An offline multi-modal AI integrated deployment system and method for a petroleum inspection handheld terminal

Abstract

The application discloses a kind of petroleum inspection handheld terminal offline multi-modal AI integrated deployment system and method, it is related to oilfield digital inspection field.Solve the defects that existing inspection system is highly dependent on cloud, offline function is incomplete, data is not encrypted protection, weak network synchronization efficiency is low, offline model cannot be iterated, terminal computing power scheduling is rigid, etc., the system includes terminal hardware adaptation layer, multi-modal data acquisition layer, edge AI inference engine layer, local business function layer, encrypted storage management layer, offline synchronization and operation and maintenance layer, realize full-link business sinking and local offline operation of explosion-proof handheld terminal;Combined with three-level model light weight, SM4 national secret full-amount encryption storage, fragmented hash breakpoint resume, offline differential incremental model update and self-adaptive computing power scheduling mechanism, complete multi-modal acquisition, local AI defect identification, risk alarm, encrypted storage, offline report generation whole-process inspection business in off-line environment, complete breakpoint resume synchronization and model incremental iteration update after networking.

Description

Technical Field

[0001] This invention relates to the field of digital inspection technology in oilfields, specifically to an offline multimodal AI integrated deployment system and method for oilfield handheld inspection terminals. Background Technology

[0002] Currently, intelligent inspection systems in oilfields are divided into two main types: cloud-based centralized systems and ordinary offline PDA terminals. Both of these have inherent drawbacks, making them difficult to adapt to remote areas without network access.

[0003] The first category is the traditional cloud-based centralized inspection system, which generally adopts an architecture of "terminal data collection + cloud inference + cloud business." Representative solutions include fixed multimodal inspection systems for oilfield gathering and transportation stations and cloud-based inspection and control platforms. This type of solution relies on the sufficient computing power of servers to achieve high-precision AI recognition and complex business logic processing, but the core functions all depend on cloud servers, and handheld terminals only perform basic functions such as data collection and uploading. For some remote well stations and long-distance pipeline sections in Daqing Oilfield, where there are blind spots in both the public network and the oilfield's private network, the terminals cannot carry out intelligent inspections when the network is completely interrupted, and can only perform manual paper records, resulting in low work efficiency and a high risk of missed or incorrect inspections.

[0004] The second category is ordinary offline handheld inspection terminals, mostly industrial-grade PDA devices. They only have basic functions such as taking pictures, positioning, and storing images locally in plaintext. They lack the ability to perform edge AI inference, local anomaly warning, and automatic report generation. They cannot complete the complete inspection loop of "task execution - defect identification - anomaly handling - record archiving" in offline mode. At the same time, the data is mostly stored in plaintext, lacking a compliant encryption mechanism. Data synchronization relies on full transmission, resulting in low synchronization efficiency and poor accuracy in weak network environments. Model updates require a full download via the internet, making them unsuitable for long-term offline scenarios.

[0005] Based on field measurements conducted at Daqing Oilfield, the core flaws of the existing solution are as follows: 1. High dependence on cloud computing; operation is completely interrupted in the absence of network access. The core functions of the existing cloud-based inspection system, such as AI defect identification, anomaly level determination, and inspection report generation, are all deployed on cloud servers. Actual testing shows that the integrity of the intelligent inspection function is less than 15% in environments without network connectivity, the interruption rate of oilfield inspection operations exceeds 80%, and intelligent inspection in remote areas is essentially impossible to implement.

[0006] 2. Fragmented edge terminal functions prevent the formation of a complete business loop. Ordinary offline handheld terminals have limited functions, with most only having the ability to collect data, perform basic positioning, and store offline photos. They lack the ability to perform edge AI inference, local anomaly warning, and automatic report generation. The quality of inspections is highly dependent on the experience of on-site personnel, and the degree of standardization is insufficient, resulting in a defect omission rate that is more than 40% higher than that of cloud-based solutions.

[0007] 3. Offline data storage lacks security protection and does not meet industry compliance requirements. Inspection data from remote oilfields contains sensitive information such as equipment parameters, pipeline distribution, and production operations. Existing terminals mostly use plaintext storage mode and lack encryption mechanisms that comply with national cryptographic standards. There is a serious risk of information leakage when terminals are lost or data is exported without authorization, which cannot meet the data security compliance management requirements of the petroleum industry.

[0008] 4. Offline data synchronization is inefficient and lacks the ability to resume interrupted data transfers. Current terminal data synchronization mostly adopts a full-transmission mode, requiring all offline data to be uploaded at once after network recovery. In the weak network environment of oilfields, problems such as transmission interruption, duplicate transmission, and data packet loss are prone to occur. Actual testing shows that the synchronization accuracy is less than 90%. Furthermore, there is no transmission progress memory mechanism, requiring a full retransmission after an interruption, resulting in low synchronization efficiency and consuming a significant amount of operation time.

[0009] 5. Model updates rely on full online downloads, making offline scene iterations difficult. The existing inspection AI model requires a full download of the model files when updated online. Remote oilfield operation terminals are often offline and cannot obtain timely algorithm iteration updates, so the model recognition accuracy cannot be continuously optimized. Furthermore, it does not support offline incremental updates on local media, resulting in high algorithm deployment and maintenance costs and long cycles.

[0010] In summary, existing solutions cannot simultaneously achieve the five core capabilities of offline full-process AI inspection, full encryption, fragmented breakpoint resume, offline incremental update, and adaptive multi-model scheduling. Pain points such as network dependence, data insecurity, inference latency, and high operation and maintenance costs restrict the implementation of intelligent inspection across the entire oilfield. Summary of the Invention

