Rail transit vehicle door intelligent operation and maintenance system and method based on edge side large language model

Abstract

The invention discloses a rail transit vehicle door intelligent operation and maintenance system and method based on an edge side large language model, and belongs to the technical field of rail transit equipment operation and maintenance. The rail transit vehicle door intelligent operation and maintenance system comprises an intelligent front-end module, an embedded large model reasoning module and a core algorithm service module; a rail transit vehicle door intelligent operation and maintenance page which is accurately matched according to a user role and accords with a post responsibility is generated, and a core algorithm service module analyzes a user role intention and can generate a chart; through the intelligent operation and maintenance page of the rail transit vehicle door, the post responsibilities of the user can be accurately matched, expert experience is mathematized, automatic closed loop of alarm is realized, accurate private knowledge retrieval is supported, hidden operation and maintenance knowledge is converted into computable, reusable and auditable services, man-made misjudgment is reduced, and the maintenance efficiency and the equipment reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance systems, specifically to an intelligent operation and maintenance system and method for rail transit doors based on a large language model on the edge side. Background Technology

[0002] In the current technological context, the deployment and performance of intelligent operation and maintenance (O&M) systems face numerous challenges. Firstly, a common problem in intelligent O&M systems is the large number of parameters in large language models, which places extremely high demands on the computing power and memory of edge devices. Edge devices are often equipped with low-power CPUs or entry-level GPUs, which cannot support such massive data processing capabilities, making them unable to directly handle the inference computation tasks of these large models. This means that intelligent O&M systems often have to rely on cloud services, which can lead to response latency issues when the network is unstable, failing to meet the needs of real-time on-site response.

[0003] Secondly, the model reasoning process required to generate structured instructions is typically time-consuming, far exceeding the real-time response requirements for decision support and task execution in the operations and maintenance field. For example, in handling equipment fault diagnosis and work order dispatch, a fast and accurate response is crucial, and excessively long response times may lead to untimely problem handling, thereby affecting the overall performance and efficiency of the system. Furthermore, while general-purpose language models can provide broad knowledge coverage and answer diverse queries, they fall short in understanding specialized domain knowledge. For instance, in specialized industries like rail transit, numerous personalized professional knowledge points are involved, such as door alignment, V-shaped anomaly identification, and sub-health alarm management. This knowledge is difficult to accurately interpret and apply using ordinary language models. If the system cannot correctly understand and apply this specialized domain knowledge, it will cause instruction execution errors or fail to effectively identify critical events such as oil and gas leaks and equipment failures, thus damaging the actual usability of the entire system.

[0004] In summary, existing technologies such as Parameter Efficient Fine-Tuning (PEFT) cannot completely solve these problems, especially in terms of direct deployment on edge devices and understanding and application of domain-specific knowledge. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide an intelligent operation and maintenance system for rail transit doors based on a large language model on the edge side to solve the problems of infeasibility of edge deployment, high interaction latency, and lack of professional operation and maintenance capabilities, thereby improving the efficiency and accuracy of operation and maintenance. On the other hand, it provides an intelligent operation and maintenance method for rail transit doors based on a large language model on the edge side.

[0006] Technical solution: The intelligent operation and maintenance system for road traffic vehicle doors described in this invention includes: The intelligent front-end module is used to receive user natural language and user role information and input them into the embedded large model inference module. It dynamically adjusts the layout, theme variables and component visibility status of the intelligent operation and maintenance page of rail transit doors according to the user role to ensure that human-computer interaction accurately matches job responsibilities. An embedded large model inference module is used to construct an embedded large model using a three-stage compression and fine-tuning strategy. First, the open-source large language model is structurally pruned to obtain a first intermediate model; the first intermediate model is distilled on an operation and maintenance task dataset to obtain a second intermediate model; two types of adapters are superimposed on the second intermediate model for fine-tuning to obtain the embedded large model. The core algorithm service module is used to parse user role intent and generate function call instructions, call the corresponding algorithm services, including rail transit equipment health algorithm, sub-health alarm closed-loop status algorithm and private knowledge question answering knowledge retrieval algorithm, parse the algorithm results to generate corresponding charts, and display them on the rail transit door intelligent operation and maintenance page dynamically generated according to user role.

[0007] The intelligent operation and maintenance method for rail transit vehicle doors described in this invention includes the following steps: The open-source large language model is structured and pruned to obtain the first intermediate model; Distill the first intermediate model onto the operation and maintenance task dataset to obtain the second intermediate model; Two types of adapters are superimposed on the second intermediate model for fine-tuning to obtain an embedded large model; The user's natural language and user role information are input into the embedded big model. The user role intent is parsed and function call instructions are generated. The corresponding algorithm services are called. The generated function call results are parsed and corresponding charts are generated and displayed on the intelligent operation and maintenance page of rail transit doors, which is dynamically generated according to the user role. The algorithm services include rail transit equipment health algorithm, sub-health alarm closed-loop state algorithm, and private knowledge question answering knowledge retrieval algorithm.

[0008] Furthermore, the first intermediate model introduces a binary mask to score the importance of the attention head, and retains or removes weights according to a preset threshold, thereby obtaining the pruned first intermediate model.

[0009] Furthermore, the second intermediate model uses the open-source large language model as the teacher and the first intermediate model as the student, and performs distillation on the operation and maintenance task dataset. The loss function combines cross-entropy loss and KL divergence loss, and the optimized result is the second intermediate model.

[0010] Furthermore, the embedded large model is obtained by superimposing the LoRA module and the Adapter module on the second intermediate model, wherein the LoRA module is used for the attention layer and the Adapter module is used for the FFN layer.

