A public transportation pre-service safety digitalization system based on big data mining

Abstract

The application discloses a kind of public traffic pre-service safety digitalization system based on big data mining, it is related to big data analysis technical field, including: pre-service access module, for accessing public traffic pre-service safety management data and driving task operation data;Batch binding module, for binding pre-service detection batch, and keeping the occurrence position sequence of each associated record;State arrangement module, for constituting batch state sequence;Batch stringing module, for constituting pre-service operation sequence;Sequence mining module, for inputting the pre-service operation sequence into improved SCINet model, outputting the abnormal representation of current pre-service detection batch;Abnormal positioning module, for analyzing the abnormal representation, determine the record position that causes current pre-service detection batch cannot enter departure authorization state;On-duty interlocking module, for updating the detection state associated with the record position.The application can realize pre-service abnormality accurate positioning and departure authorization closed-loop interlocking.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to a digital and intelligent system for pre-job safety in public transportation based on big data mining. Background Technology

[0002] Before commencing their shifts, public transport drivers typically undergo pre-shift safety procedures including identity verification, physical condition checks, reaction time assessments, vehicle inspections, and safety confirmations. Existing pre-shift safety management systems often employ individual test thresholds or fixed rules to assess data such as alcohol concentration, body temperature, blood pressure, and vehicle inspection results, determining a driver's suitability for duty based on a single test result. However, pre-shift test records, vehicle inspection records, scheduling data, and driving task execution data are usually scattered across different business systems, lacking a unified task attribution and sequence correlation, making it difficult to reflect status changes between consecutive driving tasks for the same driver. Current methods also struggle to identify local anomalies from multiple test statuses and historical results; anomaly detection results usually only point to the entire test batch, failing to pinpoint specific record locations, easily leading to missed anomaly detections or incorrectly locked normal tasks. Furthermore, there is a lack of continuous linkage between re-inspection results, manual handling results, and departure authorization status, still requiring manual verification and restriction removal by management personnel, impacting the efficiency of pre-shift safety management and the timeliness of departure scheduling.

[0003] Therefore, how to provide a digital and intelligent system for pre-job safety in public transportation based on big data mining is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a digital and intelligent public transportation pre-shift safety system based on big data mining. This invention arranges pre-shift inspection records, vehicle three-inspection records, safety confirmation records, and driving task operation records continuously according to driving tasks and occurrence order, forming a pre-shift operation sequence that simultaneously represents the current pre-shift status and historical operation status. Based on this, a node-by-node back-construction differential closed loop is set in each non-leaf-level sampling interaction node of the SCINet model. The current node is reconstructed using the lower-level relayed status, and the relayed status is corrected, ensuring that local anomalies maintain their original order during multi-level sampling. Furthermore, the back-construction differential is bound to the associated record identifier, record type, and occurrence order to determine the location of the blocked record. The inspection status and departure authorization status are updated based on the re-inspection results or handling results, thereby improving the accuracy of pre-shift safety anomaly identification and location, reducing the false locking rate, and achieving closed-loop linkage between pre-shift inspection, anomaly handling, and departure authorization.

[0005] According to an embodiment of the present invention, a digital and intelligent public transportation pre-job safety system based on big data mining includes: The pre-job access module is used to access public transportation pre-job safety management data and driving task operation data, and to aggregate public transportation pre-job safety management data and driving task operation data belonging to the same driving task; The batch binding module is used to define the pre-shift inspection process around the planned departure task, bind the associated records generated by the same driver in the pre-shift inspection process into a pre-shift inspection batch, and retain the occurrence sequence of each associated record; The status arrangement module is used to organize the pre-job testing batches according to the pre-job testing process, and arrange the testing status of each associated record according to the occurrence order to form a batch status sequence. The batch serialization module is used to connect multiple batch status sequences according to the driving task sequence of the same driver, and to connect the running status of each driving task to the batch status sequence of the same driving task to form a pre-shift operation sequence. The sequence mining module is used to input the pre-job operation sequence into the improved SCINet model. The improved SCINet model sets up a node-by-node back-construction difference closed loop in each non-leaf level sampling interaction node, back-constructs the current node based on the transmission status of the lower-level sampling interaction node, extracts the back-construction difference and corrects the transmission status, and outputs the abnormal characterization of the current pre-job detection batch. The anomaly localization module is used to analyze the anomaly characteristics by combining the occurrence sequence of the current pre-shift inspection batch, and determine the record position that causes the current pre-shift inspection batch to be unable to enter the departure authorization state. The on-duty interlocking module is used to determine the driving task whose departure authorization status needs to be adjusted based on the recorded location, and to update the detection status associated with the recorded location based on the re-inspection result or handling result of the recorded location.

[0006] Optionally, the pre-job access module specifically includes: Read public transportation operation scheduling data, extract driver identifiers, planned arrival times, planned departure times, and driving task identifiers, and establish a driving task index; Access pre-job inspection records, vehicle three-inspection records, and safety confirmation records, and retain driver identification, collection time, record type, and record content; Access driving task execution records and retain driving task identifiers, occurrence times, and execution status; By using driver identification, planned arrival time, and planned departure time, driving task identification is matched to pre-shift inspection records, vehicle three-inspection records, and safety confirmation records. Pre-shift inspection records, vehicle three-inspection records, safety confirmation records, and driving task operation records with the same driving task identification are grouped into the same driving task.

[0007] Optionally, the batch binding module specifically includes: Read the driver identifier, planned arrival time, planned departure time, and driving task identifier of the planned departure task, and set the start and end boundaries of the pre-shift inspection process with the planned arrival time and planned departure time respectively; Extract associated records from public transportation pre-job safety management data where the collection time is between the start and end boundaries and the driver's identifier is the same as the driver's identifier for the planned departure task, and write the driving task identifier into the extracted associated records. The associated records extracted according to the collection time are sequentially written into the batch number. The associated records with the same driving task identifier are bound into the pre-job inspection batch, and the occurrence order is recorded by the batch number.

[0008] Optionally, the state orchestration module includes: Read the pre-job inspection process, set multiple status positions along the execution order of the pre-job inspection process, and configure the writable record type for each status position; Records of the same type within the pre-employment testing batch are written to the same status position. When there are multiple related records in the same status position, the multiple related records are arranged according to the order of occurrence. Read the completion marker, detection conclusion, and validity marker of each associated record, combine them into a detection status, and set the status position that has not been written to the associated record to an absent status; The status positions are arranged according to the execution order of the pre-job testing process to form a batch status sequence.

[0009] Optionally, the batch serialization module specifically includes: Retrieve multiple batch status sequences with the same driver identifier, read the planned departure time of the driving task to which each batch status sequence belongs, arrange multiple batch status sequences according to the planned departure time, and write the driving task position order for each batch status sequence. Use driving task identifiers to retrieve driving task operation data, arrange the retrieved operation records according to the time of occurrence, extract the operation status from each operation record, and compile them into operation status segments; Set each batch status sequence as a driving task status segment, and append the retrieved running status segment to the end of the driving task status segment with the same driving task identifier; Multiple driving task status segments are connected end to end according to the driving task position sequence to form the pre-shift operation sequence.

[0010] Optionally, the improved SCINet model includes a sampling interaction tree, and node back-construction unit, differential extraction unit, two differential projection units and two correction branches set in each non-leaf level sampling interaction node; The sampling interaction tree consists of multiple sampling interaction nodes connected hierarchically in a binary tree structure. Each non-leaf level sampling interaction node includes an interleaved sampling unit, a dual-path interaction unit, and two lower-level node interfaces. The two lower-level node interfaces are respectively connected to two lower-level sampling interaction nodes. The input of the node back-construction unit is connected to the return of the two lower-level node interfaces respectively. The differential extraction unit is connected to the input of the non-leaf-level sampling interaction node and the output of the node back-construction unit respectively. The input of the two differential projection units is connected to the output of the differential extraction unit. The input of the two correction branches is connected to the output of the two differential projection units respectively. The output of the two correction branches is connected to the return of the two lower-level node interfaces respectively.

