Order multi-dimensional abnormality monitoring, identification and comprehensive analysis system based on artificial intelligence

Abstract

The application discloses an order multidimensional abnormality monitoring, recognition and comprehensive analysis system based on artificial intelligence, which comprises an order data acquisition module, a path deviation triggering module, a remark counter-evidence verification module, an order value deviation analysis module and a deviation source marking module; the system acquires basic data, remark text and actual travel trajectory of an order to be recognized, analyzes the remark text to generate a remark pointing object when the actual trajectory meets the deviation triggering condition relative to the recommended travel path, and extracts a deviation trajectory object from the actual trajectory; the remark explanation validity is calculated, the deviation degree of the order mileage, time length and amount relative to the reference interval of similar orders is combined to generate an order value deviation degree, and then a path deviation source label such as a reasonable deviation of the remark, a partial explanation deviation of the remark, an unreasonable deviation of the remark or suspected detour is output; thus, the remark text can be avoided to be directly used as a detour exemption basis, and the accuracy and interpretability of order path abnormality recognition are improved.

Description

Technical Field

[0001] This invention relates to the fields of order risk control, trajectory analysis, natural language semantic parsing and anomaly recognition, and in particular to an artificial intelligence-based multi-dimensional anomaly monitoring, recognition and comprehensive analysis system for orders. Background Technology

[0002] With the development of ride-hailing, online car-hailing, instant delivery, and other location-based order platforms, identifying route anomalies during order fulfillment has gradually become a crucial aspect of platform risk control. Existing order anomaly detection systems typically rely on the actual driving trajectory, origin and destination, mileage, driving time, order amount, and historical data of similar orders to determine whether the driver has taken detours, made unusual stops, inflated mileage, or delayed fulfillment. This approach can identify some orders that significantly deviate from the recommended route or exceed historical reasonable ranges, but its judgment criteria usually focus on trajectory data and numerical order metrics.

[0003] In actual order fulfillment scenarios, users often specify special fulfillment requirements through remarks when placing an order. For example, a user might request the driver to pick up passengers at locations such as the north gate of the residential complex, the exit of the underground parking garage, the back door of a hotel, or the entrance to a parking lot. They might also request the driver to wait, take a temporary detour, pick up accompanying passengers first, or make a brief stop at a specific location. These remarks can affect the actual driving trajectory of the order, causing it to deviate from the platform's default recommended route. They may also alter the order's dwell time, driving distance, or order amount.

[0004] Existing methods for identifying route anomalies, relying solely on route deviation rate, mileage discrepancy, time deviation, or amount discrepancy, are prone to misclassifying reasonable deviations caused by genuine user requests as malicious detours by drivers. Furthermore, if the system simply detects semantics like "detour," "waiting," "doorway," or "parking lot" in the order notes and immediately reduces the risk of route deviation, new misjudgment problems may arise. Drivers or parties involved in abnormal orders might use vague notes as explanations for detours; even if the actual deviated route, stop point, and order status do not match the notes, the system may still incorrectly classify the deviation as reasonable.

[0005] Therefore, the existing technology has at least the following shortcomings: when the order trajectory deviates from the path, there is a lack of a verification mechanism to check whether the order remarks text truly interprets the path deviation, especially the lack of technical means to perform consistency matching of multiple remarks; this may lead to the remarks text being incorrectly used as a basis for reducing the risk of detours, thereby affecting the accuracy and interpretability of the path anomaly identification results. Summary of the Invention

[0006] This invention provides an AI-based multi-dimensional anomaly monitoring, identification, and comprehensive analysis system for orders. When an order's actual driving trajectory deviates from its course, the system verifies whether the order's notes can explain the deviation. This reduces the likelihood of misjudging reasonable deviations as detours or using invalid notes as a basis for reducing detour risks.

[0007] The system described in this invention can be used as a path offset source identification subsystem in a multi-dimensional order anomaly monitoring, identification, and comprehensive analysis system, or it can be deployed in conjunction with driver profiling, user profiling, license plate verification, and peak / valley statistics modules. The artificial intelligence is at least used for semantic parsing of order remarks text, or to assist in determining the semantic type of remarks after training a model on historical remarks text and path offset source tags.

[0008] To achieve the above objectives, the present invention provides an artificial intelligence-based multi-dimensional anomaly monitoring, identification and comprehensive analysis system for orders. The system includes an order data acquisition module, a path offset triggering module, a remark verification module and an offset source marking module.

[0009] The order data acquisition module acquires the basic order data, order remarks text, and actual driving trajectory of the order to be identified. The path offset triggering module determines whether the order to be identified meets the path offset triggering condition based on the positional difference between the actual driving trajectory and the recommended driving path. The remarks verification module generates a remarks pointer object based on the order remarks text when the order to be identified meets the path offset triggering condition, and performs verification on the remarks pointer object and the offset trajectory object in the actual driving trajectory, including position consistency, action consistency, and order status consistency, to obtain the remarks interpretation validity. The offset source tagging module generates a path offset source tag for the order to be identified based on the remarks interpretation validity.

[0010] With this design, the system does not directly reduce the risk of route deviation after detecting the order remarks text, nor does it directly determine the detour anomaly based solely on the route deviation rate. Instead, after the route deviation is triggered, it verifies the relationship between the remarks text and the actual driving trajectory, that is, it determines whether there is a consistency relationship between the two in terms of position, action, and state. In this way, it can determine whether the remarks text is qualified to explain the route deviation, thereby improving the accuracy and interpretability of the route deviation source identification.

