Gas stealing user recommendation method based on deep learning

Abstract

The invention relates to the technical field of gas data processing, and discloses a gas stealing user recommendation method based on deep learning, and the method comprises the steps: constructing a gas stealing preliminary screening model, and screening out preliminary screening abnormal users; the gas stealing preliminary screening model comprises a continuous zero gas consumption index, a conventional statistical index, a linear goodness of fit index and an outlier index; constructing a data set based on the primarily screened abnormal users for training a deep learning model; and inputting the user name and the historical gas consumption data of the user into the trained deep learning model, and outputting a recommended gas stealing user. The objective of the invention is to improve the accuracy of recommending the gas stealing behavior of the user based on the historical gas consumption data.

Description

Technical Field

[0001] The present invention relates to the technical field of gas data processing, and particularly relates to a method for recommending gas theft users based on deep learning. Background Art

[0002] Recommending gas theft behavior based on the historical gas consumption data of users has multiple solutions in mathematics. On the one hand, gas theft behavior will definitely cause certain data phenomena; on the other hand, similar data phenomena cannot be very surely determined to be caused by gas theft behavior. This recommendation system is similar to today's audience software. The recommended content is mostly what users are interested in, and there is also a certain possibility that users are not interested in it. However, with the development of the business and after several rounds of iterative learning, the accuracy of the recommendation model is getting higher and higher. Therefore, after iterative learning based on historical gas consumption data, how to improve the accuracy of recommending users' gas theft behavior is a problem that needs to be solved. Summary of the Invention

[0003] The purpose of the present invention is to improve the accuracy of recommending users' gas theft behavior based on historical gas consumption data, and to provide a method for recommending gas theft users based on deep learning.

[0004] In order to achieve the above-mentioned invention purpose, the embodiments of the present invention provide the following technical solutions:

[0005] A method for recommending gas theft users based on deep learning includes the following steps:

[0006] Step 1, construct a preliminary screening model for gas theft to screen out preliminary screening abnormal users; the preliminary screening model for gas theft includes continuous zero gas consumption index, conventional statistical index, linear goodness-of-fit index, and outlier index;

[0007] Step 2, construct a data set based on the preliminary screening abnormal users for training a deep learning model;

[0008] Step 3, input the user name and the historical gas consumption data of the user into the trained deep learning model, and output the recommended gas theft users.

[0009] Compared with the prior art, the beneficial effects of the present invention:

[0010] The theft gas preliminary screening model constructed by the present invention is used for the initial launch of the recommendation system. Generally speaking, at this stage, there is still a lack of labeled data and a machine learning model cannot be directly established. Manually logical indicators are mainly used for recommendation. The theft gas preliminary screening model defaults to continuous zero gas consumption indicators, conventional statistical indicators, linear goodness-of-fit indicators, and outlier indicators. These indicator algorithms are combined to form the theft gas preliminary screening model, which can screen out users with abnormal preliminary screening. The accuracy of the model at this stage is not high, but it includes all theft gas behaviors as much as possible. The gas company organizes a comprehensive manual investigation based on the users with abnormal preliminary screening at this stage to verify the true theft gas behaviors and normal gas consumption behaviors.

[0011] The deep learning model aims to train and recommend theft gas behaviors based on the results of manual verification. Currently, the data dimensions of the model include the user's historical gas consumption and the user's name. The purpose of introducing the user's name is to expect the model to learn the user name features. On the one hand, as a filtering condition, it helps to exclude users who do not have the motivation for theft gas (such as government users). On the other hand, as an attention mechanism, it assists the model to focus on the gas consumption in relevant months (such as the gas consumption of school users during non-holiday periods). BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0013] Figure 1 It is the flowchart of the method of the present invention;

[0014] Figure 2 It is the schematic diagram of the deep learning network model of the present invention;

[0015] Figure 3 It is the schematic diagram of the comparative training of the deep learning model of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the present invention described and illustrated herein can be arranged and designed in a variety of different configurations. Therefore, the detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0017] It should be noted that similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present invention, terms such as "first" and "second" are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance, or implying any such actual relationship or order between these entities or operations. In addition, terms such as "connected" and "coupled" can be directly connected between components or indirectly connected via other components.

