Security check quality control supervision method and system based on multi-source heterogeneous data fusion

Abstract

The application discloses a security check quality control supervision method and system based on multi-source heterogeneous data fusion, relates to the technical field of security check quality control, and comprises the following steps: S1, collecting security check quality control monitoring data, and performing pretreatment on the security check quality control monitoring data; S2, generating passenger anchoring records, performing occupied section segmentation and cyclic reuse disassembly on tray records, performing time sequence constraint association and backtracking matching on X-ray records and video records, and generating security check quality control evidence chain data; S3, constructing and calling a cross-modal evidence deviation representation model to obtain evidence deviation representation values, calculating review suspicious point score values, and generating review judgment data; and S4, comparing the review suspicious point score values with supervision disposal thresholds, performing closed-loop judgment and recovery registration on abnormal supervision records, and outputting quality control supervision result data. The problems that in the prior art, multi-source data correlation is not accurate, evidence chains are prone to misalignment or breakage, review judgment is unstable, and reliable basis is lacked in closed-loop supervision are solved.

Description

Technical Field

[0001] This invention relates to the field of security inspection quality control technology, specifically to a security inspection quality control and supervision method and system based on multi-source heterogeneous data fusion. Background Technology

[0002] As the demands for security screening efficiency and risk control capabilities continue to rise in scenarios such as civil aviation, rail transit, and large event venues, security screening operations have gradually evolved from single-device image interpretation to informatization, intelligence, and interconnection. Existing technologies, on the one hand, construct a quality supervision system for security inspectors' image interpretation through X-ray security inspection equipment, image interpretation quality assessment equipment, and quality control information visualization devices; on the other hand, data interaction between the detection terminal and the quality control service terminal enables security image processing, image interpretation result comparison, and work quality assessment, thereby driving the evolution of security screening quality control from manual spot checks to a data-supported intelligent quality control model.

[0003] For example, application CN111126771B discloses a system and method for supervising and ensuring the quality of security inspectors' image recognition based on regional attention prediction. The system includes X-ray security inspection equipment, security data analysis device, security inspector image recognition information interaction device, network transmission device, image recognition quality assessment device, and quality control information visualization device. The X-ray security inspection equipment is communicatively connected to the security data analysis device and the security inspector image recognition information interaction device. The security data analysis device is also communicatively connected to the security inspector image recognition information interaction device and the network transmission device. The network transmission device is also communicatively connected to the image recognition quality assessment device and the quality control information visualization device.

[0004] For example, application CN118313706A discloses a security inspection quality control management system, method, apparatus, electronic device, and computer storage medium. This security inspection quality control management system includes a detection terminal and a quality control server. The detection terminal is used to acquire security inspection image data of the item to be inspected and first image judgment result data of the security personnel, and sends the security inspection image data and the first image judgment result data to the quality control server. The quality control server is used to process the security inspection image data to obtain second image judgment result data, compare the second image judgment result data with the first image judgment result data to obtain quality control comparison result data, and obtain the work quality data of the security personnel based on the quality control comparison result data.

[0005] However, existing technologies mainly focus on security inspection image processing, image judgment result comparison, and security personnel work quality assessment. While they can achieve image-level quality control analysis and image judgment quality management to a certain extent, they still have significant shortcomings in multi-source heterogeneous data fusion supervision scenarios centered on passenger information: First, existing systems typically use single images or single image judgment results as quality control objects, lacking a unified chain-building mechanism for passenger inspection information, tray flow information, X-ray image information, video surveillance information, and handling result information, making it difficult to form a complete evidence chain of "passenger-tray-X-ray-video-handling result"; Second, in complex scenarios such as tray recycling at the security inspection site, abnormal tray number recognition, and high-concurrency multi-channel inspection, existing systems lack a backtracking matching mechanism based on a time sliding window, making it difficult to solve problems such as tray mismatch, broken record chains, and inaccurate personnel-bag relationships; Third, existing technologies are more focused on image result comparison or image judgment quality evaluation, lacking technical means to integrate image verification, video tracing, flight-level aggregated supervision, and closed-loop retrieval of abnormal records into the same supervision link, making it difficult for security inspection quality control results to further support full-process supervision and closed-loop handling.

[0006] Therefore, in order to address the above issues, there is an urgent need for security inspection quality control and supervision methods and systems based on the fusion of multi-source heterogeneous data. Summary of the Invention

[0007] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a security inspection quality control and supervision method and system based on multi-source heterogeneous data fusion. This solves the problems in existing security inspection quality control and supervision, such as inaccurate correlation of multi-source data, easy mismatch or breakage of evidence chains, resulting in unstable review and judgment and a lack of reliable basis for closed-loop supervision.

[0008] Technical solution To achieve the above objectives, this invention provides the following technical solution: a security inspection quality control and supervision method based on multi-source heterogeneous data fusion, comprising: S1, collecting security inspection quality control monitoring data, and performing time synchronization, missing data completion, anomaly removal, noise suppression, and scale unification processing on the security inspection quality control monitoring data, outputting preprocessed security inspection quality control monitoring data; S2, generating passenger anchoring records based on the preprocessed security inspection quality control monitoring data, establishing a time sliding window around the passenger anchoring records, performing segmentation and cyclic reuse disassembly on the tray records, and performing temporal constraint association and backtracking matching on X-ray records and video records to generate security inspection quality control evidence chain data; S3, extracting X-ray image regions, video keyframe sequences, and temporal feature sequences based on the security inspection quality control evidence chain data, constructing and calling cross-modal evidence deviation... The representation model obtains the evidence deviation representation value, which is used to characterize the degree of deviation between image features and temporal features in the same security inspection quality control evidence chain data. The suspected area offset and the link temporal break are combined to calculate the review suspicion score, generating review judgment data. The suspected area offset characterizes the degree of offset of the center position of the suspected area relative to the center position of the security inspection image area. The link temporal break is used to characterize the degree of temporal connection anomalies between the stages of tray transfer, security inspection, video acquisition, image processing, and alarm deactivation. In S4, flight-level regulatory index records are generated based on the review judgment data. The review suspicion score is compared with the regulatory processing threshold to form abnormal regulatory records or through regulatory records. Closed-loop judgment and retrieval registration are performed on the abnormal regulatory records, outputting quality control regulatory result data.

[0009] Furthermore, the specific steps for collecting security inspection quality control monitoring data and performing time synchronization, missing data completion, anomaly removal, noise suppression, and scale standardization processing on the security inspection quality control monitoring data to output preprocessed security inspection quality control monitoring data are as follows: Real-time collection of security inspection quality control monitoring data, including passenger inspection channel number, passenger gate passage time, boarding pass scanning time, flight number, flight date, check-in sequence number, passenger document verification time, passenger image capture time, tray number, tray number recognition time, tray entry into recognition area time, tray exit from recognition area time, X-ray image capture time, and baggage entry time. The system collects and processes the following data: entry time value, baggage exit time value, video frame acquisition time value, video frame sequence number value, image judgment start time value, image judgment end time value, processing result generation time value, and alarm deactivation time value. For the collected security inspection quality control monitoring data, a precise time protocol synchronization algorithm is used to perform multi-source time reference synchronization processing; a forward nearest neighbor completion algorithm is used to perform missing data completion processing; a Hample filter anomaly detection algorithm is used to perform anomaly data identification and removal processing; a first-order exponential weighted moving average algorithm is used to perform temporal noise suppression processing; and a minimum-maximum normalization algorithm is used to perform scale unification processing, outputting the preprocessed security inspection quality control monitoring data.

[0010] Further, the specific steps for generating passenger anchoring records based on preprocessed security check quality control monitoring data are as follows: Read the preprocessed security check quality control monitoring data; merge the passenger gate passage time, boarding pass scanning time, passenger document verification time, and passenger image capture time into the same channel according to the passenger gate number; using the passenger gate passage time as a benchmark, select the boarding pass scanning time, passenger document verification time, and passenger image capture time with the smallest absolute time difference from the passenger gate passage time within the predetermined time tolerance range, and match them as the same passenger passage record; then form a passenger passage sequence according to the order of the passenger gate passage time, and write the flight number, flight date, and check-in number into the corresponding record of the passenger passage sequence to generate a passenger anchoring record.

[0011] Furthermore, the specific steps for generating security inspection quality control evidence chain data are as follows: A time-sliding window of length L is established around the passenger anchor record. The tray records are segmented into occupied sections and decomposed for cyclic reuse. Time-series constraint association and backtracking matching are performed on X-ray and video records. The following steps are also taken: A time-sliding window of length L is established around the passenger anchor record. The tray number identification time value, tray entry time value, tray exit time value, X-ray image acquisition time value, baggage entry time value, baggage exit time value, video frame acquisition time value, and video frame sequence number value are mapped to the corresponding time-sliding window. For tray records falling within the same time-sliding window, occupied sections are segmented according to the tray entry time value and tray exit time value. When the same tray number corresponds to multiple occupied sections, it is processed as multiple cyclically reused tray records. When the tray number is missing, duplicated, or inconsistent, all tray records within the time-sliding window are retained for subsequent backtracking matching. The time-sliding window records are then further processed. The system performs temporal constraint association on the recording, X-ray recording, and video recording. A candidate association sequence is constructed based on the chronological order of the time values ​​of the tray entering the recognition area, the tray number recognition, the X-ray image acquisition, and the luggage leaving the X-ray channel. Then, video recordings are spliced ​​together according to a continuously increasing rule for video frame numbers to obtain video segment records covering the X-ray image acquisition time values. The tray records, X-ray records, and video segment records are then linked to the corresponding passenger anchor records to generate candidate evidence chain records. A time-sliding window backtracking matching process is performed on the candidate evidence chain records. For records that fail to associate in real-time, have conflicts with tray reuse, or have abnormal tray numbers, the system backtracks forward and backward along the passenger inspection sequence, re-searching for tray records, X-ray records, and video segment records within the time-sliding window. Records with temporal breaks are filtered out, and association paths with continuous temporal connections and non-overlapping occupied segments are retained. The image judgment start time value, image judgment end time value, processing result generation time value, and alarm deactivation time value are written into the association path, outputting the security inspection quality control evidence chain data.

