Multimodal industrial operation and maintenance systems, methods, and applications integrating hierarchical large model architecture and adaptive reinforcement learning

Abstract

This invention discloses a multimodal industrial operation and maintenance system, method, and application that integrates a hierarchical large-scale model architecture with adaptive reinforcement learning. It relates to the fields of industrial automation and artificial intelligence. The system includes an expert layer module, a system layer module, a chief engineer layer module, an adaptive reinforcement learning engine, and a Bayesian collaborative network. The expert layer module receives equipment time-series data and word vectors from a local knowledge manual to generate preliminary operation and maintenance decisions. The system layer module receives the decision outputs from at least two expert layer modules and corresponding system-level documents, performing multimodal data fusion and conflict resolution. The chief engineer layer module receives the outputs from all system layer modules and plant-wide documents to generate a global optimization strategy. This system and method overcome the limitations of single-layer model architectures, the obstacles to multimodal data fusion, and the inefficiencies of dynamic resource allocation, improving the system's intelligence level and optimizing industrial operation and maintenance efficiency. It can be applied in hydropower plants, thermal power plants, and intelligent manufacturing production lines.

Description

Technical Field

[0001] This invention relates to the fields of industrial automation and artificial intelligence, specifically to a multimodal industrial operation and maintenance system and method, and its application that integrates a hierarchical large model architecture with adaptive reinforcement learning. Background Technology

[0002] Currently, with the rapid development of artificial intelligence technology, the intelligent upgrading of industrial operation and maintenance systems faces the following technical bottlenecks:

[0003] 1. Limitations of layered model architecture:

[0004] Existing systems (such as Siemens SIMATIC PCS 7) employ a single AI model that is centralized in the cloud or distributed at the edge, resulting in fragmented decision-making between different levels.

[0005] Device-side models (such as lightweight LSTM networks deployed on PLCs) are limited by computing resources (typical computing power <1 TOPS) and can only perform simple tasks such as threshold alarms, with a decision error rate as high as 18.7% (based on the IEC 62443 standard test set).

[0006] While cloud-based large-scale models (such as Alibaba Cloud ET Industrial Brain) possess complex decision-making capabilities, their decision-making latency reaches 800-1200ms (ISO 23837 standard test results), which cannot meet the real-time control requirements specified in IEC 61131-3 (latency requirement ≤300ms).

[0007] 2. Obstacles to multimodal data fusion:

[0008] Industrial scenarios involve heterogeneous data such as vibration signals (0-20kHz frequency domain), infrared thermal imaging (640×480 resolution), and equipment logs (unstructured text). Existing technologies (such as Huawei Cloud EI Industrial Intelligence Agent) lack a unified semantic interface, leading to failures in cross-level data association. Taking turbine fault diagnosis as an example, the false association rate between control cabinet temperature alarm data (Modbus / TCP protocol) and subsystem vibration spectrum characteristics (MQTT protocol) increased by 23%.

[0009] 3. Inefficiency of dynamic resource allocation:

[0010] Traditional methods (such as the federated learning architecture proposed in CN114217314A) employ a static weight allocation strategy, which cannot adapt to the drift of operating conditions.

[0011] Experiments show that when the hydropower station load changes abruptly (±30% of rated power), network congestion is caused by API call conflicts at the subsystem layer, and the TCP retransmission rate surges from 0.5% to 12.3% (based on Wireshark packet capture analysis).

[0012] After the device firmware is upgraded (such as the PLC being upgraded from v2.1 to v3.0), the data distribution offset (KL divergence > 0.15) causes the decision accuracy of the traditional Bayesian network (CN113570129A) to decrease by 41.2%. Summary of the Invention

[0013] The purpose of this invention is to provide a multimodal industrial operation and maintenance system, method, and application that integrates a hierarchical large model architecture and adaptive reinforcement learning, so as to improve the technical problems mentioned in the above-mentioned technical background.

[0014] This invention is achieved through the following technical solution:

[0015] A multimodal industrial operation and maintenance system integrating a hierarchical large-scale model architecture and adaptive reinforcement learning, including an expert layer module, a system layer module, a chief engineer layer module, an adaptive reinforcement learning engine, and a Bayesian collaborative network.