[0011] In one optional implementation, the basic order data includes the order origin, order destination, order mileage, order duration, and order amount. The order remarks text is the fulfillment requirements text filled in by the user when placing the order or added before the order is executed. The actual driving trajectory includes the location of trajectory points, the time of trajectory points, the speed of road segments, and the location of stop points.

[0012] In one optional implementation, the path deviation triggering module includes a deviation rate calculation unit and a triggering judgment unit. The deviation rate calculation unit is used to calculate the path deviation rate based on factors such as the deviation distance of the actual driving trajectory relative to the recommended driving path, the number of deviation segments, or the duration of deviation; the triggering judgment unit is used to determine that the order to be identified meets the path deviation triggering condition when the path deviation rate reaches a preset deviation threshold.

[0013] In one optional implementation, the remark rebuttal verification module includes a remark semantic extraction unit, which is used to extract at least one of location words, direction words, action words and waiting words from the order remark text, and determine the remark semantic type based on the extraction result; the remark semantic type includes at least one of location pointing class, direction pointing class, waiting pointing class, pick-up and drop-off pointing class and temporary stop pointing class.

[0014] In one optional implementation, the remark counter-verification module further includes a remark object generation unit; the remark object generation unit is used to generate a remark pointing object according to the remark semantic type; it includes the remark pointing location, the remark action type, and the remark applicable order status, etc.; the remark applicable order status includes at least one of the pre-boarding status, the driving status, and the pre-arrival status.

[0015] In one optional implementation, the remark counter-verification module further includes a trajectory object extraction unit; the trajectory object extraction unit is used to extract the offset trajectory object from the actual driving trajectory, and the offset trajectory object includes at least one of the following: offset road segment, offset direction, stop point position, stop duration, and offset occurrence status node.

[0016] In one optional implementation, the remark rebuttal verification module further includes a location matching unit, an action matching unit, and a status matching unit; the location matching unit is used to calculate the location matching degree between the remark pointing location and the stop point location or the offset segment; the action matching unit is used to calculate the action matching degree between the remark action type and the stop duration or offset direction; and the status matching unit is used to calculate the status matching degree between the order status to which the remark applies and the offset occurrence status node.

[0017] In one optional implementation, the remark rebuttal verification module further includes an interpretation validity calculation unit; the interpretation validity calculation unit is used to calculate the remark interpretation validity based on the position matching degree, action matching degree and state matching degree, and reduce the remark interpretation validity when at least one of the position matching degree, action matching degree and state matching degree is lower than the corresponding threshold.

[0018] In one optional implementation, the system further includes an order value deviation analysis module. This module generates an order value deviation score based on the degree of deviation of order mileage, order duration, and order amount from a baseline range for similar orders. The offset source labeling module further generates path offset source labels based on the validity of the annotation explanation and the order value deviation score.

[0019] In one alternative implementation, the path offset source label includes at least one of the following: a note reasonable offset label, a note partially explained offset label, a note invalid offset label, and a suspected detour label.

[0020] Beneficial effects After a path deviation is triggered, this invention does not directly rely on the order notes to reduce the risk of path deviation. Instead, it performs consistency verification between the note's location, the note's action type, and the order status to which the note applies and the deviation trajectory object in the actual driving trajectory. This allows it to determine whether the note's text truly explains the path deviation.

[0021] This invention transforms order remarks text into remarks pointing objects and actual driving trajectories into offset trajectory objects, thereby establishing a calculable and verifiable correspondence between text semantics and trajectory behavior, and improving the interpretability of path offset source identification.

[0022] This invention calculates the validity of a note's interpretation by combining location matching degree, action matching degree, and state matching degree, avoiding the determination of a note's validity based on a single factor and reducing the risk that vague notes are mistakenly used as a basis for exemption from liability for detours.

[0023] This invention combines order numerical deviation to generate path offset source tags, enabling the system to distinguish between different situations such as reasonable offset with remarks, offset with partial explanation of remarks, offset with invalid remarks, and suspected detours, which facilitates subsequent risk control review and handling. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of a system structure provided for an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of an order path offset identification process provided in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of a remark-based counter-proof verification module provided in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram illustrating the mapping between semantic types of remarks and objects pointed to by remarks, provided as an embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of a triple consistency check provided in an embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram of path offset source tag generation provided in an embodiment of the present invention.

[0030] It should be noted that the above figures are used to help understand the technical solution of the present invention and do not constitute a limitation on the scope of protection of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some embodiments of the present invention, and not all embodiments. Other implementation methods obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0032] This embodiment provides an artificial intelligence-based multi-dimensional anomaly monitoring, identification, and comprehensive analysis system for orders. For example... Figures 1 to 6 As shown, this system can be deployed on the order platform server, or it can be implemented collaboratively by the order platform server, trajectory server, map server, and risk control server. This system is used to identify path deviations in the actual driving trajectory of the order to be identified, and after a path deviation is triggered, it performs a counter-verification check on whether the order's remarks text can explain the path deviation.

[0033] This system can serve as a route deviation source identification subsystem within a multi-dimensional order anomaly monitoring, identification, and comprehensive analysis system. It can also be deployed in conjunction with driver profiling, user profiling, license plate verification, and peak / valley statistics modules. The system's artificial intelligence capabilities are manifested in semantic parsing of order notes, semantic type training of historical notes, and assisted identification of route deviation source tags.

[0034] The system includes an order data acquisition module, a path offset triggering module, a note rebuttal verification module, and an offset source marking module. The order data acquisition module acquires the basic order data, order note text, and actual driving trajectory of the order to be identified. The path offset triggering module determines whether the order to be identified meets the path offset triggering conditions based on the positional difference between the actual driving trajectory and the recommended driving path. The note rebuttal verification module generates a note pointing object based on the order note text when the order to be identified meets the path offset triggering conditions, and performs verification on the note pointing object and the offset trajectory object in the actual driving trajectory, including position consistency, action consistency, and order status consistency, to obtain the note interpretation validity. The offset source marking module generates a path offset source label for the order to be identified based on the note interpretation validity.

