Shovel repair video generation system and method based on size model collaboration and global cache optimization

Abstract

The application belongs to the technical field of video generation, and discloses a system and method for generating electric shovel repair videos based on size model cooperation and global cache optimization, which comprises the following steps: obtaining structured electric shovel repair data and performing knowledge preprocessing, obtaining standardized prompt words, and transmitting the standardized prompt words to a size model; performing high-noise processing on the standardized prompt words based on an electric shovel repair special large model, generating a repair process semantic structure, and adaptively detecting model switching conditions; when the model switching conditions are met, switching an electric shovel repair special small model to perform low-noise processing and generate a preliminary video frame sequence; analyzing the preliminary video frame sequence, obtaining an optimal cache path, and performing precision compression and acceleration on model cooperation processing; performing quality checking on the preliminary video frame sequence, and encoding, outputting and storing the preliminary video frame sequence after the quality checking is passed; through the architecture design and function integration at the system level, the application effectively optimizes the professionalism, efficiency, adaptability and stability of the electric shovel repair video generation.

Description

Technical Field

[0001] This invention belongs to the field of engineering machinery maintenance and video generation technology, specifically involving an electric shovel maintenance video generation system and a method for electric shovel maintenance video generation based on large-scale model collaboration and global caching optimization. Background Technology

[0002] As a core production equipment in open-pit mines, the maintenance process for the slewing bearing of electric shovels is complex, involving key processes such as central pivot installation and gear ring tightening, requiring extremely high operational standardization and precision. Current maintenance training and guidance rely primarily on electric shovel maintenance videos obtained through on-site filming, which presents the following problems: 1. High production cost: Real-world filming requires downtime for equipment, coordination with professional maintenance teams and venues, and cannot flexibly demonstrate the disassembly and assembly process of complex internal structures; 2. Difficulty in reproducing details: Key processes in maintenance (such as 0.254mm gap adjustment and 1978.5 N·m bolt tightening torque) are difficult to clearly present through actual filming, which can easily lead to misunderstandings by trainees; 3. Low generation efficiency: Traditional text-based video technology suffers from problems such as poor spatiotemporal consistency (e.g., chaotic component disassembly and assembly trajectories) and distortion of professional details (e.g., structural errors in hydraulic systems) when processing industrial-grade professional scenarios. Furthermore, the reasoning process is time-consuming and cannot quickly adapt to the needs of different maintenance scenarios. 4. Deficiencies of existing acceleration technologies: Cache-based video generation acceleration methods (such as TeaCache and PAB) only focus on local error optimization, which can easily lead to global content degradation; simple large and small model collaboration strategies lack targeted adaptation to professional scenarios and are difficult to balance the professionalism of repair videos with generation efficiency.

[0003] Furthermore, most existing mainstream video generation systems are general-purpose designs that lack customized development for electric shovel maintenance scenarios, resulting in the following prominent defects: a lack of integrated maintenance expertise, leading to low fidelity in the reproduction of component parameters and process specifications in generated videos, failing to meet training needs; a loose model collaboration mechanism, failing to optimize switching logic based on the noise characteristics of maintenance scenarios, resulting in an imbalance between semantic structure and detail presentation; a lack of targeted caching strategies, failing to consider the semantic relevance and error sensitivity of maintenance steps, making it difficult to balance inference efficiency and quality; and a lack of integrated end-to-end management functions, with fragmented development, debugging, and deployment processes, resulting in insufficient flexibility in adapting to industrial scenarios. Summary of the Invention

[0004] To address some or all of the technical problems existing in the prior art, this invention provides a system and method for generating electric shovel maintenance videos based on size model collaboration and global caching optimization.

[0005] The first aspect of this invention provides a shovel maintenance video generation system based on size model collaboration and global caching optimization. It adopts a distributed architecture design, including a hardware layer, a software layer, and an application layer. Each layer collaborates to achieve full-process management of shovel maintenance video generation. The shovel maintenance video generation system includes: The task configuration and knowledge loading module is used to obtain task configurations based on user selections, retrieve structured electric shovel maintenance data from the structured electric shovel maintenance knowledge base and perform knowledge preprocessing, obtain standardized prompt words, and transmit them to the size model; The model collaboration generation module is used to process standardized prompt words with high noise based on the large model dedicated to electric shovel maintenance, generate the semantic structure of the maintenance process, and adaptively detect model switching conditions. When the model switching conditions are met, the module switches to the small model dedicated to electric shovel maintenance for low noise processing to generate a preliminary video frame sequence. The global cache optimization module is used to analyze the initial video frame sequence, obtain the optimal cache path, and perform precision compression and acceleration for model collaborative processing. The quality verification and output module is used to verify the quality of the initial video frame sequence. After the verification is successful, the frame is encoded, output, and stored.

[0006] Furthermore, the hardware layer includes: The computing node unit is used to deploy computing servers at the central computing node to run collaborative computing tasks for electric shovel maintenance using a specific model of different sizes. Deploy computing devices on edge computing nodes to run local inference and cache scheduling; The storage node unit is used to store the electric shovel maintenance knowledge base, the model weight file of the electric shovel maintenance-specific size model, the generated video data, and the operation logs using a distributed storage cluster. The interface device unit is used to acquire input data of the electric shovel repair scene and video preview and playback data according to the preset image acquisition device; Data is processed for input / output and remote interaction using remote communication devices.

[0007] Furthermore, the application layer includes: The scene configuration unit provides a graphical user interface, allowing users to select maintenance scenes and configure generation parameters to obtain video generation tasks. The video preview and editing unit is used to preview the video in real time during the video generation process and provides simple editing functions, including key step cropping and parameter standard adjustment. The data management unit is used for the classification, storage, retrieval, export, sharing, data backup, and data recovery of electric shovel maintenance videos. The operation management unit is used for user permission management, role configuration, log querying, and operation upgrades.

[0008] Furthermore, the task configuration and knowledge loading module includes: The knowledge base unit is used to build a structured electric shovel maintenance knowledge base, which allows for adding, deleting, modifying, querying, and version management of electric shovel maintenance knowledge. The structured electric shovel maintenance knowledge base includes: a component parameter library, a process specification library, and a tool and spare parts library. The knowledge preprocessing unit integrates natural language processing algorithms and feature extraction models to preprocess the structured electric shovel maintenance data in the structured electric shovel maintenance knowledge base, converting it into standardized prompt words that can be recognized by the electric shovel maintenance-specific size model, and automatically supplementing process logic association information and safety specification descriptions. The knowledge update unit is used to incrementally update the structured electric shovel maintenance knowledge base according to the preset knowledge update method, and to simultaneously optimize the model adaptability of the feature extraction model and the electric shovel maintenance-specific size model through the incremental training mechanism. The preset knowledge update methods include: on-site data collection and maintenance record import. The task configuration unit is used to obtain task configuration based on user selection, where user selection includes: maintenance scenario selection and configuration video generation parameter selection; The knowledge loading unit is used to retrieve structured electric shovel maintenance data from the structured electric shovel maintenance knowledge base according to the task configuration, perform knowledge preprocessing, obtain standardized prompt words, and transmit them to the electric shovel maintenance-specific size model.

[0009] Furthermore, the model collaboration generation module includes: A dedicated large model unit is used to deploy a dedicated large model for electric shovel maintenance. It optimizes and trains high-noise steps, strengthens the semantic structure of the generated maintenance process, and outputs core data for electric shovel maintenance. The core data for electric shovel maintenance includes: component spatial location data, disassembly and assembly sequence data, and tool placement logic data. A dedicated small model unit is used to deploy a small model specifically for electric shovel maintenance, to render low-noise steps in detail, to integrate a maintenance tool feature library and parameter visualization components, and to optimize the electric shovel maintenance detail data, which includes: bolt tightening action data and torque scale display data. The adaptive switching unit is used to build a switching determination algorithm to calculate the relative distance data of denoised samples in adjacent time steps in real time. When the model switching conditions are met, the model switching is automatically triggered and the data is seamlessly transmitted through the shared memory pool to switch to the electric shovel maintenance special small model. The collaborative generation unit is used to summarize the core data and detailed data of electric shovel maintenance and generate a preliminary video frame sequence. The model optimization unit is used to fine-tune the model based on feedback data from maintenance scenarios, dynamically adjusting semantic understanding weights and detail rendering accuracy.

