Multi-source data fusion law enforcement record document generation method

Abstract

The invention discloses a law enforcement record document generation method based on multi-source data fusion, and belongs to the technical field of data processing, and the method comprises the following steps: S1, obtaining multi-source data of an alarm case, classifying texts, images and audios in the multi-source data, and obtaining a structured data set; s2, extracting case related entities from the structured data set, and generating an event description table; and S3, determining a corresponding event element field in the event description table, and generating a law enforcement record document according to a preset document specification template. The multi-source data fusion law enforcement record document generation method solves the problem of low reliability of a law enforcement document in practical application due to low efficiency and inconsistency of an existing law enforcement record document generation mode.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method for generating law enforcement record documents by fusing multi-source data. Background Art

[0002] The generation of law enforcement records is an essential component of public security and the judicial system. Their accuracy and standardization are directly linked to the legitimacy and efficiency of case handling. With the increasing number of police cases and the diversification of data sources, the efficient and accurate generation of compliant law enforcement records has become a key research focus. Traditional methods of document generation rely primarily on manual recording and organization. While flexible, this approach struggles to cope with the demands of processing large amounts of data from multiple sources.

[0003] Existing methods often rely on manual labor, supplemented by simple template-filling tools. However, these methods often suffer from inefficiency and error-proneness when applied to complex police cases. This is especially true when data comes from multiple sources, such as on-site investigation records and interrogation transcripts. Manual integration is not only time-consuming but also difficult to ensure the integrity of the extracted information and the professionalism of the language. This inefficiency and inconsistency limit the reliability of law enforcement documents in practical applications. Summary of the Invention

[0004] In order to overcome the defects of the existing technology, the present invention provides a method for generating law enforcement record documents by fusion of multi-source data to solve the above problems.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a method for generating law enforcement record documents by fusion of multi-source data, comprising the following steps: S1: Obtain multi-source data on police cases, classify the text, images, and audio in the multi-source data, and obtain a structured data set; S2: extracting case-related entities from the structured data set and generating an event description table; S3: Determine the corresponding event element field in the event description table, and generate a law enforcement record document according to a preset document specification template.

[0006] It is worth noting that in step S1, multi-source data with the same case number are input into a pre-established parsing module, and the pre-established parsing module is used to determine the type according to the suffix name of each data in the multi-source data, and output a pre-classification data set containing text labels, image labels and audio labels; then the data with text labels, image labels and audio labels in the pre-classification data set are extracted respectively to obtain text structured data, image structured data and audio structured data to form a structured data set.

[0007] Specifically, in the step S1, data with text tags is extracted from the pre-classification data set, and Jieba word segmentation is used to split the data with text tags into text structured data including case numbers, people, time, locations, and means.

[0008] Specifically, in the step S1, data with image tags is extracted from the pre-classification data set, and HOG feature vectors are extracted using the OpenCV library to obtain image structured data including case numbers and feature vectors.

[0009] Preferably, in the step S1, data with audio tags is extracted from the pre-classification data set, and short-time Fourier transform is performed using the Librosa library and Mel-frequency cepstral coefficients are extracted to obtain audio structured data including case numbers and sound types.

[0010] Preferably, in the step S1, the text structured data, image structured data, and audio structured data are aligned according to case numbers, and the K-means clustering algorithm is used to cluster the text structured data, image structured data, and audio structured data with different case numbers to obtain a unified structured data set including case numbers and clustering labels; According to the clustering labels in the unified structured data set and the condition that the time interval is less than a preset threshold, the association relationships of cases with different case numbers are matched to obtain a case association data set as the structured data set.

[0011] Optionally, in the step S3, an exact string match is performed on the event description table using a preset list of crime keywords to obtain a refined information set; The event element fields in the refined information set are extracted, and a preset document specification template that matches the event element fields is obtained. The person field, time field, location field, and means field in the event description table are filled according to the preset document specification template.

[0012] Optionally, in the step S3, when filling the preset document specification template, if a field is missing, it is marked as NULL; for the field marked as NULL, a filling function is called from the structured data set to supplement the corresponding information to the preset document specification template.