[0011] This invention overcomes the shortcomings of existing inspection systems, such as heavy reliance on the cloud, incomplete offline functionality, lack of data encryption protection, low synchronization efficiency in weak networks, inability to iterate offline models, and rigid terminal computing power scheduling. It provides an integrated deployment architecture and system for multimodal AI algorithms on a handheld oil inspection terminal that supports full-function operation without network connectivity. To solve the above technical problems, this invention achieves its goals through the following technical solutions: Option 1: This invention proposes an offline multimodal AI integrated deployment system for oilfield inspection handheld terminals, including an explosion-proof handheld terminal hardware carrier. The explosion-proof handheld terminal hardware carrier integrates a multimodal acquisition unit, an edge computing unit, a local solid-state storage unit, a multi-standard communication unit, a BeiDou positioning unit, and an explosion-proof power supply unit. The system deploys a six-layer local integrated operation architecture, which, from bottom to top, consists of: The terminal hardware adaptation layer is used to uniformly manage various hardware peripherals of the terminal, providing multimodal data acquisition, computing power scheduling, local storage read and write, explosion-proof and low power consumption management capabilities. It is adapted to flammable and explosive, low temperature and no network operation conditions in oil fields. It automatically adjusts the model running priority or performs function degradation according to the ambient temperature, power and computing power load status. The multimodal data acquisition layer is used to simultaneously acquire visible light images, infrared thermal imaging data, equipment temperature, pressure and vibration sensing data, RFID equipment electronic tags, and Beidou positioning coordinates from multiple heterogeneous sources. It completes data standardization preprocessing and time-series alignment, and outputs multimodal inspection data in a unified format. The edge AI inference engine layer has a built-in lightweight multimodal inspection AI model pool and dynamic computing power scheduler. It performs multi-task AI inference, such as pipeline defect identification, equipment corrosion detection, crude oil leakage judgment, and personnel violation verification, in parallel on the terminal and outputs standardized identification results including defect type, location coordinates, confidence level, and risk level. The local business function layer deploys the business logic of the entire offline inspection process, realizing inspection task caching and scheduling, AI abnormal sound and light alarm, on-site handling record retention, batch generation and local export of offline inspection reports, and independently realizing all core business functions of the cloud inspection platform in offline state. The encrypted storage management layer automatically encrypts and stores all raw inspection data, AI recognition results, business records, and report files based on the SM4 block cipher algorithm. It is equipped with hardware security unit hierarchical key management, user hierarchical authentication, and storage read / write anomaly protection mechanism. The offline synchronization and operation and maintenance layer integrates a sharded hash breakpoint resume synchronization module and an offline incremental model update module; when the network is connected, it automatically performs sharded data breakpoint resume and full data verification in the cloud; it supports importing model differential patches through local storage media to complete seamless iterative updates of the terminal's local AI model.

[0012] Furthermore, a preferred implementation is provided, wherein the edge AI inference engine layer is configured with a three-level model lightweight optimization strategy and an adaptive computing power scheduling mechanism; The three-level model lightweight optimization strategy includes structured pruning, quantization, and knowledge distillation; the adaptive computing power scheduling mechanism dynamically adjusts the CPU usage threshold based on the ambient temperature and the remaining battery power of the terminal, and allocates computing power resources according to task priority.

[0013] Furthermore, a preferred embodiment is provided, wherein the encrypted storage management layer adopts the SM4-ECB128-bit block encryption mode; the encryption key is divided into a device root key and a user working key. The device root key is stored in the hardware security unit and cannot be exported externally, and can only be updated periodically by the administrator through authorized keys; the user working key is derived from the device root key, and hierarchical access permissions are divided according to the roles of inspectors and administrators; the encryption scope covers all inspection business data, and when storage medium damage or key verification failure is detected, the data to be stored is automatically written to the backup storage partition and the encrypted data is locked.

[0014] Furthermore, a preferred implementation is provided, wherein the fragmented hash breakpoint resume synchronization module divides the encrypted inspection data to be synchronized into fragments of a configured fixed size, generates an MD5 hash checksum for each fragment, and generates a fragment index list; during the synchronization process, a synchronization progress snapshot is persisted in real time, and after network interruption recovery, the progress snapshot is read and the transmission resumes from the currently untransmitted fragment; if a single fragment fails hash verification three times consecutively, it is split into sub-fragments for retry, and if multiple retry failures occur, it is marked as delayed synchronization without blocking the overall data upload; after the cloud receives all fragments, it performs a full hash integrity check, and if the check fails, it sends an abnormal fragment list to trigger terminal retransmission.

[0015] Furthermore, a preferred implementation is provided, wherein the differential patch of the offline incremental model update module is generated by comparing the differences in parameters between the old and new AI models in the cloud, and the patch size is 10% to 30% of the full model file; the patch verification process includes three levels of verification: file integrity MD5 verification, basic model version compatibility verification, and publisher digital signature verification; the currently running model is automatically backed up before the model update, and the terminal battery level and running status are monitored during the update process. If the update is interrupted due to low battery or unexpected power failure, the model is automatically rolled back to the stable model before the update; the terminal locally caches up to 3 historical model versions.

[0016] Furthermore, a preferred embodiment is provided, wherein the method for automatically handling task conflicts in the local business function layer configuration is as follows: Multiple overlapping inspection tasks are automatically sorted according to task priority and deadline, and inspection paths are merged when points overlap; when an emergency inspection task is inserted, the progress of the current low-priority task is automatically saved, and the original inspection process is restored after the emergency task is completed; the report module supports batch aggregation of inspection data by well station, pipeline, and administrative region, and uses streaming computing to generate offline inspection reports in PDF and Excel formats.

[0017] Furthermore, a preferred embodiment is provided in which the multimodal data acquisition layer uses a hardware high-precision synchronous clock to uniformly trigger the visible light, infrared, vibration sensing, RFID, and Beidou positioning acquisition modules. The timestamp deviation of the output data of each acquisition module is controlled within ±5ms, thereby achieving spatiotemporal alignment of multimodal inspection data.