[0035] In this embodiment, the system first determines whether there is a path offset in the order that needs to be explained, and then determines whether the order remarks text actually corresponds to the path offset, rather than directly reducing the path offset risk level based on the remarks text.

[0036] It should be noted that the technical terms used in this invention include: Orders to be identified refer to orders that require path offset source identification. Basic order data refers to data describing the basic fulfillment information of the order, including the order origin, destination, mileage, duration, and amount. Order remarks text refers to the fulfillment requirements text filled in by the user when placing the order or added before order execution, including but not limited to pick-up location, waiting requirements, temporary stop requirements, detour requirements, and direction requirements.

[0037] Actual driving trajectory refers to the trajectory data collected during order execution, including trajectory point locations, trajectory point times, road segment speeds, and stop point locations. Recommended driving route refers to the suggested route generated by the platform based on the order's origin and destination, or a commonly used route determined based on similar historical orders. Path deviation rate refers to the degree of deviation between the actual driving trajectory and the recommended driving route.

[0038] The object to which the remarks are directed refers to the structured object formed by parsing the order remarks text. This includes the location indicated by the remarks, the action type of the remarks, and the order status to which the remarks apply. The location indicated by the remarks refers to the geographical location or area indicated by the order remarks text, such as the north gate of a residential complex, the exit of an underground parking garage, the back door of a hotel, or the entrance to a parking lot. The action type of the remarks refers to the type of fulfillment action expressed in the order remarks text, including waiting, picking up passengers, detouring, stopping temporarily, making a U-turn, or driving to a designated location. The order status to which the remarks apply refers to the order fulfillment status to which the action type of the remarks applies, including the pre-boarding status, the driving status, and the pre-arrival status.

[0039] Offset trajectory objects refer to trajectory objects related to path deviation extracted from the actual driving trajectory, including the offset segment, offset direction, stop point location, stop duration, and offset occurrence status node. Offset occurrence status node refers to the order status node corresponding to the path deviation or stop behavior, including the pre-boarding status node, the in-journey status node, and the pre-arrival status node.

[0040] Location matching degree refers to the degree of spatial matching between the location indicated by the remarks and the location of the stop or the off-road segment; Action matching degree refers to the degree of matching between the noted action type and the actual stopping behavior or offset direction; Status matching degree refers to the degree of matching between the order status to which the remarks apply and the status node where the offset occurred; The validity of the remarks explanation refers to the degree to which the order remarks text can reliably explain the current path offset.

[0041] The benchmark range for similar orders refers to a reasonable range of order mileage, order duration, and order amount determined based on historical orders or sliding window orders similar to the order to be identified. Order value deviation refers to the degree to which the order mileage, order duration, and order amount of the order to be identified deviate from the benchmark range for similar orders. Path offset source labels refer to the structured labels output by the system for the reasons for path offsets, including labels indicating reasonable offsets, labels explaining offsets in the remarks, labels indicating invalid offsets, and labels indicating suspected detours.

[0042] The order data acquisition module can obtain relevant data of the order to be identified from the order database of the order platform, the trajectory service system, the map service system, and the user's order page.

[0043] Basic order data can include the order origin, order destination, order mileage, order duration, and order amount. The order origin and destination can be latitude and longitude coordinates, map points of interest (POIs), or internal platform addresses. Order mileage can be the actual mileage calculated after the order is completed, or the mileage accumulated from track points. Order duration is the time from the start to the end of the order. Order amount can be the amount paid by the user or the order billing amount.

[0044] Order remarks can be text filled in by the user when placing the order, or it can be text of performance requirements added by the user before the order is executed. For example, order remarks could be "Pick me up at the north gate of the community", "I'm at the underground parking garage exit", "Wait for me for five minutes", "Pick up my friend first", or "At the back gate of the hotel".

[0045] The actual driving trajectory can include trajectory point locations, trajectory point times, road segment speeds, and stop point locations. Trajectory point locations can include latitude and longitude coordinates or road locations after map road network matching. Trajectory point times can be the trajectory sampling time. Road segment speeds can be calculated based on the distance and time interval between adjacent trajectory points. Stop point locations can be determined based on the set of trajectory points where the vehicle's displacement is less than a preset distance within a preset time.

[0046] In one implementation, the order data acquisition module can also perform cleaning processing on the actual driving trajectory; the cleaning processing includes removing positioning jump points, removing obviously abnormal speed points, merging duplicate trajectory points, and unifying trajectory point timestamps; obviously abnormal speed points can be trajectory points whose speed exceeds a preset speed limit when calculated from adjacent trajectory points; thus, through trajectory cleaning processing, the impact of positioning errors on path deviation judgment can be reduced.

[0047] The path offset triggering module can obtain the recommended driving route corresponding to the order to be identified and compare the actual driving trajectory with the recommended driving route. The recommended driving route can be generated by the platform navigation system based on the order's origin and destination, or it can be determined by the commonly used routes of similar historical orders.

[0048] The offset rate calculation unit in the path offset triggering module can calculate the path offset rate based on the offset distance of the actual driving trajectory relative to the recommended driving path, the number of offset road segments, or the offset duration.

[0049] In one implementation, the path deviation rate can be defined as the ratio of the length of the actual driving trajectory that deviates from the recommended driving path by more than a preset distance threshold to the total length of the actual driving trajectory. The path deviation rate R can be expressed as: R = L1 / L0.