[0016] The expert layer module is a lightweight DeepSeek model deployed in industrial field control equipment. M I It is used to receive device timing data and word vectors from the local knowledge manual to generate preliminary operation and maintenance decisions;

[0017] The system layer module is a medium-sized DeepSeek model deployed on an industrial control computer. M m It receives the decision outputs from at least two expert-level modules and the corresponding system-level documents, and performs multimodal data fusion and conflict resolution.

[0018] The chief engineer layer module is a complete DeepSeek model deployed on a factory-level server. M c It receives the output of all system-level modules and plant-wide documents, and generates a global optimization strategy.

[0019] The adaptive reinforcement learning engine dynamically adjusts the API call weights between levels based on three-dimensional reward and punishment signals;

[0020] Bayesian cooperative networks are used to model the state transition probabilities between levels.

[0021] Furthermore, multimodal data fusion includes:

[0022] a1. Convert the equipment knowledge manual text into 128-dimensional word vectors using the BERT-Whitening algorithm;

[0023] b1. Perform wavelet denoising and LSTM feature extraction on the sensor time-series data;

[0024] c1. The pre-trained BERT-Emotion model is used to parse the voice / text interaction content of the operation and maintenance personnel and output the emotion value Feeling∈[-5,+5].

[0025] Furthermore, the adaptive reinforcement learning engine includes:

[0026] a2. Reward and punishment fusion network: Gated recurrent unit (GRU) is used, hidden layer dimension d=64, and the activation function is Sigmoid;

[0027] b2. Strategy Optimization Module: The system layer performs the following Bellman equation model update:

[0028] ;

[0029] in, A t Indicates the selection action; Q( M m , A t ) represents the model in the current state. M m Select Action A t Value; or Represents the learning rate ( or ∈[0.05,0.2]); R ' indicates dynamic rewards; A ' represents all possible actions; c It represents the discount factor (discount factor) c =0.99, exponential decay rate 0.95 / cycle), used to balance long-term and short-term interests; t Represents temperature coefficient ( t =0.1), used for the selection and exploration of balance control strategies.

[0030] Furthermore, the three-dimensional reward and punishment signals are human rating (Rating), sentiment analysis (Feeling), and inter-model rating (Model).

[0031] The weight updates of an adaptive reinforcement learning engine satisfy the following relationship:

[0032] ,

[0033] in, w i Indicates the first i The weights of the layer model at time step t, R ( t The ) represents the reward at the current time step t; the learning rate. or ∈[0.05,0.2]; lRepresents the regularization coefficient ( l =0.85, 95% confidence interval ±0.03), used for balancing strategy exploration and development; K represents the weight dimension involved in the calculation, and the L2 (L2 is the norm of 2) penalty coefficient β=1e-4 is used to prevent weight overfitting.

[0034] Furthermore, Bayesian cooperative networks define the conditional probabilities of upper layers to lower layers as satisfying the following relationship, taking the total layer to the system layer as an example:

[0035] ,

[0036] Where c' represents all possible master-level models. i c ∈ R d The learnable parameters from the system layer to the chief engineer layer (dimension d=128). T The matrix transpose operation is represented by the feature mapping function ϕ(•), which is implemented using a two-layer CNN. The first layer uses the ReLU activation function, and the second layer uses the Tanh activation function. The kernel size is 3×3, and the stride is 1.

[0037] Furthermore, it also includes a performance evaluation module for:

[0038] I. Compile contribution indicators for each expert-level module, including decision adoption rate, reasoning time, and user satisfaction index;

[0039] II. When the contribution of an expert module is less than 0.2 for three consecutive evaluation periods (each period is one week), at least one of the following operations will be triggered:

[0040] Model parameter expansion: Upgrade the model size according to the hardware compatibility table;

[0041] Precision quantization compression: FP8 quantization is used to reduce resource consumption;

[0042] Knowledge base update: Retrieve the latest documents and fine-tune the model using the RAG (Retrieval Enhancement and Generation) technique.

