Double-channel adaptive continual learning method and device for small sample industrial scene
By employing a dual-channel adaptive continuous learning method, utilizing structural knowledge injection and semantic cognition to optimize the channels, and combining noise filtering and parameter isolation mechanisms, the problem of model generalization ability and robustness under small sample conditions in industrial scenarios is solved, achieving efficient and forget-free model updates.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING EASY TIMES DIGITAL TECH
- Filing Date
- 2025-12-19
- Publication Date
- 2026-04-28
AI Technical Summary
In the process of industrial intelligent transformation, deep learning models face problems such as insufficient generalization ability with few samples, poor robustness in high-noise environments, and catastrophic forgetting during continuous learning. In particular, the transfer performance of models degrades sharply under few sample conditions, making it difficult to adapt quickly, and traditional methods cannot achieve deep adaptation.
A dual-channel adaptive continuous learning approach is adopted. The rule subset is encoded into an industrial domain adapter and embedded into the pre-trained model through the structured knowledge injection channel. The semantic cognition optimization channel is combined to perform knowledge enhancement prompting engineering and perform incremental fine-tuning with noise filtering. Dynamic confidence filters and task-aware parameter isolation mechanisms are used to prevent forgetting. An online continuous learning and model update mechanism is deployed.
It achieves efficient and robust continuous learning under small sample conditions, reduces dependence on large-scale labeled data, improves the model's generalization ability and robustness in industrial scenarios, prevents catastrophic forgetting, and improves the accuracy and efficiency of decision-making.
Smart Images
Figure CN121390178B_ABST
Abstract
Description
Technical Field
[0001] The embodiments in this specification relate to the field of artificial intelligence technology, and in particular to a dual-channel adaptive continuous learning method for small-sample industrial scenarios. Background Technology
[0002] In the process of industrial intelligent transformation, the deployment and application of deep learning models have long been constrained by real-world limitations such as scarce labeled data, severe on-site noise interference, and dynamic evolution of production line conditions. This has led to key technical bottlenecks in the actual operation of existing models, including weak generalization ability, poor robustness, and high maintenance costs. Specifically, traditional continuous learning methods have the following three prominent problems in industrial scenarios:
[0003] The generalization ability of small samples is seriously insufficient: when new industrial production lines and new equipment are launched, there are often only a very small number (usually <50 cases) of labeled samples available. However, the general large model fine-tuning method relies on a large amount of supervised data. Under small sample conditions, the transfer performance degrades sharply and it is difficult to support rapid adaptation.
[0004] Poor robustness of models in high-noise environments: Industrial sensor data generally suffers from low signal-to-noise ratio (often <20dB), time-series drift, and sudden anomalies. Traditional training strategies fail to effectively identify and suppress noisy samples, resulting in drastic fluctuations in model performance (usually 5%~10%), which affects the reliability of decision-making.
[0005] The continuous learning process is prone to catastrophic forgetting: when the model learns new tasks online, if full parameter fine-tuning is used, it will overwrite historical knowledge, resulting in a significant decrease in the accuracy of recognizing old working conditions; if a fixed backbone network strategy is used, deep adaptation cannot be achieved, which limits the model's evolutionary ability.
[0006] Therefore, a better solution is urgently needed. Summary of the Invention
[0007] In view of this, embodiments of this specification provide a dual-channel adaptive continuous learning method for small-sample industrial scenarios. One or more embodiments of this specification also relate to a dual-channel adaptive continuous learning device for small-sample industrial scenarios, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.
[0008] According to a first aspect of the embodiments of this specification, a dual-channel adaptive continuous learning method for small-sample industrial scenarios is provided, comprising:
[0009] Obtain a small sample industrial training dataset, which includes multimodal industrial data, corresponding annotations, and associated rule subsets;
[0010] A dual-channel adaptive continuous learning engine is constructed, which includes a structural knowledge injection channel and a semantic cognitive optimization channel. The structural knowledge injection channel is used to encode a subset of rules into an industrial domain adapter and embed it into a pre-trained model. The semantic cognitive optimization channel is used to perform knowledge enhancement prompting engineering on multimodal inputs based on industrial instruction templates.
[0011] Incremental fine-tuning with noise filtering is performed, including using a dynamic confidence filter to identify and deweight samples in a small industrial training dataset, employing a task-aware parameter isolation mechanism to fine-tune only the sub-networks relevant to the new task, and combining elastic weight solidification and replay buffers to prevent catastrophic forgetting.