[0010] Furthermore, the global cache optimization module includes: The Directed Acyclic Graph (DAG) construction unit is used to perform offline analysis on the preliminary video frame sequence and construct a DAG specifically for maintenance. The maintenance steps are used as nodes, the semantic relationship between steps is used as edge weights, and the classification attributes and error sensitivity of the core steps and auxiliary steps are labeled. The path optimization unit is used to obtain the optimal cache path by employing a lexicographical minimax algorithm combined with a fixed inference budget. The precision optimization unit is used to integrate an optimized attention technology to perform precision compression and acceleration on the collaborative processing of the electric shovel repair-specific model, and is configured with a dynamic overflow compensation mechanism to dynamically compensate for quantization errors. The cache management unit is used to monitor cache hit rate and error changes in real time, dynamically adjust the cache strategy, and automatically trigger cache data updates when the cache error of the auxiliary steps exceeds the error threshold.

[0011] Furthermore, the quality verification and output module includes: The full-process control unit is used to manage the entire process of generating electric shovel maintenance videos. The full-process control includes: process design, process compilation, process distribution, process operation, and process monitoring. The process design unit is used to design a visual configuration interface, providing options for maintenance scenario templates and video generation parameter settings. The compilation unit is used to automatically compile the user-selected configuration into an executable inference task for a model specifically designed for electric shovel maintenance, and to perform syntax checks. The distribution unit is used to dynamically allocate inference tasks based on hardware resource load and distribute them between central computing nodes and edge computing nodes according to deployment scenario requirements; The execution unit is used to provide an independent operating environment for multi-task parallel execution and resource isolation. The monitoring unit is used to collect real-time operational status and video generation quality, and to provide abnormal alarms and log recording; The quality verification unit is used to perform quality verification on the initial video frame sequence according to the whole process control, and to verify the consistency of component structure, the accuracy of process parameters and the compliance of safe operation. The output storage unit is used to encode and output the preliminary video frame sequence according to the preset video format after the verification is qualified, and store it in the distributed storage cluster.

[0012] A second aspect of the present invention provides a video generation system for electric shovel maintenance based on size model collaboration and global caching optimization, comprising: Maintenance knowledge modeling module: Stores component parameters, process specifications, tool lists, and safety regulations related to electric shovel maintenance, and supports automatic generation of structured prompts; Large and small model collaboration module: includes a large model and a small model for electric shovel maintenance, equipped with an adaptive switching unit based on the second derivative of noise variation, where the large model is a sketch model and the small model is a rendered model; Global cache optimization module: includes a DAG construction unit for maintenance scenarios, a lexicographically ordered minimax path solver, and an FP8 precision optimization unit; Professional quality verification module: Built-in maintenance specification verification rule library to verify the accuracy of component structure, process parameters and safe operation in the video.

[0013] A third aspect of the present invention provides a method for generating electric shovel maintenance videos based on size model collaboration and global caching optimization, comprising the following steps: Step 1: Maintenance knowledge modeling and prompt word construction: Extract component parameters, maintenance processes, tool lists, and safety specifications of the electric shovel's slewing bearing device to generate structured prompt words; Step 2, Video sketch-rendering generation through collaboration between large and small models: A dedicated large model is used to process high-noise steps to generate semantic structure, and an adaptive switching mechanism is used to switch to a small model to optimize details; Step 3: Global cache optimization for inference acceleration: Construct a DAG specifically for maintenance scenarios, adopt a lexicographical minimax path optimization strategy to select the optimal cache path, and combine precision optimization to improve inference speed; Step 4: Professional quality verification and output: Based on the maintenance specifications, verify the consistency of components, the accuracy of processes, and the compliance with safety requirements of the video, and output qualified maintenance videos.

[0014] Furthermore, the adaptive switching mechanism in step 2 is based on the second derivative of the relative L1 distance between denoised samples in adjacent time steps. ,satisfy ,in The switch is triggered when at least 5 steps of sketch generation are completed.

[0015] Furthermore, in step 3, the DAG uses "maintenance steps" as nodes and "semantic correlation between steps" as edge weights, prioritizing the caching of semantically stable non-core process steps to avoid the accumulation of caching errors for critical process steps.

[0016] Furthermore, the structured prompts include the electric shovel model, maintenance scenario parameters, procedure specifications, and professional requirements, adapting to the maintenance needs of the slewing bearing device of the WK-35 / WK-20 electric shovel.

[0017] The main advantages of the technical solution of this invention are as follows: This invention provides a video generation system for electric shovel maintenance based on large and small model collaboration and global caching optimization. The system acquires structured electric shovel maintenance data and performs knowledge preprocessing to obtain standardized prompt words, which are then transmitted to both large and small models. The large electric shovel maintenance model performs high-noise processing on the standardized prompt words to generate a semantic structure of the maintenance process. It also adaptively detects model switching conditions; when these conditions are met, it switches to a small electric shovel maintenance model for low-noise processing, generating a preliminary video frame sequence. The preliminary video frame sequence is analyzed to obtain the optimal caching path, and the model collaboration processing is subjected to precision compression and acceleration. The preliminary video frame sequence undergoes quality verification; once verified, it is encoded, output, and stored. Through system-level architecture design and functional integration, this invention effectively optimizes the professionalism, efficiency, adaptability, and stability of electric shovel maintenance video generation. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the electric shovel maintenance video generation system based on size model collaboration and global cache optimization provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the overall architecture of the electric shovel maintenance video generation system based on size model collaboration and global caching optimization provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the workflow of the model collaboration module provided in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of path optimization provided in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the electric shovel maintenance video generation system based on size model collaboration and global caching optimization provided in Embodiment 2 of the present invention; Figure 6 This is a schematic diagram of the process for generating the sketch model (Wan-14B fine-tuning version) provided in Embodiment 3 of the present invention; Figure 7 This is a schematic diagram of the overall process for generating electric shovel maintenance videos based on size model collaboration and global caching optimization, as provided in Embodiment 3 of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] Example 1 refer to Figure 1 The electric shovel maintenance video generation system based on size model collaboration and global caching optimization provided in Embodiment 1 of the present invention includes the following structure: The task configuration and knowledge loading module is used to obtain task configurations based on user selections, retrieve structured electric shovel maintenance data from the structured electric shovel maintenance knowledge base and perform knowledge preprocessing, obtain standardized prompt words, and transmit them to the size model; The model collaboration generation module is used to process standardized prompt words with high noise based on the large model dedicated to electric shovel maintenance, generate the semantic structure of the maintenance process, and adaptively detect model switching conditions. When the model switching conditions are met, the module switches to the small model dedicated to electric shovel maintenance for low noise processing to generate a preliminary video frame sequence. The global cache optimization module is used to analyze the initial video frame sequence, obtain the optimal cache path, and perform precision compression and acceleration for model collaborative processing. The quality verification and output module is used to verify the quality of the initial video frame sequence. After the verification is successful, the frame is encoded, output, and stored.

[0021] In an embodiment of the present invention, reference is made to Figure 2 The electric shovel maintenance video generation system adopts a distributed architecture design, including a hardware layer, a software layer, and an application layer. Each layer works together to achieve full-process management of electric shovel maintenance video generation.

[0022] In this embodiment of the invention, user selection refers to the user's selection on the graphical operation interface when generating electric shovel repair videos, such as selecting the repair scenario and configuring generation parameters.

[0023] In this embodiment of the invention, the configuration generation parameters refer to the configuration parameters when generating electric shovel maintenance videos, such as resolution parameters: Frame rate parameter: 25fps.

[0024] In this embodiment of the invention, the structured electric shovel maintenance knowledge base refers to a database involved in the electric shovel maintenance process that can be efficiently analyzed, reasoned about, and updated. The structured nature is specifically reflected in the fact that the electric shovel maintenance data is stored in an associated form, with semantic association, standardized prompts, and incremental updates. For example, data storage includes each data entry in the component parameter library, which contains the component name and parameter type. Semantic association includes, for example, the torque tightening data in the process specification library is associated with the bolt data in the component parameter library.