[0018] Option 2: A method for offline multimodal AI integrated deployment of a handheld oil inspection terminal, the method being implemented based on the offline multimodal AI integrated deployment system for a handheld oil inspection terminal described in Option 1, and the method comprising the following steps: S1. The explosion-proof handheld terminal hardware carrier is connected to the oilfield private network to synchronize inspection tasks, equipment ledgers, and AI model incremental patches from the cloud inspection management platform, complete patch verification and local model updates, and cache task and ledger data to local storage. S2. The explosion-proof handheld terminal hardware carrier is integrated into the network-free field block. The terminal hardware adaptation layer schedules various acquisition modules to synchronously collect multi-source inspection data. The multi-modal data acquisition layer performs standardized preprocessing and time-series alignment on the collected data. S3, the edge AI inference engine loads a lightweight model pool, adaptively allocates computing resources based on the terminal's real-time computing power load, ambient temperature, and remaining power, executes multiple types of defect identification inference in parallel, and outputs standardized anomaly judgment results; S4. Offline business closed-loop processing stage: After the local business function layer identifies high-risk defects, it triggers light and sound alarms and records on-site handling information; after all points are inspected, offline inspection reports are generated in batches; all inspection data are automatically encrypted and written to local storage by the encrypted storage management layer. S5. After the terminal detects an available network, it performs offline synchronization and reads the local persistent synchronization progress from the operation and maintenance layer, and performs sharded hash breakpoint resume upload of encrypted inspection data; after synchronization is completed, it detects cloud model updates, downloads incremental patches online, or imports differential patches through offline media to complete local model iterative updates. Under conditions such as low temperature, low power, insufficient storage space, transmission interruption, and model update failure, the system automatically performs function degradation, data backup, and model rollback protection to continuously ensure that core inspection operations are not interrupted.

[0019] Option 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in Option 2.

[0020] Option 4: A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method described in Option 2.

[0021] The advantages of this invention are: The oilfield inspection handheld terminal offline multimodal AI integrated deployment system described in this invention relies on a full-service local closed-loop architecture. The entire unit can support more than 8 hours of uninterrupted full-function inspection when fully charged, and the integrity of business functions can reach 100% in the absence of network environment. According to field tests in oilfields, compared with traditional cloud-dependent solutions, the interruption rate of field inspection operations is reduced by 92%, completely eliminating the limitations of public network and oilfield private network coverage, and truly realizing the intelligent inspection of remote well stations and long-distance pipelines in oilfields, greatly reducing the labor intensity and safety hazards of manual paper inspection.

[0022] The offline multimodal AI integrated deployment system for oil inspection handheld terminals described in this invention uses the SM4 algorithm to perform on-the-ground encryption on all original inspection images, sensor data, AI recognition results, and inspection reports. It is equipped with hardware security unit key hierarchical management and abnormal data fallback protection mechanism. No plaintext data is stored locally on the terminal, eliminating the risk of leakage of sensitive production information caused by equipment loss or illegal copying from the root. It fully complies with the data security management specifications and commercial cryptography standards of the oil and gas industry.

[0023] The offline multimodal AI integrated deployment system for oilfield inspection handheld terminals described in this invention uses data fragmentation, MD5 hash-based fragment verification, and persistent storage of transmission progress mechanisms to eliminate the need for full retransmission after network interruption. In actual tests, the data synchronization efficiency under weak network conditions is improved by 70% compared to the traditional full upload solution, the synchronization stability accuracy can reach 99.95%, and the fragmentation and retransmission overhead is reduced by 90%, perfectly adapting to the complex communication environment of oilfield field signal fluctuations and frequent network outages.

[0024] The oil inspection handheld terminal offline multimodal AI integrated deployment system described in this invention only generates incremental patches for parameter differences in the cloud. The patch size is only 10% to 30% of the complete model file and supports local offline import and update via USB flash drive. The overall model update time is reduced by 75%, the maintenance cost of handheld terminals in remote areas without network access is reduced by 60%, and there is no need for the terminal to connect to the network to download large-capacity models, thus solving the pain point of difficulty in algorithm iteration for long-term offline devices.

[0025] The oilfield inspection handheld terminal offline multimodal AI integrated deployment system described in this invention is equipped with a three-level model lightweight scheme and working condition adaptive computing power threshold adjustment logic. The comprehensive inference time of multimodal data at a single point is no more than 50ms, and the overall response latency of terminal services is controlled within 100ms. The system can automatically adjust the computing power allocation strategy according to the ambient temperature and remaining power, and operate stably in extreme field conditions such as high and low temperatures of -20℃ to 55℃ and low power. When the computing power overflows, it automatically degrades to ensure that the core defect identification task is not interrupted.

[0026] This invention is also applicable to applications such as offline inspection operations in remote areas like Daqing Oilfield. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall architecture of the offline multimodal AI integrated deployment system for oil inspection handheld terminals described in Implementation Method 1.

[0028] Figure 2 This is a flowchart illustrating the offline inspection full-service closed loop described in Implementation Method 1.

[0029] Figure 3 This is a schematic diagram of the data synchronization process for segmented verification and breakpoint resumption as described in Implementation Method 1.

[0030] Figure 4 This is a schematic diagram of the offline incremental model update process described in Implementation Method 1. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0032] Implementation Method 1, see Figures 1 to 4 This embodiment describes an offline multimodal AI integrated deployment system for handheld oil inspection terminals. The system specifically includes: Terminal hardware adaptation layer: As the hardware carrier of the architecture, based on an industrial-grade explosion-proof handheld terminal, it integrates a visible light / infrared imaging module, an RFID / NFC read / write module, a temperature, pressure and vibration sensing module, a satellite positioning module, a local storage unit, a computing processing unit and a wireless communication unit, providing hardware computing power and data acquisition capabilities for upper-layer functions, and adapting to the explosion-proof, dust-proof and waterproof working environment of oil fields.

[0033] Multimodal data acquisition layer: Unifies the scheduling of various acquisition hardware on the terminal, completes the standardized acquisition and preprocessing of multi-source heterogeneous data such as visual images, equipment sensor data, electronic tag information, location coordinates, and inspection audio and video records during the inspection process, and outputs multimodal inspection data in a unified format to provide input for subsequent AI inference and data storage.

[0034] Edge AI Inference Engine Layer: As the core computing layer of the architecture, it has a built-in multimodal inspection AI model set and dynamic inference scheduler. It completes parallel inference of multiple tasks such as pipeline defect identification, equipment anomaly detection, leakage determination, and personnel compliance verification on the terminal, and outputs standardized identification results and confidence scores. No cloud computing resources are required throughout the process.

[0035] Local business function layer: It carries the business logic of the entire inspection process and includes five major functional modules: inspection task management, local abnormal sound and light alarm, inspection record archiving, automatic generation of offline reports, and on-site handling records. It fully supports the business closed loop of inspection operations in offline mode and replaces the core functions of traditional cloud business systems.

