this model. The core of the model lies in the accurate use of the loss function, which realizes the efficient and accurate identification of violations through the accurate calculation of operational losses. The loss detection function is: 9N=Lϵ+δe+εL+θLk

where L_ϵ is the characteristic value of the movement for each individual. e is the coordinate error value on the image data. L_k is the critical point loss function. δ, ε and θ are the weighted values of the above parameters.

To ensure the safe use of construction equipment, the overall loss characteristic of construction can be accurately calculated according to Eq. Comparing it with the maximum loss value of the construction, if the combined loss characteristics are outside the safety limits, it can be determined that there is a violation. Thus, the identification of power marketing metering field operation violations is completed.

According to the violation determination standard, when there is a situation such as the operator has brought a helmet into the workplace but just put it aside without wearing it correctly, this paper adopts the intersection and merger ratio (IoU) in order to judge the degree of overlap of the position between the recognition frames as an additional condition for determining safe operation. The intersection and merger ratio is calculated as shown in Fig. 4. That is, the ratio of the intersection and concatenation of the two target rectangular boxes. The figure shows that when the two target boxes do not intersect, the value is 0, and when the two target rectangular boxes completely overlap, the value is 1. In this paper, the degree of overlap of the position between the judgment of the person and the critical item is calculated by the formula: 10IoU=M(xp)∩M(xn)M(xp)∪M(xn)

Figure 4.

Calculation method of intersection ratio

Where M(x_p) represents the pixel area of the operator’s “person” target box area and M(x_n) represents the pixel area of the key items that the operator needs to wear, such as helmets, goggles, protective gloves, and so on.

From the above equation, we can define the behavioral assessment function for the regular operation category of “working without helmet” as: 11Jhelmet=M(xp)∩M(xhelmet)M(xp)∪M(xhelmet)

When multiple additional conditions need to be satisfied, the product of the multiplicity union ratio is defined as a logical judgment function as a criterion for detecting whether there is a violation. Since not wearing goggles and protective gloves during electric and gas welding work is a violation of the law, it is necessary to identify the above objects, and in order to ensure correct wearing, each key item needs to have a certain position relationship with the target frame of the operator, so the behavior evaluation function that defines the operation category of the specific scenario of “electric and gas welding work scene” is: 12Jwelding=M(xp)∩M(xgoggles)M(xp)∪M(xgoggles)×M(xp)∩M(xSgloves)M(xp)∪M(xSgloves)

Where M(x_goggles), M(xs_gloves) represents the pixel area of the “goggles” and “S_gloves” identification box area, respectively, and J_welding represents the degree of overlap between the operator and the goggles and protective gloves.

If J is 0, i.e., there is no intersection between the target frames, it means that the helmet, goggles, and protective gloves are not worn correctly according to the regulations.

In order to construct a scientific and reasonable electric power operation behavioral assessment model, it is necessary to define the assessment function as the function threshold I. When the behavioral assessment function J reaches threshold I, it indicates that the operator reasonably wears the key items specified under the current operation scenario, which is judged to be a safe operation, or else it is regarded as a violation of the operation. Then, the definition of threshold I directly affect the accuracy of the violation judgment. Therefore, it is necessary to know the sensitivity of small targets to IoU so as to set a reasonable I. In the COCO dataset, targets smaller than 32 × 32 pixels are defined as small targets, which are more likely to be de-emphasized during the target detection process compared to medium and large targets, and their IoU sensitivity is different.

