Urban management illegal behavior recognition system and method based on large model

Abstract

The invention relates to the technical field of city management, in particular to a city management illegal behavior recognition system and method based on a large model, and the system comprises a front-end sensing module which is used for collecting city site information; the edge calculation module is used for carrying out real-time analysis on the collected city site information, and comprises a model construction sub-module for constructing a Transform model for identifying city management illegal behaviors; using a CLIP algorithm to carry out joint training on the image and the text, and mapping the image and the text to a shared embedding space; reinforcement learning is carried out by using an RLHF algorithm, and the learning effect of the intelligent agent is optimized by combining knowledge and experience of experts; the data management module is used for labeling and managing the data; and the cloud management storage module is used for storing the data of the front-end sensing module and managing and providing a remote access interface. According to the technical scheme, urban management illegal behaviors can be efficiently and accurately identified and analyzed in real time, so that the urban management efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of urban management, and in particular to a system and method for identifying urban management violations based on a large model. Background Art

[0002] With the accelerated advancement of urbanization, the construction of urban management informatization has made significant progress, the scale of cities has continued to expand, and the quality of life of the people has gradually improved. However, this has also posed new challenges to the development of urban management. The expansion of urbanization has made the task of comprehensive law enforcement in urban management increasingly arduous and difficult, especially in the traditional law enforcement model that relies on manual patrols and monitoring. The problems of low efficiency, high cost and difficulty in achieving comprehensive coverage have become increasingly prominent.

[0003] When faced with illegal activities such as occupying roads for business, illegal construction, and environmental pollution, existing urban management methods generally rely on on-site inspections and judgments by law enforcement personnel. This method not only consumes a lot of human resources, but also makes it difficult to achieve real-time monitoring and rapid response. In addition, manual law enforcement also has problems such as strong subjectivity and inconsistent judgment standards, which can easily lead to disputes and affect the objectivity and fairness of law enforcement.

[0004] In order to solve the above-mentioned pain points in urban management, improve urban management efficiency, reduce law enforcement costs, and achieve more fair, objective, and comprehensive law enforcement, advanced technologies such as machine learning (Machine Learning) and computer vision (Computer Vision, CV) are gradually being introduced into urban management. By applying these technologies, "non-contact law enforcement" and "smart law enforcement" in urban management can be achieved, thus overcoming the limitations of traditional law enforcement models.

[0005] As an important intelligent monitoring method, computer vision technology can accurately capture, identify and analyze various target behaviors in complex urban environments. It uses deep learning and other algorithms to intelligently analyze video data captured by surveillance cameras, quickly and accurately detect illegal behaviors in urban management, and provide real-time feedback to law enforcement personnel. Compared with traditional methods, intelligent video analysis technology has multiple advantages such as timely case discovery, reduced labor costs, and improved law enforcement efficiency and accuracy.

[0006] However, although some urban management systems based on computer vision have been applied in practice, most systems still have problems such as low recognition accuracy, slow processing speed, poor real-time performance, and weak ability to adapt to the environment. Especially in the face of complex and changing urban environments, how to further improve the recognition accuracy, real-time performance, and robustness of the system, as well as how to efficiently store and manage large-scale video data, have become issues that need to be urgently addressed by current technology. Summary of the invention

[0007] The purpose of the present invention is to propose a system and method for identifying urban management violations based on a large model. This technical solution can realize efficient, accurate and real-time identification and analysis of urban management violations to improve urban management efficiency.

[0008] To achieve the above objectives, in a first aspect, the embodiments of the present disclosure provide a method for identifying illegal urban management behaviors based on a large model, comprising: Front-end perception module, used to collect city site information; Edge computing module, used for real-time analysis of collected urban site information, including: The model building submodule builds a Transformer model for identifying illegal urban management behaviors; uses the CLIP algorithm to jointly train images and texts, maps images and texts to a shared embedding space, and obtains a visual-language model; uses the RLHF algorithm for reinforcement learning to optimize the learning effect of the intelligent agent by combining the knowledge and experience of experts; Data management module, used to label and manage data; The cloud management storage module is used to store large-scale data from the front-end perception module and manage and provide a remote access interface.

[0009] Beneficial effects of the basic solution: The urban scene information collected by the front-end perception module can be used as data for model training on the one hand, and on the other hand, it ensures the immediacy of subsequent identification of illegal behaviors. The edge computing module enables the collected urban scene information to be processed and analyzed in real time, greatly reducing the delay in data transmission to the cloud, and improving the system's response speed and processing efficiency. The close integration of the front-end perception module and the edge computing module realizes the real-time collection and preliminary processing of data, greatly shortening the time from data generation to analysis result output, and improving the overall response speed of the system.