[0050] Where R represents the path deviation rate, L1 represents the length of the actual driving trajectory that deviates from the recommended driving path by more than a preset distance threshold, and L0 represents the total length of the actual driving trajectory. The value of R ranges from 0 to 1. The units of L1 and L0 can be meters, and L0 is greater than 0. When L0 is less than the preset minimum trajectory length, the system does not calculate the path deviation rate R, or transfers the order to manual review or low-confidence trajectory processing.

[0051] In another implementation, the ratio of the number of deviated road segments to the number of actual road segments traversed can be used as the path deviation rate; on the other hand, the ratio of the duration of the deviation to the total travel time of the order can also be used as the path deviation rate.

[0052] The trigger judgment unit can compare the path offset rate with the preset offset threshold. When the path offset rate reaches the preset offset threshold, it determines that the order to be identified meets the path offset trigger condition and triggers the remark counter-verification module to perform subsequent processing. When the path offset rate does not reach the preset offset threshold, the remark counter-verification can be omitted, or the order can be simply archived as a normal trajectory.

[0053] The remark rebuttal verification module may include a remark semantic extraction unit; the remark semantic extraction unit is used to perform word segmentation, entity recognition and keyword matching on the order remark text to extract at least one of location words, direction words, action words and waiting words.

[0054] Location words can include words indicating location such as residential area, building, parking lot, hotel, shopping mall, entrance, exit, garage, underground garage, parking lot exit, etc. Directional terms can include words indicating direction or location such as North Gate, South Gate, East Gate, West Gate, Back Gate, Front Gate, Left Side, Right Side, East Side, West Side, South Side, and North Side; Action verbs can include words that indicate driving or fulfilling a contract, such as picking someone up, taking a detour, stopping temporarily, turning around, driving to, delivering, going first, picking up first, and pulling over. Waiting words can include phrases such as "wait for me," "please wait a moment," "ten minutes," "I'll be there soon," "I'll contact you when I arrive," "wait a minute," "wait five minutes," etc., which indicate a request to wait or stay.

[0055] The semantic extraction unit can determine the semantic type of the remarks based on the extracted words. The semantic type of the remarks can include at least one of the following: location pointing, direction pointing, waiting pointing, pick-up / drop-off pointing, and temporary stop pointing.

[0056] In one implementation, the semantic extraction unit for order notes can be implemented using a keyword dictionary, rule templates, named entity recognition models, or semantic parsing models. The semantic parsing model can be a natural language processing model used to identify location words, direction words, action words, and waiting words in the order notes text. The keyword dictionary and rule templates can be pre-configured and updated based on historical order review results.

[0057] The remark-based counter-verification module may include a remark object generation unit. This unit generates a remark-targeting object based on the remark's semantic type. The remark-targeting object may include the remark's target location, the remark's action type, and the order status to which the remark applies.

[0058] The location indicated by the remarks can be determined based on location and direction words in the remarks text. The action type of the remarks can be determined based on action words and waiting words in the remarks text. The order status to which the remarks apply can be determined based on the action type and content of the remarks.

[0059] When the note text includes directional terms such as "North Gate," "South Gate," "East Gate," "West Gate," "Back Gate," or "Front Gate," the note object generation unit determines the corresponding gate location based on the map's point of interest boundaries at the order's origin or destination, and uses that gate location as the note's pointing location. When the note text includes location terms such as "Underground Garage," "Parking Exit," or "Parking Entrance," the note object generation unit determines the note's pointing location based on parking lot points of interest near the order's origin, road entrances, or the platform's address database.

[0060] When the remarks text includes words like "wait for me," "please wait a moment," "five minutes," or "ten minutes," the remarks object generation unit determines the remarks action type as "waiting" and the applicable order status as "before boarding" or "before arrival." When the remarks text includes action words like "pick up a friend first," "pick up someone midway," or "go somewhere first," the remarks object generation unit determines the remarks action type as "midway pick-up / drop-off" and the applicable order status as "in motion." When the remarks text includes action words like "pull over," "temporary stop," or "stop at the door," the remarks object generation unit determines the remarks action type as "temporary stop" and determines the applicable order status based on the current order location, order progress, and remarks content.

[0061] When the order remarks text cannot be parsed to reveal specific location or direction terms, or when the location pointed to by the remarks cannot be determined through map point of interest matching, the remarks verification module sets the location matching degree to a preset low value or marks the location matching degree as invalid, thus reducing the validity of the remarks interpretation. Therefore, vague remarks that cannot be located will not be directly used as a basis for reducing the risk of path deviation.

[0062] The rebuttal verification module may include a trajectory object extraction unit. The trajectory object extraction unit is used to extract offset trajectory objects from the actual driving trajectory. The offset trajectory object may include at least one of the following: offset segment, offset direction, stop point location, stop duration, and offset occurrence status node.

[0063] Offset segments can be road segments in the actual driving trajectory that deviate from the recommended driving path. In one embodiment, the trajectory object extraction unit can identify trajectory segments where the nearest distance from the actual trajectory point to the recommended driving path is continuously greater than a preset distance threshold as offset segments.

[0064] The offset direction can be determined based on the positional relationship of the offset segment relative to the recommended driving path; for example: When the entire off-road section is located north of the recommended driving route, the off-road direction can be determined to be northward. When the entire off-road section is located east of the recommended driving route, the direction of the off-road can be determined to be eastward.

[0065] The location of the stop point can be determined based on the set of trajectory points in the actual driving trajectory where the vehicle speed is lower than the preset speed and the duration exceeds the preset time; for example, when the speed of consecutive trajectory points is less than 3 kilometers per hour and the duration exceeds 2 minutes, the center position of the trajectory point set can be determined as the stop point location; the stop duration can be calculated based on the start time and end time corresponding to the stop point.