[0018] Embodiment 1:

[0019] The present invention is implemented through the following technical solutions. As Figure 1 shown, a method for recommending gas theft users based on deep learning includes the following steps:

[0020] Step 1, construct a preliminary screening model for gas theft to screen out users with abnormal preliminary screening. The preliminary screening model for gas theft includes a continuous zero gas consumption index, a conventional statistical index, a linear goodness-of-fit index, and an outlier index.

[0021] The preliminary screening model for gas theft consists of four indexes, namely, a continuous zero gas consumption index, a conventional statistical index, a linear goodness-of-fit index, and an outlier index. The function of the preliminary screening model for gas theft is to sort and screen users according to logical indexes. Each index extracts at most cnt abnormal users for making data labels for training a deep learning model.

[0022] (1) Continuous zero gas consumption index.

[0023] The continuous zero gas consumption index is based on the number of consecutive zero gas consumptions of users and the maximum daily gas consumption as the sorting basis.

[0024] In this embodiment, a time period of one year (365 days) is given, and the calculation unit is 1 day. If the gas consumption for a whole day is 0, it is recorded as one consecutive zero gas consumption. First, calculate the number of consecutive zero gas consumptions z of each user iAverage number of consecutive zero gas consumption times for all users :

[0025] where z i represents the number of consecutive zero gas consumption times of the i-th user, i = 1, 2,..., n; n represents the total number of users managed by a certain gas company; represents the average number of consecutive zero gas consumption times for all users.

[0026] Secondly, assume that n users are all users of a gas company. According to calculate the Poisson distribution of the number of zero gas consumption segments of the gas company and give a probability value P zero , calculate the lower limit T of the number of consecutive zero gas consumption times zero :

[0027] where j is an index variable with no actual physical meaning, and its value range is between [0, 100].

[0028] Then, judge for each user whether the number of zero gas consumption times z i is greater than the lower limit T zero , and regard the users with the number of consecutive zero gas consumption times z i greater than T zero as abnormal users:

[0029] where user zero represents the set of abnormal users.

[0030] Finally, according to the principle of "catching the big and letting go of the small", taking the median of the maximum daily gas consumption of users in the set user zero median(user zero ) as the threshold, further screen out the users with large gas consumption among the users to be checked, and at most recommend cnt abnormal users after sorting:

[0031] where u i represents the maximum daily gas consumption of the i-th user in the set user zero in the past year; Abnormal zero represents the set of users whose maximum daily gas consumption in one year is greater than the median median(user zero ). For example, the set user zeroThere are a total of 10 users, and each user has 365 daily gas consumptions. One maximum daily gas consumption of each user can be selected from them, so a total of 10 maximum daily gas consumptions can be obtained. After sorting them by size, the median is selected as the threshold median(user zero ); then compare the maximum daily gas consumption u zero of each user in the set user i with median(user zero ). If it is greater than median(user zero ), then include this user in the set Abnormal zero . It should be noted that if the number of people in the sequentially selected Abnormal zero exceeds cnt, then after sorting u i from large to small, select the corresponding top cnt users and include them in the set Abnormal zero .

[0032] (2) Conventional statistical indicators.

[0033] The conventional statistical indicators include the skewness index and the user's daily gas consumption index.

[0034] First, calculate the skewness index s i of each user in one year:

[0035] where x i,t represents the daily gas consumption of the i-th user on the t-th day in the past 1 year; represents the average daily gas consumption of the i-th user in the past 1 year.

[0036] Then, judge for each user whether the skewness index s i is less than 0. If it is less than 0, it means that the small-flow data in the user's daily gas consumption exceeds the quantity of the standard normal distribution, which may be caused by gas theft. Take the users with the skewness index s i less than 0 as abnormal users:

[0037] where user stas represents the set of abnormal users.