Assuming that the area of A, B, and C pixels is 6 × 6, and the diagonal deviation is 1 and 4 pixels, respectively, the IoU of small targets is given by: 13IoU=A∩BA∪B=0.53

Change for: 14IoU=A∩CA∪C=0.06

Assuming A, B, and C pixel areas of 36 × 36 and diagonal deviations of 1 and 4 pixels, respectively, the IoU of the normal-size target is given by: 15IoU=A∩BA∪B=0.90

Change for: 16IoU=A∩CA∪C=0.65

From the above example, it can be seen that the IoU of the small target decreases sharply and quickly, from 0.53 to 0.06. The IoU of the normal-sized target, on the other hand, has a decrease from 0.90 to 0.65, which is a smaller rate of change than that of the small target. Therefore, the reasonableness of the threshold I was setting is more dependent on the small targets. It is shown in the related literature that small targets occupy a small area in the image, and it is difficult for the IoU to reach 0.5. Therefore, the threshold range between 0 and 0.5 is selected in this paper.

Constructing a sound and scientific assessment method is the first condition to reach or realize the power operation violation assessment model. The traditional LEC method of hazard assessment of operating conditions measures the hazard level of operating conditions through the probability of occurrence of accidents (including hazardous events) L, the frequency of hazardous environments in which operators are exposed to E, and the probable outcome of occurrence of violations C, which are three dimensions D. The LEC method is mainly aimed at the construction conditions at the operating site, such as common hidden dangers, failure to wear helmets, and failure to implement daily inspections.

The functional relationship between the indicators in the traditional LEC method is: 17D=L⋅E⋅C

In order to better describe the hazardousness of operating conditions, the value of L is taken as 0~10, the value of E is taken as 0~10, and the value of C is taken as 1~100. Since the assessment of electric power operation violations has certain differences with the LEC assessment, this paper adjusts the assessment parameters on the basis of the LEC method, i.e., the frequency of violations is defined as F, the probability of violation leading to accidents is defined as P, and the severity causing the severity of the accident is defined as C, as the measurement parameters of the assessment method in this paper, and adjust the differences between the two methods.

Define the violation level by calculating the D-value. Define violation i, which corresponds to violation level D_i, then D_i is denoted as: 18Di=Fi⋅Pi⋅Ci

Where D_i indicates the danger level of violation i. F_i denotes the frequency of occurrence of violation event i. P_i denotes the probability that the violation event i causes an accident. C_i denotes the severity of the violation event i causing an accident.

In this paper, we utilize the improved YOLOv5s model for multi-task recognition and detection of six types of targets in the self-constructed electric power operation dataset: people, helmets, smoke, welding, goggles, and protective gloves. Based on the relationship between the person and the remaining five types of targets, the location relationship is judged, and the operator is evaluated to see if he is in a safe operating environment. If there is a violation, it is graded according to the violation grading method. The flowchart of the power operation violation assessment model is shown in Figure 5.

Figure 5.

Evaluation model of electric power operation violations

The training experiment configuration parameters are shown in Table 3. The Kinetics-skeleton skeleton behavior dataset was used as the training test set, which is a 3D-like skeleton sequence obtained by extracting skeleton features using the OpenPose algorithm on the Kinect-400 dataset and then using confidence as the Z-axis. It contains about 240,000 video clips covering 400 action classes and at least 600 video clips lasting about 10 seconds per action class. The cross-validation set contains 20,000 short video behavioral messages, each clip from a unique YouTube video that has been manually annotated over multiple rounds and covers a wide range of action data. The environment was deployed using a deep learning server to train the proposed model in this paper for a total of 60 iteration rounds on the training set. Data evaluation on the cross-validation set was performed every 10th time to count the probability of true value results in the TOPl and TOP5 classification. The dataset training results are shown in Table 4.

Test the robustness of the algorithm in different environments involving strong light, low light, sparse, and crowded. The experimental scenarios are shown in Fig. 6. Using precision rate and recall rate as performance metrics. The abnormal behavior recognition performance is shown in Table 5. As can be seen from the table, under the strong light and sparse conditions, this paper’s algorithm has the highest precision rate. Under low light and crowded conditions, this paper’s algorithm has the lowest precision rate. From the precision rate degradation, it can be seen that congestion causes more false detections than low light when other conditions remain unchanged. And the recall rate does not differ much in different conditions, indicating that low light and congestion do not cause more missed detections. The reason is that low light or crowding leads to incomplete skeleton extraction, and crowding leads to incomplete skeleton extraction after occlusion, causing more false detections. A normal behavior skeleton will be misdetected as an abnormal behavior skeleton, but the abnormal behavior skeleton will not be detected as a normal behavior skeleton, and there will be fewer leakage cases. Overall, the algorithm in this paper has little difference in the performance indicators in different environments, and the accuracy rate can reach more than 90%. It shows that the algorithm has good robustness.