[0066] The offset occurrence status node can be determined based on the order fulfillment process; for example, an offset or stop that occurs before the driver picks up the user can be marked as a pre-boarding status node; an offset or stop that occurs during the order's journey can be marked as a driving status node; and an offset or stop that occurs before approaching the order's destination can be marked as a pre-arrival status node.

[0067] The rebuttal verification module may include a position matching unit, an action matching unit, and a state matching unit.

[0068] The location matching unit can calculate the location matching degree P between the location indicated by the remarks and the location of the stop point or the offset road segment. The value of P ranges from 0 to 1. The larger the value of P, the closer the location indicated by the remarks is to the actual offset trajectory object.

[0069] In one implementation, let d represent the distance between the location indicated by the note and the location of the stop point, d1 represent a first distance threshold, d2 represent a second distance threshold, and d2 is greater than d1; When d is not greater than d1, P=1; When d is greater than d1 and not greater than d2, P = (d2-d) / (d2-d1); When d is greater than d², P = 0; The first distance threshold can be 50 meters, and the second distance threshold can be 300 meters. These can also be configured according to different city road densities and map accuracy.

[0070] The action matching unit can calculate the action matching degree A between the noted action type and the dwell duration or offset direction; the value of A ranges from 0 to 1, and the larger the value of A, the more the actual trajectory behavior matches the noted action type; The status matching unit can calculate the status matching degree S between the order status applicable to the remarks and the offset occurrence status node; the value of S ranges from 0 to 1, and the larger the value of S, the more the path offset occurrence stage matches the remark applicable stage.

[0071] In one implementation, the state matching degree S is 1 when the applicable order status of the note matches the offset occurrence status node. The state matching degree S is 0.5 when the applicable order status of the note is adjacent to the offset occurrence status node. The state matching degree S is 0 when the applicable order status of the note is unrelated to the offset occurrence status node. All matching degrees are normalized to 0 to 1 before being entered into the formula for interpreting the validity of the note.

[0072] The action matching degree A can be determined according to Table 1 below, or it can be configured equivalently within the range of 0 to 1 according to the actual business rules.

[0073] Table 1

[0074] The remark rebuttal verification module may include an interpretation validity calculation unit; the interpretation validity calculation unit is used to calculate the remark interpretation validity E based on the position matching degree P, action matching degree A, and state matching degree S; In one implementation, the validity of the annotation interpretation E is calculated according to the following formula: E=wP×P+wA×A+wS×S; Where E represents the validity of the annotation explanation, P represents the position matching degree, A represents the action matching degree, S represents the state matching degree, wP represents the position matching degree weight, wA represents the action matching degree weight, and wS represents the state matching degree weight; the values ​​of P, A, S, and E are all from 0 to 1; the values ​​of wP, wA, and wS are all from 0 to 1, and the sum of wP, wA, and wS is 1. In one implementation, wP can be 0.4, wA can be 0.3, and wS can be 0.3; different weights can be used for different note semantic types; for example, when the note text includes explicit location words, wP can be increased; when the note text includes waiting words, wA and wS can be increased.

[0075] In one implementation, a key matching item threshold can also be set; when any of the location matching degree P, action matching degree A, and state matching degree S is lower than the corresponding threshold, the interpretation validity calculation unit reduces the annotation interpretation validity degree E; for example, when the annotation text contains explicit location words, the location matching degree P can be set as a key matching item; when the annotation text contains waiting words, the action matching degree A and state matching degree S can be set as key matching items.

[0076] The system may also include an order value deviation analysis module. This module generates an order value deviation score based on the degree of deviation of order mileage, order duration, and order amount from a benchmark range for similar orders.

[0077] Similar orders can be historical orders or sliding window orders that meet preset similar conditions in terms of origin region, destination region, order time period, and estimated mileage range. For example, similar orders can be historical orders with the same or adjacent origin regions, the same or adjacent destination regions, the same order time period, and the same estimated mileage range.

[0078] The benchmark range for similar orders can be determined by the median, upper percentile, and lower percentile of the order mileage, order duration, and order amount for similar orders. Alternatively, it can be determined by a platform-preset reasonable range, a regional peak-valley range, or a historical statistical range.

[0079] When the number of similar order samples is lower than the preset minimum sample size, the order value deviation analysis module expands the sample range for similar orders. Expansion methods include extending the historical time window, expanding the starting or ending area, merging adjacent time periods, or using a platform-preset baseline interval. If the expanded sample size is still lower than the preset minimum sample size, the order value deviation is only used as an auxiliary label and does not directly generate a suspected detour label.

[0080] In one implementation, the order value deviation analysis module determines the mileage benchmark interval, the duration benchmark interval, and the amount benchmark interval, respectively. The mileage benchmark interval can be determined by the lower and upper quantile values ​​of the mileage of similar orders; the duration benchmark interval can be determined by the lower and upper quantile values ​​of the duration of similar orders; and the amount benchmark interval can be determined by the lower and upper quantile values ​​of the amount of similar orders.

[0081] The order value deviation D can be calculated based on the mileage deviation, duration deviation, and amount deviation. The value of D ranges from 0 to 1, and the larger the value of D, the more the order value deviates from the benchmark range for similar orders.

[0082] In one implementation, the order numerical deviation D is determined as follows: D = wL × DL + wT × DT + wM × DM; Where D represents the order value deviation, DL represents the mileage deviation, DT represents the duration deviation, DM represents the amount deviation, wL represents the mileage deviation weight, wT represents the duration deviation weight, and wM represents the amount deviation weight; the values ​​of DL, DT, DM, and D are all from 0 to 1; the values ​​of wL, wT, and wM are all from 0 to 1, and the sum of wL, wT, and wM is 1.