[0025] In this embodiment of the invention, structured electric shovel maintenance data refers to structured data stored in a structured electric shovel maintenance knowledge base.

[0026] In this embodiment of the invention, knowledge preprocessing refers to the process of converting structured data into standardized prompt words.

[0027] In this embodiment of the invention, standardized prompt words refer to standard words obtained based on knowledge preprocessing, which are used to accurately represent structured electric shovel maintenance data and realize automated mapping from knowledge to model input.

[0028] In this embodiment of the invention, the large and small models refer to the large model and the small model specifically for electric shovel repair.

[0029] In this embodiment of the invention, high noise processing refers to the optimization training of high noise steps based on a large-scale model specifically designed for electric shovel maintenance.

[0030] In this embodiment of the invention, the maintenance process semantic structure refers to the structured expression generated by the large-scale electric shovel maintenance model to describe the maintenance process.

[0031] In this embodiment of the invention, the model switching condition refers to the condition for switching from the large electric shovel repair model to the small electric shovel repair model.

[0032] In this embodiment of the invention, low-noise processing refers to detailed rendering of low-noise steps based on a small model specifically designed for electric shovel maintenance.

[0033] In this embodiment of the invention, the preliminary video frame sequence refers to the sequence of video frames initially obtained after processing standardized prompt words according to a size model.

[0034] In this embodiment of the invention, the optimal cache path refers to the optimal scheduling scheme selected by the lexicographical minimax algorithm in the global cache optimization module, which guides which steps in the video generation process can reuse the cache and which must be fully computed.

[0035] In this embodiment of the invention, precision compression and acceleration refers to the operation of replacing the traditional high-precision format with a lower bit-width numerical format, thereby reducing the amount of computation and memory usage, and ensuring precision through a dynamic overflow compensation mechanism.

[0036] In this embodiment of the invention, quality verification refers to verifying the consistency of component structure, the accuracy of process parameters, and the compliance of safe operation on the preliminary video frame sequence.

[0037] In this embodiment of the invention, the hardware layer includes: The computing node unit is used to deploy computing servers at the central computing node to run collaborative computing tasks for electric shovel maintenance using a specific model of different sizes. Deploy computing devices on edge computing nodes to run local inference and cache scheduling; The storage node unit is used to store the electric shovel maintenance knowledge base, the model weight file of the electric shovel maintenance-specific size model, the generated video data, and the operation logs using a distributed storage cluster. The interface device unit is used to acquire input data of the electric shovel repair scene and video preview and playback data according to the preset image acquisition device; Data is processed for input / output and remote interaction using remote communication devices.

[0038] In this embodiment of the invention, the central computing node refers to a node deployed in the cloud for running collaborative computing tasks of large and small models.

[0039] In this embodiment of the invention, an edge computing node refers to a node that is deployed locally for complex local inference and cache scheduling.

[0040] In this embodiment of the invention, a distributed storage cluster refers to a collection of storage nodes that use a distributed structure for data storage.

[0041] In this embodiment of the invention, the model weight file refers to the file containing the parameters obtained from training a large or small model.

[0042] In this embodiment of the invention, the preset image acquisition device refers to a preset image acquisition device used for on-site scene shooting input and video preview and playback.

[0043] In this embodiment of the invention, a remote communication device refers to a device used to realize data input / output and remote interaction.

[0044] For example, two computing servers are deployed at the central computing node, each configured with two NVIDIA H2O GPUs and 128GB DDR5 memory; and one computing device is deployed at the edge computing node, each configured with one NVIDIA L4 GPU and 64GB memory.

[0045] For example, a distributed storage cluster is used in the storage nodes, with a total capacity of not less than 10TB.

[0046] For example, industrial-grade image acquisition equipment is used to acquire on-site scene input data for electric shovel maintenance; high-definition display equipment is used for video preview and playback; and 5G or Wi-Fi 6E dual-mode is used for remote communication.

[0047] In this embodiment of the invention, the application layer includes: The scene configuration unit provides a graphical user interface, allowing users to select maintenance scenes and configure generation parameters to obtain video generation tasks. The video preview and editing unit is used to preview the video in real time during the video generation process and provides simple editing functions, including key step cropping and parameter standard adjustment. The data management unit is used for the classification, storage, retrieval, export, sharing, data backup, and data recovery of electric shovel maintenance videos. The operation management unit is used for user permission management, role configuration, log querying, and operation upgrades.

[0048] In this embodiment of the invention, the graphical user interface refers to an interface that graphically represents the maintenance operation, allowing users to select maintenance scenarios and generate parameter settings on this interface.

[0049] In this embodiment of the invention, configuring generation parameters refers to the parameter settings selected by the user when generating the video.

[0050] In this embodiment of the invention, the video generation task refers to the task of generating electric shovel repair videos according to user requirements.

[0051] In this embodiment of the invention, the simple editing function refers to the function of simply editing the video.

[0052] In this embodiment of the invention, the key steps refer to the steps that must exist in the electric shovel maintenance process.

[0053] In this embodiment of the invention, stable system operation is ensured through user permission management, role configuration, log querying, and operation upgrades.

[0054] For example, maintenance scenarios include pivot replacement and gear ring repair.

[0055] In this embodiment of the invention, the task configuration and knowledge loading module includes: The knowledge base unit is used to build a structured electric shovel maintenance knowledge base, which allows for adding, deleting, modifying, querying, and version management of electric shovel maintenance knowledge. The structured electric shovel maintenance knowledge base includes: a component parameter library, a process specification library, and a tool and spare parts library. The knowledge preprocessing unit integrates natural language processing algorithms and feature extraction models to preprocess the structured electric shovel maintenance data in the structured electric shovel maintenance knowledge base, converting it into standardized prompt words that can be recognized by the electric shovel maintenance-specific size model, and automatically supplementing process logic association information and safety specification descriptions. The knowledge update unit is used to incrementally update the structured electric shovel maintenance knowledge base according to the preset knowledge update method, and to simultaneously optimize the model adaptability of the feature extraction model and the electric shovel maintenance-specific size model through the incremental training mechanism. The preset knowledge update methods include: on-site data collection and maintenance record import. The task configuration unit is used to obtain task configuration based on user selection, where user selection includes: maintenance scenario selection and configuration video generation parameter selection; The knowledge loading unit is used to retrieve structured electric shovel maintenance data from the structured electric shovel maintenance knowledge base according to the task configuration, perform knowledge preprocessing, obtain standardized prompt words, and transmit them to the electric shovel maintenance-specific size model.

[0056] In this embodiment of the invention, the structured electric shovel maintenance knowledge base supports the addition, deletion, modification, and querying of knowledge, as well as version management.

[0057] In this embodiment of the invention, a natural language processing algorithm is used to perform preliminary data processing on the structured electric shovel maintenance data to ensure the quality and efficiency of subsequent data conversion.

[0058] In this embodiment of the invention, the feature extraction model refers to a model for extracting features from structured electric shovel maintenance data.

[0059] In this embodiment of the invention, standardized prompt words refer to standard words obtained based on knowledge preprocessing, which are used to accurately represent structured electric shovel maintenance data and realize automated mapping from knowledge to model input.

[0060] In this embodiment of the invention, process logic association information refers to information about the correlation between process logic.

[0061] In this embodiment of the invention, the preset knowledge update method refers to a preset method for updating the knowledge base of the structured electric shovel maintenance system.

[0062] In this embodiment of the invention, task configuration refers to the configuration for generating electric shovel maintenance videos.

[0063] For example, the component parameter library stores 32 types of core parameters, such as the central pivot diameter of 508.33±0.03mm; the process specification library stores the operation procedures and parameter thresholds of 13 key steps; and the tool spare parts library stores the appearance characteristics and operation specifications of 34 special tools.

[0064] For example, configure video generation parameter selection, resolution Frame rate 25fps.

[0065] For example, based on the task configuration, the corresponding component parameters, process specifications and tool information are retrieved from the structured electric shovel maintenance knowledge base, converted into standardized prompt words after knowledge preprocessing, and transmitted to the size model.