[0012] Furthermore, based on the security inspection quality control evidence chain data, the specific steps for extracting X-ray image regions, video keyframe sequences, and temporal feature sequences, and constructing and calling the cross-modal evidence deviation characterization model to obtain evidence deviation characterization values ​​are as follows: Historical review records are extracted from the quality control review result database. X-ray records, video clip records, and the time values ​​for tray entry into the recognition area, luggage entry into the X-ray channel, X-ray image acquisition, luggage leaving the X-ray channel, video frame acquisition, image judgment start, image judgment end, processing result generation, and alarm deactivation are read according to the security inspection quality control evidence chain data corresponding to the historical review records. The corresponding X-ray image region is extracted according to the X-ray image acquisition time value, and the video keyframe sequence covering the X-ray image acquisition time value is extracted according to the video frame acquisition time value. The time values ​​from tray entry into the recognition area to alarm deactivation are then arranged in chronological order to form a temporal feature sequence, and the review conclusion is written as a sample label. For each corresponding sample, a review training sample is generated. A cross-modal evidence deviation representation model is constructed using a cross-attention constrained cross-modal contrastive learning network. The X-ray image region from the review training sample is input into the image coding layer, and the video keyframe sequence and temporal feature sequence are input into the temporal coding layer. The image feature vector and temporal feature vector output from the same review training sample are brought closer together, while the image feature vector and temporal feature vector output from different review training samples are brought further apart. The image feature vector and temporal feature vector are then input into the deviation representation layer, and the evidence deviation representation value is output. The review training sample is input into the cross-modal evidence deviation representation model for iterative training. When the prediction error of the review conclusion is lower than the error threshold for N consecutive training cycles, the model training is considered complete. X-ray image regions, video keyframe sequences, and temporal feature sequences are obtained from the real-time security inspection quality control evidence chain data and input into the trained cross-modal evidence deviation representation model, and the evidence deviation representation value is output.

[0013] Furthermore, the specific steps for calculating the review suspicion score and generating review judgment data by combining the suspected area offset and link timing breakage are as follows: High-response region extraction is performed on the X-ray image area to obtain the suspected area location result. The pixel coordinates of the center position of the suspected area and the center position of the X-ray image area are extracted separately. The pixel distance between the two is normalized by dividing by the diagonal pixel length of the X-ray image area to obtain the suspected area offset. The data from the same security inspection quality control evidence chain, including tray transfer, X-ray screening, video acquisition, image processing, and... The alarm cancellation time values ​​at each stage are used to calculate the link timing break amount based on the timing deviation between the time values ​​at each stage. The suspected area offset is incremented by one and the natural logarithm is taken to obtain the area offset modulation term. The negative of the link timing break amount is used as the exponent to take the exponential function value of the natural constant e, and then incremented by one to obtain the timing suppression term. The evidence deviation characterization value is multiplied by the area offset modulation term and then divided by the timing suppression term to obtain the review suspicion score value. The review suspicion score value and the suspected area location result are written into the corresponding security inspection quality control evidence chain data, and the review judgment data is output.

[0014] Furthermore, the specific steps for calculating the link timing break amount based on the timing deviation between the time values ​​of each stage are as follows: The average of the time value of the luggage entering the X-ray channel and the time value of the luggage leaving the X-ray channel is determined as the median time of the machine passage. Then, the first timing deviation value between the time value of the tray entering the recognition area and the time value of the luggage entering the X-ray channel, the second timing deviation value between the time value of the X-ray image acquisition and the median time of the machine passage, the third timing deviation value between the time value of the video frame acquisition and the time value of the X-ray image acquisition, the fourth timing deviation value between the time value of the image judgment start and the time value of the X-ray image acquisition, the fifth timing deviation value between the time value of the processing result generation and the time value of the image judgment end, and the sixth timing deviation value between the time value of the alarm cancellation and the time value of the processing result generation are calculated. The absolute values ​​of the first to sixth timing deviation values ​​are taken respectively and then summed to obtain the link timing break amount.

[0015] Furthermore, based on the review judgment data, a flight-level regulatory index record is generated. The specific steps for comparing the review suspicion score with the regulatory handling threshold to form an abnormal regulatory record or a pass regulatory record are as follows: Read the review judgment data, merge the corresponding security inspection quality control evidence chain data according to the flight number and flight date to generate a flight-level regulatory index record, and write the corresponding security inspection quality control evidence chain data and the review suspicion score into the flight-level regulatory index record; compare the review suspicion score with the regulatory handling threshold. When the review suspicion score is not greater than the regulatory handling threshold, a pass regulatory record is generated; when the review suspicion score is greater than the regulatory handling threshold, an abnormal regulatory record is generated.

[0016] Furthermore, the specific steps for performing closed-loop judgment and recovery registration on abnormal supervision records and outputting quality control supervision result data are as follows: An interception marker value and suspected area location result are written to the abnormal supervision record; then, the closed-loop handling is determined based on the time value generated by the handling result and the alarm cancellation time value. When the alarm cancellation time value is missing or the time interval between the alarm cancellation time value and the time value generated by the handling result is greater than the closed-loop time limit threshold, a record to be recovered is generated, and the record to be recovered is linked to the corresponding flight-level supervision index record. A pass marker value is written to the passing supervision record. Recovery registration results are continued to be written to the record to be recovered, and the abnormal supervision record, passing supervision record, and recovery registration result are written to the quality control supervision database, outputting the quality control supervision result data.

[0017] The second aspect of this invention provides a security inspection quality control and supervision system based on multi-source heterogeneous data fusion, comprising: a data acquisition and processing module, a spatiotemporal evidence chain building module, a cross-modal fusion review module, and a closed-loop supervision and disposal module, wherein: the data acquisition and processing module is used to acquire security inspection quality control monitoring data, and perform time synchronization, missing data completion, anomaly removal, noise suppression, and scale unification processing on the security inspection quality control monitoring data, outputting preprocessed security inspection quality control monitoring data; the spatiotemporal evidence chain building module is used to generate passenger anchor records based on the preprocessed security inspection quality control monitoring data, establish a time sliding window around the passenger anchor records, perform occupancy segmentation and cyclic reuse disassembly on the tray records, and perform temporal constraint association and backtracking matching on X-ray records and video records to generate security inspection quality control evidence chain data; the cross-modal fusion review module is used to extract X-ray image regions and video keyframe sequences based on the security inspection quality control evidence chain data. Based on the temporal feature sequence, a cross-modal evidence deviation representation model is constructed and invoked to obtain evidence deviation representation values. The evidence deviation representation values ​​are used to characterize the degree of deviation between image features and temporal features in the same security inspection quality control evidence chain data. The review doubt score value is calculated by combining the suspected area offset and the link temporal break value to generate review judgment data. Among them, the suspected area offset is used to characterize the degree of offset of the center position of the suspected area relative to the center position of the security inspection image area. The link temporal break value is used to characterize the degree of abnormality in the temporal connection between each stage of tray transfer, security inspection, video acquisition, image judgment and processing and alarm cancellation. The closed-loop supervision and processing module is used to generate flight-level supervision index records based on the review judgment data, compare the review doubt score value with the supervision and processing threshold to form abnormal supervision records or through supervision records, perform closed-loop judgment and retrieval registration on abnormal supervision records, and output quality control supervision result data.

[0018] Beneficial effects The present invention has the following beneficial effects: (1) Security inspection quality control supervision method and system based on multi-source heterogeneous data fusion, by constructing a spatiotemporal association mechanism with passenger information as the core, and combining pallet occupancy segmentation, cyclic reuse disassembly and time sliding window backtracking matching processing, can achieve accurate association of passengers, pallets, X-ray records and video records in scenarios of pallet cyclic reuse, abnormal pallet number recognition and multi-channel high-concurrency inspection, thereby improving the accuracy and robustness of human bag correspondence recognition in complex environments.

[0019] (2) The security inspection quality control supervision method and system based on multi-source heterogeneous data fusion establishes security inspection quality control evidence chain data and integrates X-ray image area, video key frame sequence and temporal feature sequence on the same evidence chain. Combined with cross-modal evidence deviation characterization model, suspected area offset and link temporal break amount to generate review judgment results, which can improve the stability and consistency of review judgment and enhance the ability to identify suspected missed detection and abnormal handling scenarios.

[0020] (3) Security inspection quality control supervision method and system based on multi-source heterogeneous data fusion, by establishing the timestamp alignment relationship between image stream and video stream, and integrating images, passengers, trays, videos and handling results into the evidence chain and supervision index record, can realize one-click linkage of image review and video traceability and penetrating query of images, people, bags and videos in one integrated way, thereby improving the verifiability of security inspection quality control process and the traceability of the whole process.

[0021] (4) Security inspection quality control supervision method and system based on multi-source heterogeneous data fusion, by generating flight-level supervision index records, and performing closed-loop judgment, pending record linking and recovery registration processing on abnormal supervision records, can establish a closed-loop supervision link from abnormal discovery to personnel identification and recovery registration, and realize the logical coordination and independent supervision of security inspection execution and quality control supervision, thereby improving the full-chain precise control capability of security inspection risks. Attached Figure Description

[0022] Figure 1 A flowchart of a security inspection quality control and supervision method based on multi-source heterogeneous data fusion; Figure 2 This is a structural diagram of a security inspection quality control and supervision system based on multi-source heterogeneous data fusion. Figure 3 Flowchart for generating evidence chain for security inspection quality control; Figure 4 Layered projection diagram for reviewing suspicious points. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figures 1-4 This invention provides a technical solution: a security inspection quality control and supervision method based on multi-source heterogeneous data fusion, comprising: S1, collecting security inspection quality control monitoring data, and performing time synchronization, missing data completion, anomaly removal, noise suppression, and scale unification processing on the security inspection quality control monitoring data, and outputting preprocessed security inspection quality control monitoring data; S2, generating passenger anchoring records based on the preprocessed security inspection quality control monitoring data, establishing a time sliding window around the passenger anchoring records, performing segmentation and cyclic reuse decomposition on the tray records, and performing temporal constraint association and backtracking matching on X-ray records and video records to generate security inspection quality control evidence chain data; S3, extracting X-ray image regions, video keyframe sequences, and temporal feature sequences based on the security inspection quality control evidence chain data, constructing and calling a cross-modal evidence deviation representation model. The evidence deviation characterization value is obtained, which is used to characterize the degree of deviation between image features and temporal features in the same security inspection quality control evidence chain data. The suspected area offset and the link temporal break are combined to calculate the review doubt score value and generate review judgment data. Among them, the suspected area offset is used to characterize the degree of offset of the center position of the suspected area relative to the center position of the security inspection image area. The link temporal break is used to characterize the degree of abnormality in the temporal connection between each stage of tray transfer, security inspection, video acquisition, image judgment and processing and alarm cancellation. S4, based on the review judgment data, a flight-level supervision index record is generated. The review doubt score value is compared with the supervision and processing threshold to form an abnormal supervision record or through the supervision record. The abnormal supervision record is subject to closed-loop judgment and retrieval registration, and the quality control supervision result data is output.