[0038] Finally, according to the principle of "catching the big and letting go of the small", using the median median(user stas ) of the maximum daily gas consumption of the users in the set user stas as the threshold, further screen out the users with large gas consumption among the users to be checked. After sorting, at most cnt abnormal users are recommended:

[0039] Among them, u i represents the maximum daily gas consumption of the i-th user in the set user stas in the past year; Abnormal stas represents the set of users whose maximum daily gas consumption in one year is greater than the median median(user stas ). For example, there are 11 users in the set user stas , and each user has 365 daily gas consumptions. One maximum daily gas consumption of each user can be selected from them, so a total of 11 maximum daily gas consumptions can be obtained. After sorting them by size, the median is selected as the threshold median(user stas ); then the maximum daily gas consumption u stas of each user in the set user i is compared with median(user stas ). If it is greater than median(user stas ), then this user is included in the set Abnormal stas . It should be noted that if the number of people in the sequentially selected Abnormal stas exceeds cnt, then after sorting u i from large to small, the corresponding first cnt users are selected and included in the set Abnormal stas .

[0040] (3) Linear fitting goodness-of-fit index.

[0041] Generally speaking, if a gas consumption curve is drawn with the day as the abscissa and the cumulative gas consumption (per day) as the ordinate, this curve is most likely a straight line. Therefore, taking the first reporting period as the zero point, the date and the cumulative gas consumption can be fitted, and the goodness-of-fit r i is used as an evaluation index. The goodness-of-fit r i ranges from 0 to 1. The larger the goodness-of-fit r i , the higher the linearity.

[0042] First, calculate the goodness-of-fit r i for each user:

[0043] Among them, y i,tRepresents the actual daily gas consumption of the i-th user on the t-th day; T is the number of days, and in this embodiment, T = 365; f i,t Represents the predicted daily gas consumption of the i-th user on the t-th day under the fitting model; , SS reg , SS tot , SS res Are all intermediate parameters.

[0044] Secondly, exclude users with a goodness of fit r i Greater than 0.95, and use users with a goodness of fit r i Less than or equal to 0.95 as abnormal users:

[0045] Among them, user fit Represents the set of abnormal users.

[0046] Finally, according to the principle of "grasping the big and letting go of the small", using the median of the maximum daily gas consumption of users in the set user fit median(user fit ) as the threshold, further screen out large gas consumption users among the users to be checked, and at most recommend cnt abnormal users after sorting:

[0047] Among them, u i Represents the maximum daily gas consumption of the i-th user in the set user fit in the past year; Abnormal fit Represents the set of users whose maximum daily gas consumption in one year is greater than the median median(user fit ). For example, there are 12 users in the set user fit , and each user has 365 daily gas consumptions. One maximum daily gas consumption of each user can be selected from them, so a total of 12 maximum daily gas consumptions can be obtained. After sorting by size, the median is selected as the threshold median(user fit ); then compare the maximum daily gas consumption u fit of each user in the set user i with median(user zfit ). If it is greater than median(user fit ), then include this user in the set Abnormal fit . It should be noted that if the number of people in the sequentially selected Abnormal fit exceeds cnt, then for u iAfter sorting from large to small, select the corresponding top cnt users and include them in the set Abnormal fit 。

[0048] (4)Outlier index.

[0049] Some macro factors (such as holidays, climate, etc.) will affect a group of users at the same time. These users have similar gas consumption characteristics and together form a characteristic group. When the gas consumption characteristics of a user are relatively isolated, they can be included in the list of abnormal users.

[0050] First, calculate the cosine similarity of the gas consumption of each user and another user based on the daily gas consumption as the basic data:

[0051] Among them, Similar a,b represents the cosine similarity between user a and user b; a t represents the daily gas consumption of user a on the t-th day; b t represents the daily gas consumption of user b on the t-th day.

[0052] Second, for a certain user i, use the minimum value of its cosine similarity to the remaining all users as the outlier index sim i :

[0053] Among them, sim i represents the outlier index of the i-th user; {Similar i,n-1} represents the set of cosine similarities between the i-th user and the remaining all users; n represents the total number of users.