[0043] Furthermore, the contribution index satisfies the following relationship:

[0044] ,

[0045] Where α1=0.6, α2=0.3, α3=0.1, KL divergence is calculated by variational inference, and the threshold is <1e-3.

[0046] The multimodal industrial operation and maintenance method based on the aforementioned system includes the following steps:

[0047] S1. Deploy lightweight DeepSeek on edge devicesM I The model loads local knowledge word vectors;

[0048] S2, System Layer Model M m Receive expert decisions and perform multi-source alignment based on an attention mechanism;

[0049] S3, Chief Engineer's Model M c Generate a global strategy and issue optimization instructions through a Bayesian network;

[0050] S4. Update the weights between levels based on real-time reward and punishment signals, and adjust the Bayesian network parameters synchronously. i c .

[0051] Furthermore, in step S2, multi-source alignment includes:

[0052] S21. Impose a weighted voting mechanism on conflict decisions, with the voting weight proportional to the historical accuracy of the expert module;

[0053] S22. Concatenate the complementary decision execution features and input them into the GRU network to generate the fusion result;

[0054] In step S4, the parameter update must satisfy the following:

[0055] S41, Learning Rate or Adjust according to cosine annealing strategy:

[0056] ,

[0057] in, t This indicates the current training step count, with a period T=1000. or min = 0.05 ,or max = 0.2;

[0058] S42, Bayesian network parameters i c Updated every 50 cycles using the variational EM algorithm, with the convergence condition being an ELBO change rate of <0.1%.

[0059] Applications of multimodal industrial operation and maintenance systems or methods that integrate hierarchical large model architecture and adaptive reinforcement learning, as described in any of the preceding items, in hydropower stations, thermal power plants, and intelligent manufacturing production lines.

[0060] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0061] The multimodal industrial operation and maintenance system and method proposed in this invention integrates a hierarchical large model architecture with adaptive reinforcement learning. By establishing a hierarchical structure and dynamic adjustment mechanism, it overcomes the limitations of single-layer model architecture, the obstacles of multimodal data fusion, and the inefficiency of dynamic resource allocation, thereby improving the intelligence level of the system and optimizing the efficiency of industrial operation and maintenance. It can be applied in hydropower stations, thermal power plants, and intelligent manufacturing production lines. Attached Figure Description

[0062] Figure 1 This is a system architecture diagram of the present invention. Detailed Implementation

[0063] The present invention will be further described in detail below with reference to embodiments, but the implementation of the present invention is not limited thereto.

[0064] Example 1

[0065] To facilitate public understanding of the present invention, this embodiment takes a specific example of a multimodal industrial operation and maintenance system that integrates a hierarchical large model architecture and adaptive reinforcement learning, and further illustrates the solution with reference to the accompanying drawings.

[0066] A multimodal industrial operation and maintenance system integrating a hierarchical large model architecture and adaptive reinforcement learning includes an expert layer module, a system layer module, a chief engineer layer module, an adaptive reinforcement learning engine, and a Bayesian collaborative network.

[0067] The expert layer module is a lightweight DeepSeek model deployed in industrial field control equipment. M I It is used to receive device timing data and word vectors from the local knowledge manual to generate preliminary operation and maintenance decisions.

[0068] The system layer module is a medium-sized DeepSeek model deployed on an industrial control computer. M m It receives the decision outputs from at least two expert-level modules and the corresponding system-level documents, and performs multimodal data fusion and conflict resolution.

[0069] Multimodal data fusion includes:

[0070] a1. Convert the equipment knowledge manual text into 128-dimensional word vectors using the BERT-Whitening algorithm;

[0071] b1. Perform wavelet denoising and LSTM feature extraction on the sensor time-series data;

[0072] c1. The pre-trained BERT-Emotion model is used to parse the voice / text interaction content of the operation and maintenance personnel and output the emotion value Feeling∈[-5,+5].

[0073] The chief engineer layer module is a complete DeepSeek model deployed on a factory-level server. M c It receives the output of all system-level modules and plant-wide documents, and generates a global optimization strategy.

[0074] The adaptive reinforcement learning engine dynamically adjusts the API call weights between layers based on three-dimensional reward and punishment signals (human rating, sentiment analysis, and inter-model rating).