[0066] In an embodiment of the present invention, reference is made to Figure 3 The model collaboration generation module includes: A dedicated large model unit is used to deploy a dedicated large model for electric shovel maintenance. It optimizes and trains high-noise steps, strengthens the semantic structure of the generated maintenance process, and outputs core data for electric shovel maintenance. The core data for electric shovel maintenance includes: component spatial location data, disassembly and assembly sequence data, and tool placement logic data. A dedicated small model unit is used to deploy a small model specifically for electric shovel maintenance, to render low-noise steps in detail, to integrate a maintenance tool feature library and parameter visualization components, and to optimize the electric shovel maintenance detail data, which includes: bolt tightening action data and torque scale display data. The adaptive switching unit is used to build a switching determination algorithm to calculate the relative distance data of denoised samples in adjacent time steps in real time. When the model switching conditions are met, the model switching is automatically triggered and the data is seamlessly transmitted through the shared memory pool to switch to the electric shovel maintenance special small model. The collaborative generation unit is used to summarize the core data and detailed data of electric shovel maintenance and generate a preliminary video frame sequence. The model optimization unit is used to fine-tune the model based on feedback data from maintenance scenarios, dynamically adjusting semantic understanding weights and detail rendering accuracy.

[0067] In this embodiment of the invention, the large-scale model for electric shovel maintenance refers to a model specifically designed for optimizing high-noise steps in electric shovel maintenance.

[0068] In this embodiment of the invention, the high-noise step refers to the early inference step with a high noise level during the diffusion process. In this step, the image structure has not yet been formed, and the large model for electric shovel maintenance is responsible for macroscopic generation and layout.

[0069] In this embodiment of the invention, the core data for electric shovel maintenance refers to the data generated by optimizing and training high-noise steps based on a large-scale electric shovel maintenance model.

[0070] In this embodiment of the invention, the electric shovel repair-specific small model refers to a model specifically designed for rendering low-noise steps in electric shovel repair.

[0071] In this embodiment of the invention, the low-noise step refers to a late-stage inference step with low noise levels during the diffusion process. In this step, the image structure has been formed, and the small model for electric shovel repair is mainly responsible for detail rendering.

[0072] In this embodiment of the invention, the electric shovel repair detail data refers to the data that accurately represents the repair process from a microscopic perspective during the electric shovel repair process, which is obtained based on a special small model for electric shovel repair.

[0073] In this embodiment of the invention, the built-in switching determination algorithm refers to a built-in algorithm used to determine the size model for switching.

[0074] In this embodiment of the invention, adjacent time steps refer to maintenance steps at adjacent times.

[0075] In this embodiment of the invention, the preliminary video frame sequence refers to the frame sequence determined by summarizing the core data and detailed data of electric shovel maintenance generated based on the size model.

[0076] In this embodiment of the invention, the model is fine-tuned based on feedback data from the maintenance scenario, and the semantic understanding weight and detail rendering accuracy are dynamically adjusted to improve the adaptability of the generated electric shovel maintenance video.

[0077] For example, a large model specifically for electric shovel repair, the Wan-14B fine-tuned version.

[0078] For example, a small model specifically for electric shovel repair, the Wan-1.3B fine-tuned version.

[0079] For example, high-noise steps: the first 10-15 steps; low-noise steps: the remaining steps.

[0080] For example, the second derivative of the relative L1 distance between denoised samples at adjacent time steps is calculated in real time. When 0 < the second derivative of the relative L1 distance < 0.01 and the large model has completed at least 5 inference steps, the model switching is automatically triggered, and seamless data transmission is achieved through a shared memory pool.

[0081] In an embodiment of the present invention, reference is made to Figure 4 The global cache optimization module includes: The Directed Acyclic Graph (DAG) construction unit is used to perform offline analysis on the preliminary video frame sequence and construct a DAG specifically for maintenance. The maintenance steps are used as nodes, the semantic relationship between steps is used as edge weights, and the classification attributes and error sensitivity of the core steps and auxiliary steps are labeled. The path optimization unit is used to obtain the optimal cache path by employing a lexicographical minimax algorithm combined with a fixed inference budget. The precision optimization unit is used to integrate an optimized attention technology to perform precision compression and acceleration on the collaborative processing of the electric shovel repair-specific model, and is configured with a dynamic overflow compensation mechanism to dynamically compensate for quantization errors. The cache management unit is used to monitor cache hit rate and error changes in real time, dynamically adjust the cache strategy, and automatically trigger cache data updates when the cache error of the auxiliary steps exceeds the error threshold.

[0082] In this embodiment of the invention, the maintenance-specific directed acyclic graph refers to a graph that uses maintenance steps as nodes, semantic relationships between steps as edge weights, and labels the classification attributes and error sensitivity of core steps and auxiliary steps, and is used to identify the core steps and auxiliary steps of electric shovel maintenance.

[0083] In this embodiment of the invention, the core steps refer to the steps that must exist in the electric shovel repair process, and the auxiliary steps refer to the steps that optimize the repair quality in the electric shovel repair process.

[0084] In this embodiment of the invention, error sensitivity refers to the sensitivity of the maintenance steps to errors. The higher the error sensitivity, the greater the impact of the decline in the quality of electric shovel maintenance when errors occur.

[0085] In this embodiment of the invention, the lexicographical minimax algorithm refers to an algorithm that minimizes the maximum cost and combines lexicographical comparison under the constraint of a fixed inference budget. It is used to optimize the selection of caching strategies and ensure that cache reuse efficiency is maximized while ensuring the quality of core steps.

[0086] In this embodiment of the invention, auxiliary steps are cached and reused according to the optimal caching path, and core steps are fully computed. Cache reuse refers to directly using stored historical results during the generation of electric shovel repair videos, while full computation refers to performing complete model inference calculations from scratch during the electric shovel repair process to generate new results.

[0087] In this embodiment of the invention, the fixed inference budget refers to the upper limit of computing resource consumption that is pre-set when generating a video of electric shovel repair.

[0088] In this embodiment of the invention, the caching strategy refers to the strategy of caching data during the generation of electric shovel repair videos.

[0089] In this embodiment of the invention, the optimal cache path refers to the optimal scheduling scheme selected by the lexicographical minimax algorithm to guide which steps in the video generation process can reuse the cache and which must be fully computed.

[0090] In this embodiment of the invention, by integrating and optimizing attention technology, the model collaboration process of the electric shovel repair-specific model is compressed and accelerated with improved accuracy. A dynamic overflow compensation mechanism is configured to dynamically compensate for quantization errors, thereby improving the visual realism of repair details while increasing inference speed.

[0091] In this embodiment of the invention, the optimized attention technology refers to a technology used to accelerate precision compression.

[0092] In this embodiment of the invention, the dynamic overflow compensation mechanism refers to a mechanism for real-time monitoring and dynamic compensation of numerical overflow problems that may occur during precision compression in the process of model collaborative processing.

[0093] In this embodiment of the invention, the cache hit rate and error changes are monitored in real time, and the caching strategy is dynamically adjusted. When the cache error of the auxiliary step exceeds the error threshold, the cache data is automatically updated to ensure the validity of the cache data.

[0094] In this embodiment of the invention, the cache hit rate refers to the proportion of data that can be directly obtained from the cache when generating electric shovel repair videos.

[0095] In this embodiment of the invention, error change refers to the degree of difference between cache reuse and full computation and its trend over time.

[0096] In this embodiment of the invention, the error threshold refers to a preset upper limit of error used to determine whether to trigger a cached data update.

[0097] For example, the core steps are: pivot installation and torque tightening; the auxiliary steps are: tool preparation and site arrangement.

[0098] For example, FP8 precision compression, where FP8 refers to precision compression of 8-bit floating-point format.

