Online car-hailing abnormal order detection method and device based on auto-encoder

Abstract

The invention provides an online car-hailing abnormal order detection method based on an auto-encoder, and the method comprises the following steps: S1, training an offline recognition model: S11, obtaining order data within 180 days, the order data being collectively called by order issuing data and order completion data; s12, order data feature engineering is carried out, and order features are constructed; s13, performing order recording feature engineering; s14, training an auto-encoder model; s2, data prediction and abnormal order identification; S21, feature engineering is carried out on the completed orders in real time; s22, for each piece of order data, coding and decoding are carried out by using self-encoder network alignment, and a decoded result is obtained; according to the method, the identification capability and adaptability to various cheating means are enhanced, the accuracy of an identification result is improved, and the final purpose is that the abnormal order can be identified and detected as completely as possible while the algorithm design, starting and operation maintenance cost is reduced as much as possible.

Description

Technical Field

[0001] The present invention relates to the technical fields of data processing and intelligent transportation, and particularly relates to a method and device for detecting abnormal orders of online car-hailing based on an autoencoder. Background Art

[0002] In the existing online car-hailing industry, common abnormal order detection techniques include methods such as rule-based abnormal order detection, statistics-based abnormal order detection, and machine learning-based abnormal order detection.

[0003] Rule-based abnormal order detection is the most common method. By setting a series of business rules, such as the order price exceeding the normal range, the order distance exceeding a certain threshold, and a large number of orders being initiated by the same user in a short period of time, once a situation violating the rules occurs, it is determined as an abnormal order. Rule-based abnormal order detection is simple and easy to understand, but it also has limitations because the manifestations of abnormal orders may vary widely, and it is difficult for rules to cover all situations.

[0004] Statistics-based abnormal order detection is to perform statistical analysis on historical order data to find orders that conform to abnormal characteristics. For example, by performing statistical analysis on parameters such as order price, travel time, and distance, abnormal orders that deviate from the normal distribution are found. Statistics-based abnormal order detection has high requirements for historical data. If there is not enough data support, there may be deviations.

[0005] Machine learning-based abnormal order detection is more intelligent. By establishing a model to learn the differences between normal orders and abnormal orders, abnormal orders can be better identified. Common methods include SVM, LightGBM, etc. For example, by establishing a model, the risk probability of different order characteristics is predicted, and orders are screened according to the threshold based on the risk probability to identify abnormal orders. Machine learning-based abnormal order detection is more intelligent, can learn and optimize independently, but also requires a large amount of data support and a professional technical team for research and development and maintenance. At the same time, when training a machine learning model, a huge amount of manual annotation of abnormal orders is required, and the cost is extremely high. Therefore, a method and device for detecting abnormal orders of online car-hailing based on an autoencoder are proposed. Summary of the Invention

[0006] In view of this, embodiments of the present invention hope to provide a method and device for detecting abnormal orders of online car-hailing based on an autoencoder to solve or alleviate the technical problems existing in the prior art and at least provide a beneficial option.

[0007] The technical solution of the embodiments of the present invention is implemented as follows: A method for detecting abnormal orders of online car-hailing based on an autoencoder includes the following steps:

[0008] S1. Train the offline recognition model:

[0009] S11. Obtain the order data within 180 days, where the order data is the general term for the data of order issuance and order completion;

[0010] S12. Conduct feature engineering on the order data to construct order features;

[0011] S13. Conduct feature engineering on the order recordings;

[0012] S14. Train the autoencoder model;

[0013] S2. Data prediction and identification of abnormal orders:

[0014] S21. Conduct feature engineering on the orders after completion in real time;

[0015] S22. For each order data, use the autoencoder network alignment to encode and decode to obtain the decoded result;

[0016] S23. Compare the differences between the order data and the decoded data, generally using the Euclidean distance and the Mahalanobis distance for calculation;

[0017] S24. Set a threshold, and identify the orders with differences greater than the threshold as suspected abnormal orders;

[0018] S25. Conduct further manual review on the suspected abnormal orders or use other models for refined analysis.