Figure 6.

Experimental scene

Abnormal behavior identification is carried out using the Optical Flow Method, Motion Instability, Support Vector Machine, and Markov Random Field and compared with the identification results of this paper’s algorithm and the comparison of abnormal behavior identification performance is shown in Table 6. As can be seen from the table, the performance of this paper’s algorithm is optimal when abnormal behavior identification is performed. The reason is that Motion Instability calculates motion instability based on a trajectory to discriminate abnormal behavior, and it is difficult to track accurately when the occlusion is serious. The Optical Flow method ignores the dynamic correlation information of the moving target in time and adopts an iterative solution calculation method, which requires a longer computation time and is more affected by noise. The support vector machine method makes it more difficult to select features in the case of occlusion and is not applicable to the environment of multiple bursts. Markov random field can only reflect the simple changes of the target in the direction and velocity magnitude, and the recognition accuracy is not high in complex motion scenes. The experimental results show that the algorithm in this paper has better performance in abnormal behavior recognition.

The model accuracy is shown in Fig. 7, and the overall accuracy of the model is 0.845, as shown by the validation curve, which proves that the model in this paper correctly predicts most of the behavioral actions, but the validation curve is not smooth enough.

Figure 7.

Model overall accuracy

The precision and recall for the three types of actions are shown in Table 7. Sitting (label 0), standing (label 1), and falling (label 2) are given in the table. The precision Prcn data for the falling down action, which is the focus of the model goal, is 0.69, and the recall Rcll is 0.96. The data set Prcn represents the ratio of the number of correct predictions for the action of falling to all samples predicted as falling. Rcll represents the ratio of the number of collapses that were correctly predicted to the number of moves that were collapses in the dataset.

The confusion matrix is shown in Figure 8. Of the 250 samples considered to be actually falling, 195 were successfully predicted. The number of samples in which the falling down action was successfully recognized is high, so the low Prcn for the falling down action is actually due to the low recognition accuracy for the sitting down and standing up actions. Therefore, whether the model can achieve effective recognition of abnormal behaviors, the more concerned index is the recall rate. By observing the call rate, only a very small number of actions in the dataset labeled as falling were recognized as other actions, accounting for more than ninety percent of the total samples of the falling actions that were successfully recognized as falling. Considering the actual electric power operation scenario, as long as the abnormal behavior can be accurately identified to play the role of prevention, warning, and protection of the operating personnel, the recall rate of the downed behavior reaches 0.96, which can be achieved under this precision for the vast majority of the downed abnormal behavior identification.

Figure 8.

Confusion matrix

In order to evaluate the reasonableness of the three features, the video samples in the Fall Detection Dataset are selected as tests, and the data of the three types of features are outputted separately.

The relative height coefficient of the human head is shown in Fig. 9. The figure gives the curve of the human head’s relative height coefficient in the test video with the advancement of video frames. The head relative height coefficient stays between 0.6 and 0.8 from 0 to 200 frames of the video and has a clear tendency to increase when the video runs near 300 frames and eventually improves to near 1.0 after 300 frames. At around 350 frames of the video, the head relative height coefficient dropped off a cliff to near the value of 0 and then recovered to near 0.19. The first change in the relative head height coefficient is consistent with the increase in head height when a person shifts from sitting to standing, and the second change in the relative head height coefficient to 0 and then back to a certain height, which is consistent with the cliff-like decline in head height and the slight adjustment of the head height when the person shifts from standing to falling and tries to regain the posture after falling in the test video. The above analysis proves that the relative head height can effectively differentiate the three types of actions and highlights the difference between the abnormal behavior of falling and the other two types of routine actions.