[0075] The adaptive reinforcement learning engine includes:

[0076] a2. Reward and punishment fusion network: Gated recurrent unit (GRU) is used, hidden layer dimension d=64, and the activation function is Sigmoid;

[0077] b2. Strategy Optimization Module: Performs improved Bellman equation updates, taking the system layer as an example:

[0078] ;

[0079] in, A t Indicates the selection action, Q( M m , A t ) represents the model in the current state. M m Select Action A t Value, learning rate or ∈[0.05,0.2], R ' indicates dynamic rewards. A ' represents all possible actions, discount factor c =0.99 (exponential decay rate 0.95 / cycle), used to balance short-term and long-term interests, temperature coefficient. t =0.1, used for balancing control strategy selection and exploration.

[0080] The weight updates of an adaptive reinforcement learning engine satisfy the following relationship:

[0081] ,

[0082] in,, w i Indicates the first i The weights of the layer model at time step t, R ( t The ) represents the reward at the current time step t, and the learning rate. or ∈[0.05,0.2], regularization coefficient l=0.85 (95% confidence interval ±0.03), used for balancing strategy exploration and development, K represents the weight dimension involved in the calculation, and the L2 (L2 is the norm of 2) penalty coefficient β=1e-4 is used to prevent weight overfitting.

[0083] Bayesian cooperative networks are used to model the state transition probabilities between levels.

[0084] Bayesian cooperative networks define the conditional probability of an upper layer relative to a lower layer (e.g., the chief engineer layer relative to the system layer) as follows:

[0085] ,

[0086] Where c' represents all possible master-level models. i c ∈ R d The learnable parameters (dimension d=128) from the system layer to the chief engineer layer are used. The feature mapping function ϕ(•) is implemented using a two-layer CNN. The first layer uses the ReLU activation function, and the second layer uses the Tanh activation function. The convolution kernel size is 3×3 and the stride is 1.

[0087] Preferably, it also includes a performance evaluation module for:

[0088] I. Compile contribution indicators for each expert-level module, including decision adoption rate, reasoning time, and user satisfaction index;

[0089] II. When the contribution of an expert module is less than 0.2 for three consecutive evaluation periods (each period is one week), at least one of the following operations will be triggered:

[0090] Model parameter expansion: Upgrade the model size (7B→14B) according to the hardware compatibility table.

[0091] Precision quantization compression: FP8 quantization is used to reduce resource consumption;

[0092] Knowledge base update: Retrieve the latest documents and fine-tune the model using the RAG (Retrieval Enhancement and Generation) technique.

[0093] The contribution index satisfies the following relationship:

[0094] ,

[0095] Where α1=0.6, α2=0.3, α3=0.1, KL divergence is calculated by variational inference, and the threshold is <1e-3.

[0096] Based on the multimodal industrial operation and maintenance methods of the above system, refer to Figure 1 This includes the following steps:

[0097] Step 1: Deploy lightweight DeepSeek on edge devices M I The model loads local knowledge word vectors and time-series data.

[0098] In this step, lightweight DeepSeek models are deployed in the PLCs or edge boxes of each control cabinet. M I As an agent expert, I was responsible for initial data processing and local decision-making. I also transformed the knowledge manual related to the control cabinet into word vectors and input the collected time-series historical data into the model. M I .

[0099] Step 2: System Layer Model M m Receive expert decisions and perform multi-source alignment based on an attention mechanism.

[0100] In this step, each system is equipped with an industrial control computer to deploy a medium-sized DeepSeek model. M m As the person in charge of the agency system, I integrate the decisions of various experts and transform the relevant knowledge manuals of the system into word vectors for the model. M m use.

[0101] Multi-source alignment includes:

[0102] S21. Impose a weighted voting mechanism on conflict decisions, with the voting weight proportional to the historical accuracy of the expert module;

[0103] S22. Concatenate the complementary decision execution features and input them into the GRU network to generate the fusion result.

[0104] Step 3: Chief Engineer Level Model M c Generate a global policy and issue optimization instructions through a Bayesian network.