[0011] Optionally, the output of the anomaly characteristics of the current pre-job inspection batch includes: The pre-shift operation sequence is written into the root-level sampling interaction node of the sampling interaction tree. The interleaved sampling unit splits the node input state into two interleaved sub-sequences. The dual-path interaction unit updates the two interleaved sub-sequences and passes them to the two lower-level sampling interaction nodes through the two lower-level node interfaces until they reach the leaf-level sampling interaction node. In each non-leaf-level sampling interaction node, the node back-construction unit receives the forwarded states returned by the two lower-level sampling interaction nodes, and re-embeds the two forwarded states according to the staggered sampling positions to obtain the node back-construction state. The differential extraction unit extracts the back-construction difference between the node input state and the node back-construction state. The two differential projection units project the back-constructed differential into two differential components according to the staggered sampling positions. These components are then incorporated into two recursive states via two correction branches. The two corrected recursive states are then re-embedded and submitted to the next-level sampling interaction node. The back-construction difference generated by each non-leaf-level sampling interaction node is restored to the original position of the pre-shift operation sequence according to the staggered sampling position. The back-construction differences falling in the same original position are merged, and the back-construction difference of the current pre-shift testing batch is extracted to form the abnormal characterization of the current pre-shift testing batch.

[0012] Optionally, the step of analyzing the anomaly characterization by combining the occurrence sequence of the current pre-employment testing batch includes: Read the associated record identifier, record type and occurrence sequence of each associated record in the current pre-job testing batch, and divide the abnormal characteristics of the current pre-job testing batch into multiple record position segments according to the occurrence sequence of each associated record; The absolute value of the reconstruction difference within each recording location segment is taken and summed to obtain the location anomaly of each recording location. Bind the associated record identifier, record type, and occurrence order to the location anomaly quantity of the same record location.

[0013] Optionally, determining the record location that causes the current pre-shift inspection batch to be unable to enter the dispatch authorization state includes: From the historical pre-shift inspection batches that have entered the departure authorization state, extract the location anomalies that have the same record type and occurrence sequence, and set the maximum location anomaly as the location anomaly boundary. Compare the positional anomaly quantity of each record location in the current pre-job testing batch with the positional anomaly boundary that is the same in both record type and occurrence sequence, and retain the record locations where the positional anomaly quantity exceeds the positional anomaly boundary. Read the detection status within the reserved record positions, mark the record positions with detection status of incomplete, invalid, or abnormal as blocked record positions, and determine the blocked record positions as the record positions that prevent the current pre-shift inspection batch from entering the departure authorization state.

[0014] Optionally, the on-duty interlocking module specifically includes: Read the driving task identifier associated with each blocking record location, identify the driving task pointed to by the driving task identifier as the driving task whose departure authorization status needs to be adjusted, and switch the departure authorization status to the locked status. Receive the re-inspection results or handling results of each blocking record location, and update the detection status within the blocking record location using the re-inspection results or handling results; Check each updated blocking record location. If the detection status of all blocking record locations is complete, valid, or normal, the lock is released. If there are blocking record locations that are incomplete, invalid, or abnormal, the lock remains.

[0015] The beneficial effects of this invention are: 1. This invention organizes pre-job inspection records, vehicle three-inspection records, safety confirmation records, and driving task operation records according to driving tasks, and forms batch status sequences and pre-job operation sequences according to the inspection process and occurrence sequence, so that scattered pre-job safety data can be continuously expressed in the same sequence, which is convenient to reflect the current pre-job inspection status and the historical driving task operation status at the same time.

[0016] 2. In each non-leaf level sampling interaction node of the SCINet model, the present invention sets up a node back-construction unit, a differential extraction unit, two differential projection units and two correction branches. The current node is back-constructed by the lower level recursion state, and the recursion state is corrected by using the back-construction differential. This allows local anomalies in different sampling levels to retain their original position order and reduces the attenuation of anomaly information in the multi-level sampling and recursion process.

[0017] 3. This invention restores the differential data back to the original position of the pre-shift operation sequence and binds it with the associated record identifier, record type and occurrence position. Based on this, the location of the blocking record is determined. Then, the detection status and departure authorization status are updated in combination with the re-inspection results or handling results, thereby improving the accuracy of pre-shift anomaly location, reducing the accidental locking of normal driving tasks, and realizing closed-loop management of anomaly detection, re-inspection handling and departure authorization interlocking. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a structural diagram of a digital intelligent system for pre-job safety in public transportation based on big data mining, as proposed in this invention. Figure 2 This is a node-by-node back-construction differential closed-loop structure diagram of a public transportation pre-job safety digital intelligent system based on big data mining proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 A digital and intelligent public transportation pre-job safety system based on big data mining includes the following: The pre-job access module is used to access public transportation pre-job safety management data and driving task operation data, and to aggregate public transportation pre-job safety management data and driving task operation data belonging to the same driving task; The batch binding module is used to define the pre-shift inspection process around the planned departure task, bind the related records generated by the same driver during the pre-shift inspection process into a pre-shift inspection batch, and retain the occurrence sequence of each related record; The status orchestration module is used to organize pre-job testing batches according to the pre-job testing process, and to arrange the testing status of each associated record according to the occurrence sequence to form a batch status sequence. The batch serialization module is used to connect multiple batch status sequences according to the driving task sequence of the same driver, and to connect the running status of each driving task to the batch status sequence of the same driving task to form a pre-shift operation sequence. The sequence mining module is used to input the pre-job operation sequence into the improved SCINet model. The improved SCINet model sets a node-by-node back-construction difference closed loop in each non-leaf level sampling interaction node, reconstructs the current node based on the transmission status of the lower-level sampling interaction node, extracts the back-construction difference and corrects the transmission status, and outputs the anomaly representation of the current pre-job detection batch. The anomaly localization module is used to analyze the anomaly characteristics by combining the occurrence sequence of the current pre-shift inspection batch to determine the record location that causes the current pre-shift inspection batch to be unable to enter the departure authorization state. The on-duty interlocking module is used to determine the driving task whose departure authorization status needs to be adjusted based on the recorded location, and to update the detection status associated with the recorded location based on the re-inspection results or handling results of the recorded location.

[0021] In this embodiment, the pre-job access module specifically includes: Read public transportation operation scheduling data, extract driver identifiers, planned arrival times, planned departure times, and driving task identifiers, and establish a driving task index; Access pre-job inspection records, vehicle three-inspection records, and safety confirmation records, and retain driver identification, collection time, record type, and record content; Access driving task execution records and retain driving task identifiers, occurrence times, and execution status; By using driver identification, planned arrival time, and planned departure time, driving task identification is matched to pre-shift inspection records, vehicle three-inspection records, and safety confirmation records. Pre-shift inspection records, vehicle three-inspection records, safety confirmation records, and driving task operation records with the same driving task identification are grouped into the same driving task.

[0022] In the specific implementation process, the pre-shift access module reads public transportation operation scheduling data through a database interface. Each planned departure task is treated as a scheduling record. From each scheduling record, the driver's ID, scheduling task ID, planned arrival time, and planned departure time are read and written into the driver identifier, driving task identifier, planned arrival time, and planned departure time, respectively. The planned arrival time and planned departure time are uniformly converted to timestamps within the same time zone. Using the driver identifier as the primary search key and the planned departure time as the sorting key, scheduling records for the same driver are arranged according to their planned departure time. Each index record contains the driving task identifier, planned arrival time, and planned departure time, forming a driving task index.

[0023] Pre-job inspection records are accessed through the pre-job safety management terminal, vehicle three-inspection records are accessed through the vehicle inspection terminal or vehicle inspection data interface, and safety confirmation records are accessed through the pre-job safety confirmation interface. When accessing a single record, the interface identifier, driver identifier, collection time, original record type code, and original record content are read from the interface message. The collection time is converted to the same format as the public transportation operation scheduling data, and the original record content is completely preserved.

[0024] During system initialization, a record type mapping table is established based on the type coding specifications provided by each data interface. Each record in the record type mapping table includes an interface identifier, an original record type code, and a unified record type code, with the combination of the interface identifier and the original record type code serving as the search key. The unified record type codes for pre-job testing records represent identity verification, temperature detection, alcohol detection, blood pressure detection, blood oxygen saturation detection, heart rate detection, electrocardiogram detection, mood detection, reaction time detection, and spatial perception detection, respectively. The unified record type codes for vehicle inspection records represent vehicle appearance and safety facility inspection, braking and steering component inspection, and onboard equipment inspection, respectively. All vehicle inspection records are generated before the planned departure time. The unified record type codes for safety confirmation records represent safety training completion, risk warning confirmation, and safety matter confirmation, respectively. When accessing a single record, the record type mapping table is retrieved using the interface identifier and the original record type code. The retrieved unified record type code is written into the record type field, while the original record type code is retained.

[0025] Driving task operation records are accessed through the vehicle operation management database or the dispatch operation data interface. A driving task operation record is generated whenever the driving task operation status changes. The interface identifier, driving task identifier, status change time, and original operation status code are read from the driving task operation record, and the status change time is converted into the occurrence time.