[0012] Deploy online continuous learning and model update mechanisms, including incremental fine-tuning triggered by semantic drift detection, and manage model versions.
[0013] In one possible implementation, obtaining a small-sample industrial training dataset includes:
[0014] Collect multimodal industrial data, including text modality, visual modality, time-series modality, and structured knowledge modality;
[0015] Alignment of multimodal industrial data based on equipment identifiers and timestamps;
[0016] Industrial terms in the text are linked to rule base nodes using named entity recognition and entity linking technologies.
[0017] Construct a training sample set, where each sample includes a multimodal input, a label, and a subset of associated rules.
[0018] In one possible implementation, the structural knowledge injection channel includes encoding a subset of rules into an industry domain adapter, which is a lightweight fully connected network; embedding the industry domain adapter between Transformer layers of a pre-trained industrial large model; fine-tuning only the parameters of the industry domain adapter while freezing the backbone network parameters of the pre-trained industrial large model.
[0019] In one possible implementation, the semantic cognition optimization channel includes: constructing an industrial instruction template library, which includes fillable templates for equipment, phenomena, and operations; concatenating the industrial instruction templates with multimodal inputs to form contextualized prompts; and introducing a semantic correction algorithm to perform rule consistency verification on the model output and to provide feedback correction signals when rules are violated.
[0020] In one possible implementation, a dynamic confidence filter is used to identify and reduce noise in samples from a small industrial training dataset, including:
[0021] For each sample, calculate the confidence level of the model prediction;
[0022] If the confidence level is below 0.3 or above 0.95, it is judged as a noise sample;
[0023] Noisy samples are assigned a low first weight, and normal samples are assigned a second weight; wherein the first weight is less than the second weight.
[0024] In one possible implementation, the task-aware parameter isolation mechanism includes:
[0025] An importance matrix of key parameters for historical tasks is solidified using flexible weights.
[0026] Construct a replay buffer to store historical high-confidence samples;
[0027] Define a composite loss function, which includes the new task loss calculated based on sample weights, the replay loss calculated based on samples in the replay buffer, and the elastic weight solidification regularization term calculated based on parameter importance and original parameter values.
[0028] Only the subnetworks relevant to the new task are fine-tuned, including the industry domain adapter, cue vectors, and classification head.
[0029] One possible implementation involves deploying an online continuous learning and model update mechanism, including:
[0030] Calculate the average semantic distance between the new sample and the historical knowledge base; if the average semantic distance exceeds 0.35, trigger incremental fine-tuning; complete the fine-tuning within 24 hours and generate a new version of the model; use semantic version numbers to manage model versions, and support A / B testing and rollback.
[0031] According to a second aspect of the embodiments of this specification, a dual-channel adaptive continuous learning device for small-sample industrial scenarios is provided, comprising:
[0032] The data acquisition module is configured to acquire a small sample industrial training dataset, which includes multimodal industrial data, corresponding annotations, and associated rule subsets.
[0033] The engine building module is configured to build a dual-channel adaptive continuous learning engine, which includes a structural knowledge injection channel and a semantic cognitive optimization channel. The structural knowledge injection channel is used to encode a subset of rules into an industrial domain adapter and embed it into a pre-trained model. The semantic cognitive optimization channel is used to perform knowledge enhancement prompting engineering on multimodal inputs based on industrial instruction templates.
[0034] The model fine-tuning module is configured to perform incremental fine-tuning with noise filtering, including using a dynamic confidence filter to identify and deweight samples in a small industrial training dataset, employing a task-aware parameter isolation mechanism to fine-tune only the sub-networks relevant to the new task, and combining elastic weight solidification and replay buffers to prevent catastrophic forgetting.
[0035] The model deployment module is configured to deploy online continuous learning and model update mechanisms, including incremental fine-tuning triggered based on semantic drift detection, and managing model versions.
[0036] According to a third aspect of the embodiments of this specification, a computing device is provided, comprising:
[0037] Memory and processor;
[0038] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the dual-channel adaptive continuous learning method for small-sample industrial scenarios described above.
[0039] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the dual-channel adaptive continuous learning method for small-sample industrial scenarios described above.
[0040] According to a fifth aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the dual-channel adaptive continuous learning method for small-sample industrial scenarios described above.