[0010] The visual-language model built using the Transformer model combined with the CLIP algorithm not only improves the accuracy of image recognition, but also enhances the model's ability to understand complex scenes and nuances through joint training of text and images. The cross-modal joint training method enables the system to more accurately identify various illegal behaviors in urban management, such as illegal parking and occupying roads for business. The introduction of the RLHF algorithm further optimizes the learning effect of the intelligent agent through reinforcement learning combined with the knowledge and experience of human experts, enabling the system to continuously optimize itself and improve recognition accuracy and generalization ability.

[0011] The cloud management storage module can store large-scale data from front-end sensing devices, manage and provide remote access interfaces, and realize centralized management and sharing of data. The data management module is responsible for labeling and managing data, ensuring the quality and consistency of data, and providing a reliable foundation for subsequent analysis and training. At the same time, through effective data management, historical data can be traced and analyzed, providing strong support for decision-making in urban management. The data management module ensures the quality and standardization of data, while the cloud management storage module provides powerful data storage and access capabilities. The two work together to provide strong support for the stable operation and efficient analysis of the system.

[0012] This technical solution uses computer vision technology to realize the value mining and application of urban management law enforcement video data, realize the transformation of urban management from an empirical management model to a data-driven scientific management model, and provide forward-looking and predictive data support for urban management. For the first time, the Transformer visual recognition model is applied in urban management, providing unified AI capability services, reducing the popularization cost of intelligent law enforcement systems and smart services, and improving the level of intelligence in urban management law enforcement.

[0013] As an implementable preferred solution, the front-end perception module includes an image acquisition device and an environmental sensor; the image acquisition device is used to collect urban environment images, and the environmental sensors include humidity, noise, and air quality sensors to assist in analyzing illegal activities.

[0014] As an implementable preferred solution, the edge computing module includes a timing processing submodule, which is used to process video frame sequences using a Transformer model, identify the dynamic process of behavior, understand the start, development and end stages of behavior, and model frame sequences in the video.

[0015] As an implementable preferred solution, the edge computing module also includes a data fusion submodule for fusing video image data, audio signals and text information, and at the same time introduces a cross-modal attention mechanism to allow information interaction between different modalities to enhance feature expression.

[0016] As an implementable preferred solution, the edge computing module also includes a feature processing submodule, which uses the Transformer encoder-decoder structure to automatically extract features and add position encoding, dynamically adjusts attention weights during training, and adaptively learns the most important features for identifying illegal activities.

[0017] As an implementable preferred solution, the edge computing module also includes a spatial modeling submodule, which is used to model spatial information of different scales in the video, and utilizes a multi-head self-attention mechanism to simultaneously focus on features of different scales. By adjusting the number of heads in the self-attention layer and the perception range of each head, the microscopic local details and macroscopic overall scenes of the behavior can be captured.

[0018] As an implementable preferred solution, the edge computing module further includes a computing reasoning submodule for analyzing and identifying illegal behaviors in large-scale urban on-site information; the optimization process of the computing reasoning submodule includes the following contents: Design the student model and the teacher model, use the output of the teacher model as the soft target to guide the training of the student model, and control the smoothness of the soft target by adjusting the temperature parameter T; combine the cross entropy loss and the distillation loss, where the distillation loss calculates the KL divergence between the student model output and the teacher model soft target; Through low-rank decomposition technology, the large weight matrix is ​​decomposed into several low-rank matrix products; the model is reparameterized, and the original weight matrix is ​​replaced by the decomposed low-rank matrix. The low-rank matrix is ​​iteratively optimized during training to minimize the reconstruction error; Quantize the floating-point weights in the model into low-bit fixed-point or binary representations; use quantization-aware methods to adjust the model weights during training to minimize accuracy loss; and perform model calibration after quantization is complete; Design specific compression strategies for different modal data; apply compression strategies in the multimodal feature fusion stage, and also compress the encoders specific to different modalities.

[0019] As an implementable preferred solution, the data management module includes: The data check submodule is used to visualize and check the data quality, and to display the completeness, missing values ​​and outliers of the data in charts; generate descriptive statistical summaries of the data, including mean, median and standard deviation; The data analysis submodule is used to draw histograms to display the distribution of data in order to observe the central tendency, dispersion, skewness and kurtosis of the data; use the cumulative distribution function graph to observe the cumulative probability distribution of the data in order to understand the extreme values ​​and tail behavior of the data; use the scatter plot matrix to display the relationship between different features and identify the correlation and pattern between features.