[0026] During system initialization, a running status mapping table is established based on the status coding specifications provided by the driving task running data interface. Each record in the running status mapping table includes an interface identifier, an original running status code, and a unified running status code, with the combination of the interface identifier and the original running status code serving as the search key. The unified running status code includes at least four states: pending departure, departed, running, reached destination, and task completed. When accessing driving task running records, the running status mapping table is retrieved using the interface identifier and the original running status code. The retrieved unified running status code is written to the running status field, while the original running status code is retained.

[0027] When matching driving task identifiers with pre-job inspection records, vehicle three-inspection records, and safety confirmation records, the driving task index is first retrieved using the driver identifier. Then, the collection time of the record is compared with the planned arrival time and planned departure time in each index record. If the collection time is no earlier than the planned arrival time and no later than the planned departure time, the driving task identifier in the index record is written to the corresponding record. If the collection time falls within the time range of multiple planned departure tasks, the time interval between the collection time and each planned departure time is calculated, and the planned departure task with the smallest time interval is determined as the matching result.

[0028] After matching driving task identifiers, a driving task data group is established using the driving task identifier as the aggregation key. Pre-job inspection records, vehicle three-inspection records, and safety confirmation records with the same driving task identifier are written into the same driving task data group. Furthermore, using the driving task identifiers already carried in the driving task operation records, the driving task operation records are written into the same driving task data group. Each driving task data group retains the record source, record type, collection time or occurrence time, record content, and operational status.

[0029] In this embodiment, the batch binding module specifically includes: Read the driver identifier, planned arrival time, planned departure time, and driving task identifier of the planned departure task, and set the start and end boundaries of the pre-shift inspection process with the planned arrival time and planned departure time respectively; Extract associated records from public transportation pre-job safety management data where the collection time is between the start and end boundaries and the driver's identifier is the same as the driver's identifier for the planned departure task, and write the driving task identifier into the extracted associated records. The associated records extracted according to the collection time are sequentially written into the batch number. The associated records with the same driving task identifier are bound into the pre-job inspection batch, and the occurrence order is recorded by the batch number.

[0030] In the specific implementation process, the planned departure tasks to be processed are retrieved from the driving task index, and the driver identifier, planned arrival time, planned departure time, and driving task identifier are read. The planned arrival time and planned departure time are converted into the same time format as the time of collection of associated records. The planned arrival time is used as the starting boundary and the planned departure time is used as the ending boundary to obtain the pre-shift detection time interval corresponding to a single planned departure task.

[0031] Retrieve pre-shift inspection records, vehicle three-inspection records, and safety confirmation records from the driving task data group, and read the driver identifier and collection time for each record. If the driver identifier matches the driver identifier of the planned departure task and the collection time is no earlier than the start boundary and no later than the end boundary, the corresponding record will be identified as the associated record of the current pre-shift inspection batch; records that do not meet the driver identifier or time interval conditions will not be written into the current pre-shift inspection batch.

[0032] Write the driving task identifier of the currently planned departure task into the batch attribution field of the associated record. If the associated record already carries the same driving task identifier, keep the original driving task identifier unchanged and only write the batch attribution status; if the driving task identifier carried by the associated record is inconsistent with the currently planned departure task, do not write the corresponding record into the current pre-shift inspection batch. The record type, record content, and collection time remain unchanged.

[0033] The captured related records are arranged from earliest to latest according to their collection time. When multiple related records have the same collection time, their order is determined by the sequence of their related record identifiers. The first related record is assigned a batch number of 1, and the batch numbers are incremented sequentially for each record in the arrangement order. The batch number is then written into the related record. The batch number also serves as the position of the related record within the pre-job testing batch.

[0034] Establish a pre-job inspection batch record based on the driving task identifier, and write related records with the same driving task identifier into the pre-job inspection batch record in sequence according to the batch number. Each related record shall retain at least the related record identifier, driving task identifier, record type, collection time, record content, and batch number. The batch number shall maintain the sequential relationship of each related record during the pre-job inspection process.

[0035] In this embodiment, the state orchestration module includes: Read the pre-job inspection process, set multiple status positions along the execution order of the pre-job inspection process, and configure the writable record type for each status position; Records of the same type within the pre-employment testing batch are written to the same status position. When there are multiple related records in the same status position, the multiple related records are arranged according to the order of occurrence. Read the completion marker, detection conclusion, and validity marker of each associated record, combine them into a detection status, and set the status position that has not been written to the associated record to an absent status; The status positions are arranged according to the execution order of the pre-job testing process to form a batch status sequence.

[0036] In the specific implementation process, the pre-job testing process is read from the pre-job safety management configuration data. The pre-job testing process is recorded as a single process item, and each process record includes at least a flow procedure number, a status position identifier, and a writable unified record type code. Process records are arranged in ascending order of flow procedure number, and each process record is assigned a status position, with the unified record type code carried by the process record written to that status position. When multiple unified record type codes belong to the same process item, all unified record type codes are written to the same status position. For example, identity verification, physical condition detection, competency testing, vehicle three-inspection, and safety confirmation are each assigned a status position. The status position corresponding to physical condition detection can be written with the unified record type codes for temperature detection, alcohol detection, blood pressure detection, blood oxygen detection, heart rate detection, and electrocardiogram detection.

[0037] Retrieve pre-job testing batches, read each associated record according to the batch sequence number, read the record type from the associated records, and search for status positions in the pre-job testing process where the same record type can be written. After finding the status position, write the associated record identifier, occurrence sequence, and record content to the status position. When multiple associated records are written to the same status position, expand the multiple associated records according to their occurrence sequence, and each associated record forms an independent sequence position. Do not merge multiple associated records into the same sequence position. Each independent sequence position retains the status position identifier, associated record identifier, record type, and occurrence sequence. Status positions where no associated records are written form a missing sequence position, and write the missing status in the missing sequence position.

[0038] The completion flag, detection conclusion, and validity flag are included in the record content of the associated record. The completion flag uses binary encoding; it is written as 1 when the record collection process has ended and been successfully submitted, and as 0 when the record collection process is not completed. The validity flag also uses binary encoding; it is written as 1 when the record passes the data integrity check of the source terminal, and as 0 when the record is marked as invalid by the source terminal. The detection conclusion uses a unified conclusion encoding, which includes at least normal, abnormal, and undetermined. When the detection conclusion is undetermined, the completion flag is written as 0; the completion flag is written as 1 only when a normal or abnormal confirmed detection conclusion is obtained. When the pre-job access module accesses associated records from different data sources, it converts the original completion status, original detection conclusion, and original validity status into the above unified encoding, while retaining the original field values.

[0039] For each associated record at a written status position, the completion flag, detection conclusion, and validity flag are read sequentially, and the detection status is composed according to the fixed field order of "completion flag - detection conclusion - validity flag". When multiple associated records exist for a status position, the multiple detection statuses are arranged according to their occurrence order, and each detection status is bound to the identifier of the associated record that generated the detection status. Status positions without associated records written in the pre-job detection process are written as "absent status", and no associated record identifier is configured for "absent status".

[0040] After writing all status positions, arrange the status positions according to the flow program number. When only one associated record is written to the same status position, the associated records form a sequence position; when multiple associated records are written to the same status position, the multiple associated records form multiple consecutive sequence positions according to their occurrence order; status positions without associated records form a missing sequence position. Starting from 1, write sequence position numbers sequentially to the arranged sequence positions, and connect the sequence positions according to their sequence position numbers to form the batch status sequence of the pre-job inspection batch. Each sequence position with an associated record maintained a binding relationship between the sequence position number and the associated record identifier.

[0041] In this embodiment, the batch serial connection module specifically refers to: Retrieve multiple batch status sequences with the same driver identifier, read the planned departure time of the driving task to which each batch status sequence belongs, arrange multiple batch status sequences according to the planned departure time, and write the driving task position order for each batch status sequence. Use driving task identifiers to retrieve driving task operation data, arrange the retrieved operation records according to the time of occurrence, extract the operation status from each operation record, and compile them into operation status segments; Set each batch status sequence as a driving task status segment, and append the retrieved running status segment to the end of the driving task status segment with the same driving task identifier; Multiple driving task status segments are connected end to end according to the driving task position sequence to form the pre-shift operation sequence.

[0042] In the specific implementation process, the driver identifier, driving task identifier, and planned departure time are read from the batch information carried by the batch status sequence. Multiple batch status sequences for the same driver are collected using the driver identifier as the search key. These multiple batch status sequences are arranged from earliest to latest according to the planned departure time; if the planned departure times are the same, they are arranged according to the character order of the driving task identifier. The driving task position is set as 1 for the first batch status sequence, and the driving task position is incremented sequentially along the arrangement order, and the driving task position is written into the batch status sequence.