[0025] Specifically, the steps for collecting security inspection quality control monitoring data and performing time synchronization, missing data completion, anomaly removal, noise suppression, and scale standardization on the security inspection quality control monitoring data, and outputting the preprocessed security inspection quality control monitoring data are as follows: Real-time collection of security inspection quality control monitoring data, including passenger screening channel number, passenger gate passage time, boarding pass scanning time, flight number, flight date, check-in sequence number, passenger document verification time, passenger image capture time, tray number, tray number recognition time, tray entry into recognition area time, tray exit from recognition area time, X-ray image capture time, and baggage entry into the X-ray channel time. The time values ​​for passenger inspection are: baggage leaving the X-ray channel, video frame acquisition, video frame sequence number, image interpretation start time, image interpretation end time, processing result generation, and alarm deactivation. Among these, the passenger inspection channel number is taken from the channel mapping table stored in the channel control unit; the passenger inspection gate passage time value is taken from the trigger timestamp written by the gate control board after the gate's infrared beam is blocked; the boarding pass scanning time value, flight number, flight date, and check-in sequence number are taken from the write-back record of the barcode parsing result from the scanning terminal; the passenger document verification time value is taken from the confirmation timestamp after identity verification by the document verification terminal; and the passenger image capture time value is taken from the shutter release time of the capture camera. The frame generation time after triggering, the tray number, and the tray number recognition time value are taken from the result record after the tray number recognition camera completes character recognition. The tray entering the recognition area time value is taken from the timestamp corresponding to the trigger signal at the leading edge of the recognition area. The tray leaving the recognition area time value is taken from the timestamp corresponding to the release signal at the trailing edge of the recognition area. The X-ray image acquisition time value is taken from the writing time after the security inspection machine's imaging plate completes image frame writing. The luggage entering the X-ray channel time value is taken from the trigger timestamp of the entrance photoelectric sensor. The luggage leaving the X-ray channel time value is taken from the trigger timestamp of the exit photoelectric sensor. The video frame acquisition time value and video frame sequence number value are taken from the acquisition time field and frame sequence in the video camera's push frame header. The fields are as follows: the image judgment start time value is taken from the timestamp of the first image judgment operation recorded after the image judgment interface is opened; the image judgment end time value is taken from the timestamp of the image judgment result confirmation and submission; the disposal result generation time value is taken from the timestamp of the disposal terminal after forming the disposal conclusion; and the alarm cancellation time value is taken from the confirmation timestamp after the alarm cancellation operation is completed. After collection, passenger records, tray records, X-ray records, video records, image judgment records, and disposal records in the same inspection process are written into the same data batch according to the passenger inspection channel number to ensure that subsequent time sequence association has a unified index basis. For the collected security inspection quality control monitoring data, a precision time protocol time synchronization algorithm is used to perform multi-source time reference synchronization processing.Specifically, the system periodically sends synchronization messages to the gate control board, barcode scanning terminal, document verification terminal, capture camera, tray number recognition camera, security screening machine, video camera, image interpretation terminal, and processing terminal. It records the sending, receiving, and returning times of these synchronization messages, calculates the clock deviation of each acquisition terminal relative to the master clock, and then rewrites the passenger gate passage time, boarding pass scanning time, passenger document verification time, passenger image capture time, tray number recognition time, tray entry into recognition area time, tray exit from recognition area time, X-ray image acquisition time, baggage entry into X-ray channel time, baggage exit from X-ray channel time, video frame acquisition time, and image interpretation start time using the master clock time. The time synchronization process, which involves setting the end time of image judgment, the time of processing result generation, and the time of alarm cancellation, ensures that multi-source records fall under the same time base. The technical principle behind this process is to eliminate the order distortion caused by local clock drift at the acquisition end, guaranteeing that subsequent time sliding window construction, time sequence constraint association, and link time sequence break calculation are all based on a unified time base. A forward proximity completion algorithm is used to complete missing data. Specifically, the security inspection quality control monitoring data after time synchronization is processed is arranged into a time sequence queue according to the passenger inspection channel number. When the video frame sequence number values ​​in adjacent records are consecutive, the video frame acquisition time value monotonically increases, or a single field is missing in a single record, the nearest valid field with the same name before the missing position is retrieved and written into the queue. For missing data points, the passenger gate passage time, boarding pass scanning time, tray number recognition time, X-ray image acquisition time, processing result generation time, and alarm deactivation time are used as primary anchor fields to maintain the original acquisition results and are not included in the supplementary data. The technical principle of missing data completion is to utilize the continuity of adjacent records within a short period of time during the same inspection process to restore non-primary anchor fields, ensuring the integrity of the data link without disrupting the true sequential relationship of key time values. Anomaly detection using the Hampshire filter algorithm is employed to perform anomaly data identification and removal. Specifically, for passenger image capture time, tray number recognition time, tray entry into recognition area time, tray exit from recognition area time, and X-ray image acquisition time... A fixed-length sliding window is constructed using the tick value, video frame acquisition time value, image judgment start time value, image judgment end time value, processing result generation time value, and alarm cancellation time value. Within each sliding window, the median and absolute median difference are calculated. Records that deviate from the median of the sliding window by more than a predetermined multiple of the absolute median difference are judged as abnormal records and deleted. Records with abrupt changes in the tray number and abnormal expansion of the time interval between adjacent records are judged as abnormal records and deleted. The technical principle of abnormal data identification and removal is to use the steady-state center of the local temporal distribution to identify isolated spikes, abnormal jumps, and out-of-order insertions, so as to prevent abnormal records from entering the subsequent passenger anchoring record generation process. A first-order exponential weighted moving average algorithm is used to perform temporal noise suppression processing.Specifically, a smoothing sequence is constructed from the time values ​​of passenger image capture, tray entry into the recognition area, tray exit from the recognition area, X-ray image capture, and video frame capture. The current sampled value is combined with the previous smoothed value according to a predetermined smoothing coefficient to obtain the current smoothed result. The original sampled value is then replaced with the current smoothed result. The technical principle of temporal noise suppression is to reduce the disturbances of sensor trigger jitter, frame generation jitter, and write delay jitter on the key time axis, ensuring a more stable input basis for subsequent tray occupancy segmentation, video key frame extraction, and temporal feature sequence construction. The minimum-maximum normalization algorithm is used to perform scale unification processing, and the preprocessed security inspection quality control monitoring data is output. Specifically, the time values ​​of tray number recognition, tray entry into the recognition area, and tray exit are processed. The maximum and minimum values ​​for the following time points within the current processing batch are extracted: identification area time point, X-ray image acquisition time point, baggage entry time point, baggage exit time point, video frame acquisition time point, image judgment start time point, image judgment end time point, processing result generation time point, and alarm deactivation time point. These values ​​are then mapped to a unified numerical range. Passenger inspection channel number, flight number, flight date, check-in sequence number, tray number, and video frame sequence number values ​​retain their original format and are not involved in scale unification. The technical principle of scale unification is to eliminate the differences in magnitude between different continuous numerical fields and remove the dimensional influence between them. This ensures that subsequent cross-modal evidence deviation characterization, suspected area offset calculation, and link temporal break calculation are based on directly comparable dimensionless numerical values.

[0026] This implementation plan establishes a unified collection standard, a unified time benchmark, and a unified scale for security inspection quality control monitoring data. This ensures that passenger records, tray records, X-ray records, video records, image analysis records, and handling records generated under the constraints of passenger inspection channel numbers have a data foundation that can be directly aligned, continuously tracked, and stably compared. This significantly reduces the interference of multi-source record distortion, missing key fields, abnormal jumps, and trigger jitter on the construction of security inspection quality control evidence chain data. In turn, it improves the accuracy of passenger anchor record generation, the stability of time sliding window matching, the credibility of cross-modal evidence deviation representation, and the consistency of review doubt score calculation, providing a reliable premise for the generation of subsequent review judgment data.

[0027] Read the pre-processed security check quality control monitoring data, and merge the passenger gate passage time, boarding pass scanning time, passenger document verification time, and passenger image capture time by passenger gate number. During merging, first, establish an event cache table within the channel using the passenger gate number as the primary key, and then use the passenger gate passage time as the primary sorting field to perform ascending reordering of the boarding pass scanning time, passenger document verification time, and passenger image capture time, so that passenger check events collected separately under the same passenger gate number fall into a unified time axis. Using the passenger gate passage time as a benchmark, select the boarding pass scanning time, passenger document verification time, and passenger image capture time with the smallest absolute value of the time difference from the passenger gate passage time within a range of 0.5 seconds to 3 seconds before and after, and match them as the same passenger check record. During matching, for each passenger gate passage time, the event cache table is created using the passenger gate passage time as the primary key. The passenger screening gate calculates the absolute value of the first time difference between its time value and the scanning time values ​​of all candidate boarding passes, the second time difference between its time value and the verification time values ​​of all candidate passenger documents, and the third time difference between its time value and the image capture time value of all candidate passengers. It then selects the boarding pass scanning time value, passenger document verification time value, and passenger image capture time value with the smallest value within a range of 0.5 to 3 seconds before and after the current passenger screening record. If the same candidate record is simultaneously matched by multiple passenger screening gates, the passenger screening gate with the smallest absolute time difference from the current candidate record is retained. The remaining unmatched passenger screening gate time values ​​are used to continue searching for the next candidate record, thus preventing the same boarding pass scanning record, the same document verification record, and the same image capture record from being repeatedly linked to multiple passenger screening records.The 5-second to 3-second range covers the time fluctuation range formed by the gate release trigger delay, the barcode scanning terminal parsing delay, the document verification response delay, and the image capture trigger delay. Its technical principle is that the passenger gate passage time value is closest to the actual time the passenger crosses the passage; the boarding pass scanning time value reflects the time the flight binding action occurs; the passenger document verification time value reflects the time the identity confirmation action is completed; and the passenger image capture time value reflects the time the visual retention action is triggered. By converging these three types of events to the passenger gate passage time value, the identity information, flight information, and image information of the same passenger in a single passage process can be unified to a similar time anchor point. Then, a passenger passage sequence is formed according to the order of the passenger gate passage time values, and the flight number, flight date, and check-in sequence number are written into the corresponding record of the passenger passage sequence, generating a passenger anchoring record. During writing, the current passenger gate passage time value is used as the reference. The barcode parsing result corresponding to the successfully matched boarding pass scanning time value is used as the source for writing the flight number, flight date, and check-in sequence number. The identity verification result corresponding to the passenger document verification time value is used as the basis for validating the current passenger's security check record. If the same passenger's security check record contains multiple flight numbers, multiple flight dates, and multiple check-in sequence numbers, the set of fields with the smallest absolute value of the time difference with the passenger's security gate passage time value is retained. When generating the passenger anchoring record, the passenger security check channel number, passenger security gate passage time value, boarding pass scanning time value, passenger document verification time value, passenger image capture time value, flight number, flight date, and check-in sequence number are written into the same record unit. An anchoring index is formed by adding the passenger security check channel number to the passenger security gate passage time value, ensuring that subsequent tray record mapping, time sliding window construction, and security check quality control evidence chain data generation are all based on a unique passenger anchoring.