[0041] This specification provides a dual-channel adaptive continuous learning method and apparatus for small-sample industrial scenarios, including: constructing a dual-channel adaptive continuous learning engine, which includes a structural knowledge injection channel and a semantic cognitive optimization channel. The structural knowledge injection channel is used to encode a subset of rules into an industrial domain adapter and embed it into a pre-trained model, while the semantic cognitive optimization channel is used to perform knowledge enhancement prompting engineering on multimodal inputs based on industrial instruction templates; performing incremental fine-tuning with noise filtering, including using a dynamic confidence filter to identify and reduce the weight of samples in the small-sample industrial training dataset, and using a task-aware parameter isolation mechanism to fine-tune only the sub-networks related to new tasks, and combining elastic weight solidification and replay buffers to prevent catastrophic forgetting. This addresses the technical bottlenecks in industrial scenarios, such as weak model generalization, poor robustness, and easy forgetting during continuous learning caused by small sample sizes, high noise, and dynamic evolution. Through the dual-channel collaborative mechanism of "structural knowledge injection" and "semantic cognitive optimization," it achieves efficient, robust, and forgetting-free continuous learning. Attached Figure Description
[0042] Figure 1 This is a flowchart of a dual-channel adaptive continuous learning method for small-sample industrial scenarios provided in one embodiment of this specification;
[0043] Figure 2 This is a schematic diagram of the structure of a dual-channel adaptive continuous learning device for small-sample industrial scenarios provided in one embodiment of this specification;
[0044] Figure 3 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation
[0045] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.
[0046] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.
[0047] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."
[0048] This specification provides a dual-channel adaptive continuous learning method for small-sample industrial scenarios. This specification also relates to a dual-channel adaptive continuous learning device for small-sample industrial scenarios, a computing device, and a computer-readable storage medium, which are described in detail in the following embodiments.
[0049] See Figure 1 , Figure 1A flowchart of a dual-channel adaptive continuous learning method for small-sample industrial scenarios, according to an embodiment of this specification, is shown, specifically including the following steps.
[0050] Step 101: Obtain a small sample industrial training dataset, which includes multimodal industrial data, corresponding annotations, and associated rule subsets.
[0051] In one possible implementation, obtaining a small-sample industrial training dataset includes: collecting multimodal industrial data, including textual, visual, temporal, and structured knowledge modalities; aligning the multimodal industrial data based on equipment identifiers and timestamps; linking industrial terms in the text to rule base nodes using named entity recognition and entity linking technologies; and constructing a training sample set, where each sample includes multimodal inputs, labels, and associated rule subsets.
[0052] Among these, small-sample industrial training datasets refer to datasets containing a small number of labeled samples collected in industrial scenarios, used for model training and evaluation. Multimodal industrial data refers to industrial data from different sources and types, such as text, images, and time series data, providing comprehensive environmental information. Textual modality refers to industrial data existing in textual form, such as logs and reports, which can be used for natural language processing tasks. Visual modality refers to industrial data existing in image or video form, such as monitoring footage and equipment photos, to support visual analysis. Time-series modality refers to industrial data collected in chronological order, such as sensor readings, used for time series analysis. Structured knowledge modality refers to industrial knowledge stored in a structured form, such as databases and knowledge graphs, capable of representing domain rules and relationships. Equipment identifiers refer to codes or names that uniquely identify industrial equipment, used for data association and tracking. Timestamps refer to markers that record the time of data collection to synchronize data from different modalities. Named entity recognition refers to techniques for identifying and classifying entities such as equipment names and parameters from text, used for information extraction. Entity linking technology refers to the technique of linking entities in text to corresponding nodes in a knowledge base, which can disambiguate and enrich semantics. Industry terms refer to words or phrases specific to an industry, representing specific concepts or operations. Rule base nodes refer to units in a knowledge base that store rules or knowledge to organize and manage domain logic. Training sample sets refer to the collection of data samples used for training machine learning models, for model learning and optimization. Multimodal inputs refer to input data containing multiple data types such as text, images, and time series, providing rich features. Labels refer to the category or output value of a sample, used to supervise model training. Rule subsets refer to a set of task-specific rules extracted from the rule base, used to guide model behavior.
[0053] In practical applications, the following multi-source heterogeneous data are collected synchronously from the industrial site: text modality: equipment operating procedures, process change notices, and fault maintenance logs; visual modality: equipment appearance images, production line monitoring video frames, and AR operation interface screenshots; temporal modality: sensor sampling sequences for temperature, pressure, vibration, current, etc. (sampling frequency ≥ 1Hz); and structured knowledge modality: equipment operation and maintenance rule base, safety specification base, and historical fault case base (stored in the form of triples).