[0011] Furthermore, the equipment health algorithm calculates the equipment health by weighting three indicators: normalized door opening time, jamming frequency penalty, and maintenance time-related degradation of rail transit vehicle door equipment.

[0012] Furthermore, the sub-health alarm closed-loop state algorithm defines states such as alarm trigger, confirmed, dispatched, in progress, verified, and closed, as well as corresponding deterministic transfer functions, and automatically escalates the alarm if the target state is not reached within a preset time.

[0013] Furthermore, the private knowledge question-answering knowledge retrieval algorithm, enhanced by RAG knowledge retrieval operators, supports the solution of professional questions related to rail transit vehicle doors.

[0014] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: It obtains an embedded large model through a three-stage compression and fine-tuning strategy, compressing the model parameter count. It only requires a regular CPU to run, eliminating the need for a GPU, and can be directly deployed on resource-constrained devices such as vehicle-mounted industrial control computers and station edge servers, freeing it from cloud dependence, improving system availability and data security, and reducing hardware procurement and maintenance costs. Simultaneously, it retains core maintenance capabilities, fully supporting four core functions: structured instruction generation, role-adaptive interface, device health assessment, and private knowledge Q&A. All outputs can be parsed and executed by downstream systems, forming a closed loop, improving Function Call accuracy, reducing health calculation errors, and CPU inference latency. Furthermore, this invention mathematicallyizes expert experience, achieving automatic alarm closed-loop, supporting precise private knowledge retrieval, and transforming implicit maintenance knowledge into computable, reusable, and auditable services, reducing human error and improving maintenance efficiency and equipment reliability. Finally, this invention dynamically generates all functions based on roles through the front-end interface and operation permissions. Before a call is executed, permissions are verified to prevent unauthorized operations, comply with enterprise IT security standards, ensure accurate information delivery, avoid irrelevant interference, and improve the efficiency and satisfaction of users in different positions. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] With the rapid development of rail transit systems, train doors, as a key component of rail transit vehicles, directly affect passenger safety and the normal operation of trains. Traditional rail transit door maintenance mainly relies on manual inspections and periodic maintenance, which suffers from low efficiency and delayed response. In recent years, the development of artificial intelligence technology, especially the emergence of Large Language Models (LLM), has provided a new technological path for the intelligent maintenance of rail transit doors.

[0018] However, existing technologies have the following shortcomings: 1. Existing intelligent operation and maintenance systems for rail transit mostly rely on large models deployed in the cloud. Onboard industrial control computers and station edge servers are typically equipped only with low-power CPUs or entry-level GPUs, which cannot handle the video memory and computing power required for large model inference. This forces intelligent operation and maintenance systems to be deployed in the cloud, failing to meet real-time on-site response needs. Furthermore, they rely on network stability, resulting in issues such as the impracticality of edge deployment, high interaction latency, and weak professional capabilities. 2. Generating structured instructions (such as Function Calls) from large models is often time-consuming, failing to meet real-time operation and maintenance decision-making needs. This affects the timeliness of critical operations such as fault diagnosis and work order dispatch, reducing system usability. 3. General-purpose large models lack understanding of proprietary knowledge in vertical fields such as rail transit (e.g., door alignment, V-shaped anomalies, and sub-health alarm loops), easily leading to format errors, terminological misuse, or logical deviations. Even if responses are generated, they are difficult for downstream systems to parse or execute.

[0019] To address the aforementioned issues, this invention provides an intelligent operation and maintenance system and method for rail transit doors based on a large language model at the edge, which solves problems such as the impracticality of edge deployment, high interaction latency, weak professional capabilities, and the inadequacy of existing technologies. It achieves the technical effects of extremely lightweight edge deployment, preservation of functional integrity, deep integration of professional domain knowledge, and a balance between security and user experience.

[0020] Example 1

[0021] like Figure 1 As shown, this embodiment provides an intelligent operation and maintenance system for rail transit doors based on an edge-side large language model, including: an intelligent front-end module 1, an embedded large model module 2, and a core algorithm service module 3. The system is formally defined as S, with the following expression: S: u R VUA; Where u represents the user's natural language input space, R = {r1, r2, ..., r...} K} represents the set of user role information (such as dispatcher, maintenance engineer, system administrator), V represents the front-end view state space (including layout template, theme variables, component visibility status), and A represents the algorithm service action space (including health calculation, work order creation, knowledge retrieval, etc.).

[0022] The intelligent front-end module 1 is used to receive user natural language and user role information and input them into the embedded large model inference module. It dynamically adjusts the layout, theme variables and component visibility status of the intelligent operation and maintenance page of rail transit doors according to the user role to ensure that human-computer interaction accurately matches job responsibilities.

[0023] Specifically, the intelligent front-end module 1 is responsible for receiving user natural language and user role information, and inputting this information into the embedded large model inference module. The intelligent front-end module 1 can dynamically adjust the layout, theme variables, and component visibility states of the intelligent operation and maintenance page for rail transit doors according to the user role, ensuring that human-computer interaction accurately matches job responsibilities. The intelligent front-end module 1 adopts a responsive design architecture, supporting adaptation to various terminal devices, including desktops, tablets, and mobile devices, ensuring a consistent user experience across different usage scenarios. The intelligent front-end module 1 also integrates voice recognition functionality, enabling maintenance personnel to interact with the system via voice commands even when operating equipment with both hands.

[0024] For example, in a subway depot in a city, a maintenance engineer is inspecting an early morning train and asks a question to the intelligent maintenance assistant via a handheld terminal: User input: Has there been any recent abnormality at gate 2 of car 05? What is its health status? User role: Maintenance engineer (with equipment diagnostic and repair permissions).