[0020] As an implementable preferred solution, the data management module further includes: The feature engineering submodule is used to select appropriate feature engineering methods based on the distribution, correlation and relationship between the features and the target variable; The data enhancement submodule is used to determine the category imbalance of the data based on the data analysis results, so as to perform data enhancement to address the category imbalance problem.

[0021] On the second aspect, the disclosed embodiment also provides a method for identifying urban management violations based on a large model, which utilizes the above-mentioned urban management violation identification system based on a large model. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a schematic diagram of the architecture of the urban management violation identification system based on a large model.

[0023] Figure 2 Schematic diagram of the structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the technical solution and advantages of the present application clearer, the technical solution of the present invention will be further described in detail below in conjunction with the accompanying drawings. It will be understood that the specific embodiments described herein are only partial embodiments of the present invention, which are only used to explain the present application, rather than to limit the present application. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered isolated, and they can be combined with each other to achieve better technical effects. The same reference numerals appearing in the drawings of the following embodiments represent the same features or components, which can be applied to different embodiments.

[0025] In addition, unless otherwise defined, the technical terms or scientific terms used in the description of the present invention should have the common meanings understood by those skilled in the art in the art to which the present invention belongs.

[0026] The present invention is further described in detail below in conjunction with the accompanying drawings: Reference numerals: electronic device 500 , processor 501 , communication interface 502 , memory 503 , bus 504 .

[0027] Reference Figure 1 ,The urban management illegal behavior identification system based on a large model includes a front-end perception module, an edge computing module, a data management module, and a cloud management storage module.

[0028] The front-end perception module relies on front-end perception equipment to collect urban on-site information. The front-end perception equipment includes image acquisition devices and environmental sensors.

[0029] The image acquisition device preferably uses a high-definition camera with night vision and wide dynamic range functions to ensure that the city environment image can be captured under different lighting conditions (strong light during the day and dark light at night). The monitoring equipment installed in the city can also be connected to the front-end perception module. The city environment image includes street order, city environment, publicity and advertising, landscaping and construction management, etc.

[0030] Environmental sensors include temperature, humidity, noise, air quality and other sensors, which are used to assist in analyzing illegal activities.

[0031] The edge computing module includes a model building submodule, a time series processing submodule, a data fusion submodule, a feature processing submodule, a feature reasoning submodule, a computing reasoning submodule, and a protocol conversion submodule. It is used to adopt edge computing technology to conduct real-time analysis of urban environmental images, and detect the characteristics of urban management violations through the Transformer visual model, such as drying clothes and bed sheets on the side of the road, small vendors setting up stalls and pushing carts on the roadside, shops occupying the road beyond the store scope, merchants operating with umbrellas within a limited area and time, shops placing billboards on the road, motor vehicles and non-motor vehicles illegally occupying sidewalks, random stacking of items and construction materials, unauthorized occupation of roads for construction, illegal excavation and damage to roads, illegal occupation of sidewalks for construction, erection of illegal construction structures, unauthorized transplantation and felling of urban trees, drying clothes on trees, planting vegetables in green spaces, unauthorized setting up of business stalls in green spaces, unauthorized parking of vehicles in green spaces, etc.

[0032] The model building submodule is used to build a general large model technology base using the Transformer algorithm and build a Transformer model for identifying urban management violations. The Transformer algorithm is a deep learning model based on the self-attention mechanism, which is mainly used in the fields of natural language processing and computer vision. It consists of an encoder and a decoder, each of which contains multiple sublayers, one of which is a multi-head attention layer, which is used to learn the relationship between or within the input or output sequence. The Transformer model does not need to use recurrent neural networks or convolutional neural networks to process sequence data, but uses position encoding to represent the position information of words in the sequence, thereby improving the parallelism and efficiency of the model and better capturing the characteristics of long-distance dependencies.

[0033] The CLIP (Contrastive Language-Image Pretraining) algorithm is used to jointly train images and texts, so that the model can understand the semantic relationship between images and texts. By mapping images and texts to a shared embedding space, the text and image with the closest spatial distance are matched, while the unmatched images and texts are farther away in the space.