[0036] Encrypted storage management layer: Based on the SM4 block cipher algorithm, a local data security system is built to automatically encrypt and store all raw inspection data, AI recognition results, and business record files. At the same time, it is equipped with identity authentication and access log mechanisms to ensure the security of terminal data in offline mode.

[0037] Offline synchronization and operation and maintenance layer: Includes a breakpoint resume data synchronization module and an incremental model update module. When the network is available, it automatically completes the breakpoint resume transmission of local data to the cloud according to the fragment verification mechanism, and at the same time receives the model differential incremental packets sent by the cloud to complete the seamless update of the local model; when the network is unavailable, it automatically suspends the synchronization task and remembers the transmission progress.

[0038] (1) The terminal hardware adaptation layer is specifically as follows: Explosion-proof rating: The whole machine meets the ExibIIBT4Gb intrinsically safe explosion-proof rating. It is designed with hardware current limiting and voltage limiting protection circuits for flammable and explosive oilfield scenarios. The battery uses explosion-proof lithium polymer cells and is equipped with overcharge, over-discharge and over-temperature protection circuits, making it suitable for operation in dangerous areas such as wellheads, metering rooms and gathering and transportation stations.

[0039] Low-power scheduling logic: Built-in three-level power management strategy: ① In normal inspection mode, unnecessary peripherals (such as NFC and infrared fill light) are turned off, and only visible light acquisition and core computing units are retained; ② In standby mode, the processor frequency is reduced to the lowest operating frequency, the screen and non-core sensors are turned off, and the standby time on a full charge is no less than 72 hours; ③ When the low battery warning (20% remaining) is triggered, power degradation is automatically triggered to prioritize the core inspection and data storage functions and turn off unnecessary inference tasks.

[0040] Multi-acquisition module synchronization control strategy: A hardware-triggered synchronization mechanism is adopted, using the system's high-precision clock as a reference, and a unified trigger signal is set for the visible light camera, infrared thermal imaging, vibration sensor, and positioning module. The timestamp deviation of each module is controlled within ±5ms, ensuring the spatiotemporal alignment accuracy of multimodal data and providing a data foundation for subsequent joint inference.

[0041] (2) The edge-side multi-model parallel inference engine is specifically as follows: Design concept: To address the limited computing power of handheld terminals and the need for parallel multimodal inspection tasks, an inference engine architecture of "model pool + dynamic scheduler" is constructed. Multiple inspection AI models, such as pipeline crack identification, equipment corrosion detection, leak determination, and personnel violation identification, are deployed locally. Multi-task parallel inference is achieved through dynamic computing power scheduling, balancing inference real-time performance and terminal resource utilization.

[0042] Core implementation logic: Multi-model lightweighting optimization methods: A three-level lightweighting scheme is adopted, consisting of structured pruning, INT8 quantization, and knowledge distillation. 1. Structured pruning: Remove redundant convolutional channels in the model, with the pruning ratio controlled at 30%-40%, ensuring that the accuracy loss is less than 2%; 2. INT8 quantization: Converts 32-bit floating-point model parameters to 8-bit integers, compressing the model size to 25% of the original and increasing inference speed by more than 3 times; 3. Knowledge distillation: Using a high-precision large model in the cloud as the teacher model, the lightweight student model on the terminal is trained to make up for the accuracy loss caused by pruning and quantization.

[0043] Dynamic computing power scheduling and threshold adaptive rules: The default computing power scheduling threshold is 85% CPU utilization. For low-temperature outdoor environments (ambient temperature below -10℃), the processor's heat dissipation is improved, and the threshold is automatically increased to 90% to fully utilize hardware computing power. For low-battery scenarios (remaining battery <20%), the threshold is automatically decreased to 75% to prioritize ensuring terminal battery life. The threshold adjustment process is smooth to avoid inference interruptions caused by sudden increases or decreases in computing power. High-priority tasks (such as leak detection) are allocated computing power first, while low-priority tasks (such as device information identification) share resources in a time-sharing manner. When the terminal resource utilization exceeds the threshold, non-core inference tasks are automatically suspended to ensure the stable operation of core inspection functions.

[0044] Standardized output of inference results: The identification results of all models are uniformly encapsulated into a standardized format that includes defect type, location coordinates, confidence level, and risk level, which directly connects to the local business function layer without the need for secondary parsing in the cloud.

[0045] (3) The SM4 encrypted storage module is specifically as follows: Design concept: In view of the sensitivity and compliance requirements of oilfield inspection data, the local storage security system is built using the SM4 symmetric encryption algorithm to realize automatic encryption when writing data and automatic decryption when reading data. The whole process is transparent to the upper-level business. At the same time, it is equipped with identity authentication and anomaly fallback mechanism to prevent unauthorized access and data corruption.

[0046] Core implementation logic and calculation rules: Block encryption mechanism: The SM4-ECB block encryption mode is adopted. The inspection data to be stored is divided into 128-bit blocks, and each block is encrypted with a 128-bit key. The encrypted data is written to the local solid-state storage unit. When reading the data, reverse decryption is performed to restore the original data.

[0047] Key update and hierarchical authentication process: Keys adopt a hierarchical management mechanism, divided into device root keys and user working keys; the root key is stored in the hardware security unit (SE) and cannot be exported. It can only be updated annually by the administrator through an authorized USB key. The update process requires dual authentication of the administrator's identity and the device identifier; the working key is derived from the root key and is used for daily data encryption and decryption. It supports hierarchical authorization according to user roles (inspectors only have read and write permissions, while administrators have key management permissions). The entire operation is recorded and an encrypted access log is generated.

[0048] Encryption anomaly fallback mechanism: When storage media damage or sector read / write anomalies are detected, the data protection mechanism is automatically triggered, prioritizing the writing of the current inspection data to the backup storage partition, and generating a verification and recovery record for the damaged data; if the key is lost or the authorization verification fails, the terminal automatically locks the encrypted data, which needs to be unlocked through a key recovery process authorized by the administrator. The recovery process requires verification of the device's unique serial number and the administrator's authorization code to prevent the data from being illegally obtained.

[0049] Full data coverage encryption: The encryption scope covers all data including original inspection images, sensor data, AI recognition results, inspection reports, and task records. No plaintext data is stored on the ground, completely eliminating the risk of data leakage in offline states.