[0013] It should be noted that after the step S3 is executed, there is also a step S4: Record the generation process and data source of the law enforcement record document, and generate a verifiable document generation log.

[0014] Specifically, in the step S4, the logs collected by the log recording function are read from the database; an operation time series is generated according to the timestamps corresponding to the generation processes in the logs; The API call records in the operation time series are matched by using regular expressions; the call paths containing the data source fields are extracted from the API call records; the call paths are input into a JSONPath parsing tool; the institutional codes and table names in the data source fields are separated by the JSONPath parsing tool; a source identifier set is generated according to the institutional codes and table names; According to the institutional codes in the source identifier set, the filling contents of the corresponding fields are obtained from the cache; a hash table is established to store the mapping relationship between the filling contents and the document generation steps; wherein the keys of the hash table are field names, and the field names include person fields, time fields, location fields and means fields, and the values are the operation steps of the generation process; a filling mapping table is output according to the hash table; The operation steps of the generation process in the filling mapping table are aligned with the timestamps in the operation time series; the field contents with the same timestamp are merged; a log information set containing field names, operation steps and timestamps is generated.

[0015] The beneficial effects of the present invention are as follows: in the method for generating law enforcement record documents with multi-source data fusion, by classifying text, image and audio data, a structured data set is obtained; then the entities related to the case are extracted and an event description is generated; finally, according to the preset document specification template, the corresponding event element fields are determined, and law enforcement record documents are generated. Through intelligent means, the integration and consistency processing of multi-source data are realized instead of manual operation, the accuracy of generating law enforcement record documents is improved, and thus the reliability of law enforcement documents in practical applications is improved. Description of the Drawings

[0016] Figure 1 is a flowchart of a method for generating law enforcement record documents with multi-source data fusion in an embodiment of the present invention; Figure 2 is a flowchart of generating a verifiable document generation log in an embodiment of the present invention. Detailed Embodiments

[0017] The following further describes the specific embodiments of the present invention with reference to the drawings. It should be noted here that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation to the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0018] Such as Figure 1 and 2As shown in the figure, a method for generating law enforcement record documents through multi-source data fusion includes the following steps: S1: Obtain multi-source data of police cases, classify the text, images, and audio in the multi-source data, and obtain a structured data set. S2: Extract case-related entities from the structured data set to generate an event description table. For example, a structured data set of a certain case contains "2025-03-30-23:00, a certain street corner, a man robbing Zhang with a knife". Through Named Entity Recognition (NER), entities such as time "2025-03-30-23:00", location "a certain street corner", person "Zhang", and means "robbing with a knife" can be extracted. These entities form a preliminary information set to ensure that the core elements of the case are completely captured. It should be noted that named entity recognition can effectively distinguish complex fields such as person names and place names, improving the accuracy of information extraction. S3: Determine the corresponding event element fields in the event description table, and generate law enforcement record documents according to a preset document specification template.

[0019] In the method for generating law enforcement record documents through multi-source data fusion, a structured data set is obtained by classifying text, image, and audio data; then case-related entities are extracted and an event description is generated; finally, the corresponding event element fields are determined according to a preset document specification template to generate law enforcement record documents. Through intelligent means, the integration and consistency processing of multi-source data are realized instead of manual work, improving the accuracy of generating law enforcement record documents, and thus enhancing the reliability of law enforcement documents in practical applications.

[0020] It should be noted that in step S1, the multi-source data with the same case number is input into a pre-established parsing module. The pre-established parsing module determines the type according to the suffix name of each data in the multi-source data, and outputs a pre-classification data set containing text tags, image tags, and audio tags. Then, the data with text tags, image tags, and audio tags in the pre-classification data set are respectively extracted to obtain text structured data, image structured data, and audio structured data, so as to form a structured data set.

[0021] For example, the police in a certain place collect case clues with the same case number through cameras, alarm calls, and online reporting platforms. These data may be text alarm records, surveillance images, or on-site recordings. After the data is input into the pre-established parsing module, the type is determined according to the suffix name of the data. For example, txt represents text, jpg represents image, and mp3 represents audio. Thus, text tags, image tags, or audio tags are assigned to the data according to the suffix name, and then a pre-classification data set is output. This method facilitates the subsequent processing of different types of data and improves efficiency.