[0034] The training process of the CLIP algorithm consists of two main steps: visual encoding and text encoding. In visual encoding, the image is extracted through a CNN or ViT and mapped to an embedding space. In text encoding, the text is encoded through a Transform model and also mapped to an embedding space. The two encoders share parameters to ensure that the embedding spaces of images and texts are consistent. During training, CLIP uses a contrastive loss function to measure the difference between matched and mismatched image-text pairs, so that the model can learn the semantic association between images and texts.

[0035] By jointly training images and text, CLIP can obtain a general vision-language model, which enables the Transformer model to perform well on multiple vision and language tasks, such as image classification, image generation, text description, etc. CLIP encodes images and text into a shared embedding space through contrastive learning, and uses contrastive loss functions to learn the semantic relationship between images and text, thereby realizing the model's understanding and expression capabilities of images and text.

[0036] Reinforcement learning is performed using the RLHF (Reinforcement Learning with Human Feedback) algorithm, which combines the knowledge and experience of human experts to optimize the learning effect of the intelligent model. In RLHF, humans provide feedback on the behavior of the intelligent system, such as which behaviors are correct and which behaviors are wrong. Based on this feedback, the intelligent system can gradually improve its behavior strategy and take more sensible actions in the future. This approach alleviates the problem of a lot of trial and error in traditional reinforcement learning, allowing the intelligent system to learn tasks more efficiently and quickly.

[0037] The temporal processing submodule is used to decompose the video into frame sequences and process these frame sequences using the Transformer model. Each frame is regarded as an element in the sequence. The self-attention mechanism is used to capture the temporal relationship between different frames, identify the dynamic process of the behavior, capture long-distance dependencies, understand the start, development and end stages of the behavior, and model the frame sequence in the video.

[0038] The data fusion submodule is used to fuse the video's image data, audio signals, and possible text information (such as subtitles or billboard text in the video), design modality-specific encoders, such as an image encoder to process visual information and an audio encoder to process sound data. At the same time, a cross-modal attention mechanism is introduced to allow information interaction between different modalities and enhance the expressiveness of features.

[0039] The feature processing submodule uses the encoder-decoder structure of Transformer to automatically extract and transform features without manually designing feature extraction algorithms. Position encoding is added to the model input to ensure that the model can understand the position information of frames or features in the sequence. The model dynamically adjusts the attention weights during training and adaptively learns the most important features for illegal behavior identification.

[0040] The spatial modeling submodule is used to model spatial information of different scales in the video. It uses the multi-head self-attention mechanism to focus on features of different scales at the same time. By adjusting the number of heads in the self-attention layer and the perception range of each head, it captures the microscopic local details and macroscopic overall scenes of the behavior; it integrates features of different scales to form a comprehensive understanding of the behavior, including the specific location and environmental background of the behavior.

[0041] The computational reasoning submodule is used to analyze and identify illegal behaviors in large-scale urban scene information, and optimize performance through technologies such as model compression and acceleration.

[0042] Design a lightweight "student" model and a well-trained "teacher" model with excellent performance, and use the output of the teacher model as a soft target to guide the training of the student model. By adjusting the temperature parameter T, the smoothness of the soft target is controlled to optimize the learning accuracy and generalization ability of the student model. In the training of the student model, the cross entropy loss and distillation loss are used in combination (distillation loss is used to calculate the KL divergence between the output of the student model and the soft target of the teacher model) to ensure that the student model can effectively simulate the decision boundary and uncertainty of the teacher model while ensuring classification accuracy.

[0043] In order to further reduce the model parameters and the amount of calculation, the large weight matrix in the model is decomposed into the product of two or more smaller low-rank matrices. Through low-rank decomposition, key features are extracted, redundant information in the weight matrix is ​​removed, and the number of model parameters is reduced. The model is reparameterized and the decomposed low-rank matrix is ​​used to replace the original weight matrix. This maintains the model functionality while reducing the model complexity. During the training process, the low-rank matrix is ​​iteratively optimized to minimize the reconstruction error.

[0044] Quantize the floating-point weights in the model to low-bit fixed-point or binary representations to reduce memory usage and speed up the inference process. During the training process, adjust the model weights through quantization-aware methods to ensure that the quantized model maintains performance while minimizing accuracy loss. After quantization, calibrate the model and adjust the quantized weights to further optimize model performance.