[0083] For any order value X, its corresponding baseline range is [XL, XU], where XL represents the lower limit of the baseline range and XU represents the upper limit of the baseline range. When X is between XL and XU, the corresponding deviation is 0. When X is greater than XU, the corresponding deviation is equal to (X-XU) / XU, and the maximum value is 1. When X is less than XL, the corresponding deviation is equal to (XL-X) / XL, and the maximum value is 1. When XL or XU is 0, the preset minimum normalized base is used instead. X can be the order mileage, order duration, or order amount.

[0084] For example: When the order mileage is within the mileage benchmark range, DL is 0; When the order mileage exceeds the mileage benchmark range, DL determines it according to the above normalization rules; When the order duration is within the baseline duration range, DT is 0; When the order duration exceeds the baseline duration range, DT is determined according to the above normalization rules; When the order amount is within the base amount range, DM is 0; When the order amount exceeds the benchmark range, DM determines it according to the above normalization rules.

[0085] The offset source marking module can generate path offset source tags based on the validity of the remark interpretation (E) and the deviation of the order value (D). The path offset source tags can include at least one of the following: a reasonable offset tag, a partially explained offset tag, an offset tag with invalid remarks, and a suspected detour tag.

[0086] In one implementation, a preset validity threshold TE and a preset exception threshold TD can be set. TE indicates that when the validity of the annotation interpretation reaches this threshold, the system considers the annotation text to be able to explain the current path offset. TD indicates that when the order value deviation reaches this threshold, the system considers the order value to be significantly deviating from the benchmark range of similar orders.

[0087] When E is not less than TE and D is less than TD, the offset source marking module generates a reasonable offset label with a note. This label indicates that the current path offset can be reasonably explained by the order note text and that the order value is not obviously abnormal.

[0088] When E is not less than TE and D is not less than TD, the offset source marking module generates a note explaining the offset label. This label indicates that the note text can explain part of the path offset, but the order mileage, order duration or order amount still have abnormal deviations, and it needs to enter the low-risk warning or review.

[0089] When E is less than TE and D is less than TD, the offset source marking module generates an offset label indicating that the note is invalid. This label indicates that the note text cannot effectively explain the path offset, but the abnormality of the order value is not enough to directly identify it as a detour.

[0090] When E is less than TE and D is not less than TD, the offset source marking module generates a suspected detour label. This label can be used to trigger a pop-up review, push a risk control work order, restrict the settlement of related subsidies, or create a profile of returning drivers.

[0091] An optional example is provided to illustrate the process of generating reasonable offset labels for notes: When a user places an order, the starting point is near the south gate of a residential community, and the destination is an office building. The user writes in the order notes, "I am at the north gate of the community, please pick me up at the north gate." During the order execution, the system obtains the driver's actual driving trajectory and compares the actual driving trajectory with the recommended driving route.

[0092] The path offset triggering module calculates the path offset rate R to be 0.25, and the preset offset threshold is 0.2. Therefore, the order to be identified meets the path offset triggering condition. The semantic extraction unit extracts words such as "community," "north gate," and "pick me up" from the order's notes text and determines the semantic type of the notes as location-oriented, direction-oriented, and pick-up / drop-off-oriented. The notes object generation unit generates a notes pointing object, where the notes point to the north gate area of ​​the community near the order's origin, the notes action type is "pick up," and the notes apply to the order status as "before boarding."

[0093] The trajectory object extraction unit extracts the offset trajectory object from the actual driving trajectory, determining that the driver offset towards the north gate of the community before getting into the vehicle and stayed near the north gate for 3 minutes. The location matching unit calculates the distance between the location indicated by the note and the location of the stop to be 40 meters, which is less than the first distance threshold of 50 meters, therefore the location matching degree P is set to 1. The action matching unit determines that the action type of the note is "picking up a person," and the actual trajectory shows a stop behavior before getting into the vehicle, therefore the action matching degree A is set to 0.9. The status matching unit determines that the order status applicable to the note is consistent with the status node where the offset occurred, therefore the status matching degree S is set to 1.

[0094] If wP is 0.4, wA is 0.3, and wS is 0.3, then the validity of the remark explanation, E, is equal to 0.4 multiplied by 1 plus 0.3 multiplied by 0.9 plus 0.3 multiplied by 1, resulting in E of 0.97. The order value deviation analysis module calculates the order value deviation, D, as 0.15 based on the baseline range of similar orders. If the preset validity threshold TE is 0.7 and the preset abnormal threshold TD is 0.6, then E is not less than TE and D is less than TD. The offset source marking module generates reasonable offset labels for the remarks.

[0095] An optional example is provided to illustrate the process of generating suspected detour tags: When a user places an order, the order notes text is "Take a detour, get there quickly"; during the order execution process, the actual driving trajectory deviates significantly from the recommended driving route; the path deviation trigger module calculates the path deviation rate R as 0.36, which reaches the preset deviation threshold of 0.2, thus triggering the note rebuttal verification.

[0096] The semantic extraction unit for remarks extracts words such as "go around" and "hurry up" from the order remarks text. Since the remarks text does not provide clear location and direction words, the remarks object generation unit cannot determine the clear location of the remarks, and can only determine that the remarks action type is vaguely "go around", and the order status to which the remarks apply is unclear.