[0022] Preferably, in the step S1, data with text labels is extracted from the pre-classification data set, and Jieba segmentation is used to split the data with text labels into words to obtain text structured data including case numbers, persons, times, locations, and means.

[0023] Specifically, when extracting data with text labels from the pre-classification data set, Jieba segmentation is used to split the description statements of the data into individual words. For example, the data with text labels "A man robbed at knife point late at night" is split into "man", "late at night", "at knife point", and "robbed" to obtain text structured data with case number 001, person being the man, time being late at night, and means being at knife point and robbed. Specifically, the Jieba segmentation tool is used to perform the splitting operation on the data with text labels to obtain a set of words; then the TF-IDF weighting method is used to process the set of words to generate a feature matrix; the importance value of each word is extracted from the feature matrix to determine the weight distribution of the data with text labels; for example, the weight of "robbed" is relatively high, reflecting the core of the case.

[0024] Optionally, in the step S1, data with image labels is extracted from the pre-classification data set, and the OpenCV library is used to extract HOG feature vectors to obtain image structured data including case numbers and feature vectors.

[0025] Specifically, when dealing with data with image labels, the OpenCV library identifies the object contours in the image through HOG feature vectors. Specifically, the image data is loaded through the OpenCV library to obtain the initial image content and the original data set; the HOG feature extraction method is applied to the original data set to generate a set of feature vectors and determine the description information of the object, such as the identity of the knife and the suspect. HOG feature extraction is a method used for object detection and face recognition tasks. Vector analysis is performed on the set of feature vectors to calculate the spatial distribution of each vector and judge the contour candidate regions. If the number of contour candidate regions is greater than the preset threshold, screening is performed through the contour detection algorithm to obtain clear contour boundaries. Image recognition operations are performed according to the clear contour boundaries to classify the object categories within the boundaries to obtain a recognition result set for identifying the object contours in the image. For example, the surveillance captured a suspect holding a knife. HOG extracts the edge features of the knife, and then face recognition is performed on the suspect to obtain the identity of the suspect, which is associated with case number 001.

[0026] Specifically, in the step S1, data with audio labels is extracted from the pre-classification data set, and the Librosa library is used for short-time Fourier transform and extraction of Mel-frequency cepstral coefficients to obtain audio structured data including case numbers and sound types.

[0027] In the processing of data with audio tags, the Librosa library converts on-site recordings into spectrograms through the short-time Fourier transform and then extracts the Mel-frequency cepstral coefficients. For example, a sound type is generated based on the audio of the victim's cries for help and associated with case number 001. This feature reflects the frequency characteristics of the sound, helps to distinguish different audio sources, and enhances the richness of the data.

[0028] Specifically, the on-site environmental audio is collected by a recording device and stored as a digital audio file. The digital audio file is preprocessed, and noise reduction and framing operations are performed to obtain the processed audio signal. The Librosa library is used to perform the short-time Fourier transform on the processed audio signal to obtain a spectrogram. A Mel filter bank is generated based on the spectrogram, and a non-linear transformation is performed on the spectrogram to obtain the transformed spectrogram data. The discrete cosine transform is applied to the transformed spectrogram data to extract the Mel-frequency cepstral coefficients of the audio and obtain the spectral envelope features.