[0019] In some embodiments, in S12, the order features are constructed as follows:

[0020] S121. Order context features;

[0021] S122. Statistical features of passengers;

[0022] S123. Statistical features of drivers;

[0023] S124. Supply and demand features at the moment when the order is placed;

[0024] S125. Features of the marketing activities involved in the order;

[0025] Thus, a dataset is generated.

[0026] In some embodiments, in S121, the order context features are as follows:

[0027] S121-1. The moment when the order is placed;

[0028] S121-2. The estimated pick-up time / drop-off time / drop-off distance of the order;

[0029] S121-3, actual pick-up time / drop-off time / drop-off distance of the order;

[0030] S121-4, ratio of the estimated time / distance to the actual time / distance for order drop-off;

[0031] S121-5, current estimated trip fee / actual trip fee / actual surcharge of the order;

[0032] S121-6, estimated price per minute / price per kilometer of the order;

[0033] S121-7, city ID / tenant ID of the order;

[0034] S121-8, whether the day when the order is placed is a working day;

[0035] S121-9, duration from order placement to response / duration from response to driver's arrival at the starting point / duration from driver's arrival to passenger boarding / duration from passenger's arrival at the destination to passenger payment of the order;

[0036] S121-10, number of driver-passenger calls / times of the order;

[0037] S121-11, number of driver-passenger matching times of the order;

[0038] S121-12, number of tenants selected by the passenger when placing the order.

[0039] In some embodiments, in S122, the statistical characteristics of the passenger are as follows:

[0040] S122-1, probability that the passenger selects this tenant in the past 7 days / 30 days;

[0041] S122-2, proportion of orders in which the passenger takes a taxi within 2 km around the starting point of the current order in the past 7 days / 30 days;

[0042] S122-3, proportion of orders in which the passenger gets off within 2 km around the end point of the current order in the past 7 days / 30 days;

[0043] S122-4, minimum value, maximum value and average value of the taxi order prices of the passenger in the past 7 days / 30 days.

[0044] In some embodiments, in S123, the statistical characteristics of the driver are as follows:

[0045] S123-1, number of marketing activity rewards obtained by the driver in the past 30 days;

[0046] S123-2. Compare the reward amount obtained by the driver in the past 30 days with the positions of other drivers in the same city and tenant (whether they are drivers who won awards synchronously).

[0047] In some embodiments, in S124, the supply-demand characteristics at the moment of the order are as follows:

[0048] S124-1. The number of main orders of passengers in the current city at the current moment;

[0049] S124-2. The number of drivers who have listened to orders in the current city at the current moment;

[0050] S124-3. The supply-demand ratio of passengers to drivers in the current city at the current moment.

[0051] In some embodiments, in S125, the characteristics of the marketing activities involved in the order are as follows:

[0052] S125-1. Whether the order is the first order / last order / a single order before awarding of the activity;

[0053] S125-2. Whether there is an award for the activity involved in the order;

[0054] S125-3. The position of the award amount of the activity involved in the order among the award amounts of the same activity (whether it is the head award activity);

[0055] S125-4. The average award amount per order corresponding to the activity involved in the order;

[0056] S125-5. The number and proportion of starting price orders in the activity involved in the order;

[0057] S125-6. The interval time between the order and the previous order with the same design of the current activity. In some embodiments, in S13, the specific operation steps for performing order recording feature engineering are as follows:

[0058] S131. Process the trip recordings corresponding to the orders in the dataset and convert the audio into a spectrogram;

[0059] S132. Use an open-source speech recognition model to recognize the human voice, door opening and closing sounds, and vehicle driving wind noise in the audio;

[0060] S133. Incorporate the recognition results as features into the order features.