[0050] (4) The offline data synchronization module with breakpoint resume function is as follows: Design concept: In response to the network characteristics of remote oil fields with weak networks and frequent outages, a synchronization mechanism of "data fragmentation-hash verification-progress memory-breakpoint resumption" is constructed. Local offline data is divided into fixed-size fragments, which are transmitted and verified one by one. When the network is interrupted, the current fragment progress is remembered, and transmission continues from the breakpoint after recovery, avoiding full retransmission and improving synchronization efficiency and accuracy.

[0051] Core implementation logic and calculation formula: Data sharding and preprocessing: Divide the offline dataset to be synchronized into N consecutive shards of a fixed size (default 1MB, configurable). Calculate the MD5 hash check value for each shard and generate a shard index list. The list includes the shard number, shard size, hash check value, and data ownership information.

[0052] Segment-by-segment transmission and retransmission fallback strategy: Data is transmitted to the cloud sequentially according to the segment number. After each segment transmission is completed, the hash value is immediately compared with that of the cloud. If the verification passes, the segment transmission is marked as complete, and the synchronization progress is updated. If a single segment fails retransmission verification three times consecutively, the segment degradation mechanism is automatically triggered, splitting the segment into four equal-sized sub-segments for retransmission. If the sub-segments also fail to transmit three times consecutively, the segment is marked as pending subsequent synchronization, automatically skipped, and subsequent segment transmission continues to avoid blocking the overall synchronization task due to a single segment failure. All failed segments are recorded in the synchronization error log and automatically retried after the network quality improves.

[0053] Breakpoint progress persistence: During synchronization, the sequence number of the largest completed segment is persisted to the local storage of the terminal in real time; when the network is interrupted or the synchronization task is terminated, a progress snapshot is automatically saved; when the task resumes, the progress snapshot is read and the transmission continues from the next unfinished segment without having to start from the first segment again.

[0054] The complete process of synchronization logs and cloud verification is as follows: The terminal generates a full synchronization log locally, recording the transmission time, verification result, retry count, and operator information for each fragment. The log is stored synchronously and encrypted. After receiving all fragments, the cloud performs full hash verification and data integrity verification, comparing the total number of fragments with the total data length. If the verification passes, a synchronization confirmation receipt is generated and sent back to the terminal. If the verification fails, the cloud returns a list of abnormal fragments to the terminal, and the terminal automatically triggers retransmission to ensure zero-error data synchronization.

[0055] Synchronization accuracy calculation: After synchronization is complete, the synchronization accuracy is calculated by comparing the full fragment hashes. The formula is as follows: .

[0056] (5) Offline incremental model update module Design concept: To address the difficulty of iterating offline terminal models, an offline update mechanism of "differential incremental generation - local media import - incremental merging and verification" is constructed. The cloud generates differential incremental patches based on the old and new models, and the terminal can import the incremental patches through offline media such as USB flash drives and local area networks to automatically complete the model update without the need for a full network download.

[0057] Core Implementation Logic: Differential Incremental Generation: The cloud performs differential comparison between the old and new versions of the AI ​​model, extracting only the parts of the model parameters that have changed to generate incremental patch files. The patch size is only 10%-30% of the full model, significantly reducing the overhead of transmission and import. Detailed Judgment Logic for Patch Verification Failure: Incremental patch verification is divided into three levels of verification: 1. File integrity verification: Compare the MD5 value of the patch file. If the comparison fails, the file is considered corrupted, and you will be prompted to obtain the patch again. 2. Version compatibility check: Compare the base model version adapted by the patch with the current model version on the terminal. If the difference is too large (version difference ≥ 3 major versions), it is determined to be incompatible and the base version needs to be upgraded first. 3. Digital signature verification: Verify the digital signature of the patch publisher. If it fails, it is determined to be an illegal patch, the update is terminated, and a security log is recorded.

[0058] Update interruption protection mechanism: During the model update process, the terminal battery level and running status are monitored in real time. If the battery level is too low (<10%), there is an unexpected power outage, or the program is interrupted, the original model file before the update is automatically retained. The system will automatically roll back to the stable version on the next startup to avoid the model becoming unusable due to update interruption. The update progress is persistently recorded in real time, and the update can be resumed from the breakpoint after the interruption is recovered.

[0059] Parallel caching of multiple model versions: The terminal supports caching three historical versions of AI models locally at the same time, adopting a caching strategy of "latest priority and cyclic overwrite". Once the new version model is verified, it is set as the default enabled version, and the old version is retained for rollback. It supports manually switching the corresponding version model according to the inspection task type to adapt to the accuracy requirements of different inspection scenarios.

[0060] (6) Full-function module for local inspection business Design concept: The traditional cloud-based inspection business logic is moved to the local terminal, fully supporting the entire process of inspection task execution, anomaly alarm, record archiving, and report generation in offline mode, achieving "no interruption offline and no reduction in functionality".

[0061] Core implementation logic: Local management and conflict handling of inspection tasks: When there is a network connection, inspection tasks and equipment log information are downloaded from the cloud in advance and stored locally; in offline mode, inspections can be performed according to the task list, automatically recording inspection points and completion status, and supporting task pause and resumption. When multiple overlapping inspection tasks exist, the system automatically sorts them according to task priority and deadline, with higher priority tasks being executed first; when task points overlap, inspection paths are automatically merged to avoid duplicate work; if an emergency inspection task is temporarily inserted, the current low-priority task is automatically paused and its progress is saved, and the original task is automatically resumed after the emergency task is completed.

[0062] Local anomaly real-time alarm: When AI inference identifies high-risk defects or anomalies, the terminal immediately triggers a dual alarm with sound and light, and at the same time pops up handling guidance information to assist on-site personnel in carrying out emergency handling without waiting for cloud instructions.

[0063] Multi-point batch report generation logic: Supports batch selection of inspection points by well station, pipeline, administrative region and other dimensions, automatically aggregates inspection data, defect records and handling information of corresponding points, and generates summary reports according to a unified template; Report generation adopts streaming computing method, aggregating and generating at the same time, avoiding memory overflow under large data volume, and supports exporting to PDF and Excel formats to meet the archiving needs of different business departments in the oilfield.

