Search

CN-122020201-A - LDAR profiling quality assessment method based on image sharpening and behavior feature recognition

CN122020201ACN 122020201 ACN122020201 ACN 122020201ACN-122020201-A

Abstract

The invention discloses a LDAR profiling quality assessment method based on image sharpening and behavior feature recognition, which comprises the steps of collecting profiling multisource time sequence data to generate standardized profiling multisource time sequence data, carrying out recognition and behavior phase division on operation behaviors in a profiling process to obtain corresponding behavior phase sequences, constructing an improved Stentiford visual model to form a visual attention state sequence, executing structure-preserving type and suppression type sharpening processing to construct a combined time sequence, executing PrefixSpan algorithm to form a standardized profiling combined mode set, and carrying out matching analysis based on the combined time sequence to be assessed to generate a profiling quality assessment result. According to the invention, objective and quantitative quality assessment of documented image acquisition time and operation behavior normalization in the LDAR documenting process is realized by introducing an improved Stentiford visual model and PrefixSpan algorithm.

Inventors

  • CHEN NING

Assignees

  • 山东安诺环保技术服务有限公司

Dates

Publication Date
20260512
Application Date
20260203

Claims (9)

  1. 1. LDAR profiling quality assessment method based on image sharpening and behavior feature recognition is characterized by comprising the following steps: acquiring LDAR profiling multisource time sequence data formed in the profiling process, preprocessing the profiling multisource time sequence data, and generating standardized profiling multisource time sequence data; Identifying operation behaviors in LDAR profiling processes based on standardized profiling multisource time sequence data, generating a behavior event sequence, and performing behavior phase division on the profiling processes according to the behavior event to obtain a corresponding behavior phase sequence; Based on the standardized profiling multisource time sequence data, an improved Stentiford visual model is constructed, visual attention responses of all image areas are calculated, corresponding attention response graphs are constructed, visual attention areas are extracted, and a visual attention state sequence is formed; Based on the vision attention state sequence, dividing the profiling image data into an attention area and a non-attention area, executing structure-keeping type and suppression type sharpening processing, generating a corresponding sharpening profiling image sequence and a corresponding sharpening benefit mark, and carrying out alignment and joint coding with the behavior phase sequence to construct a joint time sequence; Mining a combined sub-sequence mode which simultaneously meets the time sequence consistency between the behavior phase change and the visual attention state change by adopting a PrefixSpan algorithm based on a combined time sequence corresponding to the high-quality LDAR profiling sample to form a standard LDAR profiling combined mode set; And carrying out matching analysis on the joint sequence corresponding to the to-be-evaluated LDAR profiling process and the standard LDAR profiling joint mode set to generate a profiling quality evaluation result.
  2. 2. The method of LDAR documenting quality assessment based on image sharpening and behavioral characteristic identification of claim 1, wherein said documenting multisource temporal data specifically comprises documented image data, work process video data and behavioral related temporal data.
  3. 3. The method for LDAR profiling quality assessment based on image sharpening and behavior feature recognition according to claim 1, wherein the preprocessing of the profiling multisource time series data specifically comprises time base unification and time series alignment, format standardization and data structure regularity, invalid data rejection and abnormal fragment filtering.
  4. 4. The method for LDAR profiling quality assessment based on image sharpening and behavioral characteristic identification of claim 1, wherein said deriving a corresponding behavioral phase sequence comprises: Based on standardized profiling multisource time sequence data, performing discrete processing on the video data in the operation process according to a unified time index, generating a video frame sequence arranged in time sequence, and performing synchronous discrete on behavior-related time sequence data, wherein each video frame corresponds to a behavior state record; extracting operation behavior visual features representing operation action types, action amplitudes and action change trends aiming at video frames corresponding to each time index, and combining the operation behavior visual features with corresponding behavior state records to form comprehensive behavior features; Based on comprehensive behavior characteristics, judging operation behaviors under each time index, and generating corresponding behavior event marks on a time axis; Determining the starting position and the ending position of a behavior event according to the change condition of the behavior event marks under adjacent time indexes, merging time fragments with consistent behavior event marks in a continuous time range, and constructing a behavior event sequence; And mapping the behavior events in the behavior event sequence into corresponding behavior stage marks, and merging the behavior events which are adjacent in time and consistent in the behavior stage marks to form a behavior stage sequence.
  5. 5. The method of LDAR profiling quality assessment based on image sharpening and behavioral characteristic identification of claim 1, wherein the forming a sequence of visual attention states comprises: reading the documented image data and the operation process video data from the standardized documented multi-source time sequence data, and arranging according to a unified time index sequence to construct a continuous documented image sequence; Based on behavior related time sequence data, an improved Stentiford visual model is constructed, the triggering frequency of operation behaviors under each time index is counted, a behavior triggering weight is generated for the corresponding time index, 3 shielding stages which are sequentially executed are set for the profile building image corresponding to the same time index according to the behavior triggering weight, and corresponding shielding intensity and shielding range are respectively configured for each shielding stage, so that multi-stage random shielding parameter configuration is formed; Aiming at the profiling image corresponding to each time index, carrying out image component semantic segmentation processing on the profiling image, dividing the profiling image into a plurality of image component areas with definite semantic attributes, marking the semantic areas of key components as structure retaining areas, and generating shielding inhibition constraints corresponding to the structure retaining areas; Based on multi-stage random shielding parameter configuration and combining shielding inhibition constraint, sequentially executing random shielding treatment of a plurality of shielding stages on the documented images corresponding to the same time index, and generating shielding images subjected to semantic segmentation constraint; Constructing two image representation forms with different view angles on the basis of the shielding image, respectively executing visual response calculation on the shielding image with each view angle to obtain corresponding shielding response value distribution, executing differential fusion processing on the shielding response value distribution with different view angles to form shielding response value distribution with enhanced view angle consistency, and constructing a shielding response evolution map; Based on the occlusion response evolution map, comprehensively considering the image content complexity distribution condition of the documented image, applying inhibition constraint to the occlusion response value distribution corresponding to the high-complexity background area, generating final visual attention response distribution, smoothly updating the visual attention response distribution under the current time index to obtain an attention response map with enhanced time consistency, extracting the visual attention area from the attention response map, and combining to form a visual attention state sequence.