[0099] In this embodiment of the invention, the quality verification and output module includes: The full-process control unit is used to manage the entire process of generating electric shovel maintenance videos. The full-process control includes: process design, process compilation, process distribution, process operation, and process monitoring. The process design unit is used to design a visual configuration interface, providing options for maintenance scenario templates and video generation parameter settings. The compilation unit is used to automatically compile the user-selected configuration into an executable inference task for a model specifically designed for electric shovel maintenance, and to perform syntax checks. The distribution unit is used to dynamically allocate inference tasks based on hardware resource load and distribute them between central computing nodes and edge computing nodes according to deployment scenario requirements; The execution unit is used to provide an independent operating environment for multi-task parallel execution and resource isolation. The monitoring unit is used to collect real-time operational status and video generation quality, and to provide abnormal alarms and log recording; The quality verification unit is used to perform quality verification on the initial video frame sequence according to the whole process control, and to verify the consistency of component structure, the accuracy of process parameters and the compliance of safe operation. The output storage unit is used to encode and output the preliminary video frame sequence according to the preset video format after the verification is qualified, and store it in the distributed storage cluster.

[0100] In this embodiment of the invention, the operating status refers to the operating status of the electric shovel repair video generation system, such as CPU / GPU utilization, inference progress, and cache hit rate.

[0101] In this embodiment of the invention, the video generation quality is determined based on the consistency of component structure and the accuracy of process parameters.

[0102] In this embodiment of the invention, when safe operation is not performed, the quality verification fails directly.

[0103] In this embodiment of the invention, the user-selected configuration is automatically compiled into an executable reasoning task for a model specifically designed for electric shovel maintenance, and a syntax check is performed to ensure the rationality of the task logic.

[0104] In this embodiment of the invention, hardware resource load is dynamically allocated for inference tasks, and distributed between central computing nodes and edge computing nodes according to the deployment scenario requirements, in order to adapt to the needs of different deployment scenarios.

[0105] In this embodiment of the invention, the deployment scenario requirement refers to the scenario requirement of deploying electric shovel maintenance video distribution.

[0106] In this embodiment of the invention, an independent operating environment is provided to enable parallel operation of multiple tasks and resource isolation, thereby avoiding interference between different tasks.

[0107] In this embodiment of the invention, the preset video format refers to a preset format used for generating electric shovel maintenance videos, such as the H.265 format.

[0108] In this embodiment of the invention, users can preview, edit, and export the encoded output results through the application layer.

[0109] In this embodiment of the invention, log recording refers to the recording of task configuration, inference duration, and quality verification results. Log recording facilitates subsequent querying and system optimization.

[0110] The electric shovel maintenance video generation system based on large and small model collaboration and global caching optimization provided in the above embodiments of the present invention acquires structured electric shovel maintenance data and performs knowledge preprocessing to obtain standardized prompt words, which are then transmitted to large and small models. High-noise processing is applied to the standardized prompt words using the large electric shovel maintenance model to generate a semantic structure of the maintenance process. Adaptive detection of model switching conditions is performed; when the switching conditions are met, the small electric shovel maintenance model is switched to perform low-noise processing to generate a preliminary video frame sequence. The preliminary video frame sequence is analyzed to obtain the optimal caching path, and the model collaboration processing is subjected to precision compression and acceleration. The preliminary video frame sequence undergoes quality verification; after passing the verification, it is encoded, output, and stored. Through system-level architecture design and functional integration, the present invention effectively optimizes the professionalism, efficiency, adaptability, and stability of electric shovel maintenance video generation.

[0111] Example 2 refer to Figure 5 The electric shovel maintenance video generation system based on size model collaboration and global caching optimization provided in Embodiment 2 of the present invention includes: Maintenance knowledge modeling module: Stores component parameters, maintenance processes, tool lists, and safety specifications for the slewing bearing of WK-35 / WK-20 electric shovels, and supports automatic generation of structured prompts; Large and small model collaboration module: includes a large model (i.e., sketch model) and a small model (i.e., rendered model) specifically for electric shovel maintenance, sharing a 3D VAE latent space, and equipped with an adaptive switching unit (switching decision-maker based on the second derivative of noise variation). Global cache optimization module: includes a dedicated DAG (directed acyclic graph) construction unit for maintenance scenarios, a lexicographically ordered minimax path solver, and an FP8 precision optimization unit, to achieve global error control and efficiency improvement in the inference process; Professional quality verification module: It has a built-in maintenance specification verification rule library and uses visual recognition technology to verify the accuracy of component structure, process parameters and safe operation in the video.

[0112] As a specific example of an embodiment of the present invention, an electric shovel maintenance video generation system based on size model collaboration and global caching optimization includes: Maintenance knowledge modeling module: Stores parameters of 18 core components of electric shovel slewing bearing, 23 maintenance procedure specifications, and 34 types of special tool data. It supports automatic generation of structured prompts by inputting keywords such as "pivot replacement" and "gear ring repair". The large and small model collaboration module: the sketch model is fine-tuned based on Wan-14B to enhance the semantic understanding of industrial mechanical structures; the rendering model is fine-tuned based on Wan-1.3B to optimize the visual reproduction of metal materials and tool details; the δ threshold of the adaptive switching unit is set to 0.01 to adapt to the semantic stability of maintenance scenarios. Global cache optimization module: DAG building units allocate edge weights according to "maintenance process priority", the path optimizer uses dynamic programming algorithm to solve the optimal cache path, and the FP8 precision optimization unit is compatible with mainstream GPU hardware; Professional quality verification module: It has 12 built-in maintenance specification verification rules and uses target detection technology to identify the component model and tool type in the video to verify the accuracy of process parameter labeling.

[0113] Example 3 refer to Figure 6 and Figure 7 The electric shovel maintenance video generation method based on size model collaboration and global cache optimization provided in Embodiment 3 of the present invention integrates the size model collaboration strategy of SRDiffusion and the global cache optimization mechanism of LeMiCa, and specifically includes the following steps: Step 1, Maintenance Knowledge Modeling and Prompt Keyword Construction, specifically includes the following steps: Step 1.1: Extract the core elements for the maintenance of the slewing bearing device of the electric shovel, including component parameters (such as the diameter of the central pivot 508.33±0.03mm), process specifications (such as liquid nitrogen cooling to -184℃), tool list (such as 100-ton jack, torque wrench, etc.), and safety specifications (such as high-altitude operation protection, hot work operation procedures, etc.). Step 1.2: Construct structured prompts in the format "WK-35 electric shovel slewing bearing device maintenance video, scenario: 60×80m compaction site (load capacity ≥36.5 tons / m)". 2 Specifically, it includes the following sub-steps: Step 1.2.1: Cool the central pivot to -184°C with liquid nitrogen (reduction of 1.14 mm). Step 1.2.2: Install into the traveling trolley hole and use a 100-ton jack for positioning assistance; Steps 1, 2, and 3: Tighten the gear ring bolts to a torque of 1978.5 N·m. Professional requirements: Disassemble after the hydraulic system valve is closed; clearance adjustment accuracy ≤ 0.254 mm; operators must wear liquid nitrogen protective gloves.

[0114] Step 2, video sketching and rendering generation for collaboration between large and small models, specifically includes the following steps: Step 2.1: Initialize the latent variables of the diffusion model. Use the large model specifically for electric shovel maintenance (Wan-14B fine-tuned version) as the sketch model. The generation process of the sketch model (Wan-14B fine-tuned version) is as follows: Figure 6 As shown. The high noise level during the diffusion process is addressed by focusing on generating the semantic structure of the maintenance workflow (such as component spatial location, disassembly / assembly sequence, and tool usage scenarios). Step 2.2, determine the model switching timing based on the adaptive switching mechanism: calculate the second derivative of the relative L1 distance between denoised samples in adjacent time steps. When satisfied (δ=0.01, adapted for maintenance scenarios) After completing at least 5 sketch generation steps, switch to the small model (Wan-1.3B fine-tuned version) as the rendering model. The entire process from generating the rendering model (Wan-1.3B fine-tuned version) to switching from the sketch model to the rendering model includes the following sub-steps: Step 2.2.1: Generate the rendering model (Wan-1.3B fine-tuning version), from the basic model to the specialized model; Step 2.2.2: Switch the decision node to connect the sketch model and the rendered model; Step 2.2.3, rendering process after model switching.