[0105] In this step, the full version of the DeepSeek model is deployed on the factory-level server. M c As acting chief engineer, I coordinated decisions across various subsystems and converted the factory-level knowledge manual into word vectors for the model. M c use.

[0106] Step 4: Update the inter-level weights based on real-time reward and punishment signals, and simultaneously adjust the Bayesian network parameters. i c .

[0107] Parameter updates must meet the following requirements:

[0108] S41, Learning Rate or Adjust according to cosine annealing strategy:

[0109] ,

[0110] in, t This indicates the current training step count, with a period T=1000. or min = 0.05 ,or max = 0.2;

[0111] S42, Bayesian network parameters i c Updated every 50 cycles using the variational EM algorithm, with the convergence condition being an ELBO change rate of <0.1%.

[0112] In this step, when operations personnel interact with the model, the system calls the APIs of that layer and its next layer (if any), combining local knowledge manuals and feedback from lower-level agents (if any) to generate corresponding responses. Simultaneously, the system can automatically generate operations reports at preset intervals when no one is actively operating the system.

[0113] During operation and maintenance interactions, a hierarchical reinforcement learning mechanism that integrates Bayesian inference is used to dynamically optimize model weights.

[0114] 1. Acquisition of multi-source reward and punishment signals.

[0115] 1.1 Manual Evaluation Signal: Operation and maintenance personnel give the generated text an instant rating of 1 to 10 points (Rating∈N). + Establish an operation log mapping relationship: Rating → <Device ID, Decision Level, Timestamp>.

[0116] 1.2 Emotional semantic signal: The interactive text / speech is parsed by the pre-trained BERT-Emotion model, and the emotional value Feeling∈[-5,+5] is output. Its softmax probability distribution satisfies: P(Feeling≤0)<0.2 (confidence ≥95%).

[0117] Inter-model scoring signal: The decision quality score of the upper-level model on the output of the lower-level model (Model∈[1,10]) is given using LLM-based semantic scoring. The upper-level model is given a clear role and its scoring is guided by a structured Prompt template.

[0118] 2. Reward integration and weight update.

[0119] Design a reward function implemented using a gated recurrent neural network (GRU):

[0120] ;

[0121] Where σ(•) is the Sigmoid activation function, which outputs a normalized reward value. R ( t )∈[0,1], weight matrix W r ∈ R 2 ×d , W m ∈ R 1×d (Hidden layer dimension d=64), bias term b∈ R d Initialize via Xavier.

[0122] Weight updates employ a regularized gradient approach:

[0123] ;

[0124] in, w i Indicates the first i The weights of the layer model at time step t, R ( t The ) represents the reward at the current time step t, and the learning rate. or The regularization coefficient ∈ [0.05, 0.2] is dynamically adjusted through Bayesian optimization. l =0.85 (95% confidence interval ±0.03), used for balancing strategy exploration and development, K represents the weight dimension involved in the calculation, and the L2 penalty coefficient β=1e-4 is used to prevent weight overfitting.

[0125] 3. Bayesian-reinforced collaboration mechanism

[0126] 1) Conditional probability modeling: Construct a three-layer Bayesian network and define the state transition probabilities, satisfying the following relationship:

[0127] ;

[0128] Where c' represents all possible master-level models. i c ∈ R d The learnable parameters from the system layer to the chief engineer layer (dimension d=128) are used. The feature mapping function ϕ(•) adopts a two-layer CNN structure. The first layer uses the ReLU activation function and the second layer uses the Tanh activation function. The convolution kernel size is 3×3 and the stride is 1.

[0129] 2) Dynamic reward update: The Q-value is adjusted based on the posterior probability, satisfying the following relationship:

[0130] ;

[0131] Where E is the conditional expectation, estimated using importance. KL ( P prior || P posterior ) represents the KL divergence between the prior distribution and the subsequent distribution, estimated by the variational method, with the coefficient constraint: α1+α2+α3=1 (default α1=0.6, α2=0.3, α3=0.1).