[0025] Embedded large model module 2 is used to obtain an embedded large model using a three-stage compression and fine-tuning strategy, including performing structured pruning on the open-source large language model to obtain a first intermediate model; distilling the first intermediate model on the operation and maintenance task dataset to obtain a second intermediate model; and superimposing two types of adapters on the second intermediate model for fine-tuning to obtain the embedded large model.

[0026] Specifically, the embedded large model module 2 employs a three-stage compression and fine-tuning strategy, including: First, structurally pruning the open-source large language model to obtain a first intermediate model; second, distilling the first intermediate model on the operation and maintenance task dataset to obtain a second intermediate model; and finally, fine-tuning the second intermediate model by overlaying two types of adapters to obtain the final embedded large model. In the structural pruning stage, a selective pruning method based on attention head importance assessment is used to reduce the number of model parameters while preserving core semantic understanding capabilities. In the knowledge distillation stage, a professional dataset containing 500,000 rail transit equipment maintenance cases and 100,000 fault diagnosis records is used, and knowledge transfer is achieved through a teacher-student network architecture. In the adapter fine-tuning stage, domain adapters and task adapters are added respectively; the former focuses on understanding rail transit professional terminology, while the latter optimizes fault diagnosis and maintenance guidance capabilities.

[0027] The Embedded Large Model Module 2 employs quantization-aware training technology, simulating quantization errors during training to ensure the model retains over 95% of its performance even after 8-bit integer quantization. It also implements a dynamic batch processing mechanism, automatically adjusting batch size based on current system load to maximize hardware utilization while maintaining response time. Furthermore, the Embedded Large Model Module 2 possesses incremental learning capabilities, continuously learning from new maintenance cases and updating its knowledge base without requiring complete retraining.

[0028] For example, the embedded large model M* receives input, identifies the intent as querying the health status of the equipment, the entity is train number 05 car, and the equipment door 2; the role is maintenance engineer, who can view detailed data and perform maintenance operations.

[0029] This generates a structured Function Call.

[0030] The model outputs valid JSON commands: { "function": "get_door_health", "args": { "train_id": "05", "door_id": "2" } } The core algorithm service module 3 is used to parse the user's role intent and generate function call instructions, call the corresponding algorithm services, including the rail transit equipment health algorithm, the sub-health alarm closed-loop status algorithm, and the private knowledge question answering knowledge retrieval algorithm, parse the algorithm results to generate corresponding charts, and display them on the rail transit door intelligent operation and maintenance page dynamically generated according to the user's role.

[0031] Specifically, the core algorithm service module 3 is used to parse user role intent and generate function call instructions, invoking corresponding algorithm services, including the rail transit equipment health algorithm, the sub-health alarm closed-loop state algorithm, and the private knowledge question-and-answer knowledge retrieval algorithm. The generated function call results are parsed and displayed in an interface dynamically generated according to the user role. The rail transit equipment health algorithm is based on multi-source sensor data fusion technology, comprehensively analyzing parameters such as vibration, sound, current, and air pressure of the doors to construct a multi-dimensional health indicator system and evaluate the health status of the door system in real time. The sub-health alarm closed-loop state algorithm adopts an anomaly detection method based on time-series patterns, which can identify potential fault symptoms in advance and track the entire process from alarm generation to processing completion. The private knowledge question-and-answer knowledge retrieval algorithm combines vector retrieval and semantic understanding technology to quickly locate relevant maintenance manuals, historical cases, and expert experience from the enterprise knowledge base.

[0032] The core algorithm service module 3 also integrates a door fault prediction model. By analyzing historical operating data and environmental factors, it predicts potential faults 7 to 14 days in advance, providing decision support for maintenance plans. The core algorithm service module 3 adopts a microservice architecture, with each algorithm service deployed independently and communicating through standard API interfaces to ensure system scalability and fault tolerance. The core algorithm service module implements an algorithm version management mechanism, supporting A / B testing and canary releases to ensure stable system operation during algorithm upgrades.

[0033] For example, calling the pre-registered device health function H(e) yields the following result: Opening time T open =2.8 seconds, number of times F has been stuck in the past 7 days jam =3 times, Δt since the last maintenance maint =45 days, health score H = 0.4 × 1.0 + 0.4 × 0.7 + 0.2 × 0.105 = 0.691. The system judges the health score to be < 0.7, triggering a "sub-health" alarm. The system automatically creates an alarm work order, and the state machine enters the alarm trigger → confirmed → work order dispatch state, and assigns it to a maintenance engineer. The intelligent front-end module 1 dynamically generates the interface based on the role (maintenance engineer) and the command (get_door_health): The system displays a health status of 69.1% (red alert) and shows details of three sub-indicators; it provides operation buttons: [View Historical Data], [Create Repair Work Order], and [Execute Gated Test]; and it injects a repair-themed CSS (high contrast, large font).

[0034] Access control: If the questioner is a dispatcher, only gate 2 of car 05 will be displayed: Sub-health (order dispatched), with no operation button.

[0035] Further professional Q&A (RAG enhancement), maintenance engineer's follow-up questions: How to handle a V-shaped abnormality in a car door? The system uses lightweight vector retrieval (K(q)) to match from a private knowledge base to a standard processing procedure: answer: Causes: Wear on the lower guide rail, loose lower control arm, and mismatch in closing torque; Handling steps: Check the straightness of the guide rail; tighten the lower swing arm bolts; recalibrate the torque parameters of the door controller; perform 5 door opening and closing tests for verification.

[0036] Note: The answer is from the company's internal maintenance manual and is not based on a model illusion.