[0043] For each batch of status sequences, driving task operation records already assigned to the same driving task are retrieved using the driving task identifier. The occurrence time and operation status of each driving task operation record are read, and the records are arranged from earliest to latest according to their occurrence time; if the occurrence times are the same, they are arranged according to the character order of the operation record identifier. The operation status is extracted one by one according to the arrangement results. For each operation status, the occurrence time and driving task identifier are retained, and these are written into consecutive status positions according to the arrangement order, forming an operation status segment.

[0044] When constructing the historical pre-shift operation sequence for model training, completed historical driving tasks are included in the complete operation state segment of the corresponding driving task. When constructing the pre-shift operation sequence for the current departure authorization judgment, the departure authorization judgment time is used as the truncation time of the current driving task's operation record. Only operation records whose occurrence time is no later than the departure authorization judgment time are included, and the operation state generated after the departure authorization judgment time of the current driving task is not read.

[0045] The batch status sequence is used as the starting part of the driving task status segment, preserving the original order of each status position within the batch status sequence. When a running status segment is retrieved, it is written starting from the first consecutive status position after the last position of the batch status sequence, making the last position of the batch status sequence adjacent to the first position of the running status segment; if no driving task running record is retrieved, the driving task status segment only retains the batch status sequence. After writing the running status segment, the pre-shift inspection status is arranged first within the driving task status segment, followed by the running status of the same driving task.

[0046] After assembling each driving task status segment, multiple driving task status segments are read in ascending order of driving task position. A continuous sequence position is assigned to the pre-shift operation sequence. The driving task status segment with position 1 is written to the starting sequence position, and then the next driving task status segment is written from the first sequence position after the last written segment, until all driving task status segments are written. During the writing process, the order of the status positions within each driving task status segment remains unchanged, resulting in a pre-shift operation sequence composed of multiple consecutively arranged driving task status segments.

[0047] In this embodiment, the improved SCINet model includes a sampling interaction tree, and node back-construction unit, differential extraction unit, two differential projection units and two correction branches set in each non-leaf level sampling interaction node; The sampling interaction tree consists of multiple sampling interaction nodes connected hierarchically in a binary tree structure. Each non-leaf level sampling interaction node includes an interleaved sampling unit, a dual-path interaction unit, and two lower-level node interfaces. The two lower-level node interfaces are connected to two lower-level sampling interaction nodes respectively. The input of the node back-construction unit is connected to the return of the two lower-level node interfaces respectively. The differential extraction unit is connected to the input of the non-leaf level sampling interaction node and the output of the node back-construction unit respectively. The input of the two differential projection units is connected to the output of the differential extraction unit. The input of the two correction branches is connected to the output of the two differential projection units respectively. The output of the two correction branches is connected to the return of the two lower-level node interfaces respectively.

[0048] In this implementation, the abnormal characteristics of the current pre-job testing batch are output, including: The pre-shift operation sequence is written into the root-level sampling interaction node of the sampling interaction tree. The interleaved sampling unit splits the node input state into two interleaved sub-sequences. The dual-path interaction unit updates the two interleaved sub-sequences and passes them to the two lower-level sampling interaction nodes through the two lower-level node interfaces until they reach the leaf-level sampling interaction node. In each non-leaf-level sampling interaction node, the node back-construction unit receives the forwarded states returned by the two lower-level sampling interaction nodes, and re-embeds the two forwarded states according to the staggered sampling positions to obtain the node back-construction state. The differential extraction unit extracts the back-construction difference between the node input state and the node back-construction state. The two differential projection units project the back-constructed differential into two differential components according to the staggered sampling positions. These components are then incorporated into two recursive states via two correction branches. The two corrected recursive states are then re-embedded and submitted to the next-level sampling interaction node. The back-construction difference generated by each non-leaf-level sampling interaction node is restored to the original position of the pre-shift operation sequence according to the staggered sampling position. The back-construction differences falling in the same original position are merged, and the back-construction difference of the current pre-shift testing batch is extracted to form the abnormal characterization of the current pre-shift testing batch.

[0049] In its implementation, the original SCINet used recursive downsampling, convolution, and an interactive structure to interleave the time series into two subsequences and extract state changes at different time scales across multiple sampling levels. The original sampling interaction nodes primarily handled the splitting of the node's input state, two-way interaction, and downward propagation. The propagated states returned by the two lower-level sampling interaction nodes were directly re-embedded and submitted upwards, lacking node-by-node verification to ensure the two propagated states could recover the current node's input state. When local states shifted during continuous splitting and interaction, the shifted state continued to propagate along the sampling interaction tree, making it difficult for the root-level output to retain the original positional order of the shifted state.

[0050] The improved SCINet model adds a node backconstruction unit, a differential extraction unit, two differential projection units, and two correction branches to each non-leaf-level sampling interaction node. After the two lower-level sampling interaction nodes return to the relay state, the node input state is first restored within the current non-leaf-level sampling interaction node. The difference between the node input state and the node backconstruction state is checked, and then the backconstruction differential is projected to the two relay states respectively and the correction is completed. Each non-leaf-level sampling interaction node independently executes "lower-level return - current backconstruction - differential extraction - dual-path correction - resubmission" once, thereby preserving the positional order and change direction of the state offset at different sampling levels, reducing the dilution of local abnormal states by multi-level relay, and enabling the abnormal representation to be located at the associated record position within the current pre-job detection batch.

[0051] Before inputting the pre-shift operation sequence into the improved SCINet model, an input state vector is generated position by position according to the sequence position of the pre-shift operation sequence. The sequence position in the batch state sequence reads the state position identifier, record type, completion mark, detection conclusion, validity mark, absence status, and driving task position; the sequence position in the operation state segment reads the unified operation state code and driving task position, and writes the state position identifier and record type specifically for the operation state respectively.

[0052] The status location identifier is converted into a 16-dimensional status location embedding through the status location embedding layer, and the record type is converted into a 16-dimensional record type embedding through the record type embedding layer. The completion flag uses 0 to indicate incomplete and 1 to indicate complete; the validity flag uses 0 to indicate invalid and 1 to indicate valid; the absence status uses 0 to indicate non-absence and 1 to indicate absence; the detection conclusion uses 0 to indicate undetermined, 1 to indicate normal, and 2 to indicate abnormal; the unified operation status code uses 1 to indicate waiting to depart, 2 to indicate departed, 3 to indicate in operation, 4 to indicate arrival at the destination, and 5 to indicate task completion.

[0053] When the sequence position in the batch status sequence does not carry the running status, the unified running status code is written as 0; when the sequence position in the running status segment does not carry the detection status, the completion mark, detection conclusion, valid mark, and missing status are all written as -1. The value -1 only indicates that the corresponding field is not applicable to the current sequence position and does not participate in the business judgment of incomplete, invalid detection, or missing status.

[0054] The 16-dimensional state position embedding, 16-dimensional record type embedding, completion marker, detection conclusion, valid marker, missing state, unified operation state code, and driving task position sequence are sequentially concatenated to form a 38-dimensional initial state vector. An input linear mapping layer then maps this 38-dimensional initial state vector to a 64-dimensional node input state, transforming the pre-shift operation sequence into an input state matrix consisting of the sequence length and 64 state channels.

[0055] The sampling interaction tree adopts a fixed three-level binary hierarchical structure. The root-level sampling interaction node serves as the first-level non-leaf-level sampling interaction node; the two lower-level sampling interaction nodes connected to the root-level node serve as the second-level non-leaf-level sampling interaction nodes; the two lower-level sampling interaction nodes connected to each second-level non-leaf-level node serve as the third-level non-leaf-level sampling interaction node; and the two lower-level node interfaces of each third-level non-leaf-level node are connected to leaf-level sampling interaction nodes. After the pre-shift operation sequence undergoes three levels of interleaved sampling, each leaf-level sampling interaction node receives the state positions of the pre-shift operation sequence that are 8 positions apart from the original position sequence.

[0056] Before writing the pre-shift running sequence into the sampling interaction tree, read the sequence length of each pre-shift running sequence within the same training batch, select the maximum sequence length, and adjust the maximum sequence length upwards to a minimum multiple of 8 (not less than 8) to obtain the batch padding length. Add all-zero state positions to the end of each pre-shift running sequence until all input state matrices reach the batch padding length. When the model infers a single pre-shift running sequence, adjust the sequence length upwards to a minimum multiple of 8 (not less than 8), and add all-zero state positions to the end of the sequence.