[0061] In some embodiments, in S14, the specific operation steps for training the autoencoder model are as follows:

[0062] S141. Randomly extract a specific number of order data with constructed features from the dataset and substitute them into the autoencoder neural network for training;

[0063] S142. Since the proportion of abnormal orders in daily online car-hailing orders is extremely small, in the selected training dataset, abnormal orders will not have an obvious impact on the autoencoder. The autoencoder can learn the data feature performance of normal orders and the performance of normal order data in the low-dimensional space during the training process.

[0064] A detection device for abnormal online car-hailing orders based on an autoencoder, comprising:

[0065] Autoencoder: Used to take order data as feature engineering through statistical and combinatorial methods, combine the parsing results of the trip recordings in the order using a speech recognition model with the feature engineering to form order features, then substitute the order features into the autoencoder for encoding and decoding to obtain a decoding result, and compare the original feature data with the decoded data to determine whether the order is abnormal;

[0066] The autoencoder is trained using a large amount of normal data and can be used as a mapping between the normal features of the order (high-dimensional space) and the features after data compression (low-dimensional space). For normal order data, the encoded and decoded data of the autoencoder is similar to the original data. For abnormal order data, more information will be lost during the encoding and decoding process of the autoencoder, and the decoded data will be quite different from the original data.

[0067] Due to the adoption of the above technical solutions in the embodiments of the present invention, it has the following advantages:

[0068] First, through this method, the present invention enhances the recognition ability and adaptability to various cheating means, improves the accuracy of the recognition result. The ultimate goal is to identify and detect abnormal orders as comprehensively as possible while minimizing the algorithm design, startup, and operation and maintenance costs.

[0069] Second, through this method, the present invention can achieve automatic feature extraction. The autoencoder can automatically learn effective features from the original data without excessive manual feature design, greatly simplifying the task.

[0070] Third, through this method, the present invention can perform efficient data compression. Through the bottleneck structure of the middle layer, the autoencoder can achieve effective data compression, reduce the computational complexity, and improve the processing speed.

[0071] Fourth, through this method, the present invention has a high abnormal detection ability. During the training process, the autoencoder will try its best to maintain the consistency between the input data and the output data. When there is a large deviation in the input data (i.e., an abnormal order), since the model cannot reconstruct these data well, there will be a large error in the output, thus achieving the purpose of abnormal detection.

[0072] V. The present invention has strong adaptability through this method. The autoencoder has low requirements for data types and can process various types of data including text, images, etc., which gives it certain advantages in processing various types of order information.

[0073] VI. The present invention can handle non - linear problems through this method. The autoencoder can fit complex non - linear relationships through a multi - layer neural network, which is very useful when dealing with order problems involving multiple factors.

[0074] VII. The autoencoder of the present invention through this method can be used as a pre - processing step for feature selection or dimensionality reduction, and then the results are input into other classification or regression models.

[0075] The above summary is only for the purpose of the specification and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the present invention will be readily apparent by reference to the drawings and the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] To more clearly illustrate the technical solutions in the embodiments of the present application or in the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0077] Figure 1 It is the schematic diagram of the autoencoder of the present invention;

[0078] Figure 2 It is the flowchart of the autoencoder of the present invention for identifying abnormal orders. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0079] In the following, only some exemplary embodiments are simply described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the present invention. Therefore, the drawings and the description are considered to be exemplary in nature rather than restrictive.

[0080] It should be noted that terms such as "first", "second", "symmetric", "array", etc. are only used for the purpose of distinguishing descriptions and position descriptions, and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, those limited by features such as "first", "symmetric", etc. may explicitly or implicitly include one or more of such features; similarly, when there is no numerical limitation on certain features in the form of words such as "two", "three", etc., it should be noted that such features also belong to those that explicitly or implicitly include one or more feature quantities;

[0081] In the present invention, unless otherwise clearly stipulated and defined, terms such as "install", "connect", "fix", etc. should be understood in a broad sense; for example, it may be a fixed connection, a detachable connection, or an integrally formed one; it may be a mechanical connection, a direct connection, a welding connection, or an indirect connection through an intermediate medium, and it may be the communication inside two components or the interaction relationship between two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood in combination with the specific circumstances according to the accompanying drawings of the specification.