[0037] Performance (actual test data): End-to-end response time: 280 ms (Intel i5-1240P CPU, no GPU); Model memory usage: approximately 420 MB (including runtime cache); Network dependency: Completely offline, no cloud connection required.

[0038] Business value is reflected in: Maintenance engineers can complete a health assessment within 5 seconds without having to log into multiple systems to check data; the system automatically identifies sub-health conditions to prevent the fault from worsening; young engineers can quickly learn expert experience through Q&A; and terminal devices do not require GPU upgrades, reusing existing industrial control hardware.

[0039] In a preferred embodiment, the system further includes an edge computing hardware module. This module employs a heterogeneous computing architecture, integrating a CPU, GPU, and NPU to provide an efficient inference environment for embedded large models. The edge computing hardware module features adaptive power management, dynamically adjusting the processor frequency and the number of cores used based on the current task complexity, maximizing battery life while ensuring performance.

[0040] In another preferred embodiment, the system integrates a multimodal perception module, including a high-definition camera, an infrared sensor, and an acoustic sensor, capable of collecting visual, thermal, and acoustic data during the operation of the vehicle door. Through sensor fusion algorithms, the multimodal perception module achieves comprehensive monitoring of the door's opening and closing status, sealing performance, and mechanical wear, providing multidimensional data support for fault diagnosis.

[0041] In another preferred embodiment, the system is equipped with a door digital twin module, which constructs a virtual simulation environment for the door system based on a physical model and historical data. Maintenance personnel can simulate different fault scenarios and repair plans within the digital twin environment, intuitively assess repair effectiveness, and reduce operational risks. The digital twin module also supports a "what-if analysis" function to predict the impact of different repair strategies on the future performance and lifespan of the door, assisting in making optimal repair decisions.

[0042] When the system is in operation, the intelligent front-end module 1 first identifies the user's identity and role, such as maintenance engineer, dispatcher, or manager, and automatically adjusts the interface layout and function permissions accordingly. The user inputs maintenance requests via natural language or voice, such as "Check the abnormal noise problem of the left door of carriage 3." After understanding the user's intent, the embedded large model module 2 generates function call instructions, invoking the corresponding core algorithm services. The system calls the rail transit equipment health algorithm to analyze the historical operating data of the door, calls the sub-health alarm closed-loop state algorithm to check for relevant alarm records, and uses a private knowledge question-and-answer knowledge retrieval algorithm to find maintenance cases of similar noise problems. Finally, the system displays a door health status assessment report, possible fault cause analysis, and recommended maintenance solutions on the interface, along with relevant maintenance manuals and historical case references.

[0043] The technical solution provided in this invention achieves offline intelligent analysis capabilities by deploying an embedded large model at the edge, enabling normal operation even in subway tunnel environments with unstable networks. Simultaneously, the system's adaptive interface design and multimodal interaction significantly improve the work efficiency of maintenance personnel and reduce training costs. The knowledge reasoning and experience extraction functions based on the large model effectively inherit expert maintenance experience, addressing the talent gap problem in the rail transit industry.

[0044] Example 2

[0045] like Figure 2 As shown in the figure, this embodiment provides an intelligent operation and maintenance method for rail transit doors based on an edge-side large language model, which includes the following steps: S1. Perform structured pruning on the open-source large language model to obtain the first intermediate model.

[0046] For example, a suitable open-source large language model for the rail transit domain is selected as the base model, such as the Llama-2 or Bloom model with 7B parameters. This base model is then subjected to structured pruning, introducing a binary mask to score the importance of attention heads. Specifically, a binary mask variable is assigned to each attention head, with a value of 1 indicating retention and a value of 0 indicating removal. An importance score is obtained by calculating the contribution of each attention head to the model output. A preset threshold is set (e.g., attention heads with an importance score below 0.3 are pruned), and weights are retained or removed based on this threshold, thus obtaining the pruned first intermediate model. This step reduces the number of model parameters while maintaining the model's basic performance on general tasks.

[0047] S2. Distill the first intermediate model on the operation and maintenance task dataset to obtain the second intermediate model.

[0048] Specifically, the first intermediate model obtained in step S1 is distilled on the operation and maintenance task dataset. This distillation process uses the original open-source large language model as the teacher model and the first intermediate model as the student model. The loss function used in the distillation process combines cross-entropy loss and KL divergence loss. The model parameters are optimized by training on a professional dataset containing knowledge of rail transit door maintenance, fault diagnosis cases, and operation and maintenance dialogues, ultimately yielding the second intermediate model. This step allows the model to maintain a small size while focusing on knowledge and capabilities in the field of rail transit door operation and maintenance.

[0049] S3. Two types of adapters are superimposed on the second intermediate model for fine-tuning to obtain the final embedded large model.

[0050] Specifically, based on the second intermediate model, two types of adapters are superimposed for fine-tuning: the LoRA module and the Adapter module. The LoRA module is used in the attention layer, using low-rank decomposition technology to improve the model's performance on specific tasks without increasing a large number of parameters. In its implementation, a low-rank update ΔW=AB is added to the weight matrix W of each attention layer, where B and A are small trainable matrices. The Adapter module is applied after the feedforward neural network (FFN) layer, employing a bottleneck structure design, adding a bypass structure consisting of a down-projection layer, a non-linear activation function, and an up-projection layer after the original network layer. By fine-tuning these two types of adapters on a dataset of instructions related to rail transit door operation and maintenance, the final embedded large model is obtained. This step enables the model to possess professional capabilities in rail transit door operation and maintenance scenarios while maintaining the model's miniaturization and efficiency.