[0057] A masking tag is written for each all-zero state position, and an unmasked tag is written for the valid state positions in the pre-job running sequence. The three-level non-leaf level sampling interaction structure remains unchanged during training and inference. All-zero state positions with masking tags are not included in the calculation of backconstruction loss, backconstruction difference merging, and anomaly representation truncation.

[0058] Each non-leaf-level sampling interaction node receives a node input state consisting of multiple sequence positions and 64 state channels. The interleaved sampling unit extracts the state positions with odd positions from the node input state to form a first interleaved subsequence, and extracts the state positions with even positions to form a second interleaved subsequence. The interleaved sampling unit simultaneously records the original positions of each state position in the first and second interleaved subsequences within the node input state.

[0059] The dual-channel interaction unit comprises four one-dimensional convolutional sub-layers. A first interleaved subsequence is input into the first one-dimensional convolutional sub-layer to obtain a first scaling factor. An element-wise exponential mapping is performed on the first scaling factor to obtain a first scaling coefficient. The first scaling coefficient is multiplied by the corresponding position and state channel value of the second interleaved subsequence to obtain a second scaled subsequence. The second interleaved subsequence is then input into the second one-dimensional convolutional sub-layer to obtain a second scaling factor. An element-wise exponential mapping is performed on the second scaling factor to obtain a second scaling coefficient. The second scaling coefficient is multiplied by the corresponding position and state channel value of the first interleaved subsequence to obtain a first scaled subsequence.

[0060] The first scaled subsequence is input into the third one-dimensional convolutional sub-layer to obtain the first additive correction value, which is then added element-wise to the second scaled subsequence. The second scaled subsequence is input into the fourth one-dimensional convolutional sub-layer to obtain the second additive correction value, which is then added element-wise to the first scaled subsequence. The first scaled subsequence with the added second additive correction value is set to the first recursive state, and the second scaled subsequence with the added first additive correction value is set to the second recursive state.

[0061] Each of the four one-dimensional convolutional sub-layers uses a one-dimensional convolutional window of length 3, with a stride of 1. A zero-state position is padded at both ends of the sequence to maintain consistent subsequence length before and after convolution. Each one-dimensional convolutional sub-layer has 64 input channels and 128 output channels. After channel normalization of the convolutional output, it is split into two 64-channel components along the state channel. Sigmoid activation is applied to the second 64-channel component, which is then multiplied element-wise with the first 64-channel component to obtain the 64-channel convolutional result. The convolutional results of the first and second one-dimensional convolutional sub-layers are used as the first and second scaling factors, respectively. The convolutional results of the third and fourth one-dimensional convolutional sub-layers are used as the first and second additive correction factors, respectively. After processing by the dual-channel interaction unit, the two forward states are written to two lower-level sampling interaction nodes via two lower-level node interfaces.

[0062] Leaf-level sampling interaction nodes no longer perform interleaved sampling. After receiving the relayed state, the leaf-level sampling interaction node performs a local encoding on adjacent state positions using a one-dimensional convolution of length 3, and returns the local encoding result through the lower-level node interface. The number of state channels returned by the leaf-level sampling interaction node remains at 64.

[0063] After the two lower-level sampling interaction nodes complete their processing, they write the two forwarded states to the return terminals of the two lower-level node interfaces of the current non-leaf-level sampling interaction node. The node reconstructing unit reads the original positions saved by the interleaved sampling unit, places the first forwarded state bit by bit back into the odd-numbered positions of the node input state, and places the second forwarded state bit by bit back into the even-numbered positions of the node input state, thus reassembling them into the node reconstructed state. The sequence length, number of state channels, and positional arrangement of the node reconstructed state are consistent with the node input state of the current non-leaf-level sampling interaction node.

[0064] The differential extraction unit subtracts the node input state and the node reconstructed state element-wise according to the same sequence position and the same state channel to obtain the reconstructed difference. The reconstructed difference retains the positive and negative directions. A positive difference indicates that the node input state was not fully recovered in the node reconstructed state, and a negative difference indicates that the node reconstructed state produced an over-response at the corresponding position. The differential extraction unit does not take the absolute value of the reconstructed difference to avoid losing the direction of state change within the model.

[0065] The first differential projection unit reads the differential states at odd-numbered positions in the backconstructed differential, while the second differential projection unit reads the differential states at even-numbered positions. Both differential projection units employ one-dimensional convolutions with 64 channels and a convolution window length of 1 to convert the read differential states into two differential components with the same number of channels as the two forward states. The convolution window length is set to 1, ensuring that differential projection only adjusts the state channels within the same sequence position, without changing the sequence position of the backconstructed differential.

[0066] The two correction branches receive the two differential components respectively, and add the first differential component to the first recursive state position by position, and add the second differential component to the second recursive state position by position, thus obtaining two correction recursive states. The node back-construction unit reads the interleaved sampling positions again, and writes the two correction recursive states back to the odd and even positions respectively, thus obtaining the correction output state of the current non-leaf level sampling interaction node. The correction output state is written to the previous level sampling interaction node via the return terminal of the current non-leaf level sampling interaction node.

[0067] Each non-leaf-level sampling interaction node saves its input state, staggered sampling position, and back-construction difference. After the leaf-level sampling interaction node returns, node back-construction and difference correction are performed starting from the non-leaf-level sampling interaction node closest to the leaf-level sampling interaction node, and are submitted level by level to the root-level sampling interaction node. After the root-level sampling interaction node completes the correction, it outputs the root-level correction state, which is consistent with the original position sequence of the pre-shift running sequence.

[0068] The improved SCINet model is trained using historical pre-shift operation sequences. Training samples are selected from historical pre-shift detection batches that have entered the departure authorization state, have complete driving task operation records, and have saved the departure authorization judgment time. For each training sample, multiple batch state sequences and corresponding complete operation state segments of the same driver prior to the target historical pre-shift detection batch are concatenated according to the driving task position order. The target historical pre-shift detection batch is set to simulate the current pre-shift detection batch, only accessing the operation state of the driving task corresponding to the target historical pre-shift detection batch before the departure authorization judgment time, excluding the states of departed, in operation, reached destination, and task completed after the departure authorization judgment time.

[0069] The driving task state segments from previous historical pre-job detection batches and the driving task state segments corresponding to the target historical pre-job detection batch are concatenated according to the driving task position order, and a training input state matrix is ​​generated according to the aforementioned state encoding method. This ensures that the data truncation method simulating the current pre-job detection batch in the training samples is consistent with the data truncation method of the current pre-job detection batch during actual model inference.

[0070] During training, the training input state matrix is ​​first written into the sampling interaction tree to obtain the node backconstruction state, backconstruction difference, and root-level corrected state for each non-leaf-level sampling interaction node. The training loss includes root-level backconstruction loss and node-by-node backconstruction loss. The root-level backconstruction loss is the average of the element-wise absolute differences between the root-level corrected state and the root-level node input state; the node-by-node backconstruction loss is the average of the element-wise absolute differences between the node input state and the node backconstruction state for each non-leaf-level sampling interaction node, and then summed over all non-leaf-level sampling interaction nodes. The training loss is obtained by adding the root-level backconstruction loss and the node-by-node backconstruction loss.

[0071] The training parameters in the state position embedding, record type embedding, input linear mapping layer, two-way interaction unit, leaf-level sampling interaction node, and two-way differential projection unit are updated using backpropagation. This implementation uses the Adam optimizer with an initial learning rate of 0.001, a training batch size of 32, and a maximum training epoch of 100 rounds. After each training epoch, the training loss is calculated using the validation set; training ends when the validation set training loss does not decrease for 10 consecutive epochs, and the model parameters with the lowest validation set training loss in those epochs are retained.

[0072] After training, the pre-job operation sequence to be analyzed is converted into an input state matrix using the same state encoding method as the training samples, and then input into the trained improved SCINet model. The model generates back-construction differences in each non-leaf-level sampling interaction node. For back-construction differences generated at different levels in the sampling interaction tree, the corresponding staggered sampling positions are first read, and the position is restored level by level from the current node position to the root-level sampling interaction node position, and then arranged according to the original positional order of the pre-job operation sequence.

[0073] When the same original bit sequence receives backconstruction differences from multiple non-leaf-level sampling interaction nodes, the backconstruction differences within the same state channel are summed element-by-element, and then divided by the number of backconstruction differences written to the same original bit sequence to obtain the merged backconstruction difference of the original bit sequence. During the merging process, the positive and negative directions of the backconstruction differences and the 64 state channels are preserved, and absolute value processing is not performed.