  6. 6. The method for LDAR profiling quality assessment based on image sharpening and behavioral characteristic identification of claim 5, wherein said constructing an occlusion response evolution graph comprises: Taking the time index and the shielding stage as two-dimensional indexes, reading the view angle consistency enhanced shielding response numerical distribution corresponding to each time index and each shielding stage, and establishing a two-dimensional position mapping relation between each time index and the shielding response numerical distribution corresponding to each shielding stage to form a shielding response organization table; Performing phase difference calculation on the shielding response value distribution corresponding to the adjacent shielding phases under the same time index in the shielding response organization table to obtain response variation amplitude distribution of each image area between the adjacent shielding phases, and performing time difference calculation on the shielding response value distribution corresponding to the adjacent time index under the same shielding phase to obtain response variation amplitude distribution of each image area between the adjacent time indexes; Based on the response variation amplitude distribution and the response variation amplitude distribution, consistency screening treatment is carried out on the response variation tracks of each image area, the response variation tracks meeting the stage mutation conditions and the time mutation conditions are screened out, and the response variation tracks which are not screened out are combined with the two-dimensional sequence of the shielding stage according to the time index, so that a shielding response evolution map is constructed.
  7. 7. The method for LDAR profiling quality assessment based on image sharpening and behavioral characteristic identification of claim 1, wherein the constructing a joint timing sequence comprises: reading the filing image data arranged according to the time indexes from the standardized filing multisource time sequence data, and reading a visual attention state sequence corresponding to each time index; For the profiling image corresponding to each time index, dividing the profiling image into a concerned region and a non-concerned region according to the visual concerned region set, and respectively generating corresponding region mask identifications for the concerned region and the non-concerned region; Extracting structure retention reference characteristics aiming at the region of interest under the constraint of the region mask mark, and executing structure retention type sharpening processing on the region of interest according to the structure retention reference characteristics to generate a sharpened image fragment of the region of interest; Aiming at a non-attention area, executing inhibited sharpening processing under the constraint of an area mask mark, synthesizing an attention area sharpened image segment and an image segment processed by the non-attention area to generate a sharpening file-establishing image under a corresponding time index, and generating a sharpening benefit mark corresponding to the time index according to the structure maintenance reference characteristic variable quantity of the attention area before and after sharpening; and arranging the sharpening file-building images corresponding to the time indexes and sharpening yield identifications, and performing time index alignment and joint coding with the behavior phase sequence to obtain a joint time sequence.
  8. 8. The method for LDAR profiling quality assessment based on image sharpening and behavioral characteristic identification of claim 1, wherein the forming a canonical LDAR profiling joint pattern set comprises: Reading a combined time sequence corresponding to the high-quality LDAR profiling sample, executing PrefixSpan algorithm, and representing each time index entry in the combined time sequence as a combined event entry containing a behavior stage identifier, a visual attention state identifier and a sharpening benefit identifier to form a combined event set; Modeling a joint time sequence as a joint event diagram based on a joint event set, taking each joint event item in the joint event diagram as a prefix starting node, performing prefix projection processing on the joint time sequence to generate a corresponding prefix projection sequence set, and expanding the prefix joint event items to form a candidate joint subsequence mode; Under the constraint of the joint event diagram, calculating the occurrence frequency support degree of the candidate joint subsequence mode in the joint event set, the stability measurement value of the visual attention state identifier and the consistency measurement value of the sharpening benefit identifier respectively, and carrying out collaborative fusion to generate a composite support degree; Synchronously mining inter-phase transfer modes crossing behavior phases in the combined event graph, and constructing a support threshold curve which changes along with behavior phase propulsion according to the relative positions of the combined events in a behavior phase sequence; And carrying out segmentation judgment on the composite support degree of the candidate combined sub-sequence modes based on a support degree threshold curve, and summarizing the combined sub-sequence modes which simultaneously meet the composite support degree condition, the behavior phase alignment constraint, the combined structure similarity constraint, the semantic similarity constraint and the combined event map path constraint to form a standard LDAR documenting combined mode set.
  9. 9. The method of LDAR documenting quality assessment based on image sharpening and behavioral characteristic identification of claim 1, wherein said generating a documenting quality assessment result comprises: Reading a joint time sequence corresponding to a LDAR profiling process to be evaluated, representing each time index entry in the joint time sequence as a joint event entry, and arranging according to the time index order to form a joint event sequence to be evaluated; Constructing a to-be-evaluated joint event diagram based on a to-be-evaluated joint event sequence, taking joint event items as diagram nodes, taking time sequence association relations of the joint event items under adjacent time indexes and phase transfer relations among different behavior phases as diagram edges, and recording corresponding behavior phase position information and sharpening benefit identifications in the diagram nodes; Under the constraint of the to-be-evaluated combined event diagram, according to the behavior phase sequence constraint of each combined subsequence mode in the standardized LDAR profiling combined mode set, candidate matching fragment positioning is executed on the to-be-evaluated combined event sequence, and candidate matching fragments with consistent behavior phase sequence are screened to form candidate matching relations; Aiming at each candidate matching relation, based on the to-be-evaluated combined event diagram, respectively calculating the behavior phase consistency degree, the visual attention state consistency degree, the sharpening income consistency degree and the inter-phase transfer consistency degree, and comprehensively weighting to obtain a pattern matching score; comparing the pattern matching score with a judging threshold value of the support degree threshold value curve at the corresponding stage position, determining the matching passing state of each behavior stage, and generating a documenting quality evaluation result based on the matching passing state and the corresponding pattern matching score.