[0082] Glossary:

[0083] 1. Tenant: The online car-hailing operation company to which the driver belongs. When a passenger places an online car-hailing order on the aggregation platform, the passenger can choose which tenants' orders they want to take, and the pricing methods and standards of different tenants are different;

[0084] 2. Aggregation platform: An online car-hailing platform composed of multiple tenants. When taking a car on this platform, passengers can choose online car-hailing services from different tenants according to their preferences;

[0085] 3. Driver-passenger matching times: The total number of online car-hailing orders that the online car-hailing driver corresponding to the order and the passenger have matched and completed on the current aggregation platform in the past 90 days;

[0086] 4. Order features: Some feature data in the order placement process, such as the pick-up / drop-off distance, pick-up / drop-off time, estimated price, waiting time for the passenger after placing the order until the response, and waiting time for the driver after arriving at the destination until the passenger;

[0087] 5. Order placement process: The process of placing an online car-hailing order, which is divided into passenger placing an order -> order response / driver receiving the order -> driver arriving at the order starting point -> passenger getting on the car -> starting the journey -> arriving at the order end point -> driver confirming the fare -> passenger paying the fare;

[0088] 6. Drop-off time: The estimated time for the driver to drive from the order starting point to the end point;

[0089] 7. Drop-off distance: The estimated distance for the driver to drive from the order starting point to the end point;

[0090] 8. Pick-up time: The estimated time for the driver to drive from the pick-up starting point to the order starting point;

[0091] 9. Abnormal order: A false order completed by the driver using technical or rule means in order to boost the system rating, get good reviews, obtain rewards, etc. For example, the driver uses a purchased mobile phone number to place orders for himself multiple times, then drives empty, and then gives himself good reviews;

[0092] 10. Order dispatch: When a passenger places an order and selects the tenant they want to take a taxi from, the system will send the order to these tenants to match a driver;

[0093] 11. Order completion: After a passenger places an online car-hailing order and a driver is matched, the passenger is delivered, and the passenger completes the payment for the order;

[0094] 12. Trip fee: The fee that the passenger should pay for this trip, calculated based on the driving time and distance of the order;

[0095] 13. Surcharge: The driver's parking fee before driving, the highway toll when passing through the highway during driving, and some other additional fees that the passenger needs to pay;

[0096] 14. Listening for orders: The online car-hailing driver is online and in a state where they can receive orders, idle, and is waiting for the system to dispatch or broadcast orders;

[0097] 15. Main order: In the aggregation platform, the online car-hailing order placed by the passenger is the main order, and this main order includes the passenger's starting point and destination;

[0098] 16. Response: After the passenger places an order and the system matches a driver for the order, the process of notifying the passenger and the driver of the matching result;

[0099] 17. System order dispatch: After the online car-hailing system matches a driver for an online car-hailing order, the process of dispatching the order to the driver and notifying the driver to pick up the passenger;

[0100] 18. Supply-demand ratio: The ratio of the number of online car-hailing drivers who can receive orders to the number of passengers who are currently taking a taxi.

[0101] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0102] As Figure 1 - Figure 2 shown, the embodiments of the present invention provide a method for detecting abnormal online car-hailing orders based on an autoencoder, including the following steps:

[0103] S1. Train an offline recognition model:

[0104] S11. Obtain order data within 180 days, where order data is the collective term for order dispatch and order completion data;

[0105] S12. Perform order data feature engineering to construct order features;

[0106] S13. Perform order recording feature engineering;

[0107] S14. Train the autoencoder model;

[0108] S2. Data prediction and abnormal order identification:

[0109] S21. Perform feature engineering on the orders after completion in real time;

[0110] S22. For each order data, use the autoencoder network alignment to encode and decode to obtain the decoded result;

[0111] S23. Compare the differences between the order data and the decoded data, generally using Euclidean distance and Mahalanobis distance to calculate;

[0112] S24. Set a threshold, and identify the orders with differences greater than the threshold as suspected abnormal orders;

[0113] S25. Conduct further manual review on the suspected abnormal orders or perform refined analysis using other models.