[0028] In this implementation plan, by uniformly merging, temporally consolidating, and resolving conflicts of passenger gate passage time values, boarding pass scanning time values, passenger document verification time values, and passenger image capture time values ​​under the constraint of the same passenger inspection channel number, the originally scattered passenger identity confirmation information, flight binding information, and visual retention information can be stably aggregated into the same passenger anchor record. This ensures that the passenger anchor record has a clear time anchor point, clear channel affiliation, clear flight direction, and clear identity association, thereby significantly reducing the risk of anchor offset caused by the cross-mixing of adjacent passenger records in high-concurrency continuous inspection scenarios, and improving the accuracy of subsequent tray record mapping, the effectiveness of time sliding window constraints, and the stability of security inspection quality control evidence chain data construction.

[0029] Specifically, the steps for generating security inspection quality control evidence chain data are as follows: A time sliding window is established around passenger anchor records; occupied segmentation and cyclic reuse are performed on tray records; temporal constraint association and backtracking matching are performed on X-ray and video records. Figure 3As shown, a time sliding window of length L is established around the passenger anchor record, where L ranges from 4 seconds to 15 seconds. The time sliding window uses the passenger's passage through the gate in the previous passenger anchor record as the center time, extending forward by 1 to 5 seconds and backward by 3 to 10 seconds, thus covering the continuous event intervals corresponding to tray entering the recognition area, X-ray passing through the machine, video acquisition, image processing, and alarm deactivation. The time values ​​of tray number recognition, tray entering the recognition area, tray leaving the recognition area, X-ray image acquisition, baggage entering the X-ray channel, baggage leaving the X-ray channel, video frame acquisition, and video frame sequence number are mapped to the corresponding time sliding window. During the scanning process, the system first performs a same-channel search based on the passenger inspection channel number. Then, it determines the window assignment based on the time deviation of each time value relative to the passenger inspection gate's passage time value. Only records whose time deviation falls within the current time sliding window are retained, thus restricting tray records, X-ray records, and video records to the same passenger event neighborhood. For tray records falling within the same time sliding window, the occupied segments are divided according to the tray's entry into the identification area time value and the tray's exit from the identification area time value. The entry into the identification area time value is used as the starting point of the occupation, and the exit from the identification area time value is used as the ending point. The continuous interval between the starting point and the ending point is recorded as a single tray occupied segment. All tray occupied segments are then sorted in ascending order of the starting time. When the same pallet number corresponds to multiple occupied sections, it is processed as multiple cyclically reused pallet records. During processing, the occupied sections with different times under the same pallet number are written into independent record units, and then the split confirmation is performed based on the time interval between the occupied sections. When the time interval between adjacent occupied sections is within the range of 0.5 seconds to 20 seconds, it is considered a cyclic reuse phenomenon of the same pallet number in a continuous inspection scenario. When the pallet number is missing, duplicated, or inconsistent, all pallet records within the retention time sliding window are entered into subsequent backtracking matching. Missing means the pallet number field is empty, duplicate means the same pallet number appears in different occupied sections, and inconsistent means adjacent records within the same occupied section are not matched. The corresponding tray number characters change, thus avoiding premature deletion of potentially valid records under abnormal tray number conditions; temporal constraint association is performed on tray records, X-ray records, and video records within the time sliding window. Candidate association sequences are constructed according to the order of the tray entering the recognition area, the tray number recognition time, the X-ray image acquisition time, and the luggage leaving the X-ray channel. During construction, it is first verified that the tray entering the recognition area is no later than the tray number recognition time, then the tray number recognition time is no later than the X-ray image acquisition time, and then the X-ray image acquisition time is no later than the luggage leaving the X-ray channel. Only records that meet all the sequential constraints are combined and written into the candidate association sequence.The video records are then spliced ​​according to the rule of continuously increasing video frame sequence values. During splicing, a difference of 1 between video frame sequence numbers is used as the continuity criterion, and a monotonically increasing video frame acquisition time value is used as the time criterion. Continuous frames are combined into a single video segment record, resulting in a video segment record covering the X-ray image acquisition time value. Coverage is determined by the condition that the acquisition time value of the first frame of the video segment record is no later than the acquisition time value of the X-ray image, and the acquisition time value of the last frame of the video segment record is no earlier than the acquisition time value of the X-ray image. The tray record, X-ray record, and video segment record are then linked to the corresponding passenger anchor record. During linking, the passenger anchor record is used as the main index unit, and the passenger inspection channel number plus the passenger inspection gate passage time value is used as the link. The system sequentially writes tray records, X-ray records, and video clip records that meet the constraints of window co-operation, time sequence, and coverage into the same link unit to generate candidate evidence chain records. A time-sliding window backtracking matching process is then performed on the candidate evidence chain records. For records that fail to be associated in real-time, have tray reuse conflicts, or have abnormal tray numbers, the system backtracks forward and backward along the passenger inspection sequence, re-searching for tray records, X-ray records, and video clip records within the time-sliding window. During backtracking, one to five passenger anchor records on each side of the previous passenger anchor record are used as the expansion range. Then, the entire time-sliding window within the expansion range is used to re-execute record recall, thereby recovering the missing records caused by boundary errors, abnormal tray numbers, and reuse conflicts. Records with temporal breaks are filtered out according to the following rules: the time value of the tray entering the recognition area is later than the time value of the X-ray image acquisition; the time value of the X-ray image acquisition is later than the time value of the luggage leaving the X-ray channel; the video clip recording does not cover the time value of the X-ray image acquisition; the time value of the image judgment start is earlier than the time value of the X-ray image acquisition by a large deviation range; and the time value of the alarm cancellation is earlier than the time value of the processing result generation. Records with excessive deviation ranges are all considered to have temporal breaks. The large deviation range is defined as 2 to 30 seconds. Related paths with continuous temporal connections and non-overlapping occupied segments are retained. Continuity is determined by the absolute value of the time deviation between adjacent records not exceeding 5 seconds, and non-overlapping occupied segments are determined by the start time of the later occupied segment not being earlier than the end time of the previous occupied segment. The image judgment is then opened... The start time value, image judgment end time value, handling result generation time value, and alarm cancellation time value are written into the associated path, outputting security inspection quality control evidence chain data. The technical principle of this process lies in using passenger anchor records to limit the starting point of the person-bag association, using a time sliding window to limit the multi-source record search range, using tray occupancy segmentation to identify the boundary of a single tray occupancy, using cyclic reuse to disassemble and separate the multiple uses corresponding to repeated tray numbers, using time-sequence constraints to ensure the sequential relationship between tray flow, X-ray screening, and video acquisition, and using a time sliding window for backtracking and matching to repair link breaks caused by real-time matching failures. This forms a security inspection quality control evidence chain data that simultaneously possesses passenger orientation, tray uniqueness, image verifiability, and video traceability.

[0030] In this implementation plan, by using passenger anchor records as a unified starting point for association, the tray number, the time value of the tray entering the identification area, the time value of the tray leaving the identification area, the time value of the X-ray image acquisition, the time value of the baggage entering the X-ray channel, the time value of the baggage leaving the X-ray channel, the time value of the video frame acquisition, the video frame sequence value, the time value of the image judgment start, the time value of the image judgment end, the time value of the handling result generation, and the time value of the alarm cancellation are all converged into the same security inspection quality control evidence chain data. This can maintain the stability of the link attribution under complex conditions such as tray recycling, abnormal tray number, and real-time association failure, significantly reducing the probability of misattachment, omission, and breakage among passenger records, tray records, X-ray records, and video records. The resulting security inspection quality control evidence chain data simultaneously has clear temporal boundaries, clear spatial attribution, clear review entry, and clear handling continuity relationship, providing continuous and reliable data support for subsequent cross-modal evidence deviation representation, review doubt score calculation, and abnormal supervision record generation.