[0051] S4. Input the user's natural language and user role information into the embedded big model, parse the user role intent and generate function call instructions, call the corresponding algorithm services, parse the generated function call results and generate corresponding charts, and display them on the intelligent operation and maintenance page of rail transit doors dynamically generated according to the user role; wherein, the algorithm services include rail transit equipment health algorithm, sub-health alarm closed-loop state algorithm and private knowledge question answering knowledge retrieval algorithm.

[0052] Specifically, the user's natural language and user role information (such as maintenance engineer, dispatcher, manager, etc.) are input into the embedded big data model. The embedded big data model first parses the user's role intent, then generates corresponding function call instructions, and invokes the corresponding algorithm services.

[0053] The embedded large model parses user input into structured intents and parameters, and invokes the corresponding algorithm service based on the intent: Rail Transit Equipment Health Algorithm: This algorithm calculates the health status of rail transit vehicle doors by weighting three key indicators. These three indicators include: normalized door opening time, jamming frequency penalty, and maintenance timeliness decay.

[0054] Door opening time normalization normalizes the ratio of actual door opening time to standard door opening time, resulting in a door opening time index T, with a value ranging from 0 to 1. The closer the value is to 1, the closer the door opening time is to the standard value. Jamming frequency penalty calculates a penalty coefficient F based on the frequency of door jamming within a certain period of time. The higher the jamming frequency, the larger the penalty coefficient. Maintenance timeliness decay calculates a decay coefficient M based on the number of days since the last maintenance. The longer the time, the larger the decay coefficient.

[0055] (2) Sub-health alarm closed-loop state algorithm: This algorithm defines the complete lifecycle state transition mechanism of alarms, including alarm trigger, confirmed, dispatched, in progress, verified, and closed states, as well as corresponding deterministic transfer functions. When the system detects that the health of the equipment is below the threshold or an anomaly occurs, an alarm is triggered; after the maintenance personnel confirm the alarm, the state changes to confirmed; after the system generates a work order and assigns it to the relevant personnel, the state changes to dispatched; after the maintenance personnel start handling the problem, the state changes to in progress; after the problem is solved and verified, the state changes to verified; and finally, after the problem is confirmed to be completely resolved, the state changes to closed. In addition, when the alarm does not reach the target state within 72 hours, the system will automatically upgrade the alarm level and notify higher-level management personnel.

[0056] (3) Private Knowledge Question Answering Knowledge Retrieval Algorithm: This algorithm enhances knowledge retrieval capabilities through RAG (Retrieval Enhanced Generation) technology, supporting the answering of professional questions related to rail transit doors. The system first vectorizes the user's question, then retrieves relevant documents from the knowledge base, and inputs the retrieved documents along with the original question into the embedded large model to generate an accurate answer. The knowledge base includes door structure principles, common fault analysis, maintenance manuals, historical cases, etc. The results of function calls are parsed and displayed in a dynamically generated interface based on the user's role. For example, for maintenance engineers, the interface highlights equipment health details and maintenance suggestions; for dispatchers, the interface focuses on displaying alarm status and processing progress; for managers, the interface provides statistical analysis and trend prediction. This role-based dynamic interface generation improves the relevance and effectiveness of information display.

[0057] Through the above four steps, the technical solution provided by this invention enables the deployment of a lightweight large language model on the edge side, providing intelligent operation and maintenance support for rail transit doors, and effectively improving operation and maintenance efficiency and equipment reliability.

[0058] Optionally, the first intermediate model introduces a binary mask to score the importance of the attention head, and retains or removes weights according to a preset threshold, thereby obtaining the pruned first intermediate model.

[0059] Specifically, let the original open-source large language model be M0:x y, the original parameter set is This yields the final embedded large model M*.

[0060] Where x∈X represents a sequence of tokenized natural language and structured metadata inputs for rail transit operation and maintenance scenarios, including user queries, role information, system status, dialogue history, etc., used to drive model inference and decision-making. y∈Y represents the target output sequence that the embedded large model M* should generate, which can be a natural language response, a structured function call, or a state machine action instruction, represented in the form of a token sequence, serving as the basis for supervised training or execution.

[0061] S11. Define the pruning operator. For the attention head h at layer l, its importance score is: ; in, The value represents the head importance score, h represents the attention head index, and l represents the layer index. This represents the development set loss function. Represents the head-weighted submatrix. This represents the pruning operator. This represents the gradient operator.

[0062] S12. Introduce a binary mask. : ; in, Indicates that the mask is reserved. Represents the characteristic function, The layer pruning threshold is indicated.

[0063] S13, The weights after pruning are: ; ; in, This indicates the Query weight after pruning. Represents a mask diagonal matrix. Represents the layer mask vector, Indicates the original query weight. This represents the weight of Value after pruning. This represents the original Value weight.

[0064] S14. Obtain the first intermediate model M1, i.e., and the parameter set after pruning. .

[0065] Optionally, the second intermediate model uses the open-source large language model as the teacher and the first intermediate model as the student, and performs distillation on the operation and maintenance task dataset. The loss function combines cross-entropy loss and KL divergence loss, and the optimized result is the second intermediate model.

[0066] Specifically, using the open-source large language model M0 as the teacher and the first intermediate model M1 as the student, distillation is performed on the operation and maintenance task dataset D, with the loss function being: ; in, This represents the total distillation loss, and α represents the loss weight, where α = 0.3. Let (x, y) represent the expectation operator, (x, y) represent the input-label pair, and D represent the operation and maintenance task dataset. M1(x) represents the cross-entropy loss and is the student model output. Represents teacher logits, Let logits represent the student, T represent the distillation temperature (T=2.0), σ represent the softmax function, and KL(p||q) represent the KL divergence. After optimization, the second intermediate model M2 is obtained.