[0045] Specific compression strategies are designed for different modal data such as video, audio, and text. In the multimodal feature fusion stage, the corresponding compression strategy is applied to reduce the amount of fused data. The fused feature vector retains key information and effectively reduces the dimension and storage requirements. The modality-specific encoder is also compressed to reduce the number of encoder parameters, which not only maintains the encoding efficiency but also further reduces the consumption of computing resources.

[0046] For the optimization of the computational reasoning submodule, starting from the efficient model training of knowledge distillation, the knowledge transfer is deployed to the student model without destroying the model structure. The model is deeply compressed through low-rank decomposition and quantization technology to reduce memory resource usage, which not only improves the processing speed of the model but also maintains the analysis capability. At the same time, in view of the characteristics of multimodal data, a special compression and fusion strategy is designed and implemented to ensure that various types of information can be extracted and fused efficiently and accurately in the video analysis of complex urban environments. The above technical features are closely combined to form an efficient, low-cost, and high-precision video analysis solution.

[0047] The data management module is used to annotate data and quickly access large-scale data through the search engine. The data management module includes a data inspection submodule, a data analysis submodule, a feature engineering submodule, a data enhancement submodule, and a data annotation submodule.

[0048] The data inspection submodule is used to visualize the data, which is convenient for intuitively checking the quality of the data, including data integrity, missing values, outliers, etc. It is also used to use bar charts, pie charts and other charts to display the integrity of the data and the distribution of missing values. Boxplot is used to identify outliers in the data, and they are quickly located through quartiles and outlier markers. Descriptive statistical summaries of the data are generated, including mean, median, standard deviation, etc., so as to quickly understand the basic characteristics of the data.

[0049] The data analysis submodule is used to draw a histogram to show the distribution of data, so as to observe the central tendency, dispersion, skewness and kurtosis of the data. The cumulative distribution function (CDF) graph is used to observe the cumulative probability distribution of the data and understand the extreme values ​​and tail behavior of the data. The relationship between different features is displayed through a scatter plot matrix to identify the correlation and pattern between features. By visualizing the distribution of data, the relationship between different categories or different features can be more clearly understood.

[0050] The feature engineering submodule is used to select appropriate feature engineering methods, such as feature scaling, feature selection, and feature synthesis, based on the distribution, correlation, and relationship between the features and the target variable to improve the performance of the model.

[0051] The data enhancement submodule is used to determine the category imbalance of the data based on the data analysis results, so as to perform data enhancement for the category imbalance problem. For example, the sample distribution of the minority category can be obtained more intuitively through visualization, so as to purposefully generate more samples of this category to improve the balance of the training data.

[0052] The data annotation submodule is used to quickly and accurately annotate the visualized data. For example, in computer vision tasks, the visualized image can be used with interactive annotation tools to more accurately select the target objects in the data and annotate the bounding boxes.

[0053] The cloud management storage module is used to build a data lake, store large-scale data from front-end sensing devices, and manage and provide a secure remote access interface to achieve real-time monitoring and analysis of urban management conditions.

[0054] The disclosed embodiment also provides a system for identifying illegal urban management behaviors based on a large model. Figure 2 The system uses the above-mentioned large-model-based urban management violation identification method.

[0055] The embodiment of the present disclosure also provides a storage medium, in which a computer program is stored. When the computer program is executed by a processor, all the steps of the above-mentioned method for identifying urban management violations based on a large model can be implemented.

[0056] Those of ordinary skill in the art can understand that all or part of the processes in the method for identifying urban management violations based on a large model can be implemented by instructing related hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of various embodiments of the method for identifying urban management violations based on a large model. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0057] The embodiment of the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned method for identifying illegal urban management behaviors based on a large model when executing the program. In the embodiment of the present application, the processor is the control center of the computer system, which can be a processor of a physical machine or a processor of a virtual machine.

[0058] Reference Figure 2, the electronic device 500 includes: at least one processor 501, at least one communication interface 502, at least one memory 503 and at least one bus 504. Among them, the bus 504 is used to realize the connection and communication between these components, the communication interface 502 is used to communicate with other node devices for signaling or data, and the memory 503 stores machine-readable instructions executable by the processor 501. When the electronic device 500 is running, the processor 501 communicates with the memory 503 through the bus 504, and the machine-readable instructions are called by the processor 501 to perform the steps of the above-mentioned method for identifying illegal urban management behaviors based on a large model. The above content is only an embodiment of the present invention, and the common sense such as the known specific structure and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field know all the ordinary technical knowledge in the technical field of the invention before the application date or priority date, can know all the existing technologies in the field, and have the ability to apply the conventional experimental means before the date. The ordinary technicians in the relevant field can improve and implement the scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. The urban management illegal behavior identification system based on large model is characterized by: include: Front-end perception module, used to collect city site information; Edge computing module, used for real-time analysis of collected urban site information, including: The model building submodule builds a Transformer model for identifying illegal urban management behaviors; uses the CLIP algorithm to jointly train images and texts, maps images and texts to a shared embedding space, and obtains a visual-language model; uses the RLHF algorithm for reinforcement learning to optimize the learning effect of the intelligent agent by combining the knowledge and experience of experts; Data management module, used to label and manage data; The cloud management storage module is used to store large-scale data from the front-end perception module and manage and provide a remote access interface.