[0031] Specifically, the steps for extracting X-ray image regions, video keyframe sequences, and temporal feature sequences based on security inspection quality control evidence chain data, and constructing and calling a cross-modal evidence deviation characterization model to obtain evidence deviation characterization values ​​are as follows: Historical review records are extracted from the quality control review result database. During extraction, historical review records are first grouped and organized by flight number, flight date, and passenger inspection channel number, and then sorted in ascending order by the time value of the processing result generation. Only valid records containing review conclusions, security inspection quality control evidence chain data, X-ray records, and video clip records are retained. The quality control review result database stores the review conclusions corresponding to historical security inspection quality control evidence chain data, suspected area location results, review suspicion score values, abnormal supervision records, and passing supervision records. The system reads X-ray records, video clips, and time values ​​for tray entry into the recognition area, baggage entry into the X-ray channel, X-ray image acquisition, baggage exit from the X-ray channel, video frame acquisition, image judgment start, image judgment end, processing result generation, and alarm deactivation based on the security inspection quality control evidence chain data corresponding to historical review records. During reading, the passenger anchoring record in the security inspection quality control evidence chain data is used as the main index unit, and the passenger inspection channel number plus the passenger gate passage time value is used as the sample retrieval key. Image data, video data, and time-series data in the same chain are grouped into the same training sample cache area. The system extracts the corresponding X-ray image region based on the X-ray image acquisition time value, using the image frame corresponding to the X-ray image acquisition time value as the center frame. Within the center frame, the baggage imaging area corresponding to the current security inspection quality control evidence chain data is located. The system then uses the outermost part of the baggage imaging area in the center frame as the reference point. A rectangle serves as the base region, and its boundaries are expanded outwards by 5 to 40 pixels to obtain the corresponding X-ray image region. The technical principle of this boundary expansion is to preserve the structural context around the suspected region, enabling subsequent image coding layers to simultaneously learn local abnormal texture features and surrounding background contour features. A video keyframe sequence covering the X-ray image acquisition time value is extracted based on the video frame acquisition time value. During extraction, the X-ray image acquisition time value is used as the video alignment center time. 2 to 12 frames are selected before the center time and 2 to 12 frames are selected after the center time. These frames are then arranged in ascending order according to the video frame number value to form a continuous set of keyframes, thus obtaining a video keyframe sequence covering the X-ray image acquisition time value. The technical principle of the frame range selection is to simultaneously preserve the continuous behavioral information of the tray entering the field of view, the passenger approaching the security check area, and the luggage leaving the security check area, enabling the temporal coding layer to learn the behavioral evolution features corresponding to the current re-examination object.The time values ​​from the tray entering the recognition area to the alarm cancellation time are then arranged in chronological order to form a time-series feature sequence. This sequence starts with the tray entering the recognition area and ends with the alarm cancellation time. The time values ​​for luggage entering the X-ray channel, X-ray image acquisition, luggage leaving the X-ray channel, video frame acquisition, image judgment start, image judgment end, and processing result generation are sequentially written into a unified time-series vector. The difference between two adjacent time values ​​in the unified time-series vector is simultaneously written to the adjacent time difference position, thus ensuring the time-series feature sequence contains both absolute time information and stage interval information. The review conclusion is then written as a sample label. For each corresponding sample, re-examination training samples are generated. During the writing process, samples with re-examination conclusions of "abnormal" are marked as Class I training samples, and samples with re-examination conclusions of "pass" are marked as Class II training samples. This provides a basis for positive and negative sample division for subsequent cross-modal contrastive learning. A cross-modal evidence deviation representation model is constructed using a cross-attention constrained cross-modal contrastive learning network. During construction, an image coding layer is set in the image branch, a temporal coding layer is set in the temporal branch, a cross-attention fusion layer is set between the image coding layer and the temporal coding layer, and a deviation representation layer is set after the cross-attention fusion layer. The convolutional kernel size of the image coding layer ranges from 3×3 to 7×7, and the convolutional stride ranges from 1 to 2. The image feature vector dimension ranges from 64 to 512, the hidden state dimension of the temporal coding layer ranges from 32 to 256, the number of attention heads in the cross-attention fusion layer ranges from 2 to 8, the output dimension of the deviation representation layer ranges from 1 to 32, and the evidence deviation representation value ranges from 0 to 2. The X-ray image region from the review training samples is input into the image coding layer. Before input, the X-ray image region is scaled to a fixed size, ranging from 128×128 pixels to 512×512 pixels, and then the pixel values ​​are mapped to the 0-1 range to ensure different samples enter a unified image input scale. The video keyframe sequence and temporal sequence... The feature sequence is input into the temporal coding layer. During input, the video keyframe sequence is first expanded into frame-by-frame visual features. Then, the frame-by-frame visual features and the temporal feature sequence are aligned according to their temporal positions to form a joint temporal input tensor. The time step length of the joint temporal input tensor ranges from 5 to 25 steps. For the same reviewed training sample, the image feature vector and temporal feature vector are processed to be closer together. This is achieved by minimizing the feature distance between the image feature vector and the temporal feature vector within the same sample. The feature distance is represented by a normalized cosine distance or Euclidean distance. The cosine distance ranges from 0 to 2, and the normalized Euclidean distance ranges from 0 to 1.The image feature vectors and temporal feature vectors output from different review training samples undergo a widening process. This widening process is achieved by maximizing the feature distance between different samples. The image feature vectors and temporal feature vectors are then input into a deviation representation layer, which outputs an evidence deviation representation value. A value closer to 0 indicates higher consistency between image and temporal features, while a value closer to 2 indicates a greater degree of deviation. The technical principle behind this process lies in utilizing image information, video behavior information, and handling timing information from the same security inspection quality control evidence chain data. The inherent correspondence between information is used to construct cross-modal consistency constraints, and the differences between different security inspection quality control evidence chains are used to construct cross-modal separation constraints. This enables the deviation representation layer to learn cross-modal deviation features that can reflect the risk of abnormality in the re-inspection. The re-inspection training samples are input into the cross-modal evidence deviation representation model for iterative training. During training, a batch gradient update method is used, with the number of samples in a batch ranging from 8 to 128, the learning rate ranging from 0.0001 to 0.01, the number of sample traversals in a single training cycle being 1, and the number of training cycles N ranging from 10 to... The model is considered complete after 300 iterations. The error is considered complete when the prediction error of the review conclusion is below a threshold value for N consecutive training cycles. The error threshold ranges from 0.001 to 0.05. The prediction error of the review conclusion is represented by the average loss value between the predicted label and the true label. X-ray image regions, video keyframe sequences, and temporal feature sequences are obtained from the real-time security inspection quality control evidence chain data and input into the trained cross-modal evidence deviation representation model. The same truncation, alignment, scaling, and vector construction rules as the review training samples are used during input. The output is the evidence deviation model. The technical principle behind this process is to construct labeled cross-modal training samples from historical review records. This allows the cross-modal evidence deviation representation model to learn the joint differences between normal and abnormal security inspection quality control evidence chain data in terms of image regions, video behavior, and temporal evolution. Real-time security inspection quality control evidence chain data is then mapped to the same feature space, resulting in an evidence deviation representation value that directly characterizes the degree of deviation in the current evidence chain. This provides a unified representation basis for subsequent calculations of suspected area offset, link temporal breakage, and review suspicion score.

[0032] In this implementation plan, by uniformly incorporating X-ray image regions, video keyframe sequences, and temporal feature sequences into the same review training sample, the image anomaly information, behavioral change information, and handling evolution information in the security inspection quality control evidence chain data can be compressed into the same evidence deviation representation value. This enables the degree of deviation between different security inspection quality control evidence chain data to have a unified judgment scale, significantly reducing the risk of local misjudgment caused by relying solely on security inspection image regions, improving the ability of review conclusions to identify cross-modal anomaly associations, and thus enhancing the stability of the evidence deviation representation value in representing the real review risk. This provides a reliable input basis for subsequent calculation of suspected area offset, calculation of link temporal breakage, and generation of review suspicion score values.

[0033] Specifically, the steps for calculating the review suspicion score and generating review judgment data by combining the suspected area offset and link timing breakage are as follows: High-response region extraction is performed on the X-ray image region to obtain the suspected area location result. The pixel coordinates of the center position of the suspected area and the center position of the X-ray image region are extracted separately. The pixel distance between the two is normalized by dividing by the diagonal pixel length of the X-ray image region to obtain the suspected area offset. During high-response region extraction, grayscale recalibration is first performed on the X-ray image region, mapping the grayscale values ​​to a closed interval of zero to one. Then, a fixed-size sliding window is used to scan the X-ray image region point by point. The sliding window side length ranges from eight to thirty-two pixels. For each sliding window, local grayscale abrupt change value, local texture density value, and local boundary closure value are calculated and sequentially written into the response intensity map. Subsequently, threshold segmentation is performed on the response intensity map, retaining the pixel set whose response intensity is higher than the current response intensity average threshold by 1.2 to 2.5 times. Then, connected component aggregation is performed on the retained pixel set, deleting pixels with areas lower than the minimum face value. For isolated regions with a product threshold, the minimum area threshold ranges from twelve to eighty pixels. The connected region with the largest area and highest average response intensity is retained as the location result of the suspected region. When extracting the pixel coordinates of the center position of the suspected region and the center position of the X-ray image region, the coordinates of the upper left and lower right vertices of the bounding rectangle of the suspected region are extracted first. Then, the average of the horizontal and vertical coordinates of both is taken as the pixel coordinates of the center position of the suspected region. Subsequently, the coordinates of the upper left and lower right vertices of the X-ray image region are extracted, and the average of the horizontal and vertical coordinates of both is taken as the pixel coordinates of the center position of the X-ray image region. When normalizing by dividing the pixel distance between the two by the diagonal pixel length of the X-ray image region, the Euclidean pixel distance is first calculated based on the two sets of center position pixel coordinates. Then, the diagonal pixel length is calculated based on the width and height pixel values ​​of the X-ray image region. Finally, the Euclidean pixel distance is divided by the diagonal pixel length to obtain the normalized suspected region offset, which ranges from zero to one.The technical principle of this processing lies in the fact that high-response region extraction can centrally express local gray-level anomalies, local texture anomalies, and local boundary anomalies within an X-ray image region as a single suspected region location result. The pixel coordinates of the center position can transform the suspected region location result into a spatial location representation, and the normalization of diagonal pixel length can eliminate the dimensional influence caused by differences in the size of different X-ray image regions. This makes the suspected region offset a dimensionless spatial offset representation that can be directly used in the calculation of the review suspicion score. The process reads the time values ​​of each stage in the same security inspection quality control evidence chain data: tray flow, X-ray screening, video acquisition, image processing, and alarm deactivation. During reading, the time values ​​of tray entering the recognition area and tray leaving the recognition area are categorized into the tray flow stage; the time values ​​of luggage entering the X-ray channel, X-ray image acquisition, and luggage leaving the X-ray channel are categorized into the X-ray screening stage; the video frame acquisition time value is categorized into the video acquisition stage; and the time values ​​of image processing start, image processing end, and processing result generation are categorized into the image processing stage. The alarm deactivation time value is categorized into the alarm deactivation stage. Based on the time values ​​of each stage... The timing deviation between the two points is used to calculate the timing breakage of the link. The calculation first determines the median time of baggage entering and leaving the X-ray channel as the average of the times. Then, the following timing deviations are calculated: the first between the time of tray entering the recognition area and the time of baggage entering the X-ray channel; the second between the time of X-ray image acquisition and the median time of passage; the third between the time of video frame acquisition and the time of X-ray image acquisition; the fourth between the time of image judgment start and the time of X-ray image acquisition; and the fifth between the time of processing result generation and the time of image judgment end. The sixth time-series deviation value between the sequence deviation value, the alarm release time value, and the processing result generation time value is calculated. Then, the absolute values ​​of the first to sixth time-series deviation values ​​are taken and summed to obtain the link time-series breakage amount. The link time-series breakage amount ranges from 0 to an unbounded positive number. The technical principle of this processing is that there is a stable sequential relationship between different stages. If there is a recording delay, recording insertion, or recording jump between tray transfer, X-ray machine, video acquisition, image processing, and alarm release, the cumulative deviation value between the time values ​​of multiple stages will increase significantly. The degree of time-series breakage of the entire evidence chain can be quantified by using the cumulative deviation. The region offset modulation term is obtained by adding one to the offset of the suspected region and taking the natural logarithm. The technical principle of taking the natural logarithm is to compress the linear amplification effect of large offset samples on the result, so that the spatial deviation effect remains monotonically increasing and not excessively unbalanced. The inverse of the link time sequence break is used as the exponent to take the exponential function value of the natural constant e, and then added to one to obtain the time sequence suppression term. The technical principle of the time sequence suppression term is to map the link time sequence break into a nonlinear suppression factor, so that the security inspection quality control evidence chain data with continuous time sequence maintains high credibility, and the security inspection quality control evidence chain data with obvious time sequence breaks is more strongly suppressed in subsequent scoring.Multiplying the evidence deviation representation value by the regional offset modulation term and then dividing by the temporal suppression term yields the re-examination doubt score. The calculation formula is: the re-examination doubt score equals the evidence deviation representation value multiplied by the natural logarithm of the suspected regional offset plus one, then divided by the inverse power of the link temporal breakage amount plus the natural constant e. A larger re-examination doubt score indicates more significant re-examination doubts arising from the superposition of cross-modal evidence deviation, spatial location deviation, and temporal continuity disruption in the current security inspection quality control evidence chain data. The re-examination doubt score and the suspected regional location result are then written into the corresponding security inspection quality control evidence chain data. Upon input, the review suspicion score is written to the scoring field, the suspected area location result is written to the location field, and the passenger inspection channel number plus the passenger inspection gate passage time value is used as the writing index to output the review judgment data. The technical principle of this processing is to characterize cross-modal consistency anomalies through evidence deviation characterization values, spatial deviation intensity through suspected area offset, and temporal continuity impairment through link temporal breakage quantity. Then, these three types of quantities are uniformly mapped to the same scoring link, thereby compressing image anomalies, behavioral anomalies, and temporal anomalies into review suspicion score values ​​that can be directly used for subsequent regulatory judgment.