[0029] For example, in a street robbery case, the police used a portable recording pen to record the hisses and footsteps at the scene and stored them as a digital audio file in mp3 format. The on-site environment is complex, and the audio may be mixed with wind or vehicle noise. When preprocessing the digital audio file, the noise reduction operation can effectively reduce background interference. Specifically, the spectral subtraction technique is used to filter out the continuous low-frequency noise in the environment. For example, the frequencies below 100Hz of the wind are weakened, and the main frequency range of the human voice is retained. The framing operation divides the audio into frames of 25 milliseconds each, with an interval of 10 milliseconds between each frame, ensuring the continuity of the signal for easy analysis. This preprocessing makes the subsequent feature extraction more focused on the sounds related to the case. When the Librosa library is used to perform the short-time Fourier transform, the processed audio signal is converted into a spectrogram, which reflects the frequency distribution of the sound over time. For example, the screams of the victim in the robbery case appear as high-frequency protrusions around 1000Hz on the spectrogram, while the footsteps are concentrated below 200Hz. It should be noted that this transformation can intuitively display the dynamic characteristics of the audio and lay a foundation for subsequent analysis. When generating the Mel filter bank based on the spectrogram, the non-linear perception of the human ear for different frequencies is simulated, and the spectrogram is divided into multiple perceptual intervals. Exemplarily, after the Mel transformation of the screams, the high-frequency part is amplified and the features are more prominent. This non-linear transformation makes the perceptual characteristics of the sound more in line with the actual hearing and facilitates the extraction of key information. When the discrete cosine transform is applied to the transformed spectrogram data to extract the Mel-frequency cepstral coefficients, the spectral envelope features are obtained. These Mel-frequency cepstral coefficients describe the overall contour of the sound rather than the instantaneous details. For example, the coefficients of the screams may show that the high energy is concentrated in the first few dimensions, while the footsteps are more flat. This feature extraction helps to distinguish different sound sources.

[0030] It should be noted that in the step S1, the text structured data, image structured data, and audio structured data are aligned according to the case number, and the K-means clustering algorithm is used to cluster the text structured data, image structured data, and audio structured data with different case numbers to obtain a unified structured data set containing the case number and clustering labels; According to the clustering labels in the unified structured data set and the condition that the time interval is less than the preset threshold, the association relationships of cases with different case numbers are matched to obtain a case association data set as the structured data set.

[0031] For example, the cases with case numbers 001 and 002 are involved in similar knife robberies, and the clustering labels corresponding to their text structured data, image structured data, and audio structured data after clustering are of the same category. This can discover potential patterns between cases and improve the detection efficiency. It can be understood that the clustering labels can also assist the police in quickly screening similar cases. According to the clustering labels and the time interval condition, the association relationship is matched. For example, the time interval between the cases with case numbers 001 and 002 is only 2 hours, and according to the text structured data, image structured data, and audio structured data, the locations are 500 meters apart, so it is determined that the same suspect committed the crime, and a case association data set is generated. This method mines the relevance through the space-time dimension and enhances the ability to solve cases.

[0032] It should be noted that in the step S3, an exact string match is performed on the event description table using a preset list of crime keywords to obtain a refined information set; The event element fields in the refined information set are extracted, and according to the event element fields, a preset document specification template that matches the event element fields is obtained, and the person field, time field, location field, and means field in the event description table are filled in according to the preset document specification template.

[0033] In this embodiment, an exact string match is performed using a preset list of crime keywords; assuming that the preset list of crime keywords includes "robbery", "theft", "assault", etc., a match is made for "knife robbery" in the preliminary information set, and "robbery" is hit to generate a refined information set. This process eliminates irrelevant descriptions and focuses on the essence of the case. Specifically, strict rules can be set during the matching, such as only recognizing complete words to avoid mis-matching "knife" with other crime keywords. After extracting the event element fields from the refined information set, a preset document specification template that matches the event element fields is obtained according to the event element fields. For example, if the event element field is "robbery", then a preset document specification template that matches the robbery case is obtained, and then the corresponding information in the event description table is filled into the preset document specification template.

[0034] Preferably, in step S3, when filling in a preset document specification template, if a field is missing, it is marked as NULL; for the field marked as NULL, a filling function is called from the structured data set to supplement the corresponding information to the preset document specification template.

[0035] In this embodiment, when filling in a preset document specification template, the case-related entities extracted from the text structured data are first used as the main information for filling. When data is missing, the image structured data and audio structured data are then used as the basis for supplementing the missing fields. For example, for the case "A man robbed Zhang with a knife", it is filled as: Person NULL, Time 2025-03-30-23:00, Location A street corner, Method Robbing with a knife. At this time, the person field is missing and is marked as NULL; subsequently, a filling function will be called to supplement the person information from the structured data. For example, by identifying "Suspect Wang" through data with image tags, and then filling Wang into the person column of the preset document specification template.