[0064] On-site handling record retention: Supports on-site personnel to enter abnormal handling measures and on-site remarks, and synchronously associates them with corresponding inspection points to form a complete inspection and handling record, which is uploaded to the cloud along with the data.

[0065] The complete operation process of the system described in this embodiment is divided into two stages: offline operation and online synchronization. The two stages are seamlessly switched, as detailed below: Task preparation phase (networked environment): The terminal connects to the network and downloads the inspection task, equipment basic ledger and the latest model incremental patch from the cloud inspection management platform to complete the model update and task caching, and prepare for offline operation.

[0066] Offline inspection phase (no network environment): Inspection personnel carry terminals into remote areas without network coverage and start the inspection task; the terminal sequentially completes multimodal data collection, edge AI inference recognition, real-time anomaly alarm, encrypted data storage, and task progress update; after the inspection of a single point is completed, it automatically jumps to the next point, without the need for network support throughout the process.

[0067] Offline report generation phase: After all inspection tasks are completed, the terminal automatically summarizes all inspection data locally and generates a standardized inspection report that includes an inspection overview, defect statistics, anomaly list, and handling suggestions. It supports on-site offline viewing and export.

[0068] Online synchronization and update phase (after network recovery): When the terminal returns to a networked area, a data synchronization task is automatically triggered, and the locally encrypted inspection data is uploaded to the cloud through a breakpoint resume mechanism; at the same time, cloud model updates are detected, and incremental patches are downloaded to complete the local model iteration; the synchronization and update process does not affect the normal use of the terminal.

[0069] The technical solution of the present invention will be further explained clearly and completely below in conjunction with multiple application scenarios such as Daqing Oilfield. The described embodiments are only some application scenarios of the present invention, and not all embodiments.

[0070] Example 1: Implementation of pipeline inspection at standard well site on RK3588 platform 1. Overview of Implementation Scenarios This embodiment is applied to the inspection of surface pipelines and well station equipment in a remote block of an oil production plant in Daqing Oilfield. The block covers approximately 12 square kilometers and includes 32 oil wells and 18 kilometers of gathering and transportation pipelines. It is located in a blind spot of both the public network and the oilfield's dedicated network, and daily inspections rely entirely on manual foot patrols, resulting in low efficiency and a high rate of missed defects. The industrial-grade explosion-proof handheld inspection terminal based on the architecture of this invention enables fully functional intelligent inspections in network-free environments.

[0071] 2. Terminal hardware configuration The handheld terminal used in this embodiment is built on the RK3588 edge computing core board, meets the ExibIIBT4Gb explosion-proof level, and is suitable for oilfield field operation environment; the hardware configuration includes: 13-megapixel visible light camera + infrared thermal imaging module, high-frequency RFID read and write module, built-in temperature, pressure and vibration sensor, Beidou + GPS dual-mode positioning, 128GB local solid-state storage, 10000mAh explosion-proof battery, and supports multiple communication methods such as 4G / WiFi / USB.

[0072] 3. Specific Implementation Steps S1: System initialization and task preparation (networked environment, within the oilfield station) Before setting off, the inspection personnel connected the terminal to the oilfield's private network at the oil production station to complete system initialization and task preparation. The terminal synchronizes the daily inspection tasks from the cloud-based inspection management platform, including the inspection points of 32 oil wells and 18 kilometers of pipelines, equipment ledgers, inspection item standards, and other information, and caches them to local storage. The terminal detects cloud-based AI model updates and receives incremental patches for three types of models: pipeline crack detection, equipment corrosion detection, and crude oil leak detection. Each patch is approximately 12MB in size (corresponding to a full model size of 86MB). After verification, the local model is automatically merged and updated, with the update process taking approximately 8 seconds. The system automatically verifies the integrity of the local model, task data, and encryption key, and enters the inspection state after confirming that everything is correct.

[0073] S2: Multimodal data acquisition in areas without network access (during field inspection) Inspection personnel carry their terminals into areas without network access, initiate inspection tasks, and the terminals enter offline operation mode. They then collect data sequentially from the inspection points. When the pipeline inspection point is reached, the terminal collects images of the pipeline appearance through a visible light camera, collects pipeline temperature distribution data through an infrared module, reads the equipment number information of the pipeline electronic tag through RFID, and records the current inspection location coordinates simultaneously through the Beidou module. Upon arrival at the well site equipment location, the terminal collects images of the equipment's appearance and operating temperature data, collects equipment vibration parameters through built-in sensors, and reads basic information from the equipment's electronic tag. The multimodal data acquisition layer performs standardized preprocessing on all acquired data, unifying the data format and dimensions, and outputs it to the edge AI inference engine layer.

[0074] S3: Multi-model Parallel AI Inference at the Edge The terminal-local inference engine receives standardized multimodal data and performs multi-task parallel inference through a dynamic scheduler: Pipeline location: The pipeline crack identification model and crude oil leak identification model are called in parallel to perform inference on visible light images and infrared temperature data respectively, and output the defect type, location, confidence level and risk level; Equipment location: Parallel call of equipment corrosion detection model and equipment operation anomaly judgment model, combined with image and vibration sensor data to output equipment status judgment results; During the inference process, the dynamic scheduler monitors the terminal's computing load in real time. When the CPU utilization rate exceeds 85%, it automatically reduces the computing power allocation for non-core tasks to ensure the real-time response of core defect identification tasks. In this embodiment, the average time for single-point integrated inference is 42ms, which meets the requirements for real-time on-site inspection.

[0075] S4: Local anomaly alarm and encrypted data storage When AI inference identifies high-risk defects (such as pipeline leaks or severe cracks), the terminal immediately triggers an audible and visual alarm, and the screen displays the corresponding defect handling instructions to assist inspection personnel in carrying out preliminary on-site handling, while recording the handling measures and on-site notes. All raw collected data, AI recognition results, inspection records, and handling information are automatically encrypted using the SM4 encryption module with 128-bit block encryption. The encrypted data is written to the local storage unit, with no plaintext data being stored throughout the process. The encryption key is stored in the terminal hardware security area and is bound to the terminal device identifier, making it impossible to export externally.