[0034] The specific formula for calculating the score of the re-examination of doubtful points is as follows: ; In the formula, F represents the review score of suspicious points, D represents the evidence deviation value, P represents the offset of the suspected area, and T represents the link time sequence break.

[0035] In this implementation plan, by converging the evidence deviation characterization value, the suspected area offset, and the link temporal breakage amount into the same review doubt score value, the cross-modal deviation characteristics, spatial anomaly characteristics, and temporal distortion characteristics in the security inspection quality control evidence chain data can be transformed into a unified and comparable risk characterization result. This enables the review anomaly degree among different security inspection quality control evidence chain data to have a consistent judgment scale, significantly reducing the risk of local bias caused by relying solely on X-ray image areas, improving the ability of review judgment data to distinguish real abnormal links, and thus providing a stable and reliable quantitative basis for the generation of abnormal supervision records, closed-loop judgment, and retrieval registration.

[0036] In this embodiment, Table 1 is a data table of review doubt score values, listing the calculated review doubt score values ​​for five groups of samples. Specifically: Group 1: Evidence deviation characterization value is 0.82, suspected area offset is 12.4, link timing break is 1.35, and review doubt score value is 1.6900. Group 2: Evidence deviation characterization value is 1.06, suspected area offset is 18.7, link timing break is 2.10, and review doubt score value is 2.8148. Group 3: Evidence deviation characterization value is 0.67, suspected area offset is 9.8, link timing break is 0.92, and review doubt score value is 1.1400. Group 4: Evidence deviation characterization value is 1.24, suspected area offset is 25.3, link timing break is 2.86, and review doubt score value is 3.8347. Group 5: Evidence deviation score is 0.93, suspected area offset is 15.6, link timing break is 1.74, and review doubt score is 2.2226.

[0037] Table 1. Data on Scoring Values ​​for Re-examination of Suspicious Points

[0038] like Figure 4 As shown in the figure, the distribution relationship of five groups of samples in three dimensions—suspect area offset, link temporal break, and evidence deviation characterization value—is displayed in a three-dimensional layered projection. The horizontal axis represents the suspect area offset, which characterizes the degree of offset of the center position of the suspect area relative to the center position of the security inspection image area; the vertical axis represents the link temporal break, which characterizes the degree of temporal connection abnormality between each stage of tray transfer, security inspection, video acquisition, image analysis and processing, and alarm cancellation; the vertical height represents the evidence deviation characterization value, which characterizes the degree of deviation between image features and temporal features in the same security inspection quality control evidence chain data; the values ​​marked at the top of each projection are the review doubt score values. In the figure, the suspected area offset, link timing break, and evidence deviation characterization value of Group 3 are all at a low level, and the review doubt score marked at the top is the smallest; Group 4 is at a high level in all three dimensions, and the review doubt score marked at the top is the largest; Groups 1, 5, and 2 are distributed in the middle area, reflecting that as the suspected area offset, link timing break, and evidence deviation characterization value increase together, the review doubt score shows a gradual increasing trend. Figure 4 This paper intuitively demonstrates the coupling relationship between the evidence deviation characterization value, the suspected area offset, and the link temporal breakage amount on the review doubt score value in this invention. It can reflect the differences of different security inspection quality control evidence chain data in the three dimensions of cross-modal evidence deviation, spatial offset, and temporal breakage, and verify the feasibility of the review judgment mechanism of this invention.

[0039] Specifically, the steps for calculating the link timing break based on the timing deviation between time values ​​at each stage are as follows: The average of the time values ​​when baggage enters and leaves the X-ray channel is determined as the median time of passage. To determine this, first read the baggage entry and exit times from the same security inspection quality control evidence chain data, then add the two values ​​together and divide by two to obtain the median time of passage located at the center of the passage interval. Under the condition that the conveyor belt speed remains stable and baggage passes continuously along the X-ray channel, the median time of passage approximately represents the exposure reference time when the baggage is located in the imaging center region, thus providing a unified reference for the X-ray image acquisition time value. Next, calculate the first temporal deviation between the time the tray enters the recognition area and the time the luggage enters the X-ray channel. This is calculated by subtracting the time the tray enters the recognition area from the time the luggage enters the X-ray channel. The first temporal deviation is used to characterize the tightness of connection between the tray's starting point and the X-ray machine's starting point. Then, calculate the second temporal deviation between the X-ray image acquisition time and the median time of the X-ray machine passage. This is calculated by subtracting the median time of the X-ray machine passage from the X-ray image acquisition time. The second temporal deviation is used to characterize whether the X-ray image acquisition time is accurate. Located near the center of a reasonable imaging range; calculate the third temporal deviation between the video frame acquisition time value and the X-ray image acquisition time value. This is calculated by subtracting the X-ray image acquisition time value from the video frame acquisition time value to obtain the time deviation of the video acquisition action relative to the X-ray image acquisition action. The third temporal deviation value characterizes the degree of coverage synchronization between the video acquisition stage and the X-ray image acquisition stage. Calculate the fourth temporal deviation between the image interpretation start time value and the X-ray image acquisition time value. This is calculated by subtracting the X-ray image acquisition time value from the image interpretation start time value to obtain the response deviation of the manual image interpretation action relative to the X-ray image acquisition action. The fourth temporal deviation value characterizes whether the image interpretation intervention timing is appropriate. The timing of security image generation is kept stable; the fifth time-series deviation value between the time value of the processing result generation and the time value of the image judgment end is calculated. The calculation is performed by subtracting the time value of the image judgment end from the time value of the processing result generation to obtain the delay deviation of the processing action relative to the image judgment end action. The fifth time-series deviation value is used to characterize the degree of result connection within the image judgment processing stage; the sixth time-series deviation value between the alarm cancellation time value and the time value of the processing result generation is calculated. The calculation is performed by subtracting the time value of the processing result generation from the time value of the alarm cancellation to obtain the closed-loop deviation of the alarm cancellation action relative to the processing result formation action. The sixth time-series deviation value is used to characterize the degree of closed-loop completion of the abnormal processing end stage;After taking the absolute values ​​of the first to sixth time-series deviations, normalization is performed according to the reference duration of the corresponding stage, resulting in the following normalized deviation values: first, second, third, fourth, fifth, and sixth. The first time-series deviation corresponds to the tray transfer reference duration, the second to the X-ray machine reference duration, the third to the video acquisition reference duration, the fourth to the image judgment response reference duration, the fifth to the processing generation reference duration, and the sixth to the alarm deactivation reference duration. These normalized deviation values ​​are then sequentially accumulated to obtain the link timing breakage amount. The smaller the value, the more continuous the connection between the tray transfer stage, X-ray machine stage, video acquisition stage, image processing stage, and alarm cancellation stage. The larger the value of the link time sequence break, the more obvious the time misalignment, recording lag, and link jump between these stages. The technical principle of this processing is to uniformly map the key stage time values ​​in the security inspection quality control evidence chain data to the same time series, then use the cumulative result of the time deviation between adjacent stages to quantify the degree of continuity damage to the entire evidence chain, and eliminate the dimensional differences caused by different passenger inspection channel numbers, different equipment response times, and different processing times through normalization processing. This makes the link time sequence break comparable across different security inspection quality control evidence chain data, thereby converging the originally scattered time sequence anomalies into a single link time sequence break, providing a stable time sequence anomaly characterization basis for subsequent review and calculation of suspicious point scores.

[0040] In this implementation plan, by uniformly compressing the time values ​​of tray entering the recognition area, luggage entering the X-ray channel, X-ray image acquisition, video frame acquisition, image judgment start, image judgment end, processing result generation, and alarm deactivation into a single link time sequence break value, the originally scattered stage misalignment phenomenon in the security inspection quality control evidence chain data can be transformed into a single comparable time sequence continuity representation result. This enables the time sequence anomaly degree between different security inspection quality control evidence chain data to have a unified judgment scale, significantly improves the ability to identify the integrity of the evidence chain, enhances the sensitivity of the review doubt score value to the risk of real link distortion, and provides a stable and reliable time sequence quantitative basis for the generation of subsequent review judgment data.

[0041] Specifically, the steps for generating flight-level regulatory index records based on review judgment data, and comparing review suspicion scores with regulatory handling thresholds to form abnormal regulatory records or pass regulatory records are as follows: Read the review judgment data, first performing initial grouping by flight number, then performing secondary grouping by flight date, ensuring all review judgment data corresponding to the same flight number and date enters the same merging interval. The review judgment data must include at least security inspection quality control evidence chain data, review suspicion scores, and suspected area location results; merge the corresponding security inspection quality control evidence chain data by flight number and flight date, using the flight number plus flight date as the index key, and merging passenger anchoring records and other records under the same index key. The data from the security checkpoint, X-ray recordings, video clips, and the time values ​​for the processing results are written sequentially. Then, the review and suspicion score is synchronously linked to the corresponding link unit to generate a flight-level monitoring index record. During generation, all security check and quality control evidence chain data corresponding to the same flight number are compressed into the same flight monitoring view, ensuring that subsequent anomaly screening, closed-loop determination, and retrieval registration are all based on continuous flight-level monitoring. The corresponding security check and quality control evidence chain data and the review and suspicion score are written into the flight-level monitoring index record. During writing, the passenger security checkpoint number plus the passenger security checkpoint gate passage time value is used as the link sub-index, and the flight number plus the flight date is used as the flight main index, thus ensuring that a single evidence chain can be recorded in the flight-level index record. The system quickly locates suspicious points; it compares the review suspicion score with the regulatory handling threshold. During comparison, it first reads the review suspicion score corresponding to each security inspection quality control evidence chain data in the current flight-level regulatory index record, and then compares the review suspicion score with the regulatory handling threshold line by line. The regulatory handling threshold ranges from 1.0 to 4.5. When the review suspicion score is not greater than the regulatory handling threshold, a pass regulatory record is generated. During generation, a pass flag field is written under the current link sub-index, maintaining the current security inspection quality control evidence chain data in a normal regulatory state. When the review suspicion score is greater than the regulatory handling threshold, an abnormal regulatory record is generated. During generation, an abnormal flag field is written under the current link sub-index, and the suspected area is then located. The results, review score values, and corresponding security inspection quality control evidence chain data are all linked to the same anomaly monitoring unit for direct access in subsequent closed-loop judgment and retrieval registration. The technical principle of this process is that by first compressing the scattered review judgment data into a flight-level monitoring index record constrained by the flight number and flight date, and then using the comparison results between the review score value and the monitoring and handling threshold to perform unified monitoring and diversion on the security inspection quality control evidence chain data within the same flight, the anomaly judgment result of a single evidence chain can be elevated to a flight-level continuous monitoring result. This expands the monitoring perspective from single record verification to full-chain monitoring of the same flight, improving the centralization of anomaly screening, the targeting of monitoring and handling, and the continuity of subsequent closed-loop tracking.