[0115] The specific implementation process of generating the rendering model in step 2.2.1 is shown in the table below: ; The specific implementation process of switching the decision node and connecting the sketch model and the rendered model as described in step 2.2.2 is shown in the table below: ; The specific implementation process of the rendering workflow after model switching in step 2.2.3 is shown in the table below:

[0116] Step 3, global cache optimization for inference acceleration, specifically includes the following steps: Step 3.1: Construct a dedicated directed acyclic graph (DAG) for maintenance video generation, using "maintenance steps" as nodes and "semantic association between steps" as edge weights, and calculate the global error impact of each step offline (e.g., the weight of the core component installation step is higher than that of the auxiliary tool placement step). Step 3.1, which involves constructing a dedicated DAG for maintenance video generation, specifically includes the following steps: Step 3.1.1, breakdown of maintenance steps and definition of nodes, including: 1. Extract the key steps of the entire maintenance process of the electric shovel's slewing bearing device, and break them down into 12 core nodes according to time sequence and logical association, including: (1) Auxiliary steps (S0-S2): S0 (Tool preparation: 100-ton jack, torque wrench LDA-09S in place), S1 (Site layout: 60×80m site compaction, power supply interface installation), S2 (Spare parts inventory: central pivot K1801.12.01, anti-loosening shim NL16 checked); (2) Core process steps (S3-S10): S3 (pivot cooling: liquid nitrogen cooling to -184℃), S4 (pivot hoisting: crane positioning and installation), S5 (gap adjustment: measuring the standard gap of 0.254mm), S6 (bolt tightening: applying a torque of 1978.5N·m), S7 (gear ring installation: fixing the lower ring rail), S8 (roller assembly: free rotation test after lubrication), S9 (hydraulic system reset: valve opening and pressure adjustment), S10 (trial run test: verification of no jamming in the rotary system); (3) Finishing step (S11): S11 (site cleanup: tools put away, waste disposal).

[0117] 2. Node attribute labeling: Add attribute labels to each node, including "Step type" (auxiliary / core), "Professional priority" (core steps are set to priority level 1-5, such as S6 bolt tightening is set to level 5, and S0 tool preparation is set to level 1), and "Error sensitivity" (core steps have an error sensitivity ≥ 0.8, and auxiliary steps have an error sensitivity ≤ 0.3).

[0118] Step 3.1.2, semantic correlation analysis between steps and edge weight calculation, includes: 1. Construct a semantic association matrix: Based on the maintenance process logic, define an association strength scoring standard (0-1 points), where "direct dependency relationship" (e.g., S3→S4: the pivot can only be hoisted after cooling) scores 0.9-1.0, "indirect association relationship" (e.g., S1→S3: site layout does not directly affect cooling, but ensures the operating environment) scores 0.3-0.5, and "no association relationship" (e.g., S0→S11: tool preparation and site cleanup have no direct logic) scores 0.0-0.2; 2. Calculate the edge weight value: The weighted formula of "semantic association strength × core step priority weight" is adopted, that is, edge weight W (i→j) = semantic association strength S(i,j) × [(node ​​j professional priority + node j error sensitivity) / 2]; Example: S3 (core step, priority 5, error sensitivity 0.9) → S4 (core step, priority 5, error sensitivity 0.9), semantic association strength 1.0, edge weight W=1.0×(5+0.9) / 2=2.95; Example: S0 (auxiliary step, priority 1, error sensitivity 0.2) → S1 (auxiliary step, priority 1, error sensitivity 0.2), semantic association strength 0.4, edge weight W=0.4×(1+0.2) / 2=0.24.

[0119] Step 3.1.3, global errors affect offline calculations, including: 1. Sample data collection: 70 sets of electric shovel repair scenario prompts (covering core scenarios such as pivot replacement and gear ring repair), 5 different seed repair videos were generated for each set, and 350 complete generated trajectory data were obtained; 2. Caching Error Quantization: For each step node i, simulate two schemes: "caching this step → reusing intermediate results" and "not caching → full computation". Calculate the error value using LPIPS (Perceptual Similarity Indicators) and SSIM (Structural Similarity Indicators). The global error impact E(i) = 1 - [SSIM × 0.6 + LPIPS × 0.4] (the weights are set based on the visual fidelity requirements of the maintenance scenario). 3. Edge error labeling: Label the global error impact E(i) of node i to the corresponding incoming edge (all edges pointing to node i). Label the incoming edges of core step nodes with "error upper limit ≤ 0.05" and the incoming edges of auxiliary step nodes with "error upper limit ≤ 0.2".

[0120] Step 3.1.4, DAG topology generation, including: 1. Arrange nodes according to "time order + logical dependency", ensuring no circular dependencies (e.g., only S(i)→S(j) are allowed, where i <j); 2. Draw the DAG: Use circular nodes to represent steps (core step nodes are 20mm in diameter and filled with red; auxiliary step nodes are 12mm in diameter and filled with gray), and solid lines to represent edges (line width is positively correlated with edge weight; line width is 2mm when W≥2.0 and 1mm when W<1.0). Label the edge weight values ​​and error limits to form a DAG specifically for maintenance scenarios.

[0121] Step 3.2 adopts the lexicographical minimax path optimization strategy to select the optimal caching path under a fixed inference budget, prioritizing the caching of semantically stable steps (such as tool preparation and site layout) to avoid the accumulation of caching errors in critical process steps (such as pivot installation and torque tightening). Step 3.2, which employs the lexicographically minimax path optimization strategy, specifically includes the following steps: Step 3.2.1, setting the inference budget, including determining the inference budget B (number of forward computations of the model) based on the target speedup: 1. If the target speedup ratio is 2.9× (adapting to the speedup level of the Latte model), and the total number of diffusion steps generated by combining the maintenance video is T=50, set B=14 (that is, allow 14 full calculations, and reuse the remaining steps through caching). 2. Budget allocation rules: Core step nodes shall be allocated at least one full calculation, while auxiliary step nodes may reduce the number of calculations through caching.

[0122] Step 3.2.2, candidate path generation, includes: 1. Based on the DAG topology, traverse all paths from S0 (starting point) to S11 (ending point) and filter candidate paths that meet the following criteria: "the number of nodes in the path is ≥12", "the core steps are fully covered by nodes", and "the total number of calculations is ≤B". 2. Path filtering criteria: (1) It must include all core process steps (S3-S10) and cannot be skipped; (2) The number of full calculations = the number of "uncached nodes" in the path, which must be less than or equal to the inference budget B; (3) Total path error = the maximum error value of all edges in the path, and initially screen paths with a total error ≤ 0.2.

[0123] Step 3.2.3, lexicographically minimata path selection, includes: 1. Sort candidate paths in descending order of error sequence: Extract all edge error values ​​of each path and sort them in descending order to form an error sequence (e.g., the error sequence of path P1 is [0.18, 0.09, 0.05, ...]). 2. Lexicographical comparison: (1) Prioritize comparing the first element of the error sequence (maximum error) and select the path with the smallest maximum error; (2) If the maximum error is the same, compare the second element (the second largest error), and so on, until the optimal path is selected; 3. Caching strategy determination: In the optimal path, the edges corresponding to the auxiliary step nodes (S0-S2, S11) are marked as "cached", and the edges corresponding to the core process step nodes (S3-S10) are marked as "non-cached". That is, the intermediate results of the auxiliary steps with stable semantics are cached, and the core steps are fully calculated to avoid error accumulation.

[0124] Step 3.2.4, cache scheduling execution, includes: 1. During inference, execute along the optimal path: When the auxiliary step node is reached, reuse the cached intermediate features (such as visual features prepared by the S0 tool, which can be directly called in subsequent scenarios). 2. Cache update mechanism: After every 10 sets of maintenance videos are generated, the actual value of the path error is calculated. If the cache error of a certain auxiliary step node exceeds 0.2, it is removed from the cache list and replaced with a full calculation to ensure that the error is controllable.

[0125] Step 3.3: Integrate SageAttention with FP8 precision optimization to further improve inference speed while ensuring visual fidelity of repair details.