[0036] Specifically, after executing step S3, it further includes step S4: Record the generation process of the law enforcement record document and the data source, and generate a verifiable document generation log.

[0037] It should be noted that in step S4, the log collected by the log recording function is read from the database; an operation time series is generated according to the timestamp corresponding to the generation process in the log; The API call records in the operation time series are matched using a regular expression; the call path containing the data source field is extracted from the API call records; the call path is input into a JSONPath parsing tool; the institution code and table name in the data source field are separated by the JSONPath parsing tool; a source identifier set is generated according to the institution code and table name; According to the institution code in the source identifier set, the filling content of the corresponding field is obtained from the cache; a hash table is established to store the mapping relationship between the filling content and the document generation steps; where the key of the hash table is the field name, and the field name includes the person field, time field, location field, and method field, and the value is the operation step of the generation process; a filling mapping table is output according to the hash table; Align the operation steps of the generation process in the filling mapping table with the timestamps in the operation time series; merge the field contents with the same timestamp; generate a log information set containing the field name, operation step, and timestamp.

[0038] When collecting logs using the logging function, the log format is set to JSON format. Exemplarily, the JSON format may include fields such as "timestamp", "operation type of the generation process", "user ID", etc., which facilitates structured storage and subsequent parsing. In this embodiment, the logging function records each operation in the form of key-value pairs. For example, a log is {"timestamp": "2025-03-31-14:23:45.123", "action": "init", "user": "admin"}, and this format is clear and easy to expand. When storing the JSON-formatted log into the database, it can be understood that the database table can be designed to include three columns: "log ID", "timestamp", and "log content". The log content is directly stored in the text field. For example, the above JSON string is directly stored in the database to ensure data integrity. After reading the log from the database, an operation time series is generated based on the millisecond-level timestamp. Specifically, assuming that the timestamps of three logs read are "2025-03-31-14:23:45.123", "2025-03-31-14:23:45.456", and "2025-03-31-14:23:46.789", they are arranged in order to form an operation time series, which is convenient for tracking the operation order.

[0039] When using regular expressions to match API call records, for example, the regular expression "GET / api / data.*" can extract records such as "GET / api / data / source". This method efficiently filters out irrelevant logs and focuses on the API call path. After extracting the call path of the data source field from the API call record, the JSONPath parsing tool is input to separate the organization code and the table name. For example, for the call path " / api / data / org123 / table456", after JSONPath parsing, "org123" is extracted as the organization code and "table456" is extracted as the table name, and the separation process is intuitive and accurate. A set of source identifiers is generated based on the organization code and the table name. For example, "org123_table456" is used as the unique identifier to reflect the data source.

[0040] Fill-in content is obtained from the cache according to the set of source identifiers. Preferably, for the set of source identifiers "org123_table456", the key-value pair stored in the cache is "org123_table456: person's name", which quickly fills the field. When establishing a hash table to store the mapping relationship between the fill-in content and the document generation steps, the key is the field name such as "person", and the value is the operation step of the generation process such as "step_001". For example, "person: step_001", "time: step_002" form a mapping table, clearly corresponding to the operation steps.

[0041] After filling the mapping table according to the hash table, align it with the operation time series. It should be noted that assuming that the operation step "step_001" in the generation process corresponds to the timestamp "2025-03-31-14:23:45.123", after alignment, the field contents with the same timestamp are merged. For example, "Person: Zhang" and "Time: 14:23" are merged into one record. After arranging the merged records in ascending order of timestamps, a set of log information is generated, such as "2025-03-31-14:23:45.123, Person: Zhang, step_001" and "2025-03-31-14:23:46.789, Location: A certain street, step_003", fully tracing the field names, operation steps, and timestamps. This method ensures the orderliness and accuracy of data traceability through the operation time series and mapping relationship, facilitating subsequent analysis and auditing.

[0042] Specifically, this solution integrates multi-modal technologies such as natural language processing, computer vision, and speech recognition, enabling it to perceive, understand, and respond to complex law enforcement scenarios. It solves the deficiencies of existing methods in multi-modal information fusion processing and significantly improves the intelligent level of law enforcement record document generation.