[0097] The trajectory object extraction unit extracts the offset segments from the actual driving trajectory, determining that the actual offset segments deviate far from the recommended driving path and that additional long-distance detours occur at the status nodes during driving. The location matching unit sets the location matching degree P to 0.2 due to the lack of explicit annotations pointing to the location. The action matching unit determines that the vague detour annotations cannot correspond to a specific offset direction, setting the action matching degree A to 0.3. The status matching unit determines that the applicable order status for the annotations is unclear, setting the status matching degree S to 0.3.

[0098] If wP is 0.4, wA is 0.3, and wS is 0.3, then the validity of the remark explanation, E, is equal to 0.4 multiplied by 0.2 plus 0.3 multiplied by 0.3 plus 0.3 multiplied by 0.3, resulting in E of 0.26. The order value deviation analysis module calculates that the order mileage is significantly higher than the benchmark range for similar orders, and the order duration and order amount are also higher than the benchmark range for similar orders, resulting in an order value deviation of D of 0.78. If the preset validity threshold TE is 0.7 and the preset abnormal threshold TD is 0.6, then E is less than TE and D is not less than TD. The offset source marking module generates suspected detour tags.

[0099] An optional example is provided to illustrate the process of generating offset labels in the notes section: The user's note text was "Pick up a friend first"; the system identified the note action type as mid-journey pick-up and drop-off, and the note applied to the order status as in motion; the actual driving trajectory showed that the vehicle made a brief stop at the in-motion status node, and the stop point was located in a reasonable midway area between the order's start and end points; therefore, the system calculated the location matching degree P to be 0.8, the action matching degree A to be 0.8, and the status matching degree S to be 1. At this point, if wP is 0.4, wA is 0.3, and wS is 0.3, then the validity of the annotation explanation E is 0.86, reaching the preset validity threshold of 0.7. However, the order value deviation analysis module finds that the order mileage and order amount are still higher than the benchmark range of similar orders, and the order value deviation D is 0.68, reaching the preset abnormal threshold of 0.6. At this point, the offset source marking module generates an annotation partial explanation offset label. This label indicates that the annotation can explain part of the path offset, but the order value still has abnormal deviation, so it needs to be reviewed or low-risk warning is required.

[0100] An optional example is provided to illustrate the process of generating offset labels when a note is invalid: The user's note text read "Pick me up at the north gate of the community"; the system identified the note as pointing to the north gate area of ​​the community, the action type as picking someone up, and the order status as "before getting on the vehicle"; the actual driving trajectory deviated, but the deviated section was located near the commercial street on the south side of the community, and the actual stop occurred at the "driving" status node, which is inconsistent with the note text. At this point, the location matching unit calculates that the distance between the location indicated by the remarks and the actual stopping point is 800 meters, which is greater than the second distance threshold of 300 meters, so the location matching degree P is 0; the action matching unit determines that the actual stopping behavior is inconsistent with the picking-up behavior, so the action matching degree A is 0.2; the state matching unit determines that the order state applicable to the remarks is the pre-boarding state, while the offset occurrence state node is the driving state, so the state matching degree S is 0. The system calculates the validity of the annotation explanation (E) to be 0.06, which is lower than the preset validity threshold of 0.7. The order value deviation analysis module calculates the order value deviation (D) to be 0.3, which does not reach the preset anomaly threshold of 0.6. Therefore, the offset source marking module generates an offset label indicating that the annotation is invalid. This label indicates that the annotation text cannot effectively explain the path deviation, but the degree of order value anomaly is not enough to directly identify it as a detour. The system can record this order as a low-risk archive or enter it for sampling review.

[0101] Furthermore, the system of this invention can be deployed on an order platform server, or it can be implemented collaboratively by multiple servers. For example, the order data acquisition module can be deployed on the order platform server, the path offset triggering module can be deployed on the trajectory service server, the remark rebuttal verification module can be deployed on the semantic analysis server or the risk control server, and the offset source marking module can be deployed on the risk control server.

[0102] In one embodiment, the system may further include a memory and a processor; the memory is used to store order data, trajectory data, a remarks dictionary, a semantic model, threshold configuration, historical similar order data, and path offset source tags; the processor is used to execute the functions of the order data acquisition module, the path offset triggering module, the remarks counter-verification module, the order value deviation analysis module, and the offset source tagging module.

[0103] The system can receive basic order data, order remarks, and actual driving trajectories uploaded from the order platform via a network interface, and output path offset source tags via a risk control interface. These path offset source tags can be used for order review, driver appeals, subsidy settlement audits, abnormal order alerts, and driver profile updates.

[0104] After acquiring the basic order data, order remarks text, and actual driving trajectory of the order to be identified, the order data acquisition module sends the actual driving trajectory to the path offset triggering module. Once the path offset triggering module determines that the order to be identified meets the path offset triggering conditions, it sends the path offset triggering result to the remarks verification module.

[0105] The remark-based verification module generates a remark-targeting object from the order remark text and extracts the offset trajectory object from the actual driving trajectory. Then, the remark-based verification module calculates the position matching degree, action matching degree, and state matching degree, and calculates the validity degree of the remark interpretation based on the three.

[0106] The order value deviation analysis module generates an order value deviation score based on the degree of deviation of order mileage, order duration, and order amount from the benchmark range of similar orders. The offset source marking module generates a path offset source label based on the validity of the remarks explanation and the order value deviation score, and writes the path offset source label into the order risk control record.

[0107] In one optional implementation, the remarks semantic extraction unit can extract location words, direction words, action words, and waiting words, etc., through an artificial intelligence semantic parsing model; the semantic parsing model can be trained by historical order remarks text and manually labeled semantic types; the training samples include order remarks text and corresponding location words, direction words, action words, waiting words, and remarks semantic type labels, etc.