[0074] The driving task identifier of the current pre-shift inspection batch is read from the pre-shift operation sequence to locate the start and end positions of the current pre-shift inspection batch within the pre-shift operation sequence. The merged back-construction difference of each original position between the start and end positions is extracted, maintaining the original sequence position of each merged back-construction difference in the batch state sequence, thus forming an anomaly representation for the current pre-shift inspection batch. Each sequence position in the anomaly representation corresponds one-to-one with the same sequence position in the current pre-shift inspection batch, and the back-construction deviation of the corresponding sequence position in different state channels is retained.

[0075] In this embodiment, the anomaly characterization based on the occurrence sequence analysis of the current pre-employment testing batch includes: Read the associated record identifier, record type and occurrence sequence of each associated record in the current pre-job testing batch, and divide the abnormal characteristics of the current pre-job testing batch into multiple record position segments according to the occurrence sequence of each associated record; The absolute value of the reconstruction difference within each recording location segment is taken and summed to obtain the location anomaly of each recording location. Bind the associated record identifier, record type, and occurrence order to the location anomaly quantity of the same record location.

[0076] In the specific implementation process, the anomaly representation of the current pre-job testing batch is stored using a two-dimensional array. The row direction of the two-dimensional array corresponds to the original position of the current pre-job testing batch in the pre-job running sequence, and the column direction corresponds to the state channel output by the improved SCINet model. Each array element stores a back-construction difference, and retains the positive and negative directions of the back-construction difference. The all-zero state positions filled in during the model input stage do not belong to the current pre-job testing batch and are not written into the anomaly representation of the current pre-job testing batch.

[0077] When the state orchestration module generates a batch state sequence, it retains the associated record identifier, record type, occurrence sequence, and current state position for each associated record. When multiple associated records are written to the same state position, they are expanded into consecutive record positions according to their occurrence sequence, with each associated record occupying a separate record position. After the anomaly representation of the current pre-job inspection batch is generated, the anomaly localization module uses the associated record identifier to read the corresponding sequence position number and retrieves the back-construction difference row at the same sequence position in the anomaly representation using the sequence position number.

[0078] Read all associated records in the current pre-employment testing batch and sort them in ascending order of occurrence. For each associated record, read the associated record identifier, record type, occurrence order, and corresponding sequence position. Using the sequence position as the position index, extract one line of back-construction difference from the anomaly characteristics of the current pre-employment testing batch and set the extracted line of back-construction difference as the record position segment of the corresponding associated record.

[0079] Each record location segment retains the associated record identifier, record type, occurrence order, sequence position, and reconstructed difference of each state channel. Each associated record corresponds to one record location segment, and each record location segment corresponds to an original position in the anomaly characterization of the current pre-job detection batch.

[0080] When calculating location anomalies, the backconstruction differences of each state channel within the recorded location segment are read sequentially. The absolute value of each backconstruction difference is taken and accumulated, and the accumulated result is used as the location anomaly for the corresponding recorded location. The backconstruction differences retain their positive and negative directions within the improved SCINet model. The absolute value is taken when calculating location anomalies to prevent backconstruction differences with opposite directions from canceling each other out. In this embodiment, when the node input state includes 64 state channels, the absolute values ​​of the 64 backconstruction differences corresponding to the same recorded location are taken and then summed.

[0081] After calculating the location anomaly, the associated record identifier is used as the binding key to write the associated record identifier, record type, occurrence sequence, and location anomaly amount into the same record location result. The associated record identifier points to the original associated record in the pre-job inspection batch, the record type uses the unified record type code generated by the pre-job access module, the occurrence sequence uses the sequence number within the batch written by the batch binding module, and the location anomaly amount is the cumulative result of the absolute values ​​of all back-construction differences within the record location segment.

[0082] All recorded location results are arranged in the order of occurrence, ensuring that the order of the recorded location results matches the order of the associated records within the current pre-shift inspection batch. Each recorded location result can retrieve the original record content, inspection status, and driving task identifier through the associated record identifier, and serves as input for location anomaly boundary comparison and blocking record location determination.

[0083] In this embodiment, determining the record location that prevents the current pre-shift inspection batch from entering the dispatch authorization state includes: From the historical pre-shift inspection batches that have entered the departure authorization state, extract the location anomalies that have the same record type and occurrence sequence, and set the maximum location anomaly as the location anomaly boundary. Compare the positional anomaly quantity of each record location in the current pre-job testing batch with the positional anomaly boundary that is the same in both record type and occurrence sequence, and retain the record locations where the positional anomaly quantity exceeds the positional anomaly boundary. Read the detection status within the reserved record positions, mark the record positions with detection status of incomplete, invalid, or abnormal as blocked record positions, and determine the blocked record positions as the record positions that prevent the current pre-shift inspection batch from entering the departure authorization state.

[0084] In the specific implementation process, before the improved SCINet model is put into operation, offline playback is performed on historical pre-shift detection batches that have entered the departure authorization state. For each historical pre-shift detection batch that has entered the departure authorization state, the batch state sequence and driving task operation data formed by the same driver in the historical pre-shift detection batch and before are retrieved. According to the processing method of the batch concatenation module, a historical pre-shift operation sequence with the historical pre-shift detection batch as the target batch is constructed. Driving tasks before the historical pre-shift detection batch are connected to the complete operation state segment, while driving tasks corresponding to the historical pre-shift detection batch are only connected to the operation state generated before the historical departure authorization judgment time. The constructed historical pre-shift operation sequence is input into the trained improved SCINet model, and the reconstructed difference of the position sequence of the historical pre-shift detection batch is extracted to generate the position anomaly of the corresponding historical pre-shift detection batch. The position anomalies are then collected according to the record type and occurrence position sequence. After each record type and occurrence position sequence combination in the pre-shift detection process forms at least one historical position anomaly, the corresponding position anomaly boundary is established, and the position anomaly boundary comparison of the current pre-shift detection batch is initiated.

[0085] Retrieves historical pre-shift inspection batches that have entered the departure authorization state from the offline playback results, retaining only those batches whose departure authorization state has been unlocked and entered the authorization state, and which have generated location anomalies. For each retained historical pre-shift inspection batch, reads the associated record identifier, record type, occurrence sequence, and location anomaly quantity, and combines the record type and occurrence sequence as the anomaly location retrieval key. The record type uses the unified record type code generated by the pre-shift access module, and the occurrence sequence uses the batch intra-batch sequence number written by the batch binding module.

[0086] The anomaly location module reads all recorded location results from the current pre-shift inspection batch. For a single recorded location, it extracts the record type and occurrence sequence, and uses the record type and occurrence sequence to form an anomaly location retrieval key. It then searches for historical locations with the same anomaly location retrieval key from historical pre-shift inspection batches that have entered the departure authorization state. During the retrieval process, anomalies with different record types or occurrence sequences are not mixed.

[0087] When establishing a set of location anomalies, the retrieved historical locations are sequentially written into the same set. Each set record retains the historical pre-job inspection batch identifier, associated record identifier, and location anomaly value. The first location anomaly value is read as the initial maximum value, and then the remaining location anomalies are read one by one. If the currently read location anomaly value is greater than the saved maximum value, the currently read location anomaly value replaces the saved maximum value. After comparing all location anomalies, the final saved maximum value is written to the location anomaly boundary corresponding to the anomaly location retrieval key.

[0088] For each record location in the current pre-job inspection batch, a location anomaly boundary retrieval is performed. The location anomaly boundary is located using the record type and occurrence order carried by the record location. The location anomaly value of the record location is directly compared with the location anomaly boundary value. If the location anomaly value is greater than the location anomaly boundary value, the record location is retained; if the location anomaly value is not greater than the location anomaly boundary value, the record location is not added to the set of record locations to be verified. The comparison process maintains the same data precision for both the location anomaly value and the location anomaly boundary value; no rounding or level conversion is performed before the comparison.

[0089] After comparing the number of location anomalies, the associated record identifier carried by the retained record location is used to retrieve the detection status bound to the same associated record identifier from the current pre-job detection batch. The detection status consists of a completion flag, a detection conclusion, and a validity flag. When the completion flag is 0, the detection status is identified as incomplete; when the validity flag is 0, the detection status is identified as invalid; when the completion flag is 1, the validity flag is 1, and the detection conclusion is abnormal, the detection status is identified as abnormal.