[0054] Then, sort the users according to the outlier index sim i from large to small, and extract the top 2×cnt users as abnormal users:

[0055] Among them, user sim represents the set of the top 2×cnt users after sorting the outlier index sim i for all users; top(sim i , 2×cnt) represents the top 2×cnt users after sorting sim i from large to small.

[0056] Finally, according to the principle of "grasping the big and letting go of the small", use the median of the single-day maximum gas consumption of the users in the set user sim as median(user sim) is the threshold value. Further screen out the users with large gas consumption among the users to be investigated. After sorting, at most cnt abnormal users are recommended:

[0057] Among them, u i represents the maximum daily gas consumption of the i-th user in the set user sim in the past year; Abnormal sim represents the set of users whose maximum daily gas consumption in one year is greater than the median median(user sim ). For example, there are 13 users in the set user sim . Each user has 365 daily gas consumptions. One maximum daily gas consumption of the user can be selected from them, so a total of 13 maximum daily gas consumptions can be obtained. After sorting by size, the median is selected as the threshold median(user sim ); then compare the maximum daily gas consumption u sim of each user in the set user i with median(user sim ). If it is greater than median(user sim ), then add this user to the set Abnormal sim . It should be noted that if the number of users selected in Abnormal sim in sequence exceeds cnt, then after sorting u i from large to small, select the corresponding first cnt users and add them to the set Abnormal sim .

[0058] (5) The initial screening model for gas theft.

[0059] Take the union of the users selected by the above four indicators and count them as the initially screened abnormal users A:

[0060] Among them, A represents the initially screened abnormal users.

[0061] Step 2, construct a data set based on the initially screened abnormal users for training a deep learning model.

[0062] Suppose there are a total of B users (B = n), among which A users are initially screened as abnormal users. Then, through manual verification of these A initially screened abnormal users, it is found that among the A users, a1 are indeed gas theft users, and the remaining a2 are normal gas users. Then, the gas consumption data and their labels of the B users are used as the dataset for training the deep learning model. Among them, a1 are positive samples with the label of "gas theft users", and a2+(B - A) are negative samples with the label of "normal gas users". However, since it is the initial screening by the model and manual verification, there may still be gas theft users among the a2+(B - A) normal gas users. Therefore, it is necessary to perform the classification of the deep learning model in step 2 to improve the classification accuracy.

[0063] Please refer to Figure 2 , the deep learning model includes a backbone network and a classification network. In the backbone network, on the one hand, the user name is input into the bert language model. After extracting a 768-dimensional feature vector, it is then passed through a fully connected layer to generate a 32-dimensional user information feature vector. On the other hand, the daily gas consumption of a user for 28 days in a month is taken as a column and arranged into a user gas consumption table of 12 columns and 28 dimensions. The month to which each column of the user gas consumption table belongs forms a 7-dimensional feature vector through the embeding layer. The 7-dimensional feature vector is concatenated with the user gas consumption table to form a 12-column and 35-dimensional feature vector. Each column is 1 token. After passing these 12 tokens through 6 layers of the transformer layer and then through a fully connected layer, a 32-dimensional user gas consumption feature vector is extracted. The 32-dimensional user information feature vector and the 32-dimensional user gas consumption feature vector are concatenated to form a 64-dimensional feature vector, and then through a fully connected layer, a 32-dimensional feature vector for this user is output.

[0064] In the classification network, the 32-dimensional feature vector output by the backbone network passes through a fully connected layer to be reduced to an 8-dimensional feature vector, then passes through another fully connected layer to be reduced to a 2-dimensional feature vector, and finally passes through the softmax layer to output the user gas consumption classification result, and the classification result is "1" for gas theft users or "0" for normal gas users.

[0065] When training the deep learning model, since there are usually few training samples in the initial stage of business development, the entire training process is carried out in two stages. The first stage is contrast training. When performing contrast training, the classification network is disconnected. As Figure 3 shown, compare the 32-dimensional feature vectors of every two users output by the backbone network to make the users with the same label as similar as possible and the users with different labels dissimilar. The loss function used for comparison is the cosine similarity f similary :

[0066] Where a and b represent the 32-dimensional feature vectors of user a and user b output by the backbone network, respectively. User a and user b belong to any two of B users, so theoretically the model will learn B×(B-1) times cosine similarity.