[0054] Data alignment and annotation: Based on device ID and timestamp, visual frames, time-series segments and corresponding text descriptions are aligned; through named entity recognition (NER) and entity linking technology, industrial terms in the text (such as "centrifugal pump" and "reflux ratio") are linked to rule base nodes.
[0055] Constructing the training sample set:
[0056]
[0057] in, The total number of samples is the multimodal input (including visual, temporal, and textual features). For labels (such as fault type). This is a subset of the rules for association.
[0058] As a concrete example: In a chemical plant equipment monitoring scenario, the system collects text-based operation and maintenance logs, visual-based infrared thermal images, time-series-based temperature and pressure sensor data, and a structured knowledge-based safety rule base. It aligns multimodal data based on the device identifier Pump001 and the timestamp 2023-10-01 08:00:00. Using named entity recognition technology, it extracts industrial terms such as "centrifugal pump overheating" from the logs and links them to the rule base node "prohibit idling" using entity linking technology. A training sample set is constructed, with each sample including multimodal inputs such as text descriptions, image frames, sensor sequences, labels such as "normal" or "faulty," and associated rule subsets such as "pump operation specifications."
[0059] This application improves the model's generalization ability and robustness under small sample conditions by integrating multimodal data and structured knowledge, reduces the dependence on a large amount of labeled data, and enhances the accuracy and efficiency of decision-making in industrial scenarios.
[0060] Step 102: Construct a dual-channel adaptive continuous learning engine. The dual-channel adaptive continuous learning engine includes a structural knowledge injection channel and a semantic cognitive optimization channel. The structural knowledge injection channel is used to encode a subset of rules into an industrial domain adapter and embed it into the pre-trained model. The semantic cognitive optimization channel is used to perform knowledge enhancement prompting engineering on multimodal inputs based on industrial instruction templates.
[0061] This step constructs two parallel channels to process structured knowledge and semantic cognition respectively.
[0062] The dual-channel adaptive continuous learning engine can refer to a machine learning system that integrates a structural knowledge injection channel and a semantic cognitive optimization channel to coordinate rule injection and semantic optimization to support continuous model adaptation. The structural knowledge injection channel can be a processing module that encodes a subset of rules into an industrial domain adapter and embeds it into a pre-trained model, enabling the conversion of industrial rules into model parameters. The semantic cognitive optimization channel can be a processing module that performs knowledge-enhanced hint engineering on multimodal inputs based on industrial instruction templates, improving the accuracy of the model's semantic understanding. The industrial domain adapter can be a lightweight neural network component that encodes industrial rules and embeds them between model layers, injecting domain knowledge in a fine-tuning manner. The pre-trained model can be a basic machine learning model pre-trained on a large amount of data to provide general feature representations. The industrial instruction template can be a text template containing populated industrial elements such as equipment and operations, capable of generating context-sensitive hints. Knowledge-enhanced hint engineering can be a technique that optimizes model input by adding knowledge-related hints, enhancing the model's utilization of industrial knowledge.
[0063] As a concrete example: In the scenario of smart grid fault diagnosis, a dual-channel adaptive continuous learning engine is constructed. The structural knowledge injection channel encodes a subset of rules, such as "cut off the power when the transformer is overloaded," into an industrial domain adapter and embeds it into a pre-trained Transformer model. The semantic cognition optimization channel performs knowledge enhancement prompt engineering on multimodal inputs such as text reports, current waveform images, and voltage time series data based on industrial instruction templates, such as "when [equipment] exhibits [phenomenon], [operation] should be performed." This generates prompts such as "when the transformer experiences excessive temperature, the load should be checked." The engines work together to complete incremental model updates using only a small number of new samples, thereby improving diagnostic accuracy.
[0064] Through a dual-channel collaborative mechanism, the model can more effectively utilize industry rules and semantic prompts, enhance generalization ability and decision reliability under small sample sizes, reduce dependence on large amounts of labeled data, and support continuous learning without forgetting historical knowledge.
[0065] In one possible implementation, the structural knowledge injection channel includes encoding a subset of rules into an industry domain adapter, which is a lightweight fully connected network; embedding the industry domain adapter between Transformer layers of a pre-trained industrial large model; fine-tuning only the parameters of the industry domain adapter while freezing the backbone network parameters of the pre-trained industrial large model.
[0066] In practical applications, the Knowledge Injection Channel (BIC) receives a subset of rules as input. (e.g., "pump cannot run dry" "reactor temperature ≤ 300℃"); encode the rules as industrial adapters, which are lightweight fully connected networks; embed the adapters between the Transformer layers of a pre-trained large industrial model (e.g., an industrial cognitive base model), fine-tune the adapter parameters, and freeze the backbone network.