[0090] If the detection status of a reserved record location is incomplete, invalid, or abnormal, the record location is written to the blocked record location set. Each blocked record location retains the associated record identifier, record type, occurrence sequence, location anomaly amount, location anomaly boundary, and detection status. If the detection status is neither incomplete, invalid, nor abnormal, the reserved record location is not written to the blocked record location set.

[0091] After verifying the detection status of all retained record positions, the blocking record positions are arranged in ascending order of occurrence. The associated record identifier carried by the blocking record position is bound to the original associated record in the current pre-shift inspection batch, and the blocking record position is determined as the record position that prevents the current pre-shift inspection batch from entering the departure authorization state, so that the on-duty interlocking module can read the driving task identifier and execute the departure authorization state lock.

[0092] In this embodiment, the on-duty interlocking module is specifically as follows: Read the driving task identifier associated with each blocking record location, identify the driving task pointed to by the driving task identifier as the driving task whose departure authorization status needs to be adjusted, and switch the departure authorization status to the locked status. Receive the re-inspection results or handling results of each blocking record location, and update the detection status within the blocking record location using the re-inspection results or handling results; Check each updated blocking record location. If the detection status of all blocking record locations is complete, valid, or normal, the lock is released. If there are blocking record locations that are incomplete, invalid, or abnormal, the lock remains.

[0093] In practice, the on-duty interlocking module receives the blocking record locations output by the abnormal location module, reads the associated record identifiers carried by each blocking record location according to the occurrence sequence, and uses the associated record identifiers to retrieve the driving task identifiers written in the associated records. When multiple blocking record locations carry the same driving task identifier, the identical driving task identifiers are merged, and the departure authorization status adjustment is performed only once for the same driving task.

[0094] The departure authorization status uses at least two status codes: authorized status and locked status. The driving task pointed to by the driving task identifier is read, the current departure authorization status of the driving task is written to the locked status, and the blocking record position that triggered the lock is recorded. After the departure authorization status is written to the locked status, the dispatch confirmation interface and the vehicle key release interface do not perform departure confirmation or key release when reading the locked status.

[0095] When a blocked record location is re-inspected, the re-inspection terminal submits the driving task identifier, associated record identifier, re-inspection completion marker, re-inspection conclusion, re-inspection validity marker, and re-inspection time. When a blocked record location is manually handled, the handling terminal submits the driving task identifier, associated record identifier, handling completion marker, handling conclusion, handling validity marker, and handling time. The on-duty interlocking module uses the driving task identifier and associated record identifier to locate the blocked record location, preventing the re-inspection result or handling result from being written to other driving tasks or other associated records.

[0096] When a re-inspection result is used to replace the original detection status within the blocking record location, the re-inspection completion marker is written to the completion marker, the re-inspection conclusion is written to the detection conclusion, and the re-inspection validity marker is written to the validity marker. When a treatment result is used to replace the original detection status within the blocking record location, the treatment completion marker is written to the completion marker, the treatment conclusion is written to the detection conclusion, and the treatment validity marker is written to the validity marker. The completion marker, detection conclusion, and validity marker use the same unified encoding as the status orchestration module, and the updated three fields are recombined to represent the detection status of the blocking record location.

[0097] When multiple re-examination results or treatment results are received successively at the same blocking record location, the received results are arranged according to the re-examination time or treatment time, and the latest result is determined as the result to be written.

[0098] The writing of re-inspection or handling results, the reading of all blocking record positions for the same driving task, and the updating of departure authorization status are all performed within the same database transaction. At the start of the transaction, an exclusive lock is acquired on the departure authorization status record corresponding to the driving task identifier, the results to be written are written to the corresponding blocking record position, and all blocking record positions associated with the same driving task are reread.

[0099] A completed record location marked with 1, a valid record location marked with 1, and a normal detection result are considered as completed, valid, or normal. A completion mark of 0 indicates incompleteness, a valid record location marked with 0 indicates invalidity, and an abnormal detection result indicates abnormality.

[0100] When all blocking record positions associated with the same driving task have reached the completion, valid detection, or normal detection status, the departure authorization status is updated from locked to authorized in the database transaction. If any blocking record position is still incomplete, invalid detection, or abnormal detection, the locked status remains unchanged. After writing the departure authorization status, the database transaction is committed. If the transaction fails, the current detection status update and departure authorization status update are rolled back. The scheduling confirmation interface and vehicle key release interface read the departure authorization status after the database transaction is committed.

[0101] Example 1: To verify the feasibility of this invention in practice, it was applied to the pre-job safety management of a public transportation operating company in a certain city. This company has three operating stations with 412 drivers, executing approximately 190 scheduled departures daily. The original management method primarily relied on item-by-item judgment based on alcohol consumption, body temperature, vehicle inspection, and safety confirmation results, which resulted in problems such as scattered data, insufficient correlation of abnormal records, and a high number of mistakenly locked tasks.

[0102] During implementation, the system connects to the operations scheduling database, pre-shift safety management terminal, vehicle inspection terminal, and dispatch operation database. Pre-shift safety management data includes identity verification, body temperature, alcohol concentration, blood pressure, blood oxygen, heart rate, emotional state, reaction ability, and safety confirmation records; vehicle inspection records include the inspection results of vehicle appearance and safety facilities, braking and steering components, and onboard equipment.

[0103] The pre-shift inspection process is based on the planned arrival time to the planned departure time. Related records generated by the same driver are arranged and bound into pre-shift inspection batches according to the collection time. The pre-shift inspection process lasts an average of 42 minutes, with each batch containing an average of 16.4 related records. Status arrangement includes identity verification, physical condition detection, competency detection, vehicle three-inspection, and safety confirmation status positions. When multiple records exist at the same status position, they are formed into independent sequence positions.

[0104] The six most recent completed driving task status segments of the same driver are connected to the current driving task status segment to form the pre-shift operation sequence. Historical driving tasks are connected to the complete operation status, while the current driving task is connected to the operation status formed before the departure authorization judgment time.

[0105] The improved SCINet model adopts a three-level binary hierarchical sampling structure, with 64 channels for node input states. Each non-leaf-level sampling interaction node is equipped with a node backconstruction unit, a difference extraction unit, two differential projection units, and two correction branches. After the lower-level sampling interaction node returns to the relay state, it is re-embedded according to the staggered sampling positions to obtain the node backconstruction state. The backconstruction difference between the node input state and the node backconstruction state is extracted, and the backconstruction difference is used to correct the two relay states.

[0106] The model was trained using historical data from five consecutive months, resulting in 28,640 historical pre-employment operation sequences, of which 80% were used as the training set and 20% as the validation set. The Adam optimizer was employed with an initial learning rate of 0.001 and a training batch size of 32. The model achieved the lowest validation loss in the 73rd round.

[0107] A total of 5860 planned departure tasks within a consecutive 30-day period were selected for testing, comprising 96108 associated records. After manual review, 624 driving tasks were confirmed to have departure disruptions, containing 824 disruption record locations. The same test data was processed using the original threshold interlocking method, the traditional SCINet model method, and the method of this invention, respectively. The results are shown in Table 1 below.

[0108] Table 1. Comparison of the Effectiveness of Pre-Employment Safety Assessment in Public Transportation

[0109] Among them, the accuracy rate represents the proportion of driving tasks correctly judged as authorized or locked out of the total number of test tasks; the blocking record recall rate represents the proportion of the number of blocking record locations correctly located out of the total number of real blocking record locations; and the false lock rate represents the proportion of driving tasks that meet the departure conditions but are incorrectly locked out out of the total number of normal driving tasks.

[0110] As shown in Table 1, the original threshold interlocking method is difficult to identify the correlation between different detection states and continuous driving tasks, with a blocking record recall rate of only 81.92%. The traditional SCINet model improves the ability to extract sequence features, but local anomalies are still weakened in the multi-level sampling process. This invention restores the back-construction difference of different sampling levels to the original position of the associated records by closing the loop node by node, thereby improving the judgment accuracy to 98.58%, the blocking record recall rate to 97.33%, and reducing the false locking task rate to 1.22%.

[0111] The system is deployed on a server with a 16-core processor, 64GB of memory and a 12GB graphics card. The average processing time for a single pre-shift operation sequence is 38ms, and the average time for writing the departure authorization status is 21ms. This meets the requirements for real-time anomaly location and departure authorization interlocking for public transportation pre-shift safety management.