Figure 9.

The relative height of the human head

The relative height coefficient of the human knee joint is shown in Fig. 10. The figure gives the curve of the relative height coefficient of the human knee joint in the test video with the advancement of the video frames. The relative height of the knee joint remains between 0.28 and 0.35 at frames 0 to 200, which corresponds to the relative height coefficient of the knee joint in the test video when the tester maintains the sitting action. At 250 to 350 frames, which corresponds to the tester’s standing behavior, the relative knee height shows a certain decrease. After 350 frames, the relative knee height decreased rapidly and remained at a very low value. The changes in the relative knee height between the first 200 frames and after 350 frames are consistent with the trend of the relative knee height in the sitting down and falling down scenarios. Observing the overall curve trend of the graph, it is proved that the method of this paper can effectively distinguish the abnormal behavior between two types of routine actions, namely, falling and sitting down, but cannot distinguish between two types of routine actions, namely, standing up and sitting down.

Figure 10.

The relative height of the body’s knee

The human leg angle cosine is shown in Fig. 11. As given in the figure, the curve of the cosine of the human leg angle in the test video with the advancement of the video frames.

Figure 11.

The cosine of the upper leg of the human body

In the figure from 0 to 200 frames, the human body is kept in a sitting position, and the cosine value is maintained between -0.3 and 0. At 250-350 frames, the cosine drops sharply to -1 as the human body changes from sitting to standing. It should be noted that the cosine rises briefly before the human body falls to the ground and then decreases and stabilizes after 330 frames.

The cosine remains near 0 from 0 to 200 frames, which is consistent with the body sitting down at a near-right angle between the legs. The cosine drops sharply around -1 after the human body changes to a standing position, which is consistent with the human body’s legs being close to a flat angle at this time, and the tester made a bending motion to make the angle between the legs less than 180 degrees. The final state of the cosine value is kept at -1 because the testers in the sample kept their legs straight when they fell to the ground, but this is not representative because the human body’s leg movements cannot be determined when they fall to the ground. The above analysis demonstrates that the large and small leg cosine values are effective in distinguishing between standing and sitting postures, but it is not representative to use it to describe the abnormal movement of falling to the ground.

In summary, in this section, the changes of three types of eigenvalues in the process of human posture changes are analyzed, and the method of this paper can effectively differentiate three types of actions, the method of this paper can effectively differentiate routine actions from falling down abnormal behaviors, and the large and small leg cosine values can effectively differentiate between standing up and sitting down. The fusion of the three types of features can accurately distinguish three different types of actions and achieve the goal of accurately identifying abnormal behaviors. However, considering that there are still some abnormal changes in the process of feature value change, more action classifications and features need to be added to describe the situation when each action is transformed.

This subsection tests the algorithm performance by using the COCO image dataset with the Fall Detection Dataset video dataset.

Video dataset testing: three videos are selected for testing. Video #1 has the same scene and different angles as video #2, and video #3 has different scenes with video #1 and video #2. The total number of frames of the test videos is 351, 261, and 427 frames, respectively. The actual running data of the algorithm is shown in Table 8. From the information in the table, it can be seen that the recognition algorithm in this paper is applied to the above three groups of test videos with better results, and the recognition accuracy for the downward movement in the videos all reaches more than 96%.

The test data of the electric power operation scene is shown in Table 9. Since there is less material of abnormal behaviors in the electric power operation scenario, the number of detected frames is the total number of samples. The number of frames of successfully recognized actions is used to evaluate the performance of the algorithm. Finally, the overall success rate of the algorithms of this chapter in real operation scenarios without obscuring the key points of the operators reaches 0.942.