[0114] In this embodiment, specifically, in S12, the order features are constructed as follows:

[0115] S121. Order context features;

[0116] S122. Statistical features of passengers;

[0117] S123. Statistical features of drivers;

[0118] S124. Supply and demand features at the time when the order is placed;

[0119] S125. Features of the marketing activities involved in the order;

[0120] Thus, a data set is generated.

[0121] In this embodiment, specifically, in S121, the order context features are as follows:

[0122] S121-1. The time when the order is placed;

[0123] S121-2. The estimated pick-up time / drop-off time / drop-off distance of the order;

[0124] S121-3. The actual pick-up time / drop-off time / drop-off distance of the order;

[0125] S121-4, Ratio of the estimated driving time / distance of the order to the actual time / distance;

[0126] S121-5, Current estimated trip fee / actual trip fee / actual surcharge of the order;

[0127] S121-6, Estimated average price per minute / average price per kilometer of the order;

[0128] S121-7, City ID / Tenant ID where the order is located;

[0129] S121-8, Whether the day when the order is placed is a working day;

[0130] S121-9, Duration from order placement to response / duration from response to driver's arrival at the starting point / duration from driver's arrival to passenger boarding / duration from passenger's arrival at the destination to passenger's payment;

[0131] S121-10, Number of driver-passenger calls / times of the order;

[0132] S121-11, Number of driver-passenger matching times of the order;

[0133] S121-12, Number of tenants selected by the passenger when placing the order.

[0134] In this embodiment, specifically, in S122, the statistical characteristics of the passenger are as follows:

[0135] S122-1, Probability that the passenger has selected this tenant in the past 7 days / 30 days;

[0136] S122-2, Proportion of orders where the passenger took a taxi within 2 km around the starting point of the current order in the past 7 days / 30 days;

[0137] S122-3, Proportion of orders where the passenger got off within 2 km around the ending point of the current order in the past 7 days / 30 days;

[0138] S122-4, Minimum, maximum and average prices of the passenger's taxi orders in the past 7 days / 30 days.

[0139] In this embodiment, specifically, in S123, the statistical characteristics of the driver are as follows:

[0140] S123-1, How many marketing activity rewards the driver has obtained in the past 30 days;

[0141] S123-2, Position of the driver's obtained reward amount compared with other drivers in the same city and tenant in the past 30 days (whether the driver is a synchronously rewarded driver).

[0142] In this embodiment, specifically, in S124, the supply-demand characteristics at the moment of the order are as follows:

[0143] S124-1. The number of main orders of passengers in the current city at the current moment;

[0144] S124-2. The number of drivers who have listened to orders in the current city at the current moment;

[0145] S124-3. The supply-demand ratio of passengers to drivers in the current city at the current moment.

[0146] In this embodiment, specifically, in S125, the characteristics of the marketing activities involved in the order are as follows:

[0147] S125-1. Whether the order is the first order / last order / a single order before awarding of the activity;

[0148] S125-2. Whether there is an award for the activity involved in the order;

[0149] S125-3. The position of the award amount of the activity involved in the order among the award amounts in the same activity (whether it is the head award activity);

[0150] S125-4. The average award amount per order corresponding to the activity involved in the order;

[0151] S125-5. The number and proportion of starting price orders in the activity involved in the order;

[0152] S125-6. The interval time between the order and the previous order with the same design of the current activity.

[0153] In this embodiment, specifically, in S13, the specific operation steps for order recording feature engineering are as follows:

[0154] S131. Process the trip recording corresponding to the order in the dataset and convert the audio into a spectrogram;

[0155] S132. Use an open-source speech recognition model to recognize the human voice, door opening and closing sounds, and vehicle driving wind noise in the audio;

[0156] S133. Incorporate the recognition results as features into the order features.