[0043] The above has described the embodiments of the present invention in detail with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, without departing from the principles and spirit of the present invention, various changes, modifications, substitutions, and variations to these embodiments still fall within the protection scope of the present invention.

Claims

1. A method for generating law enforcement record documents through multi-source data fusion, characterized in that, It includes the following steps: S1: Obtain multi-source data of police cases, classify the text, images, and audio in the multi-source data to obtain a structured data set; S2: Extract case-related entities from the structured data set to generate an event description table; S3: Determine the corresponding event element fields in the event description table, and generate a law enforcement record document according to a preset document specification template.

2. The method for generating a law enforcement record document through multi-source data fusion according to claim 1, wherein: In step S1, input the multi-source data with the same case number into a pre-established parsing module, and use the pre-established parsing module to judge the type according to the suffix name of each data in the multi-source data, and output a pre-classification data set containing text tags, image tags, and audio tags; then extract the data with text tags, image tags, and audio tags in the pre-classification data set respectively to obtain text structured data, image structured data, and audio structured data, so as to form a structured data set.

3. The method for generating a law enforcement record document through multi-source data fusion according to claim 2, wherein: In step S1, extract the data with text tags from the pre-classification data set, and use Jieba word segmentation to split the data with text tags to obtain text structured data including case number, person, time, location, and means.

4. A method for generating law enforcement record documents with multi-source data fusion according to claim 3, characterized in that: In step S1, extract the data with image tags from the pre-classification data set, and use the OpenCV library to extract HOG feature vectors to obtain image structured data including case number and feature vectors.

5. A method for generating law enforcement record documents through multi-source data fusion according to claim 4, characterized in that: In step S1, extract the data with audio tags from the pre-classification data set, and use the Librosa library to perform short-time Fourier transform and extract Mel frequency cepstral coefficients to obtain audio structured data including case number and sound type.

6. The method for generating law enforcement record documents through multi-source data fusion according to claim 5, wherein: In step S1, align the text structured data, image structured data, and audio structured data according to the case number, and use the K-means clustering algorithm to cluster the text structured data, image structured data, and audio structured data with different case numbers to obtain a unified structured data set including case number and clustering labels; According to the clustering labels in the unified structured data set and the condition that the time interval is less than a preset threshold, match the association relationships of cases with different case numbers to obtain a case association data set as the structured data set.

7. A method for generating law enforcement record documents through multi-source data fusion according to claim 6, characterized in that: In step S3, perform exact string matching on the event description table using a preset list of crime keywords to obtain a refined information set; Extract the event element fields in the refined information set, obtain a preset document specification template that matches the event element fields according to the event element fields, and fill in the person field, time field, location field, and means field in the event description table according to the preset document specification template.

8. A method for generating law enforcement record documents through multi-source data fusion according to claim 7, characterized in that: In step S3, when filling in the preset document specification template, if a field is missing, mark it as NULL; for the field marked as NULL, call a filling function from the structured data set to supplement the corresponding information to the preset document specification template.

9. A method for generating law enforcement record documents through multi-source data fusion according to claim 8, characterized in that, After step S3 is executed, it further includes step S4: Record the generation process and data source of the law enforcement record document, and generate a verifiable document generation log.

10. A method for generating law enforcement record documents with multi-source data fusion according to claim 9, characterized in that: In step S4, read the log collected by the log recording function from the database; generate an operation time series according to the time stamp corresponding to the generation process in the log; Use regular expressions to match the API call records in the operation time series; extract the call path containing the data source field from the API call records; input the call path into the JSONPath parsing tool; separate the institution code and table name in the data source field through the JSONPath parsing tool; generate a source identifier set according to the institution code and table name; According to the institution code in the source identifier set, obtain the filling content of the corresponding field from the cache; establish a hash table to store the mapping relationship between the filling content and the document generation steps; where the key of the hash table is the field name, and the field name includes person fields, time fields, location fields, and means fields, and the value is the operation step of the generation process; output a filling mapping table according to the hash table; Align the operation steps of the generation process in the filling mapping table with the time stamps in the operation time series; merge the field contents with the same time stamp; Generate a log information set including field names, operation steps, and time stamps.

Images