[0042] In this implementation plan, by elevating the review judgment data to the flight-level regulatory index record by flight number and flight date, the review doubt score, suspected area location result, and corresponding security inspection quality control evidence chain data, which were originally scattered in the data of a single security inspection quality control evidence chain, can be uniformly incorporated into the same regulatory perspective for continuous management. This enables the distribution of abnormal risks within the same flight range to have a centralized presentation capability, significantly improves the efficiency of abnormal regulatory record identification, enhances the clarity of the boundaries between regulatory records and abnormal regulatory records, and provides a stable and orderly index foundation for subsequent closed-loop judgment and retrieval registration.

[0043] Specifically, the steps for performing closed-loop judgment and retrieval registration on abnormal monitoring records and outputting quality control monitoring results data are as follows: Interception marker values ​​and suspected area location results are written to the abnormal monitoring records. During writing, the corresponding flight-level monitoring index record is first located using the flight number and flight date, then the corresponding link sub-index is located using the passenger inspection channel number and passenger inspection gate passage time value. The interception marker value is written to the status field, and the suspected area location result is written to the location field, ensuring a one-to-one correspondence between the abnormal monitoring records and the corresponding security inspection quality control evidence chain data. Then, the time value and alarm cancellation time value are generated according to the handling result to determine whether the handling closure loop has been completed. During the determination, first... The system reads the time values ​​for the processing result generation and alarm cancellation corresponding to the abnormal monitoring record, then calculates the time interval between the alarm cancellation time value and the processing result generation time value. Abnormal monitoring records with a time interval not exceeding the closed-loop time limit threshold are considered to have completed the processing loop, while those with a time interval exceeding the closed-loop time limit threshold are considered to have not completed the processing loop. The closed-loop time limit threshold ranges from 30 seconds to 1800 seconds. When the alarm cancellation time value is missing or the time interval between the alarm cancellation time value and the processing result generation time value exceeds the closed-loop time limit threshold, a record to be retrieved is generated. This record is generated using the flight number and flight information of the current abnormal monitoring record. The date, passenger security checkpoint number, passenger gate passage time, re-inspection suspicious point score, and suspected area location result are used as the basic fields for the records to be recovered. These records are then linked to the corresponding flight-level monitoring index records. During linking, the flight number plus flight date is used as the primary flight index, and the passenger security checkpoint number plus passenger gate passage time is used as the sub-index, ensuring a continuous tracking relationship between the records to be recovered and the original abnormal monitoring records within the flight-level monitoring index records. A passing flag value is written to the passing monitoring record, and this value is also written to the status field, marking the current security check quality control evidence chain data as a normal release status. This ensures a clear distinction between regulatory records and abnormal regulatory records in terms of status. For records awaiting recovery, recovery registration results are continuously written. During writing, the recovery registration time, recovery processing status, and recovery completion marker are sequentially written to the corresponding fields of the record awaiting recovery. Records awaiting recovery that have not yet been completed are marked as continuously awaiting recovery, while records awaiting recovery that have been completed are marked as recovery completed. Abnormal regulatory records, passing regulatory records, and recovery registration results are written to the quality control and supervision database. During writing, archiving is first performed by flight number and flight date, then link repositioning is performed by passenger inspection channel number and passenger inspection gate passage time value, outputting quality control and supervision result data.The technical principle behind this process is to unify the interception status, suspected area location results, disposal completion status, and recovery registration results of abnormal monitoring records under the same flight-level monitoring index record. This ensures that the abnormal chain triggered by the review suspicion score no longer stops at a single judgment result, but continues to extend towards the disposal completion judgment, the confirmation of the pending recovery status, and the archiving of the recovery completion result. This transforms the review judgment data into quality control monitoring result data that is continuously traceable, status-updable, and result-writeable.

[0044] In this implementation plan, by unifying abnormal supervision records, supervision records, pending recovery records, and recovery registration results under the flight-level supervision index record, the abnormal chain triggered by the review of suspicious point scores can be extended from a one-time judgment result to a continuous and updatable supervision process. This ensures that the completion status of abnormal handling, pending recovery status, and recovery completion status remain consistent under the same index relationship, significantly improving the completeness, status continuity, and traceability of quality control supervision results data, and enhancing the closed-loop reliability of the transformation of abnormal supervision records into actual handling results.

[0045] like Figure 2As shown, the second aspect of this invention provides a security inspection quality control and supervision system based on multi-source heterogeneous data fusion, including: a data acquisition and processing module, a spatiotemporal evidence chain building module, a cross-modal fusion review module, and a closed-loop supervision and disposal module. The data acquisition and processing module is used to acquire security inspection quality control monitoring data and perform time synchronization, missing data completion, anomaly removal, noise suppression, and scale unification processing on the security inspection quality control monitoring data, outputting pre-processed security inspection quality control monitoring data. The spatiotemporal evidence chain building module is used to generate passenger anchor records based on the pre-processed security inspection quality control monitoring data, establish a time sliding window around the passenger anchor records, perform segmentation and cyclic reuse disassembly of the tray records, and perform temporal constraint association and backtracking matching on X-ray and video records to generate security inspection quality control evidence chain data. The cross-modal fusion review module is used to extract X-ray image regions and video keyframes based on the security inspection quality control evidence chain data. The system constructs and calls a cross-modal evidence deviation representation model to obtain evidence deviation representation values ​​based on sequence and temporal feature sequences. These values ​​characterize the degree of deviation between image features and temporal features within the same security inspection quality control evidence chain. It then calculates a review suspicion score by combining the suspected area offset and the link temporal break, generating review judgment data. The suspected area offset characterizes the degree of offset between the center position of the suspected area and the center position of the security inspection image area. The link temporal break characterizes the degree of temporal connection anomalies between the stages of tray transfer, security screening, video acquisition, image processing, and alarm deactivation. A closed-loop supervision and handling module generates flight-level supervision index records based on the review judgment data. It compares the review suspicion score with the supervision and handling threshold to form abnormal supervision records or, through these records, performs closed-loop judgment and retrieval registration on the abnormal supervision records, outputting quality control supervision result data.

[0046] In this implementation plan, by sequentially connecting security inspection quality control monitoring data, security inspection quality control evidence chain data, review judgment data, and quality control supervision result data, this invention can unify passenger inspection information, tray flow information, security inspection image information, video acquisition information, image judgment and processing information, and flight supervision information into the same technical link. This transforms multi-source heterogeneous data from a decentralized collection state into a form of supervision data that can be continuously linked, uniformly judged, and tracked in a closed loop. This significantly improves the ability of review suspicion score values ​​to characterize real abnormal links, enhances the accuracy of abnormal supervision record generation, the orderly organization of flight-level supervision index records, and the completeness of quality control supervision result data.

[0047] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0048] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A security inspection quality control and supervision method based on multi-source heterogeneous data fusion, characterized in that, Includes the following steps: S1 collects security inspection quality control monitoring data, and performs time synchronization, missing data completion, anomaly removal, noise suppression and scale unification processing on the security inspection quality control monitoring data, and outputs the preprocessed security inspection quality control monitoring data; S2, based on the preprocessed security inspection quality control monitoring data, generates passenger anchor records, establishes a time sliding window around the passenger anchor records, performs segmentation and cyclic reuse disassembly on the tray records, performs temporal constraint association and backtracking matching on the X-ray records and video records, and generates security inspection quality control evidence chain data; S3 extracts X-ray image regions, video keyframe sequences, and temporal feature sequences from the security inspection quality control evidence chain data. It then constructs and calls a cross-modal evidence deviation characterization model to obtain evidence deviation characterization values. These values ​​characterize the degree of deviation between image features and temporal features within the same security inspection quality control evidence chain data. Combining the suspected area offset and the link temporal breakage, a review doubt score is calculated, generating review judgment data. Specifically, the suspected area offset characterizes the degree of offset of the center position of the suspected area relative to the center position of the security inspection image region; the link temporal breakage characterizes the degree of temporal connection anomalies between the stages of tray transfer, security inspection, video acquisition, image processing, and alarm deactivation. The specific steps for calculating the review suspicion score by combining the suspected area offset and the link timing breakage to generate review judgment data are as follows: High-response region extraction is performed on the X-ray image region to obtain the location result of the suspect region. The pixel coordinates of the center position of the suspect region and the center position of the X-ray image region are extracted respectively. The pixel distance between the two is divided by the diagonal pixel length of the X-ray image region to obtain the offset of the suspect region. Read the time values ​​of each stage in the same security inspection quality control evidence chain data, including tray transfer, X-ray machine passage, video acquisition, image processing and alarm cancellation, and calculate the link timing break based on the timing deviation between the time values ​​of each stage. Add one to the suspected region offset and take the natural logarithm to obtain the region offset modulation term; take the negative of the link timing break as the exponent with respect to the natural constant e and add one to obtain the timing suppression term; multiply the evidence deviation characterization value with the region offset modulation term and divide by the timing suppression term to obtain the review suspicion score value. Write the review of suspicious points and the location results of suspected areas into the corresponding security inspection quality control evidence chain data, and output the review judgment data. S4 generates flight-level regulatory index records based on review judgment data, compares the review doubt score with the regulatory handling threshold to form abnormal regulatory records or through the regulatory records, determines whether the handling corresponding to the abnormal regulatory records has completed the closed loop based on the alarm cancellation time value, and performs recovery registration according to the judgment result, and outputs quality control regulatory result data.