[0132] 3) Collaborative decision-making update: An improved Bellman equation is adopted, satisfying the following relationship:

[0133] ;

[0134] in, A t Indicates the selection action, Q( M m , A t ) represents the model in the current state. M m Select Action A t Value, learning rate or ∈[0.05,0.2], R ' indicates dynamic rewards. A ' represents all possible actions, discount factor γ=0.99 (exponential decay rate 0.95 / cycle), used to balance short-term and long-term interests, temperature coefficient. t =0.1, used for balancing control strategy selection and exploration.

[0135] 4. Online learning optimization

[0136] Execution per cycle:

[0137] 1) Update Bayesian network parameters through variational inference i c (KL divergence threshold < 1e-3);

[0138] 2) Adjust the gradient of the importance sampling strategy to satisfy the following relationship:

[0139] ;

[0140] 3) Perform a soft update of the model weights, satisfying the following relationship:

[0141] ,

[0142] Among them, the mixing coefficient r =0.85.

[0143] Compared with the prior art, the beneficial effects of this invention are as follows:

[0144] 1. Efficiency improvements brought about by architectural innovation

[0145] Multi-level decision optimization: through the expert layer ( M l ), system layer ( M m ), Chief Engineer Level ( M c The three-level model architecture enables collaboration between device-level local decision-making and system-level global optimization, avoiding the response delay problem caused by the single-dimensional decision-making of traditional single-layer models.

[0146] Dynamic resource allocation efficiency: Based on the weight update formula (including the regularization term λ) of reinforcement learning and Bayesian network collaboration, the inefficiency of traditional static resource allocation is solved.

[0147] 2. Breakthrough in intelligent technology level

[0148] Multimodal data fusion capability: It transforms multimodal inputs such as knowledge manual text, time series data, and voice interaction into unified word vectors, breaking through the limitations of traditional single-modal analysis.

[0149] Adaptive learning mechanism: Based on the synergy of Bayesian network state transition probability modeling (Dirichlet prior distribution) and reinforcement learning dynamic reward and punishment mechanism, real-time policy optimization is achieved.

[0150] 3. Optimization of operation and maintenance costs and reliability

[0151] Reduced operational costs: Deployment through lightweight models (expert level) M I ) and automated report generation (triggered by preset time), reducing the need for manual intervention.

[0152] Enhanced system reliability: Fault tolerance mechanisms in layered architecture (such as system layer) M m The decision fusion function is combined with the anomaly detection capability of dynamic reinforcement learning.

[0153] 4. Performance analysis driven by interpretability

[0154] Reasoning path visualization: DeepSeek model's active thought process display function can record expert level ( M l From the chief engineer level ( Mc The decision-making chain.

[0155] Quantifiable indicators of expert effectiveness and the establishment of a three-tiered evaluation system:

[0156] 1) Response Quality: Calculate the user satisfaction index of the expert module based on the ratings from operations and maintenance personnel and sentiment analysis.

[0157] 2) Decision-making efficiency: Performance parameters of statistical expert models, such as inference time and resource utilization rate;

[0158] 3) Knowledge Coverage: Assess document update needs by comparing word vectors in the knowledge manual with the cosine similarity of expert outputs.

[0159] The multimodal industrial operation and maintenance system and method of the present invention, which integrates hierarchical large model architecture and adaptive reinforcement learning, can be applied in hydropower stations, thermal power plants, and intelligent manufacturing production lines.