[0157] In this embodiment, specifically, in S14, the specific operation steps for training the autoencoder model are as follows:

[0158] S141. Randomly extract a specific number of order data with constructed features from the dataset and substitute them into the autoencoder neural network for training;

[0159] S142. Since the proportion of abnormal orders in daily online car-hailing orders is extremely small, in the selected training dataset, abnormal orders will not have an obvious impact on the autoencoder. The autoencoder can learn the data feature representations of normal orders and the representations of normal order data in the low-dimensional space during the training process.

[0160] A detection device for abnormal online car-hailing orders based on an autoencoder, comprising:

[0161] An autoencoder: used to take order data as feature engineering through statistical and combinatorial methods, combine the parsing results of the trip recording in the order by using a speech recognition model with the feature engineering to form order features, then substitute the order features into the autoencoder for encoding and decoding to obtain a decoding result, and compare the original feature data with the decoded data to determine whether the order is abnormal;

[0162] The autoencoder is trained using a large amount of normal data and can serve as a mapping between the normal features of the order (high-dimensional space) and the features after data compression (low-dimensional space). For normal order data, the data after encoding and decoding by the autoencoder is similar to the original data. For abnormal order data, more information will be lost during the encoding and decoding process of the autoencoder, and the decoded data is quite different from the original data.

[0163] In this embodiment, specifically, when training the model, the specific features constructed in the feature engineering can be different, or some features can be added, modified, or deleted, and the construction method can also be adjusted accordingly;

[0164] The analysis of the trip speech is added to the input features of the autoencoder, and this part can also be replaced with others, such as the analysis of the driving trajectory and the analysis of the driving speed;

[0165] The structure of the autoencoder can be different, and the basic autoencoder can be replaced with a sparse autoencoder, a convolutional autoencoder, a variational autoencoder, etc.

[0166] As described above, the above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various changes or substitutions, and these should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claimed rights.

Claims

1. A method for detecting abnormal online car-hailing orders based on an autoencoder, characterized in that: The following steps are involved: S1. Training offline recognition model: S11. Obtain order data within 180 days, which includes order issuance and order completion data; S12. Perform feature engineering of order data and construct order features; S13. Perform feature engineering of order recordings; S14. Train the autoencoder model; S2. Data prediction and abnormal order identification: S21. Perform feature engineering on completed orders in real time; S22. For each order data, use the autoencoder network alignment to encode and decode, and obtain the decoded result; S23. Compare the difference between the order data and the decoded data, generally using Euclidean distance or Mach-Cherov distance to calculate; S24. Set a threshold and identify orders with a difference greater than the threshold as suspected abnormal orders; S25. Conduct further manual review of suspected abnormal orders or use other models for detailed analysis.

2. The method for detecting abnormal online car-hailing orders based on an autoencoder according to claim 1 is characterized in that: In S12, order features are constructed, and the specific classification is as follows: S121, order context features; S122. Statistical characteristics of passengers; S123, driver's statistical characteristics; S124, supply and demand characteristics at the time of the order; S125. Characteristics of the marketing activities involved in the order; This generates a dataset.

3. The method and device for detecting abnormal online car-hailing orders based on an autoencoder according to claim 2, characterized in that: In S121, the order context features are specifically as follows: S121-1, the time when the order is placed; S121-2, estimated pick-up time / drop-off time / drop-off distance of the order; S121-3, actual pick-up time / drop-off time / drop-off distance of the order; S121-4, the ratio of the estimated time / distance for order delivery to the actual time / distance; S121-5, the current estimated trip fee / actual trip fee / actual surcharge of the order; S121-6, estimated average price per minute / average price per kilometer; S121-7, the city ID / tenant ID of the order; S121-8, whether the order is placed on a working day; S121-9, the time from order placement to response / the time from order response to the driver’s arrival at the starting point / the time from the driver’s arrival to the passenger’s boarding / the time from the passenger’s arrival at the destination to the passenger’s payment; S121-10, the number / time of calls between the driver and the passenger of the order; S121-11, the number of driver-passenger matching times for the order; S121-12. The number of tenants selected by the passenger when placing the order.