[0067] Furthermore, samples with the same label in the data set used to train the deep learning model are defined as similar samples, and samples with different labels are defined as dissimilar samples. If a1 gas theft users and a2 normal gas users or users who do not need to be verified are marked among B users, without considering sample balance, in order to enhance the sample, a1×a1+a2×a2 positive samples and a1×(B-a1) negative samples can be formed. Because gas theft users are a case with a low probability after all, the samples randomly selected from the total samples can be considered as samples of normal gas users.

[0068] The second stage is classification training. During classification training, the backbone network and the classification network are opened at the same time to perform user classification training. The loss function used is the cross entropy loss function f classify :

[0069] Where N represents the number of samples; y i represents the true label of the i-th sample (y i =0 indicates normal gas users, y i =1 indicates a gas stealing user); represents the predicted probability.

[0070] Step 3: Input the user name and the user's historical gas usage data into the trained deep learning model, and output the recommended gas theft users.

[0071] When applying, the user's name and the user's gas consumption data for the past year are input into the trained deep learning model. When , the model recommends the user as a gas stealer, otherwise it recommends the user as a normal gas user.

[0072] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for recommending gas theft users based on deep learning, characterized in that, It includes the following steps: Step 1, construct a preliminary screening model for gas theft to screen out users with abnormal preliminary screening; the preliminary screening model for gas theft includes continuous zero gas consumption index, conventional statistical index, linear goodness-of-fit index, and outlier index; Step 2, construct a data set based on the users with abnormal preliminary screening for training a deep learning model; Step 3, input the user name and the historical gas consumption data of the user into the trained deep learning model, and output the recommended users of gas theft.

2. The method for recommending gas theft users based on deep learning according to claim 1, wherein In the said Step 1, the step of screening out abnormal users through the continuous zero gas consumption index is: Given a historical time period of T days, if the gas consumption for a day is zero, it is recorded as one instance of consecutive zero gas consumption; calculate the number of times z of consecutive zero gas consumption for each user i and the average number of times of consecutive zero gas consumption for all users : where z i is the number of consecutive zero gas consumption times of the i-th user, i = 1, 2, ..., n; n is the total number of users; is the average number of consecutive zero gas consumption times of all users; According to calculate the Poisson distribution of the number of segments of zero gas consumption, and set the probability value P zero , calculate the lower limit T of the number of consecutive zero gas consumption times zero : Where j is an index variable, and its value range is [0, 100]; The number of consecutive zero gas consumption times z i greater than T zero of the users is regarded as abnormal users: Among them, user zero represents the set of abnormal users; Taking the median of the maximum daily gas consumption of users in the set user zero as the threshold, at most cnt abnormal users are recommended after sorting: zero ) Among them, u i represents the maximum daily gas consumption of the i-th user in the set user zero in the historical time period; Abnormal zero represents the set of users whose maximum daily gas consumption in the historical time period is greater than the median median(user zero ).

3. The method for recommending gas theft users based on deep learning according to claim 2, characterized in that In the said Step 1, the step of screening out abnormal users through the conventional statistical index is: Calculate the skewness index s for each user in the historical time period i : where x i,t represents the daily gas consumption of the i-th user on the t-th day in the past year, where t = 1, 2,..., T; represents the average daily gas consumption of the i-th user in the past year; Take the skewness index s i Users less than 0 as abnormal users: Among them, user stas represents the set of abnormal users; Taking the median of the maximum daily gas consumption of users in the set user stas median(user stas ) as the threshold, at most cnt abnormal users are recommended after sorting: where, u i represents the maximum daily gas consumption of the i-th user in the set user stas during the historical time period; Abnormal stas represents the set of users whose maximum daily gas consumption during the historical time period is greater than the median median(user stas ).