[0067] In one possible implementation, the semantic cognition optimization channel includes: constructing an industrial instruction template library, which includes fillable templates for equipment, phenomena, and operations; concatenating the industrial instruction templates with multimodal inputs to form contextualized prompts; and introducing a semantic correction algorithm to perform rule consistency verification on the model output and to provide feedback correction signals when rules are violated.
[0068] In practical applications, the Semantic Optimization Channel: Input: Multimodal input ; Construct an industrial instruction template library (e.g., "When [equipment] exhibits [phenomenon], [operation] should be performed"); Design a Knowledge-Enhanced Prompt Engineering framework: concatenate instruction templates with inputs to form contextualized prompts; Introduce semantic correction algorithms to perform rule consistency checks on model outputs, and if rules are violated, feedback correction signals are provided.
[0069] Step 103: Perform incremental fine-tuning with noise filtering, including using a dynamic confidence filter to identify and deweight samples in a small industrial training dataset, employing a task-aware parameter isolation mechanism to fine-tune only the sub-networks relevant to the new task, and combining elastic weight solidification and replay buffers to prevent catastrophic forgetting.
[0070] Incremental fine-tuning with noise filtering refers to a machine learning method that integrates noise filtering mechanisms during model updates to improve the robustness and stability of the model under noisy industrial data. Dynamic confidence filters refer to a component that dynamically evaluates sample quality based on model prediction confidence, automatically identifying and handling low-quality or anomalous samples. A sample can refer to a single data instance in an industrial training dataset, containing input and output information, which can be used for model training and evaluation. Noise identification refers to the process of detecting noise or erroneous samples from data to improve data quality and model reliability. Weight reduction refers to techniques that reduce the weight or influence of noisy samples during training to minimize the negative impact of noise on model updates. Task-aware parameter isolation mechanisms refer to a mechanism for identifying and isolating model parameters related to new tasks, avoiding the forgetting of historical knowledge caused by global parameter updates. New tasks can refer to newly emerging learning tasks or scenarios during continuous learning to expand the model's adaptability. Subnetworks can refer to specific parts of the network structure in the model, such as specific layers or modules, used for updating target parameters. Elastic weight solidification refers to a continuous learning technique that protects important parameters through regularization, preventing the model from forgetting historical task knowledge. A replay buffer refers to a cache that stores historical samples, used to replay old samples when training on a new task to maintain performance on older tasks. Catastrophic forgetting refers to the phenomenon where a model forgets previously learned knowledge when learning a new task, describing a key challenge in continuous learning.
[0071] As a concrete example: In a smart manufacturing quality inspection scenario, incremental fine-tuning with noise filtering is performed. A dynamic confidence filter is used to identify noise in a small industrial training dataset containing 50 samples, such as identifying 5 low-confidence samples and downweighting them. A task-aware parameter isolation mechanism is adopted to fine-tune only the sub-networks related to new tasks, such as defect classification, including two fully connected layers. Elastic weights are used to solidify the importance of calculated parameters, and a replay buffer is used to store 20 historical high-confidence samples for mixed training. The entire process is completed within 2 hours, effectively preventing catastrophic forgetting and improving the model's accuracy on new tasks to 92%.
[0072] By integrating noise filtering and parameter isolation mechanisms, the model can handle industrial noise data more robustly, reduce interference during training, and ensure that historical knowledge is preserved, thereby improving the efficiency and reliability of continuous learning and supporting the long-term stable operation of industrial systems in dynamic environments.
[0073] In one possible implementation, a dynamic confidence filter is used to identify and reduce the weight of samples in a small industrial training dataset, including: for each sample, calculating the model prediction confidence; if the confidence is lower than 0.3 or higher than 0.95, it is determined to be a noise sample; assigning a low first weight to the noise sample and a second weight to the normal sample; wherein the first weight is less than the second weight.
[0074] In practical applications, a dynamic confidence filter is used for each new sample. Calculate the confidence level of the model prediction ;like (Low threshold, such as 0.3) or (High threshold, such as 0.95), then it is judged as a noise sample (the former is an anomaly, the latter is out of distribution); noisy samples are assigned low weights. Normal sample weights .
[0075] In one possible implementation, the task-aware parameter isolation mechanism includes: using elastic weights to solidify the importance matrix of key parameters of historical tasks; constructing a replay buffer to store historical high-confidence samples; defining a composite loss function, which includes a new task loss calculated based on sample weights, a replay loss calculated based on samples in the replay buffer, and an elastic weight-solidified regularization term calculated based on parameter importance and original parameter values; and fine-tuning only the sub-networks related to the new task, which include an industry domain adapter, cue vectors, and a classification head.