Description

LDAR profiling quality assessment method based on image sharpening and behavior feature recognition Technical Field The invention relates to the technical field of behavior feature recognition, in particular to a LDAR profiling quality assessment method based on image sharpening and behavior feature recognition. Background Along with the continuous expansion of the scale of industrial devices and the continuous improvement of safety supervision requirements, LDAR technology has been widely applied to equipment operation management in petrochemical industry, chemical industry and other industries. The traditional LDAR workflow generally relies on manual carrying detection equipment to carry out inspection on the parts such as a valve, a flange and the like, and collects detection results, shoots a filing image and records related information in the inspection process so as to form complete filing data. However, the filing process is often affected by factors such as operation habits of operators, shooting angles, ambient illumination, operation rhythm differences and the like, so that great differences exist between the quality of the filing image and the normalization of the filing process. The prior art attempts to introduce image processing and behavior analysis methods to assist in the analysis of LDAR shift data. However, the prior art scheme focuses on the local analysis of a single data type, lacks joint modeling of the profiling image, the operation process video and the behavior time sequence information, is difficult to accurately reflect whether the profiling image is acquired in a correct behavior stage, and cannot perform system evaluation on the overall consistency of the profiling process. Therefore, how to provide a LDAR documenting quality assessment method based on image sharpening and behavior feature recognition is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a LDAR profiling quality assessment method based on image sharpening and behavior feature recognition, which comprehensively utilizes a behavior time sequence analysis method and a PrefixSpan algorithm to systematically model and assess a LDAR profiling process. According to the invention, through carrying out joint analysis on the profiling image, the operation process video and the behavior related time sequence data, an improved Stentiford visual model is introduced, the effective information acquisition condition of the profiling image in different operation stages is depicted, a behavior stage sequence is constructed based on the behavior feature recognition result, and the time sequence mode feature of the standardized profiling process is extracted through PrefixSpan algorithm, so that objective matching and quantitative evaluation of the profiling process to be evaluated are realized. The method can effectively judge whether the documented image is acquired in a correct behavior stage, reduces the dependence of manual experience, improves consistency, traceability and objectivity of documented quality assessment, and is suitable for automatic quality control in a large-scale LDAR operation scene. According to the embodiment of the invention, the LDAR profiling quality assessment method based on image sharpening and behavior feature recognition comprises the following steps: acquiring LDAR profiling multisource time sequence data formed in the profiling process, preprocessing the profiling multisource time sequence data, and generating standardized profiling multisource time sequence data; Identifying operation behaviors in LDAR profiling processes based on standardized profiling multisource time sequence data, generating a behavior event sequence, and performing behavior phase division on the profiling processes according to the behavior event to obtain a corresponding behavior phase sequence; Based on the standardized profiling multisource time sequence data, an improved Stentiford visual model is constructed, visual attention responses of all image areas are calculated, corresponding attention response graphs are constructed, visual attention areas are extracted, and a visual attention state sequence is formed; Based on the vision attention state sequence, dividing the profiling image data into an attention area and a non-attention area, executing structure-keeping type and suppression type sharpening processing, generating a corresponding sharpening profiling image sequence and a corresponding sharpening benefit mark, and carrying out alignment and joint coding with the behavior phase sequence to construct a joint time sequence; Mining a combined sub-sequence mode which simultaneously meets the time sequence consistency between the behavior phase change and the visual attention state change by adopting a PrefixSpan algorithm based on a combined time sequence corresponding to the high-quality LDAR profiling sample to form a standard LDAR profiling combi