[0126] Step 3.3, the FP8 accuracy optimization using SageAttention, specifically includes the following steps: Step 3.3.1, FP8 precision adaptation configuration, including: 1. Model layer precision setting: The attention layer of the rendered model (Wan-1.3B fine-tuning version) is calculated with FP8 precision, while other layers (such as fully connected layers and normalization layers) are kept with bfloat16 precision to balance speed and accuracy; 2. Quantization parameter configuration: The “dynamic range quantization” strategy of SageAttention is adopted. Based on the feature distribution generated by the maintenance video, the exponent of FP8 is set to 5 bits and the mantissa to 2 bits, covering the numerical range of maintenance scene features (-10.0~10.0). 3. Overflow handling: Set a dynamic overflow threshold. When the feature value exceeds the FP8 representation range, automatically switch to bfloat16 precision calculation to avoid numerical distortion.

[0127] Step 3.3.2, Attention computation acceleration, includes: 1. Based on the SageAttention block computing mechanism, the latent features of the repair video are divided into blocks according to "time step × spatial dimension" (block size is set to 32×32), and attention calculation is performed in parallel. 2. Cache intermediate attention results: Cache the attention output of semantically stable auxiliary steps (such as S1 site layout) so that they can be reused directly in subsequent inferences of the same scene, reducing redundant calculations; 3. Precision Compensation Mechanism: After the attention calculation of core steps (such as S6 bolt tightening), a "precision calibration layer" is added. The residual connection corrects the small errors introduced by FP8 quantization, ensuring the visual fidelity of maintenance details (such as torque scale and bolt position).

[0128] Step 3.3.3, verification of the balance between inference speed and quality, includes: 1. Real-time monitoring metrics: Real-time monitoring of "Inference Time", "LPIPS Value", and "SSIM Value" during the inference process, with threshold settings: (1) If LPIPS>0.2 (perceived quality deterioration), automatically reduce the FP8 quantization ratio (e.g., only use FP8 for some attention heads); (2) If the reasoning time does not reach the target speedup ratio, increase the block size appropriately (maximum set to 64×64). 2. Dynamic adjustment: Based on the validation results, the FP8 precision configuration and block parameters are adaptively adjusted to ensure that while improving the inference speed (target: speeding up by more than 1.5 times compared to the base model), the visual fidelity of the repair details meets the requirements (SSIM≥0.75, LPIPS≤0.18).

[0129] The entire video generation process in this embodiment of the invention is as follows: Figure 7 As shown.

[0130] Step 4, Professional Quality Verification and Output, specifically includes the following steps: Step 4.1: Construct quality verification indicators based on electric shovel maintenance specifications, including component structural consistency (such as the fit accuracy between the central pivot and the trolley hole), process parameter accuracy (such as cooling temperature and bolt torque), and safety specification compliance (such as the wearing of protective equipment). Step 4.2: Perform inter-frame spatiotemporal consistency verification on the generated video, correct problems such as component drift and discontinuous movements, and finally output a maintenance video that meets industry standards.

[0131] As a specific example of an embodiment of the present invention, the electric shovel maintenance video generation method based on size model collaboration and global caching optimization is applied to the central pivot replacement process of the slewing bearing device of the WK-35 electric shovel. The specific implementation process is as follows: 1. Maintenance knowledge modeling and prompt word construction: Extract the core parameters of the central pivot replacement (diameter 508.33±0.03mm, cooling temperature -184℃, tightening torque 3957N·m), tool list (100-ton jack, liquid nitrogen container, torque wrench), and safety specifications (liquid nitrogen gloves, high-altitude work protection) to generate structured prompt words; 2. Collaborative generation of large and small models: 2.1 Initialize the latent variable, use the Wan-14B fine-tuned version as the sketch model, process the first 12 high-noise steps, and generate the semantic structure of site layout, tool placement, and pivot cooling. 2.2 Calculation results Once the switching conditions are met, switch to the Wan-1.3B fine-tuned rendering model, process the remaining low-noise steps, and optimize details such as pivot installation positioning, bolt tightening action, and gap measurement scale. 3. Global caching optimization: 3.3 When constructing the DAG, set "pivot cooling", "positioning and installation", and "torque tightening" as high-weight nodes and do not cache them; cache semantically stable steps such as "tool preparation" and "site clearing", and achieve a 2.9x caching speedup through path optimization; 3.4 Enable FP8 precision optimization to further improve inference efficiency; 4. Quality Verification and Output: Verify the visual accuracy of the fit between the central pivot and the trolley hole, the accuracy of the cooling temperature label, and the correctness of the torque wrench torque display. Correct the inter-frame pivot position drift problem and output a 15-minute training video on central pivot replacement.

[0132] The electric shovel maintenance video generation method based on large-scale model collaboration and global caching optimization provided in the above embodiments of the present invention addresses the problems of high production cost, difficulty in restoring professional details, and low generation efficiency of existing electric shovel slewing bearing device maintenance videos. It integrates large-scale model collaboration and global caching optimization mechanisms. First, it extracts the core elements of electric shovel maintenance (component parameters, process specifications, tool list, etc.) and constructs structured prompt words. Then, it uses a dedicated large model to process the high-noise steps of the diffusion process and generates the semantic structure of the maintenance process (sketching stage). Through an adaptive switching mechanism (based on the second derivative threshold of the denoised samples), it switches to a small model to optimize professional details (rendering stage). It constructs a dedicated directed acyclic graph (DAG) for the maintenance scene, uses a lexicographically ordered minimax path optimization strategy to achieve global caching acceleration, and integrates accuracy optimization to further improve inference efficiency. Finally, it verifies the component consistency, process accuracy, and safety compliance of the video based on maintenance specifications. This invention achieves precise visualization of maintenance videos for electric shovels such as WK-35 / WK-20, with a generation efficiency more than 6 times higher than traditional methods. Professional details (such as 0.254mm gap adjustment and 1978.5N·m bolt torque) are accurately reproduced, making it suitable for maintenance training, remote guidance and other scenarios, and significantly reducing the cost of actual shooting and equipment downtime losses.

[0133] Compared with the prior art, the present invention has the following technical advantages and beneficial effects: 1. Precise reproduction of professional details: Through maintenance knowledge modeling and the construction of special prompt words, the visualization accuracy of core processes of electric shovel maintenance (such as liquid nitrogen cooling, gap adjustment, and torque control) is ensured, and the component structure and operation process comply with the maintenance specifications of WK-35 / WK-20 electric shovels; 2. Significantly improved generation efficiency: Combining large and small model collaboration (3.2x speedup) and global caching optimization (2.9x speedup), the overall reasoning speed is more than 6 times faster than traditional text-based video technology, and the generation time of a 5-minute repair video is reduced from 1 hour to less than 10 minutes; 3. Excellent spatiotemporal consistency: The sketch model ensures semantic coherence of the maintenance process, the rendering model optimizes details to maintain fidelity, and the global caching strategy suppresses error accumulation, avoiding problems such as component drift and inconsistent actions; 4. High practicality: It can flexibly generate videos of different maintenance scenarios (such as component replacement and fault diagnosis), support close-ups of key steps and annotation of process parameters, and is suitable for practical application scenarios such as maintenance training and remote guidance, reducing the cost of actual shooting and equipment downtime losses.

Claims

1. A size model-based collaborative and global cache-optimized shovel repair video generation system, the shovel repair video generation system is designed with a distributed architecture, including a hardware layer, a software layer and an application layer, and each layer collaborates to realize the whole process management of the shovel repair video generation, characterized in that, The electric shovel maintenance video generation system includes: The task configuration and knowledge loading module is used to obtain task configurations based on user selections, retrieve structured electric shovel maintenance data from the structured electric shovel maintenance knowledge base and perform knowledge preprocessing, obtain standardized prompt words, and transmit them to the size model; The model collaboration generation module is used to process standardized prompt words with high noise based on the large model dedicated to electric shovel maintenance, generate the semantic structure of the maintenance process, and adaptively detect model switching conditions. When the model switching conditions are met, the module switches to the small model dedicated to electric shovel maintenance for low noise processing to generate a preliminary video frame sequence. The global cache optimization module is used to analyze the initial video frame sequence, obtain the optimal cache path, and perform precision compression and acceleration for model collaborative processing. The quality verification and output module is used to verify the quality of the initial video frame sequence. After the verification is successful, the frame is encoded, output, and stored.