[0108] In another optional implementation, the semantic extraction unit for remarks can also be implemented through a rule dictionary; the rule dictionary is pre-configured with a location word dictionary, a direction word dictionary, an action word dictionary, and a waiting word dictionary; the system parses the order remarks text through keyword matching and rule templates.

[0109] In another alternative implementation, the location indicated by the remarks can be determined by matching map points of interest; when the remarks text includes location descriptions such as "North Gate" or "Parking Lot Exit", the system can determine the location indicated by the remarks based on map points of interest, building boundaries, road entrances and parking lot entrances near the order's starting point or ending point.

[0110] In another alternative implementation, the baseline range for similar orders can be updated using a sliding time window; for example, the system can use the most recent seven days, the most recent thirty days, or the most recent one thousand similar orders as samples to calculate the median and quantile range of mileage, duration, and amount.

[0111] In another alternative implementation, the route offset source tag can be further fed back to the driver profile or order risk control model; when a driver has multiple invalid offset tags or suspected detour tags, the system can increase the priority of route offset review for the driver's subsequent orders.

[0112] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention; any equivalent substitutions or combinations made to the modules, rules, thresholds, weights, calculation methods or deployment methods within the technical concept of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A system for multi-dimensional anomaly monitoring, identification, and comprehensive analysis of orders based on artificial intelligence, characterized in that: This includes an order data acquisition module, a path offset triggering module, a remarks counter-verification module, and an offset source marking module; The order data acquisition module is used to acquire the basic order data, order remarks text, and actual driving trajectory of the order to be identified; The path offset triggering module is used to determine whether the order to be identified meets the path offset triggering condition based on the positional difference between the actual driving trajectory and the recommended driving path; The remark rebuttal verification module is used to generate a remark pointing object based on the order remark text when the order to be identified meets the path offset triggering condition, and to verify the remark pointing object with the offset trajectory object in the actual driving trajectory, including position consistency, action consistency and order status consistency, to obtain the remark interpretation validity. The remark pointing object includes the remark pointing location, the remark action type, and the remark applicable order status. The remark applicable order status includes at least one of the pre-boarding status, driving status, and pre-arrival status. The offset trajectory object includes at least one of the offset segment, offset direction, stop point location, stop duration, and offset occurrence status node. The remark-based counter-evidence verification module also includes a position matching unit, an action matching unit, and a state matching unit; The location matching unit is used to calculate the location matching degree between the noted location and the stop point location or the offset road segment; The action matching unit is used to calculate the action matching degree between the noted action type and the dwell duration or the offset direction; The state matching unit is used to calculate the state matching degree between the applicable order state of the remarks and the offset occurrence state node; the remarks counter-proof verification module also includes an interpretation validity calculation unit. The interpretation validity calculation unit is used to calculate the annotation interpretation validity based on the location matching degree, action matching degree, and state matching degree, and to reduce the annotation interpretation validity when at least one of the location matching degree, action matching degree, and state matching degree is lower than the corresponding threshold; The offset source marking module is used to generate the path offset source label of the order to be identified based on the validity of the annotation interpretation; the system also includes an order value deviation analysis module; The order value deviation analysis module is used to generate the order value deviation degree based on the degree of deviation of the order mileage, order duration and order amount from the benchmark range of similar orders; The offset source marking module is also used to generate the path offset source label based on the validity of the remark explanation and the deviation of the order value; the path offset source label includes at least one of the following: reasonable offset label, partially explained offset label, invalid offset label, and suspected detour label; The offset source marking module is used to generate the reasonable offset label for the remark when the validity of the remark interpretation reaches a preset validity threshold and the deviation of the order value does not reach a preset abnormal threshold. When the validity of the annotation explanation reaches the preset validity threshold and the deviation of the order value reaches the preset abnormal threshold, the annotation partial explanation offset tag is generated. When the validity of the annotation interpretation is lower than the preset validity threshold and the deviation of the order value does not reach the preset abnormal threshold, an invalid annotation offset label is generated. When the validity of the annotation interpretation is lower than the preset validity threshold and the deviation of the order value reaches the preset abnormal threshold, the suspected detour label is generated.

2. The system according to claim 1, characterized in that, The basic order data includes the order origin, order destination, order mileage, order duration, and order amount. The order remarks text is the fulfillment requirements text filled in by the user when placing the order or added before the order is executed. The actual driving trajectory includes the location of trajectory points, the time of trajectory points, the speed of road segments, and the location of stop points.

3. The system according to claim 1, characterized in that, The path offset triggering module includes an offset rate calculation unit and a trigger judgment unit; The offset rate calculation unit is used to calculate the path offset rate based on the offset distance, number of offset road segments, or offset duration of the actual driving trajectory relative to the recommended driving path; The triggering judgment unit is used to determine that the order to be identified meets the path offset triggering condition when the path offset rate reaches a preset offset threshold.

4. The system according to claim 1, characterized in that, The remark rebuttal verification module includes a remark semantic extraction unit; The semantic extraction unit for remarks is used to extract at least one of location words, direction words, action words and waiting words from the order remarks text, and to determine the semantic type of remarks based on the extraction results. The semantic type of remarks includes at least one of location pointing class, direction pointing class, waiting pointing class, pick-up and drop-off pointing class and temporary stop pointing class.

5. The system according to claim 4, characterized in that, The remark rebuttal verification module also includes a remark object generation unit; The memo object generation unit is used to generate the memo pointing object according to the memo semantic type.

6. The system according to claim 5, characterized in that, The remark-based counter-evidence verification module also includes a trajectory object extraction unit; The trajectory object extraction unit is used to extract the offset trajectory object from the actual driving trajectory.

Images