[0076] S5: Automatic generation of offline inspection reports After all inspection points are completed for the day, the terminal automatically summarizes the inspection data locally and generates a standardized inspection report: The report includes core information such as a daily inspection overview, the number of completed points, defect type statistics, risk level distribution, anomaly handling list, and rectification suggestions. The report supports offline preview and local export, allowing inspection personnel to view the inspection results on-site without waiting for cloud generation. In this embodiment, the daily report takes about 15 seconds to generate, which fully meets the needs of on-site operations.

[0077] S6: Resume data transmission and synchronization after network recovery Once the inspection personnel return to the oilfield station and their terminals connect to the oilfield's private network, a data synchronization task is automatically triggered. The synchronization module divides the local inspection data to be uploaded into 1MB chunks, and generates a chunk index list and the MD5 hash value of each chunk. The fragments are transmitted to the cloud one by one according to their fragment numbers. After each fragment is transmitted, the hash value is compared with that of the cloud. If the verification is successful, the local synchronization progress is updated. If the network experiences fluctuations and interruptions during synchronization, the terminal automatically saves the current fragment progress; after the network is restored, the progress is read and transmission continues from the breakpoint without having to re-upload all the data. In this embodiment, the total amount of daily inspection data is about 360MB. Two network interruptions were simulated during the synchronization process, and the total synchronization time was about 4 minutes. The synchronization accuracy rate reached 99.97%, which is 68% more efficient than the traditional full-volume transmission scheme.

[0078] S7: Operations and Iteration During routine maintenance of the terminal, if it is necessary to update the AI ​​model, the offline update can be completed by importing incremental patches through an authorized USB flash drive without the terminal being connected to the internet. The old model is automatically backed up during the update process, and the update is automatically rolled back if it fails, ensuring that the terminal inspection function is not affected.

[0079] Example 2: Differentiated Adaptation Implementation Across Multiple Hardware Platforms To verify the hardware versatility of the architecture, it was deployed on three mainstream edge computing platforms to test inference performance and battery life. The test environment was uniformly set to normal temperature and pressure, full charge, and continuous inspection mode.

[0080]

[0081] Test results show that this architecture can be adapted to handheld terminal hardware with different computing power levels. It can still meet the basic real-time requirements of inspection on low computing power platforms, and can support more models to run in parallel on high computing power platforms, demonstrating good hardware compatibility and scalability.

[0082] Example 3: Comparison of Synchronization Performance with Different Fragment Sizes To verify the impact of fragment size on synchronization efficiency, the synchronization performance of three typical fragment sizes was tested under the same weak network environment (average bandwidth 2Mbps, packet loss rate 5%), with a total test data volume of 360MB.

[0083]

[0084] Test results show that 1MB fragmentation is the optimal configuration, achieving the best balance between synchronization efficiency and accuracy. The fragmentation size can be flexibly adjusted for different network quality environments to adapt to the complex field network conditions in oil fields.

[0085] Example 4: Parallel Comparison Test of Similar Terminals on the Market Two mainstream handheld oil inspection terminals on the market were selected as comparison samples and benchmarked against the terminal of this invention under the same test scenario. The test environment was the same remote block without network in Daqing Oilfield, with continuous 8-hour inspection operation.

[0086]

[0087] The test results clearly show that the present invention is significantly superior to the two existing mainstream solutions in terms of functional integrity, intelligent inspection capability, data security, and synchronization efficiency in a network-free environment, and comprehensively solves the core pain points of offline inspection in remote oil fields.

[0088] Example 5: Implementation of System Backup Operation under Extreme Operating Conditions To verify the system's degraded operation and fault tolerance capabilities in response to extreme working conditions that may occur in oilfields: Low temperature and low power conditions: Ambient temperature -25℃, terminal remaining power 15%. The system automatically lowers the computing power scheduling threshold to 75%, shuts down non-core inference tasks, and only retains high-priority identification functions such as pipeline leaks and equipment failures. The terminal can run continuously for more than 3 hours, and the core inspection function is unaffected.

[0089] Insufficient storage space: When the remaining local storage space is less than 5%, the system automatically triggers the old data cleanup mechanism, prioritizing the deletion of historical inspection data that has been synchronized and retaining the latest unsynchronized data; if the cleanup is still insufficient, the system automatically reduces the image acquisition resolution to ensure that the inspection operation is not interrupted.

[0090] Overloaded computing power: When more than 5 AI models are launched at the same time, causing the computing power to be fully loaded, the system will automatically pause low-priority models (such as equipment nameplate recognition) according to task priority to ensure the real-time operation of the core defect recognition model and prevent system lag or crashes.

[0091] Module fault tolerance: When the encryption module malfunctions, it automatically switches to temporary cache mode. Data is first written to a temporary secure partition, and encryption is automatically restored after the function is restored, without interrupting the current inspection operation. When the synchronization module malfunctions, it automatically saves the data and progress, and retryes on the next startup to avoid data loss.

[0092] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0093] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. An offline multimodal AI integrated deployment system for a handheld oil inspection terminal, characterized in that, The system includes an explosion-proof handheld terminal hardware carrier, which integrates a multi-modal acquisition unit, an edge computing unit, a local solid-state storage unit, a multi-standard communication unit, a BeiDou positioning unit, and an explosion-proof power supply unit. The system deploys a six-layer local integrated operating architecture, which, from bottom to top, consists of: The terminal hardware adaptation layer is used to uniformly manage various hardware peripherals of the terminal, providing multimodal data acquisition, computing power scheduling, local storage read and write, explosion-proof and low power consumption management capabilities. It is adapted to flammable and explosive, low temperature and no network operation conditions in oil fields. It automatically adjusts the model running priority or performs function degradation according to the ambient temperature, power and computing power load status. The multimodal data acquisition layer is used to simultaneously acquire visible light images, infrared thermal imaging data, equipment temperature, pressure and vibration sensing data, RFID equipment electronic tags, and Beidou positioning coordinates from multiple heterogeneous sources. It completes data standardization preprocessing and time-series alignment, and outputs multimodal inspection data in a unified format. The edge AI inference engine layer has a built-in lightweight multimodal inspection AI model pool and dynamic computing power scheduler. It performs multi-task AI inference, such as pipeline defect identification, equipment corrosion detection, crude oil leakage judgment, and personnel violation verification, in parallel on the terminal and outputs standardized identification results including defect type, location coordinates, confidence level, and risk level. The local business function layer deploys the business logic of the entire offline inspection process, realizing inspection task caching and scheduling, AI abnormal sound and light alarm, on-site handling record retention, batch generation and local export of offline inspection reports, and independently realizing all core business functions of the cloud inspection platform in offline state. The encrypted storage management layer automatically encrypts and stores all raw inspection data, AI recognition results, business records, and report files based on the SM4 block cipher algorithm. It is equipped with hardware security unit hierarchical key management, user hierarchical authentication, and storage read / write anomaly protection mechanism. The offline synchronization and operation and maintenance layer integrates a sharded hash breakpoint resume synchronization module and an offline incremental model update module; when the network is connected, it automatically performs sharded data breakpoint resume and full data verification in the cloud; it supports importing model differential patches through local storage media to complete seamless iterative updates of the terminal's local AI model.