[0112] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A digital and intelligent public transportation pre-job safety system based on big data mining, characterized in that, include: The pre-job access module is used to access public transportation pre-job safety management data and driving task operation data, and to aggregate public transportation pre-job safety management data and driving task operation data belonging to the same driving task; The batch binding module is used to define the pre-shift inspection process around the planned departure task, bind the associated records generated by the same driver in the pre-shift inspection process into a pre-shift inspection batch, and retain the occurrence sequence of each associated record; The status arrangement module is used to organize the pre-job testing batches according to the pre-job testing process, and arrange the testing status of each associated record according to the occurrence order to form a batch status sequence. The batch serialization module is used to connect multiple batch status sequences according to the driving task sequence of the same driver, and to connect the running status of each driving task to the batch status sequence of the same driving task to form a pre-shift operation sequence. The sequence mining module is used to input the pre-job operation sequence into the improved SCINet model. The improved SCINet model sets up a node-by-node back-construction difference closed loop in each non-leaf level sampling interaction node, back-constructs the current node based on the transmission status of the lower-level sampling interaction node, extracts the back-construction difference and corrects the transmission status, and outputs the abnormal characterization of the current pre-job detection batch. The anomaly localization module is used to analyze the anomaly characteristics by combining the occurrence sequence of the current pre-shift inspection batch, and determine the record position that causes the current pre-shift inspection batch to be unable to enter the departure authorization state. The on-duty interlocking module is used to determine the driving task whose departure authorization status needs to be adjusted based on the recorded location, and to update the detection status associated with the recorded location based on the re-inspection result or handling result of the recorded location.

2. The intelligent digital system for pre-job safety in public transportation based on big data mining as described in claim 1, characterized in that, The pre-employment access module is specifically as follows: Read public transportation operation scheduling data, extract driver identifiers, planned arrival times, planned departure times, and driving task identifiers, and establish a driving task index; Access pre-job inspection records, vehicle three-inspection records, and safety confirmation records, and retain driver identification, collection time, record type, and record content; Access driving task execution records and retain driving task identifiers, occurrence times, and execution status; By using driver identification, planned arrival time, and planned departure time, driving task identification is matched to pre-shift inspection records, vehicle three-inspection records, and safety confirmation records. Pre-shift inspection records, vehicle three-inspection records, safety confirmation records, and driving task operation records with the same driving task identification are grouped into the same driving task.

3. The intelligent digital system for pre-job safety in public transportation based on big data mining as described in claim 1, characterized in that, The batch binding module is specifically as follows: Read the driver identifier, planned arrival time, planned departure time, and driving task identifier of the planned departure task, and set the start and end boundaries of the pre-shift inspection process with the planned arrival time and planned departure time respectively; Extract associated records from public transportation pre-job safety management data where the collection time is between the start and end boundaries and the driver's identifier is the same as the driver's identifier for the planned departure task, and write the driving task identifier into the extracted associated records. The associated records extracted according to the collection time are sequentially written into the batch number. The associated records with the same driving task identifier are bound into the pre-job inspection batch, and the occurrence order is recorded by the batch number.

4. The intelligent digital system for pre-job safety in public transportation based on big data mining as described in claim 1, characterized in that, The state orchestration module includes: Read the pre-job inspection process, set multiple status positions along the execution order of the pre-job inspection process, and configure the writable record type for each status position; Records of the same type within the pre-employment testing batch are written to the same status position. When there are multiple related records in the same status position, the multiple related records are arranged according to the order of occurrence. Read the completion marker, detection conclusion, and validity marker of each associated record, combine them into a detection status, and set the status position that has not been written to the associated record to an absent status; The status positions are arranged according to the execution order of the pre-job testing process to form a batch status sequence.

5. The intelligent digital system for pre-job safety in public transportation based on big data mining as described in claim 1, characterized in that, The batch serial module is specifically: Retrieve multiple batch status sequences with the same driver identifier, read the planned departure time of the driving task to which each batch status sequence belongs, arrange multiple batch status sequences according to the planned departure time, and write the driving task position order for each batch status sequence. Use driving task identifiers to retrieve driving task operation data, arrange the retrieved operation records according to the time of occurrence, extract the operation status from each operation record, and compile them into operation status segments; Set each batch status sequence as a driving task status segment, and append the retrieved running status segment to the end of the driving task status segment with the same driving task identifier; Multiple driving task status segments are connected end to end according to the driving task position sequence to form the pre-shift operation sequence.

6. The intelligent digital system for pre-job safety in public transportation based on big data mining as described in claim 1, characterized in that, The improved SCINet model includes a sampling interaction tree, and node back-construction unit, differential extraction unit, two differential projection units and two correction branches set in each non-leaf level sampling interaction node; The sampling interaction tree consists of multiple sampling interaction nodes connected hierarchically in a binary tree structure. Each non-leaf level sampling interaction node includes an interleaved sampling unit, a dual-path interaction unit, and two lower-level node interfaces. The two lower-level node interfaces are respectively connected to two lower-level sampling interaction nodes. The input of the node back-construction unit is connected to the return of the two lower-level node interfaces respectively. The differential extraction unit is connected to the input of the non-leaf-level sampling interaction node and the output of the node back-construction unit respectively. The input of the two differential projection units is connected to the output of the differential extraction unit. The input of the two correction branches is connected to the output of the two differential projection units respectively. The output of the two correction branches is connected to the return of the two lower-level node interfaces respectively.

7. A digital and intelligent public transportation pre-job safety system based on big data mining as described in claim 6, characterized in that, The output of the anomaly characteristics of the current pre-job testing batch includes: The pre-shift operation sequence is written into the root-level sampling interaction node of the sampling interaction tree. The interleaved sampling unit splits the node input state into two interleaved sub-sequences. The dual-path interaction unit updates the two interleaved sub-sequences and passes them to the two lower-level sampling interaction nodes through the two lower-level node interfaces until they reach the leaf-level sampling interaction node. In each non-leaf-level sampling interaction node, the node back-construction unit receives the forwarded states returned by the two lower-level sampling interaction nodes, and re-embeds the two forwarded states according to the staggered sampling positions to obtain the node back-construction state. The differential extraction unit extracts the back-construction difference between the node input state and the node back-construction state. The two differential projection units project the back-constructed differential into two differential components according to the staggered sampling positions. These components are then incorporated into two recursive states via two correction branches. The two corrected recursive states are then re-embedded and submitted to the next-level sampling interaction node. The back-construction difference generated by each non-leaf-level sampling interaction node is restored to the original position of the pre-shift operation sequence according to the staggered sampling position. The back-construction differences falling in the same original position are merged, and the back-construction difference of the current pre-shift testing batch is extracted to form the abnormal characterization of the current pre-shift testing batch.

8. The intelligent digital system for pre-job safety in public transportation based on big data mining as described in claim 1, characterized in that, The process of analyzing the anomaly characteristics by combining the occurrence sequence of the current pre-employment testing batch includes: Read the associated record identifier, record type and occurrence sequence of each associated record in the current pre-job testing batch, and divide the abnormal characteristics of the current pre-job testing batch into multiple record position segments according to the occurrence sequence of each associated record; The absolute value of the reconstruction difference within each recording location segment is taken and summed to obtain the location anomaly of each recording location. Bind the associated record identifier, record type, and occurrence order to the location anomaly quantity of the same record location.

9. A digital and intelligent public transportation pre-job safety system based on big data mining as described in claim 1, characterized in that, The determination of the record location that prevents the current pre-shift inspection batch from entering the dispatch authorization state includes: From the historical pre-shift inspection batches that have entered the departure authorization state, extract the location anomalies that have the same record type and occurrence sequence, and set the maximum location anomaly as the location anomaly boundary. Compare the positional anomaly quantity of each record location in the current pre-job testing batch with the positional anomaly boundary that is the same in both record type and occurrence sequence, and retain the record locations where the positional anomaly quantity exceeds the positional anomaly boundary. Read the detection status within the reserved record positions, mark the record positions with detection status of incomplete, invalid, or abnormal as blocked record positions, and determine the blocked record positions as the record positions that prevent the current pre-shift inspection batch from entering the departure authorization state.

10. A digital and intelligent public transportation pre-job safety system based on big data mining according to claim 1, characterized in that, The on-duty interlocking module is specifically as follows: Read the driving task identifier associated with each blocking record location, identify the driving task pointed to by the driving task identifier as the driving task whose departure authorization status needs to be adjusted, and switch the departure authorization status to the locked status. Receive the re-inspection results or handling results of each blocking record location, and update the detection status within the blocking record location using the re-inspection results or handling results; Check each updated blocking record location. If the detection status of all blocking record locations is complete, valid, or normal, the lock is released. If there are blocking record locations that are incomplete, invalid, or abnormal, the lock remains.

Images