[0160] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A multimodal industrial operation and maintenance system integrating a hierarchical large-scale model architecture and adaptive reinforcement learning, characterized by: It includes an expert layer module, a system layer module, a chief engineer layer module, an adaptive reinforcement learning engine, and a Bayesian collaborative network. The expert layer module is a lightweight DeepSeek model deployed in industrial field control equipment. M I It is used to receive device timing data and word vectors from the local knowledge manual to generate preliminary operation and maintenance decisions; The system layer module is a medium-sized DeepSeek model deployed on an industrial control computer. M m It receives the decision outputs from at least two expert-level modules and the corresponding system-level documents, and performs multimodal data fusion and conflict resolution. The chief engineer layer module is a complete DeepSeek model deployed on a factory-level server. M c It receives the output of all system-level modules and plant-wide documents, and generates a global optimization strategy. The adaptive reinforcement learning engine dynamically adjusts the API call weights between levels based on three-dimensional reward and punishment signals; Bayesian collaborative networks are used to model the state transition probabilities between hierarchical levels; Multimodal data fusion includes: a1. Convert the equipment knowledge manual text into 128-dimensional word vectors using the BERT-Whitening algorithm; b1. Perform wavelet denoising and LSTM feature extraction on the sensor time-series data; c1. Parse the voice / text interaction content of the operation and maintenance personnel using a pre-trained BERT-Emotion model, and output the sentiment value Feeling∈[-5,+5]; The adaptive reinforcement learning engine includes: a2. Reward and punishment fusion network: Gated recurrent unit (GRU) is used, hidden layer dimension d=64, and the activation function is Sigmoid; b2. Strategy Optimization Module: The system layer performs the following Bellman equation model update: ； in, A t Indicates the selection action; Q( M m , A t ) represents the model in the current state. M m Select Action A t Value; η Represents the learning rate; R ' indicates dynamic rewards; A ' represents all possible actions; γ It represents a discount factor, used to balance short-term and long-term interests; τ Represents the temperature coefficient, used for the selection and exploration of balance control strategies; The three-dimensional reward and punishment signals are human rating (Rating), sentiment analysis (Feeling), and inter-model rating (Model). The weight updates of an adaptive reinforcement learning engine satisfy the following relationship: ， in, w i Indicates the first i The weights of the layer model at time step t, R ( t ) represents the reward at the current time step t; λ represents the regularization coefficient, used for balancing strategy exploration and development; K represents the weight dimension involved in the calculation, and the L2 penalty coefficient β=1e-4 is used to prevent weight overfitting.

2. The multimodal industrial operation and maintenance system integrating hierarchical large model architecture and adaptive reinforcement learning as described in claim 1, characterized in that, It also includes a performance evaluation module for: I. Compile contribution indicators for each expert-level module, including decision adoption rate, reasoning time, and user satisfaction index; II. When the contribution of an expert module is less than 0.2 for three consecutive evaluation periods, at least one of the following operations will be triggered: Model parameter expansion: Upgrade the model size according to the hardware compatibility table; Precision quantization compression: FP8 quantization is used to reduce resource consumption; Knowledge base update: Retrieve the latest documents and fine-tune the model using the RAG (Retrieval Enhancement and Generation) technique.

3. The multimodal industrial operation and maintenance system integrating hierarchical large model architecture and adaptive reinforcement learning as described in claim 2, characterized in that, The contribution index satisfies the following relationship: ， Where α1=0.6, α2=0.3, α3=0.1, KL divergence is calculated by variational inference, and the threshold is <1e-3.

4. A multimodal industrial operation and maintenance method based on the system described in claim 1, characterized in that, Including the following steps: S1. Deploy lightweight DeepSeek on edge devices M I The model loads local knowledge word vectors; S2, System Layer Model M m Receive expert decisions and perform multi-source alignment based on an attention mechanism; S3, Chief Engineer's Model M c Generate a global strategy and issue optimization instructions through a Bayesian network; S4. Update the weights between levels based on real-time reward and punishment signals, and adjust the Bayesian network parameters synchronously. θ c .

5. The multimodal industrial operation and maintenance method according to claim 4, characterized in that, In step S2, multi-source alignment includes: S21. Impose a weighted voting mechanism on conflict decisions, with the voting weight proportional to the historical accuracy of the expert module; S22. Concatenate the complementary decision execution features and input them into the GRU network to generate the fusion result; In step S4, the parameter update must satisfy the following: S41, Learning Rate η Adjust according to cosine annealing strategy: ， in, t This indicates the current training step count, with a period of T=1000; S42, Bayesian network parameters θ c Updated every 50 cycles using the variational EM algorithm, with the convergence condition being an ELBO change rate of <0.1%.

6. The application of the multimodal industrial operation and maintenance system or method that integrates hierarchical large model architecture and adaptive reinforcement learning as described in any one of claims 1 to 5 in hydropower stations, thermal power plants, and intelligent manufacturing production lines.

Images