2. The size model based collaboration and global cache optimized video generation system for shovel repair of claim 1, wherein, The hardware layer includes: The computing node unit is used to deploy computing servers at the central computing node to run collaborative computing tasks for electric shovel maintenance using a specific model of different sizes. Deploy computing devices on edge computing nodes to run local inference and cache scheduling; The storage node unit is used to store the electric shovel maintenance knowledge base, the model weight file of the electric shovel maintenance-specific size model, the generated video data, and the operation logs using a distributed storage cluster. The interface device unit is used to acquire input data of the electric shovel repair scene and video preview and playback data according to the preset image acquisition device; Data is processed for input / output and remote interaction using remote communication devices.

3. The electric shovel maintenance video generation system based on size model collaboration and global caching optimization according to claim 1, characterized in that, The application layer includes: The scene configuration unit provides a graphical user interface, allowing users to select maintenance scenes and configure generation parameters to obtain video generation tasks. The video preview and editing unit is used to preview the video in real time during the video generation process and provides simple editing functions, including key step cropping and parameter standard adjustment. The data management unit is used for the classification, storage, retrieval, export, sharing, data backup, and data recovery of electric shovel maintenance videos. The operation management unit is used for user permission management, role configuration, log querying, and operation upgrades.

4. The electric shovel maintenance video generation system based on size model collaboration and global caching optimization according to claim 1, characterized in that, The task configuration and knowledge loading module includes: The knowledge base unit is used to build a structured electric shovel maintenance knowledge base, which allows for adding, deleting, modifying, querying, and version management of electric shovel maintenance knowledge. The structured electric shovel maintenance knowledge base includes: a component parameter library, a process specification library, and a tool and spare parts library. The knowledge preprocessing unit integrates natural language processing algorithms and feature extraction models to preprocess the structured electric shovel maintenance data in the structured electric shovel maintenance knowledge base, converting it into standardized prompt words that can be recognized by the electric shovel maintenance-specific size model, and automatically supplementing process logic association information and safety specification descriptions. The knowledge update unit is used to incrementally update the structured electric shovel maintenance knowledge base according to the preset knowledge update method, and to simultaneously optimize the model adaptability of the feature extraction model and the electric shovel maintenance-specific size model through the incremental training mechanism. The preset knowledge update methods include: on-site data collection and maintenance record import. The task configuration unit is used to obtain task configuration based on user selection, where user selection includes: maintenance scenario selection and configuration video generation parameter selection; The knowledge loading unit is used to retrieve structured electric shovel maintenance data from the structured electric shovel maintenance knowledge base according to the task configuration, perform knowledge preprocessing, obtain standardized prompt words, and transmit them to the electric shovel maintenance-specific size model.

5. The electric shovel maintenance video generation system based on size model collaboration and global caching optimization according to claim 1, characterized in that, The model collaboration generation module includes: A dedicated large model unit is used to deploy a dedicated large model for electric shovel maintenance. It optimizes and trains high-noise steps, strengthens the semantic structure of the generated maintenance process, and outputs core data for electric shovel maintenance. The core data for electric shovel maintenance includes: component spatial location data, disassembly and assembly sequence data, and tool placement logic data. A dedicated small model unit is used to deploy a small model specifically for electric shovel maintenance, to render low-noise steps in detail, to integrate a maintenance tool feature library and parameter visualization components, and to optimize the electric shovel maintenance detail data, which includes: bolt tightening action data and torque scale display data. The adaptive switching unit is used to build a switching determination algorithm to calculate the relative distance data of denoised samples in adjacent time steps in real time. When the model switching conditions are met, the model switching is automatically triggered and the data is seamlessly transmitted through the shared memory pool to switch to the electric shovel maintenance special small model. The collaborative generation unit is used to summarize the core data and detailed data of electric shovel maintenance and generate a preliminary video frame sequence. The model optimization unit is used to fine-tune the model based on feedback data from maintenance scenarios, dynamically adjusting semantic understanding weights and detail rendering accuracy.

6. The electric shovel maintenance video generation system based on size model collaboration and global caching optimization according to claim 1, characterized in that, The global cache optimization module includes: The Directed Acyclic Graph (DAG) construction unit is used to perform offline analysis on the preliminary video frame sequence and construct a DAG specifically for maintenance. The maintenance steps are used as nodes, the semantic relationship between steps is used as edge weights, and the classification attributes and error sensitivity of the core steps and auxiliary steps are labeled. The path optimization unit is used to obtain the optimal cache path by employing a lexicographical minimax algorithm combined with a fixed inference budget. The precision optimization unit is used to integrate an optimized attention technology to perform precision compression and acceleration on the collaborative processing of the electric shovel repair-specific model, and is configured with a dynamic overflow compensation mechanism to dynamically compensate for quantization errors. The cache management unit is used to monitor cache hit rate and error changes in real time, dynamically adjust the cache strategy, and automatically trigger cache data updates when the cache error of the auxiliary steps exceeds the error threshold.

7. The electric shovel maintenance video generation system based on size model collaboration and global caching optimization according to claim 1, characterized in that, The quality verification and output module includes: The full-process control unit is used to manage the entire process of generating electric shovel maintenance videos. The full-process control includes: process design, process compilation, process distribution, process operation, and process monitoring. The process design unit is used to design a visual configuration interface, providing options for maintenance scenario templates and video generation parameter settings; The compilation unit is used to automatically compile the user-selected configuration into an executable inference task for a model specifically designed for electric shovel maintenance, and to perform syntax checks. The distribution unit is used to dynamically allocate inference tasks based on hardware resource load and distribute them between central computing nodes and edge computing nodes according to deployment scenario requirements; The execution unit is used to provide an independent operating environment for multi-task parallel execution and resource isolation. The monitoring unit is used to collect real-time operational status and video generation quality, and to provide abnormal alarms and log recording; The quality verification unit is used to perform quality verification on the initial video frame sequence according to the whole process control, and to verify the consistency of component structure, the accuracy of process parameters and the compliance of safe operation. The output storage unit is used to encode and output the preliminary video frame sequence according to the preset video format after the verification is qualified, and store it in the distributed storage cluster.

8. A video generation system for electric shovel maintenance based on size model collaboration and global caching optimization, characterized in that, include: Maintenance knowledge modeling module: Stores component parameters, process specifications, tool lists, and safety regulations related to electric shovel maintenance, and supports automatic generation of structured prompts; Large and small model collaboration module: includes a large model and a small model for electric shovel maintenance, equipped with an adaptive switching unit based on the second derivative of noise variation, where the large model is a sketch model and the small model is a rendered model; Global cache optimization module: includes a DAG construction unit for maintenance scenarios, a lexicographically ordered minimax path solver, and an FP8 precision optimization unit; Professional quality verification module: Built-in maintenance specification verification rule library to verify the accuracy of component structure, process parameters and safe operation in the video.

9. A method for generating electric shovel maintenance videos based on size model collaboration and global caching optimization, characterized in that, Includes the following steps: Step 1: Maintenance knowledge modeling and prompt word construction: Extract component parameters, maintenance processes, tool lists, and safety specifications of the electric shovel's slewing bearing device to generate structured prompt words; Step 2, Video sketch-rendering generation through collaboration between large and small models: A dedicated large model is used to process high-noise steps to generate semantic structure, and an adaptive switching mechanism is used to switch to a small model to optimize details; Step 3: Global cache optimization for inference acceleration: Construct a DAG specifically for maintenance scenarios, adopt a lexicographical minimax path optimization strategy to select the optimal cache path, and combine precision optimization to improve inference speed; Step 4: Professional quality verification and output: Based on the maintenance specifications, verify the consistency of components, the accuracy of processes, and the compliance with safety requirements of the video, and output qualified maintenance videos.

10. The method for generating electric shovel maintenance videos based on size model collaboration and global caching optimization according to claim 9, characterized in that, The adaptive switching mechanism in step 2 is based on the second derivative of the relative L1 distance between denoised samples in adjacent time steps. ,satisfy ,in The switch is triggered when at least 5 steps of sketch generation are completed.

Images