2. The security inspection quality control and supervision method based on multi-source heterogeneous data fusion according to claim 1, characterized in that: The specific steps for collecting security inspection quality control monitoring data, performing time synchronization, missing data completion, anomaly removal, noise suppression, and scale standardization on the security inspection quality control monitoring data, and outputting preprocessed security inspection quality control monitoring data are as follows: Real-time collection of security inspection quality control monitoring data includes passenger inspection channel number, passenger gate passage time, boarding pass scanning time, flight number, flight date, check-in sequence number, passenger document verification time, passenger image capture time, tray number, tray number recognition time, tray entering recognition area time, tray leaving recognition area time, X-ray image capture time, baggage entering X-ray channel time, baggage leaving X-ray channel time, video frame capture time, video frame sequence number, image analysis start time, image analysis end time, processing result generation time, and alarm deactivation time. For the collected security inspection quality control monitoring data, the following steps are performed: a precise time protocol time synchronization algorithm is used to perform multi-source time reference synchronization processing; a forward nearest neighbor completion algorithm is used to perform missing data completion processing; a Hample filter anomaly detection algorithm is used to perform anomaly data identification and removal processing; a first-order exponential weighted moving average algorithm is used to perform time series noise suppression processing; and a minimum-maximum normalization algorithm is used to perform scale unification processing, outputting the preprocessed security inspection quality control monitoring data.

3. The security inspection quality control and supervision method based on multi-source heterogeneous data fusion according to claim 2, characterized in that: The specific steps for generating passenger anchoring records based on preprocessed security inspection quality control monitoring data are as follows: Read the pre-processed security check quality control monitoring data, and merge the passenger gate passage time, boarding pass scanning time, passenger document verification time, and passenger image capture time into the same channel according to the passenger gate number. Using the passenger gate passage time as a benchmark, select the boarding pass scanning time, passenger document verification time, and passenger image capture time with the smallest absolute value of the time difference from the passenger gate passage time within the predetermined time tolerance range, and match them as the same passenger passage record. Then, form a passenger passage sequence according to the order of the passenger gate passage time, and write the flight number, flight date, and check-in sequence number into the corresponding record of the passenger passage sequence to generate a passenger anchoring record.

4. The security inspection quality control and supervision method based on multi-source heterogeneous data fusion according to claim 3, characterized in that: The specific steps for establishing a time sliding window around passenger anchor records, performing segmentation and cyclic reuse decomposition of tray records, and performing temporal constraint association and backtracking matching on X-ray and video records to generate security inspection quality control evidence chain data are as follows: A time-sliding window of length L is established around the passenger anchor record. The time values ​​of tray number recognition, tray entry into the recognition area, tray exit from the recognition area, X-ray image acquisition, baggage entry into the X-ray channel, baggage exit from the X-ray channel, video frame acquisition, and video frame sequence number are mapped to the corresponding time-sliding window. For tray records falling within the same time-sliding window, the occupied segments are divided according to the time values ​​of tray entry into the recognition area and tray exit from the recognition area. When the same tray number corresponds to multiple occupied segments, it is processed as multiple cyclically reused tray records. When the tray number is missing, duplicated, or inconsistent, all tray records within the time-sliding window are retained for subsequent backtracking matching. Temporal constraint association is performed on the tray record, X-ray record, and video record within the time sliding window. A candidate association sequence is constructed according to the order of the time value of the tray entering the recognition area, the time value of the tray number recognition, the time value of the X-ray image acquisition, and the time value of the luggage leaving the X-ray channel. Then, the video record is spliced ​​according to the rule of continuous increment of the video frame number value to obtain a video segment record covering the time value of the X-ray image acquisition. The tray record, X-ray record, and video segment record are then linked to the corresponding passenger anchor record to generate a candidate evidence chain record. The system performs time-sliding window backtracking matching on candidate evidence chain records. For records that fail to be associated in real time, have conflicts with tray reuse, or have abnormal tray numbers, it backtracks forward and backward along the passenger inspection sequence to re-retrieve tray records, X-ray records, and video clip records within the time-sliding window. Records with temporal breaks are filtered out, and association paths with continuous time connections and non-overlapping occupied sections are retained. The system writes the image judgment start time value, image judgment end time value, processing result generation time value, and alarm cancellation time value into the association path and outputs security inspection quality control evidence chain data.

5. The security inspection quality control and supervision method based on multi-source heterogeneous data fusion according to claim 4, characterized in that: The specific steps for extracting X-ray image regions, video keyframe sequences, and temporal feature sequences based on security inspection quality control evidence chain data, and constructing and calling a cross-modal evidence deviation characterization model to obtain evidence deviation characterization values ​​are as follows: Historical review records are extracted from the quality control review result database. X-ray records, video clips, and time values ​​for tray entry into the recognition area, luggage entry into the X-ray channel, X-ray image acquisition, luggage departure from the X-ray channel, video frame acquisition, image judgment start, image judgment end, processing result generation, and alarm deactivation are read according to the security inspection quality control evidence chain data corresponding to the historical review records. The corresponding X-ray image area is extracted based on the X-ray image acquisition time value, and the video keyframe sequence covering the X-ray image acquisition time value is extracted based on the video frame acquisition time value. The time values ​​from tray entry into the recognition area to alarm deactivation are then arranged in chronological order to form a temporal feature sequence. The review conclusion is used as a sample label and written into the corresponding sample to generate review training samples. A cross-modal evidence deviation representation model is constructed using a cross-attention-constrained cross-modal contrastive learning network. X-ray image regions from the review training samples are input into the image coding layer, and video keyframe sequences and temporal feature sequences are input into the temporal coding layer. Image feature vectors and temporal feature vectors output from the same review training sample are brought closer together, while those from different review training samples are stretched apart. These image feature vectors and temporal feature vectors are then input into the deviation representation layer, which outputs the evidence deviation representation value. The cross-modal evidence deviation representation model is iteratively trained using review training samples. The model is considered complete when the prediction error of the review conclusion is below the error threshold for N consecutive training cycles. X-ray image regions, video keyframe sequences, and temporal feature sequences are obtained from real-time security inspection quality control evidence chain data and input into a trained cross-modal evidence deviation representation model, which outputs evidence deviation representation values.

6. The security inspection quality control and supervision method based on multi-source heterogeneous data fusion according to claim 1, characterized in that: The specific steps for calculating the link timing breakage based on the timing deviation between time values ​​at each stage are as follows: The average of the times when luggage enters and leaves the X-ray channel is determined as the median time of passage. Then, the following time-series deviations are calculated: the first deviation between the time when the tray enters the recognition area and the time when luggage enters the X-ray channel; the second deviation between the time when the X-ray image is acquired and the median time of passage; the third deviation between the time when the video frame is acquired and the time when the X-ray image is acquired; the fourth deviation between the time when the image judgment begins and the time when the X-ray image is acquired; the fifth deviation between the time when the processing result is generated and the time when the image judgment ends; and the sixth deviation between the time when the alarm is deactivated and the time when the processing result is generated. The absolute values ​​of the first to sixth time-series deviations are taken and summed to obtain the link timing breakage.

7. The security inspection quality control and supervision method based on multi-source heterogeneous data fusion according to claim 1, characterized in that: The specific steps for generating flight-level regulatory index records based on review judgment data, and comparing the review suspicion score with the regulatory handling threshold to form abnormal regulatory records or through regulatory records are as follows: Read the review judgment data, merge the corresponding security inspection quality control evidence chain data according to flight number and flight date, generate flight-level supervision index record, and write the corresponding security inspection quality control evidence chain data and review doubt score value into the flight-level supervision index record; The review score for suspicious points is compared with the regulatory handling threshold. If the review score is not greater than the regulatory handling threshold, a regulatory record of approval is generated; if the review score is greater than the regulatory handling threshold, an abnormal regulatory record is generated.

8. The security inspection quality control and supervision method based on multi-source heterogeneous data fusion according to claim 7, characterized in that: The specific steps for determining whether the handling of the abnormal monitoring record has completed the closed loop based on the alarm cancellation time value, and for executing the retrieval registration and outputting the quality control monitoring result data according to the determination result are as follows: The interception mark value and the location result of the suspected area are written into the abnormal supervision record. Then, the time value generated by the handling result and the alarm cancellation time value are used to determine whether the handling loop has been completed. When the alarm cancellation time value is missing or the time interval between the alarm cancellation time value and the time value generated by the handling result is greater than the closed loop time limit threshold, a record to be recovered is generated and the record to be recovered is linked to the corresponding flight-level supervision index record. Write a pass flag value to the record that passes the supervision; continue to write the recovery registration result to the record that is to be recovered, and write the abnormal supervision record, the pass supervision record and the recovery registration result to the quality control supervision database, and output the quality control supervision result data.

9. A security inspection quality control and supervision system based on multi-source heterogeneous data fusion, employing the security inspection quality control and supervision method based on multi-source heterogeneous data fusion as described in any one of claims 1-8, characterized in that, include: The module includes a data acquisition and processing module, a spatiotemporal evidence chain building module, a cross-modal fusion review module, and a closed-loop supervision and disposal module, among which: The data acquisition and processing module is used to collect security inspection quality control monitoring data, and to perform time synchronization, missing data completion, anomaly removal, noise suppression and scale unification processing on the security inspection quality control monitoring data, and output the preprocessed security inspection quality control monitoring data. The spatiotemporal evidence chain building module is used to generate passenger anchor records based on preprocessed security inspection quality control monitoring data, establish a time sliding window around the passenger anchor records, perform segmentation and cyclic reuse disassembly of the tray records, perform temporal constraint association and backtracking matching on X-ray records and video records, and generate security inspection quality control evidence chain data. The cross-modal fusion review module is used to extract X-ray image regions, video keyframe sequences, and temporal feature sequences based on security inspection quality control evidence chain data. It constructs and calls a cross-modal evidence deviation characterization model to obtain evidence deviation characterization values. The evidence deviation characterization values ​​are used to characterize the degree of deviation between image features and temporal features in the same security inspection quality control evidence chain data. It calculates review suspicion score values ​​by combining the suspected area offset and the link temporal break amount, and generates review judgment data. Among them, the suspected area offset is used to characterize the degree of offset of the center position of the suspected area relative to the center position of the security inspection image area; the link temporal break amount is used to characterize the degree of abnormality in the temporal connection between each stage of tray transfer, security inspection, video acquisition, image processing, and alarm cancellation. The closed-loop supervision and handling module is used to generate flight-level supervision index records based on review judgment data, compare the review doubt score with the supervision and handling threshold to form abnormal supervision records or, based on the alarm cancellation time value, determine whether the handling corresponding to the abnormal supervision record has completed the closed loop, and perform retrieval registration according to the judgment result, and output quality control supervision result data.

Images