[0067] Optionally, the embedded large model is obtained by superimposing a LoRA module and an Adapter module on the second intermediate model, wherein the LoRA module is used for the attention layer and the Adapter module is used for the FFN layer.

[0068] The body, with two types of adapters superimposed on the second intermediate model M2: LoRA module (for attention layer): ΔW=AB; Where ΔW represents the weight increment, A∈ , denotes a left low-rank matrix; B∈ , represents a right low-rank matrix. Let d denote the real matrix space, d denote the hidden layer dimension, r denote the low rank, and r=64.

[0069] Adapter module (used after the FFN layer): Adapter(h) = h + W u GeLU(W d h); Where Adapter(h) represents the hidden state after adaptation, h represents the output of the FFN layer, and W u∈ Let W represent the projection matrix. d ∈ Let denote the lower projection matrix, and GeLU denotes the Gaussian error linear unit.

[0070] The final parameter set of the embedded large model M* is: ; in, Represents the final parameter set, Represents the parameter set of the distillation model. Indicates the LoRA layer index. This indicates the Adapter layer. The percentage of trainable parameters is less than one percent, and the total number of parameters is less than or equal to 5.8. 10 8 .

[0071] Optionally, the equipment health algorithm calculates the equipment health by weighting three indicators: normalized door opening time, jamming frequency penalty, and maintenance time-related degradation of rail transit vehicle door equipment.

[0072] Specifically, for rail transit vehicle equipment e ∈ε, defines device health: ; Where H(e) represents the device health status, and e represents the door device. φ1(e) represents the indicator weight, and φ1(e) represents the opening time factor. T open ( e () indicates the actual opening time. T nom ( e ) represents the nominal door opening time, and φ2(e) represents the lag penalty factor. F jam ( e ) indicates the frequency of jamming, F max Let represent the maximum tolerance frequency, max(0,·) represent the non-negative cutoff, φ3(e) represent the attenuation-preserving factor, and exp represent the exponential function. λ Indicates the attenuation rate, Δ t maint ( e () indicates the time since the last maintenance.

[0073] φ1(e)=min( T open ( e ) / T nom ( e ), 1) (Opening time normalization).

[0074] φ2(e)=max(0,1- F jam ( e ) / F max (Stuck frequency penalty).

[0075] φ3(e)=exp(- λ Δ t maint ( e (Maintenance efficiency degradation) λ=0.05.

[0076] .

[0077] The device health function is registered as an algorithm service and can be invoked by the embedded large model M* via Function Call.

[0078] Optionally, the sub-health alarm closed-loop state algorithm defines states such as alarm trigger, confirmed, dispatched, in progress, verified, and closed, as well as corresponding deterministic transfer functions, and automatically escalates the alarm if the target state is not reached within 72 hours.

[0079] Specifically, define the closed-loop status of sub-health alerts: ; Where ∑ represents the closed-loop system, S represents the set of states, A represents the set of actions, s0 represents the initial state, and S goal Indicates the target state. S goal ={closed}, A={confirm, assign, execute, verify}), T(s,a) represents the deterministic transfer function, S={alarm triggered, confirmed, dispatched, in process, verified, closed}.

[0080] The system automatically executes strategy π; if S is not achieved within 72 hours... goal If so, the alarm will be upgraded.

[0081] Optionally, the private knowledge question-answering knowledge retrieval algorithm, through RAG-enhanced knowledge retrieval operators, supports the answering of professional questions related to rail transit vehicle doors.

[0082] Specifically, the knowledge retrieval operator expression is as follows: ; Where K(q) represents the search results, q represents the user query, argmax represents the maximum index, d represents the knowledge base document, and D kb Represents a knowledge base, Enc q (q) indicates querying the encoder, Enc d (d) indicates the document encoder.

[0083] For example, the question might be: "What causes the V-shaped anomaly in the car door? How should it be handled?" A: Causes: Wear on the lower guide rail, loose lower swing arm, mismatch in closing torque; Solutions: ① Check the straightness of the guide rail; ② Tighten the swing arm; ③ Recalibrate the door controller; ④ Perform 5 door opening and closing tests for verification.

[0084] Optionally, a dynamic front-end generation mechanism for character awareness can also be included, as follows: Define the role-view mapping function: G:R C V; Where G represents the role-view mapping, R represents the role collection, V represents the view object, C represents FunctionCall, C=(f,a), f represents a string, and a represents the function parameter.

[0085] Front-end execution engine updates view: v t+1 =ψ(v t ,r,c); Among them, v t+1 Indicates the new view state, ψ represents the view update engine, and v t 'r' represents the current view, 'r' represents the current role, and 'c' represents the current caller.

[0086] For example, for the layout function f=render_dashboard, the system loads a predefined template: TemplateID=LayoutMap(r),CSSVars=ThemMap(r); Different roles correspond to different information densities and operational permissions, ensuring that human-computer interaction accurately matches job responsibilities. Specifically, TemplateID represents the template identifier, LayoutMap represents the layout mapping table, CSSVars represents the CSS variable set, and ThemMap represents the theme mapping table.

[0087] Example 3

[0088] This embodiment provides a computer device, including one or more processors, a memory, and one or more programs. These programs are stored in the memory and configured to be executed by one or more processors to implement the steps of an intelligent operation and maintenance method for rail transit vehicle doors.

[0089] The computer device can be an edge computing server, an embedded computing device, or a dedicated operation and maintenance terminal, possessing sufficient computing power to support the operation of large language models at the edge. The processor of this computing device can be one or more combinations of a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), or a field-programmable gate array (FPGA). Memory includes, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, solid-state drives, or disk storage.