4. The method for recommending gas theft users based on deep learning according to claim 3, characterized in that In the said Step 1, the step of screening out abnormal users through the linear goodness-of-fit index is: Calculate the goodness of fit r for each user i : Among them, y i,t represents the actual daily gas consumption of the i-th user on the t-th day; f i,t represents the predicted daily gas consumption of the i-th user on the t-th day under the fitting model; , SS reg , SS tot , SS res are all intermediate parameters; The goodness-of-fit r i Less than or equal to 0.95 is used as an abnormal user: Among them, user fit represents the set of abnormal users; Taking the median of the maximum daily gas consumption of users in the set user fit median(user fit ) as the threshold, at most cnt abnormal users are recommended after sorting: Among them, u i represents the maximum daily gas consumption of the i-th user in the set user fit during the historical time period; Abnormal fit represents the set of users whose maximum daily gas consumption during the historical time period is greater than the median median(user fit ).

5. The method for recommending gas theft users based on deep learning according to claim 4, wherein In the said Step 1, the step of screening out abnormal users through the outlier index is: Calculate the cosine similarity of gas consumption between each user and another user: Among them, Similar a,b represents the cosine similarity between user a and user b; a t represents the daily gas consumption of user a on the t-th day; b t represents the daily gas consumption of user b on the t-th day; For user i, the minimum value of its cosine similarity to all the remaining users is used as the outlier index sim of this user i : where, sim i represents the outlier index of the i-th user; {Similar i,n-1} represents the set of cosine similarities between the i-th user and all the remaining users; Sort the users according to the outlier index sim i in descending order, and extract the top 2×cnt users as outlier users: Among them, user sim represents the set of the top 2×cnt users after sorting the outlier index sim i for all users; top(sim i , 2×cnt) represents the top 2×cnt users after sorting sim i in descending order; Taking the median of the maximum daily gas consumption of users in the set user sim median(user sim ) as the threshold, at most cnt abnormal users are recommended after sorting: Among them, u i represents the maximum daily gas consumption of the i-th user in the set user sim during the historical time period; Abnormal sim represents the set of users whose maximum daily gas consumption during the historical time period is greater than the median median(user sim ).

6. The method for recommending gas theft users based on deep learning according to claim 5, characterized in that In the said Step 1, the users with abnormal preliminary screening are: Where A represents the users with abnormal preliminary screening.

7. The method for recommending gas theft users based on deep learning according to claim 6, wherein In the said Step 2, the step of constructing a data set based on the users with abnormal preliminary screening includes: Suppose there are n users in total, among which A users with abnormal preliminary screening are selected, and then the A users with abnormal preliminary screening are verified manually. It is verified that among the A users, a1 are users of gas theft, and the remaining a2 are normal gas consumption users; then the gas consumption data and their labels of the n users are used as the data set for training the deep learning model, where a1 are positive samples, and a2+(B - A) are negative samples.

8. The method for recommending gas theft users based on deep learning according to claim 1, characterized in that The said deep learning model includes a backbone network and a classification network; the operation process of the backbone network is: Input the user name into the bert language model of the backbone network. After extracting a 768-dimensional feature vector, it is then passed through a fully connected layer to generate a 32-dimensional user information feature vector; take the daily gas consumption of 28 days in a month of a user as a column, and arrange it into a 12-column 28-dimensional user gas consumption table. The month to which each column of the user gas consumption table belongs forms a 7-dimensional feature vector through the embeding layer. Concatenate the 7-dimensional feature vector with the user gas consumption table to form a 12-column 35-dimensional feature vector, with each column being 1 token. After passing these 12 tokens through 6 layers of the transformer layer and then through a fully connected layer, extract 1 32-dimensional user gas consumption feature vector; concatenate the 32-dimensional user information feature vector and the 32-dimensional user gas consumption feature vector to form a 64-dimensional feature vector, and then output a 32-dimensional feature vector for this user after passing through a fully connected layer; The operation process of the classification network is: Pass the 32-dimensional feature vector output by the backbone network through a fully connected layer to reduce the dimension to an 8-dimensional feature vector, then pass through a fully connected layer to reduce the dimension to a 2-dimensional feature vector, and finally pass through the softmax layer to output the user gas consumption classification result, and the classification result is a user of gas theft or a normal gas consumption user.

Images