2. The urban management illegal behavior identification system based on large model according to claim 1 is characterized by: The front-end perception module includes an image acquisition device and an environmental sensor; the image acquisition device is used to collect urban environment images, and the environmental sensors include humidity, noise, and air quality sensors to assist in analyzing illegal activities.

3. The urban management illegal behavior identification system based on large model according to claim 2 is characterized by: The edge computing module includes a timing processing submodule, which is used to process video frame sequences using a Transformer model, identify the dynamic process of behavior, understand the start, development and end stages of the behavior, and model the frame sequence in the video.

4. The urban management illegal behavior identification system based on large model according to claim 3 is characterized by: The edge computing module also includes a data fusion submodule for fusing video image data, audio signals and text information, and introducing a cross-modal attention mechanism to allow information interaction between different modalities to enhance feature expression.

5. The urban management illegal behavior identification system based on large model according to claim 4 is characterized by: The edge computing module also includes a feature processing submodule, which uses the Transformer encoder-decoder structure to automatically extract features and add position encoding, dynamically adjusts attention weights during training, and adaptively learns the most important features for illegal behavior identification.

6. The urban management illegal behavior identification system based on large model according to claim 5 is characterized by: The edge computing module also includes a spatial modeling submodule, which is used to model spatial information of different scales in the video, and uses a multi-head self-attention mechanism to simultaneously focus on features of different scales. By adjusting the number of heads in the self-attention layer and the perception range of each head, the microscopic local details and macroscopic overall scenes of the behavior can be captured.

7. The urban management illegal behavior identification system based on large model according to claim 1 is characterized by: The edge computing module also includes a computing reasoning submodule for analyzing and identifying illegal behaviors in large-scale urban on-site information; the optimization process of the computing reasoning submodule includes the following contents: Design the student model and the teacher model, use the output of the teacher model as the soft target to guide the training of the student model, and control the smoothness of the soft target by adjusting the temperature parameter T; combine the cross entropy loss and the distillation loss, where the distillation loss calculates the KL divergence between the student model output and the teacher model soft target; Through low-rank decomposition technology, the large weight matrix is ​​decomposed into several low-rank matrix products; the model is reparameterized, and the original weight matrix is ​​replaced by the decomposed low-rank matrix. The low-rank matrix is ​​iteratively optimized during training to minimize the reconstruction error; Quantize the floating-point weights in the model into low-bit fixed-point or binary representations; use quantization-aware methods to adjust the model weights during training to minimize accuracy loss; and perform model calibration after quantization is complete; Design specific compression strategies for different modal data; apply compression strategies in the multimodal feature fusion stage, and also compress the encoders specific to different modalities.

8. The urban management illegal behavior identification system based on large model according to claim 1 is characterized by: The data management module comprises: The data check submodule is used to visualize and check the data quality, and to display the completeness, missing values ​​and outliers of the data in charts; generate descriptive statistical summaries of the data, including mean, median and standard deviation; The data analysis submodule is used to draw histograms to display the distribution of data in order to observe the central tendency, dispersion, skewness and kurtosis of the data; use the cumulative distribution function graph to observe the cumulative probability distribution of the data in order to understand the extreme values ​​and tail behavior of the data; use the scatter plot matrix to display the relationship between different features and identify the correlation and pattern between features.

9. The urban management illegal behavior identification system based on large model according to claim 8 is characterized by: The data management module further includes: The feature engineering submodule is used to select appropriate feature engineering methods based on the distribution, correlation and relationship between the features and the target variable; The data enhancement submodule is used to determine the category imbalance of the data based on the data analysis results, so as to perform data enhancement to address the category imbalance problem.

10. A method for identifying illegal urban management behaviors based on a large model, characterized by: The urban management violation identification system based on a large model as described in claims 1-9 is used.

Images