[0090] The intelligent operation and maintenance method for rail transit doors implemented by this computer equipment includes the following steps: First, the open-source large language model undergoes structured pruning to obtain a first intermediate model. During structured pruning, the computer analyzes the importance of neurons in each layer of the model, removing redundant or less contributing neuronal connections while maintaining the overall performance of the model. The pruning process employs an iterative pruning strategy, with fine-tuning the model after each pruning to ensure that the model's accuracy does not significantly decrease.

[0091] Secondly, the first intermediate model is distilled onto the maintenance task dataset to obtain the second intermediate model. During the distillation process, the computer equipment uses a teacher-student model architecture, where the first intermediate model acts as the teacher model, guiding the smaller student models to learn key knowledge. The maintenance task dataset used for distillation contains specialized data such as rail transit door fault diagnosis, maintenance guidance, and component replacement records.

[0092] Then, two types of adapters are superimposed on the second intermediate model for fine-tuning to obtain the final embedded large model. The two types of adapters are a task adapter and a domain adapter. The task adapter is optimized for different operational tasks such as fault diagnosis and predictive maintenance, while the domain adapter focuses on adapting knowledge from the rail transit door domain. The fine-tuning process uses a gradient descent algorithm with a low learning rate to avoid model overfitting.

[0093] Next, the computer device inputs the user's natural language and user role information into the embedded big data model, parses the user's role intent, and generates function call instructions. User roles may include different identities such as maintenance engineers, dispatchers, and operation and maintenance managers. The computer device will customize different function permissions and information display methods according to different roles. The embedded big data model analyzes user input through an intent recognition module, extracts key information, and maps it to the corresponding function calls.

[0094] The computer equipment invokes corresponding algorithm services based on the generated function call instructions, including the rail transit equipment health status algorithm, the sub-health alarm closed-loop status algorithm, and the private knowledge question-and-answer knowledge retrieval algorithm. Among these, the rail transit equipment health status algorithm, based on multi-sensor data fusion technology, comprehensively analyzes indicators such as door operating status, component wear degree, and fault frequency to generate equipment health scores and early warning information. The sub-health alarm closed-loop status algorithm tracks the entire process from alarm generation to problem resolution, recording the processing status and time nodes of each stage to ensure traceability of the maintenance process.

[0095] In a preferred embodiment, the private knowledge question-answering knowledge retrieval algorithm supports answering professional questions about rail transit vehicle doors using a knowledge retrieval operator enhanced by RAG (Retrieval Enhanced Generation). The algorithm first extracts keywords and semantic features from the user's question, then retrieves relevant document fragments from a professional knowledge base, and finally combines the retrieved knowledge with the reasoning capabilities of a large language model to generate accurate and professional answers. The knowledge base contains professional materials such as door design specifications, maintenance manuals, and fault case libraries, supporting multi-level indexing and semantic search to ensure retrieval efficiency and accuracy.

[0096] Finally, the computer device parses the function call results and displays them in a dynamically generated interface based on the user's role. The interface layout, functional modules, and data display methods differ for different roles. For example, the interface for maintenance engineers focuses on detailed fault diagnosis information and maintenance guidance, while the interface for managers emphasizes statistical data and trend analysis. The interface supports multiple interaction methods, including touch operation, voice commands, and gesture recognition, adapting to the needs of different work scenarios.

[0097] The computer equipment is also equipped with data security protection mechanisms, encrypting and storing sensitive data during transmission, and implementing role-based access control to ensure data security and user privacy. Furthermore, the computer equipment supports offline operation, providing basic fault diagnosis and maintenance guidance even in environments with unstable network connections.

[0098] Through the aforementioned technologies, this computer device can efficiently run large language models at the edge, providing comprehensive support for the intelligent operation and maintenance of rail transit doors, significantly improving maintenance efficiency, reducing maintenance costs, and extending equipment lifespan.

[0099] Example 4

[0100] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of an intelligent operation and maintenance method for rail transit doors based on an edge-side large language model.

[0101] The computer-readable storage medium may include, but is not limited to, a hard disk, an optical disk, ROM, RAM, EPROM, EEPROM, flash memory chip, or any type of medium for storing electronic instructions. A computer program may be stored in one or more storage areas of the computer-readable storage medium, and when the computer program is loaded and executed by a processor, it implements the following steps of the intelligent operation and maintenance method for rail transit doors: First, the open-source large language model undergoes structured pruning to obtain a first intermediate model. In this step, the computer program uses a binary mask to score the importance of attention heads and retains or removes weights based on preset thresholds, thus obtaining the pruned first intermediate model. This structured pruning technique effectively reduces model size and computational complexity, enabling large language models to run efficiently on edge computing devices.

[0102] Secondly, the first intermediate model is distilled on the operations and maintenance task dataset to obtain the second intermediate model. In this step, the computer program uses an open-source large language model as the teacher and the first intermediate model as the student to perform distillation on the operations and maintenance task dataset. The loss function combines cross-entropy loss and KL divergence loss, and the optimized result is the second intermediate model. This knowledge distillation process allows the smaller model to learn the knowledge representation capabilities of the larger model, further improving the model's performance on specific tasks.

[0103] Next, two types of adapters are superimposed on the second intermediate model for fine-tuning to obtain the final embedded large model. Specifically, the computer program superimposes a LoRA module and an Adapter module on the second intermediate model, where the LoRA module is used for the attention layer and the Adapter module is used for the FFN layer, to obtain the final embedded large model. This dual-adaptor fine-tuning method can achieve domain-specific adaptation with a small number of trainable parameters while keeping the main parameters of the model unchanged.