2. The offline multimodal AI integrated deployment system for oilfield inspection handheld terminals according to claim 1, characterized in that, The edge AI inference engine layer is equipped with a three-level model lightweight optimization strategy and an adaptive computing power scheduling mechanism; The three-level model lightweight optimization strategy includes structured pruning, quantization, and knowledge distillation; the adaptive computing power scheduling mechanism dynamically adjusts the CPU usage threshold based on the ambient temperature and the remaining battery power of the terminal, and allocates computing power resources according to task priority.

3. The offline multimodal AI integrated deployment system for oilfield inspection handheld terminals according to claim 1, characterized in that, The encrypted storage management layer adopts the SM4-ECB128-bit block encryption mode; the encryption key is divided into a device root key and a user working key. The device root key is stored in the hardware security unit and cannot be exported externally. It can only be updated periodically by the administrator through the authorized key; the user working key is derived from the device root key and has hierarchical access permissions according to the roles of inspectors and administrators; the encryption scope covers all inspection business data. When storage media damage or key verification failure is detected, the data to be stored is automatically written to the backup storage partition and the encrypted data is locked.

4. The offline multimodal AI integrated deployment system for oilfield inspection handheld terminals according to claim 1, characterized in that, The segmented hash breakpoint resume synchronization module divides the encrypted inspection data to be synchronized into segments of a fixed size, generates an MD5 hash check code for each segment, and generates a segment index list. The synchronization process persists a snapshot of the synchronization progress in real time. After network interruption is recovered, the progress snapshot is read and the transmission is resumed from the currently untransmitted fragment. If a single fragment fails hash verification three times consecutively, it is split into sub-fragments for retry. If multiple retry failures occur, it is marked as delayed synchronization without blocking the overall data upload. After receiving all fragments, the cloud performs a full hash integrity check. If the check fails, a list of abnormal fragments is sent to trigger terminal retransmission.

5. The offline multimodal AI integrated deployment system for oilfield inspection handheld terminals according to claim 1, characterized in that, The differential patch of the offline incremental model update module is generated by comparing the differences in parameters between the old and new AI models in the cloud. The patch size is 10% to 30% of the full model file. The patch verification process includes three levels of verification: file integrity MD5 verification, basic model version compatibility verification, and publisher digital signature verification. The currently running model is automatically backed up before the model update. The terminal battery level and running status are monitored during the update process. If the update is interrupted due to low battery or unexpected power failure, the model is automatically rolled back to the stable model before the update. The terminal locally caches up to 3 historical model versions.

6. The offline multimodal AI integrated deployment system for oilfield inspection handheld terminals according to claim 1, characterized in that, The method for automatically handling task conflicts in the local business function layer configuration is as follows: Multiple overlapping inspection tasks are automatically sorted according to task priority and deadline, and inspection paths are merged when points overlap; when an emergency inspection task is inserted, the progress of the current low-priority task is automatically saved, and the original inspection process is restored after the emergency task is completed. The reporting module supports batch aggregation of inspection data by well station, pipeline, and administrative region, and uses streaming computing to generate offline inspection reports in PDF and Excel formats.

7. The offline multimodal AI integrated deployment system for oilfield inspection handheld terminals according to claim 1, characterized in that, The multimodal data acquisition layer uses a high-precision hardware synchronous clock to uniformly trigger the visible light, infrared, vibration sensing, RFID, and Beidou positioning acquisition modules. The timestamp deviation of the output data of each acquisition module is controlled within ±5ms, realizing the spatiotemporal alignment of multimodal inspection data.

8. A method for offline multimodal AI integrated deployment of a handheld terminal for oilfield inspection, characterized in that, The method is implemented based on the offline multimodal AI integrated deployment system for oil inspection handheld terminals as described in any one of claims 1 to 7, and the method includes the following steps: S1. The explosion-proof handheld terminal hardware carrier is connected to the oilfield private network to synchronize inspection tasks, equipment ledgers, and AI model incremental patches from the cloud inspection management platform, complete patch verification and local model updates, and cache task and ledger data to local storage. S2. The explosion-proof handheld terminal hardware carrier is integrated into the network-free field block. The terminal hardware adaptation layer schedules various acquisition modules to synchronously collect multi-source inspection data. The multi-modal data acquisition layer performs standardized preprocessing and time-series alignment on the collected data. S3, the edge AI inference engine loads a lightweight model pool, adaptively allocates computing resources based on the terminal's real-time computing power load, ambient temperature, and remaining power, executes multiple types of defect identification inference in parallel, and outputs standardized anomaly judgment results; S4. Offline business closed-loop processing stage: After the local business function layer identifies high-risk defects, it triggers light and sound alarms and records on-site handling information; after all points are inspected, offline inspection reports are generated in batches. All inspection data is automatically encrypted and written to local storage by the national cryptographic encryption storage management layer. S5. After the terminal detects an available network, it reads the local persistent synchronization progress of offline synchronization and operation and maintenance layer, and performs fragmented hash breakpoint resume transmission to upload encrypted inspection data. After synchronization is complete, check for updates to the cloud model, download incremental patches online, or import differential patches via offline media to complete local model iterative updates; Under conditions such as low temperature, low power, insufficient storage space, transmission interruption, and model update failure, the system automatically performs function degradation, data backup, and model rollback protection to continuously ensure that core inspection operations are not interrupted.

9. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-7.

10. A computer device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method of any one of claims 1-7.

Images