4. The method for detecting abnormal online car-hailing orders based on an autoencoder according to claim 2, characterized in that: In S122, the passenger's statistical characteristics are specifically as follows: S122-1. The probability of the passenger circling this tenant in the past 7 days / 30 days; S122-2, the proportion of taxi orders that passengers have taken within 2km of the starting point of the current order in the past 7 days / 30 days; S122-3. The proportion of orders in which passengers got off within 2 km of the current order destination in the past 7 days / 30 days; S122-4. The minimum, maximum and average prices of taxi orders taken by passengers in the past 7 days / 30 days.

5. The method and device for detecting abnormal online car-hailing orders based on an autoencoder according to claim 2, characterized in that: In S123, the driver's statistical characteristics are as follows: S123-1. How many marketing activity rewards did the driver receive in the past 30 days? S123-2. The driver obtains the reward amount in the past 30 days and compares it with other drivers in the same city and tenant (whether they are drivers who won the award at the same time).

6. The method for detecting abnormal online car-hailing orders based on an autoencoder according to claim 2, characterized in that: In S124, the supply and demand characteristics of the order at the time are as follows: S124-1. The number of main orders of passengers in the current city at the current time; S124-2, the number of drivers who have listened to orders in the current city at the current time; S124-3. The supply and demand ratio of passengers and drivers in the current city at the current moment.

7. The method and device for detecting abnormal online car-hailing orders based on an autoencoder according to claim 2, characterized in that: In S125, the characteristics of the marketing activities involved in the order are as follows: S125-1. Is the order the first order / last order / order before the award ceremony of the activity? S125-2. Whether there are any awards for the activities involved in the order; S125-3, the position of the activity bonus amount involved in the order in the same activity (whether it is the first bonus activity); S125-4, the average reward amount corresponding to the activities involved in the order; S125-5. The number and proportion of starting price orders in the activities involved in the order; S125-6. The interval between the order and the previous order with the same design as the current activity.

8. The method and device for detecting abnormal online car-hailing orders based on an autoencoder according to claim 1, characterized in that: In S13, the specific operation steps for order recording feature engineering are as follows: S131, processing the trip recordings corresponding to the orders in the data set, and converting the audio into a spectrogram; S132. Use an open source speech recognition model to recognize human voices, door opening and closing sounds, and vehicle driving noise in the audio; S133. Use the recognition result as a feature and incorporate it into the order feature.

9. The method for detecting abnormal online car-hailing orders based on an autoencoder according to claim 1, characterized in that: In S14, the specific operation steps of training the autoencoder model are as follows: S141. Randomly extract a specific number of order data with constructed features from the data set, and substitute them into the autoencoder neural network for training; S142. Since abnormal orders account for a very small proportion of daily online ride-hailing orders, abnormal orders will not have a significant impact on the autoencoder in the selected training data set. The autoencoder can learn the data feature performance of normal orders and the performance of normal order data in low-dimensional space during the training process.

10. A detection device for abnormal online car-hailing orders based on an autoencoder, characterized in that: include: Autoencoder: It is used to use the order data as feature engineering by statistical and combined methods. The speech recognition model is used to combine the analysis results of the trip recording in the order with feature engineering to form order features. The order features are then substituted into the autoencoder for encoding and decoding to obtain the decoding results. The original feature data is compared with the decoded data to determine whether the order is abnormal. The autoencoder is trained with a large amount of normal data, which can be used as a mapping between the normal features of an order (high-dimensional space) and the features after data compression (low-dimensional space). For normal order data, the encoded and decoded data of the autoencoder are similar to the original data. For abnormal order data, the autoencoder will lose more information during the encoding and decoding process, and the decoded data will be quite different from the original data.

Images