[0104] Finally, the user's natural language and user role information are input into the embedded big data model. The user's role intent is parsed, and function call instructions are generated to invoke the corresponding algorithm services. The generated function call results are parsed and displayed in an interface dynamically generated according to the user role. These algorithm services include a rail transit equipment health algorithm, a sub-health alarm closed-loop state algorithm, and a private knowledge question-and-answer knowledge retrieval algorithm.

[0105] In a preferred embodiment, the equipment health algorithm implemented by the computer program calculates the equipment health score by weighting three indicators: normalized door opening time, jamming frequency penalty, and maintenance timeliness decay of rail transit vehicle doors. This algorithm comprehensively considers the door's operating status, failure frequency, and maintenance status, providing data support for operation and maintenance decisions.

[0106] In another preferred embodiment, the computer-implemented sub-health alarm closed-loop state algorithm defines states such as alarm triggering, confirmed, dispatched, in progress, verified, and closed, as well as corresponding deterministic transfer functions, and automatically escalates the alarm if the target state is not reached within 72 hours. This state management mechanism ensures that problems can be handled in a timely manner, preventing minor faults from escalating into major issues.

[0107] In another preferred embodiment, a computer-programmed private knowledge question-answering and knowledge retrieval algorithm, through RAG-enhanced knowledge retrieval operators, supports the answering of specialized questions regarding rail transit vehicle doors. This algorithm can retrieve relevant information from a professional knowledge base and, combined with the understanding and generation capabilities of a large language model, provide users with accurate and professional answers to their questions.

[0108] When executed, the computer program stored on this computer-readable storage medium can fully utilize edge computing resources to achieve intelligent operation and maintenance of rail transit doors, improving maintenance efficiency, reducing operation and maintenance costs, and extending equipment life. Through the natural language understanding and function invocation capabilities of the embedded large-scale model, users with different roles can obtain the necessary operation and maintenance information and services through natural language interaction, greatly enhancing the human-computer interaction experience and work efficiency.

Claims

1. A smart operation and maintenance system for rail transit doors based on an edge-side large language model, characterized in that, include: The intelligent front-end module is used to receive user natural language and user role information and input them into the embedded large model inference module. It dynamically adjusts the layout, theme variables and component visibility status of the intelligent operation and maintenance page of rail transit doors according to the user role to ensure that human-computer interaction accurately matches job responsibilities. An embedded large model inference module is used to construct an embedded large model using a three-stage compression and fine-tuning strategy. First, the open-source large language model is structurally pruned to obtain a first intermediate model; the first intermediate model is distilled on an operation and maintenance task dataset to obtain a second intermediate model; two types of adapters are superimposed on the second intermediate model for fine-tuning to obtain the embedded large model. The core algorithm service module is used to parse user role intent and generate function call instructions, call the corresponding algorithm services, including rail transit equipment health algorithm, sub-health alarm closed-loop status algorithm and private knowledge question answering knowledge retrieval algorithm, parse the algorithm results to generate corresponding charts, and display them on the rail transit door intelligent operation and maintenance page dynamically generated according to user role.

2. A method for intelligent operation and maintenance of rail transit vehicle doors based on an edge-side large language model, characterized in that, Includes the following steps: The open-source large language model is structured and pruned to obtain the first intermediate model; Distill the first intermediate model onto the operation and maintenance task dataset to obtain the second intermediate model; Two types of adapters are superimposed on the second intermediate model for fine-tuning to obtain an embedded large model; The user's natural language and user role information are input into the embedded big model. The user role intent is parsed and function call instructions are generated. The corresponding algorithm services are called. The generated function call results are parsed and corresponding charts are generated and displayed on the intelligent operation and maintenance page of rail transit doors, which is dynamically generated according to the user role. The algorithm services include rail transit equipment health algorithm, sub-health alarm closed-loop state algorithm, and private knowledge question answering knowledge retrieval algorithm.

3. The intelligent operation and maintenance method for rail transit vehicle doors according to claim 2, characterized in that, The first intermediate model is obtained by introducing a binary mask to score the importance of the attention head and retaining or removing weights according to a preset threshold.

4. The intelligent operation and maintenance method for rail transit vehicle doors according to claim 2, characterized in that, The second intermediate model uses the open-source large language model as the teacher and the first intermediate model as the student. Distillation is performed on the operation and maintenance task dataset. The loss function combines cross-entropy loss and KL divergence loss, and the second intermediate model is obtained after optimization.

5. The intelligent operation and maintenance method for rail transit vehicle doors according to claim 2, characterized in that, The embedded large model is obtained by superimposing the LoRA module and the Adapter module on the second intermediate model, wherein the LoRA module is used for the attention layer and the Adapter module is used for the FFN layer.

6. The intelligent operation and maintenance method for rail transit vehicle doors according to claim 2, characterized in that, The equipment health algorithm calculates the equipment health by weighting three indicators: normalized door opening time, jamming frequency penalty, and maintenance time-related degradation of rail transit vehicle doors.

7. The intelligent operation and maintenance method for rail transit vehicle doors according to claim 2, characterized in that, The sub-health alarm closed-loop state algorithm defines states such as alarm trigger, confirmed, dispatched, in progress, verified, and closed, as well as corresponding deterministic transfer functions, and automatically escalates the alarm if the target state is not reached within a preset time.

8. The intelligent operation and maintenance method for rail transit vehicle doors according to claim 2, characterized in that, The proprietary knowledge question-answering knowledge retrieval algorithm, enhanced with RAG knowledge retrieval operators, supports the answering of professional questions related to rail transit vehicle doors.

9. A computer device, characterized in that, It includes one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of the intelligent operation and maintenance method for rail transit doors as described in any one of claims 2-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent operation and maintenance method for rail transit doors as described in any one of claims 2-8.

Images