[0076] In practical applications, Task-Aware Parameter Isolation employs Elastic Weight Consolidation (EWC) to calculate the importance matrix of key parameters from historical tasks; it constructs a replay buffer to store a small number of historical high-confidence samples; and it defines a composite loss function.
[0077]
[0078] in, This represents the total composite loss value for the current training step. In the new task loss, Represents the index of the sample in the current new task batch. sample Confidence weights This represents the basic loss function (usually cross-entropy loss). This represents the input data and corresponding labels in the current new task. In the replay loss, This represents the replay loss coefficient (a weighted hyperparameter used to balance the learning of new tasks with the review of old knowledge). This represents the index of the sampled data in the playback buffer. This represents historical high-confidence samples and their labels randomly sampled from the replay buffer (used to prevent "catastrophic forgetting" when the model is learning new tasks). In the EWC regularization term, This represents the regularization coefficient (used to control the penalty for changes in model parameters). Indicates the index of the model network parameters. The diagonal elements representing the parametric importance matrix, This represents the parameters of the model currently being trained. This represents the historically optimal parameter.
[0079] Fine-tune only the sub-networks relevant to the new task (such as adapters, cue vectors, and classification heads), and freeze the rest of the parameters.
[0080] Step 104: Deploy an online continuous learning and model update mechanism, including incremental fine-tuning triggered by semantic drift detection and managing model versions.
[0081] One possible implementation involves deploying an online continuous learning and model update mechanism, including: calculating the average semantic distance between new samples and the historical knowledge base; triggering incremental fine-tuning if the average semantic distance exceeds 0.35; completing the fine-tuning within 24 hours and generating a new version of the model; and managing model versions using semantic version numbers to support A / B testing and rollback.
[0082] Among these, online continuous learning and model update mechanisms can refer to a system component that automatically learns and updates the model at runtime, enabling continuous adaptation and optimization of the model. New samples can refer to newly acquired data instances during continuous learning, providing new learning information to update the model. A historical knowledge base can refer to a repository containing past learning tasks and knowledge, used for comparison and maintaining knowledge consistency. Average semantic distance can refer to the average difference between new samples and the historical knowledge base in the semantic space, used to assess changes in data distribution and trigger learning. Incremental fine-tuning can refer to a learning method that makes small-scale parameter adjustments based on an existing model, used to adapt to new data without complete retraining. A new version model can refer to a new model iteration generated through fine-tuning, capable of providing improved performance or adapting to new tasks. Semantic versioning can refer to a naming system that uses version numbers to represent the meaning of model changes, used to manage model versions and track changes. A model version can refer to the state or iteration of a model at a specific point in time, supporting version control and rollback. A / B testing can refer to an experimental method for comparing the performance of two model versions, used to evaluate which version is more effective. Rollback refers to the process of reverting a system to a previous stable version, which can handle issues with the new version and ensure system stability.
[0083] In practical applications, semantic drift detection involves calculating the average semantic distance between a new sample and the historical knowledge base. Incremental update triggered: if If so, the fine-tuning process is initiated; Model self-optimization: Fine-tuning is completed within 24 hours, generating a new version of the model; Version management: Semantic version numbers are used, supporting A / B testing and rollback.
[0084] As a concrete example: In the scenario of predictive maintenance of industrial equipment, an online continuous learning and model update mechanism is deployed to calculate the average semantic distance between newly acquired vibration sensor samples and the historical knowledge base. When the distance exceeds 0.35, the incremental fine-tuning process is automatically triggered. The system completes the fine-tuning within 24 hours and generates a new version model, such as V2.1.0. Semantic version numbers are used to manage model versions, supporting A / B testing to compare the performance of the new and old models on real-time data, and allowing rollback to the stable version V2.0.5 in case of problems.
[0085] Through automated online learning and version management, the model can quickly adapt to changes in the industrial environment, reduce the need for manual intervention, improve system reliability and maintenance efficiency, and ensure that the model update process is stable and controllable.
[0086] Because the model not only uses data-driven semantic prompts for fine-tuning when learning new tasks, but also explicitly injects structured knowledge such as equipment operation and maintenance rules and safety specifications through industrial adapters, its parameter update direction is always constrained by industrial logic. As a result, it can complete efficient transfer with only a small number of samples (<50 cases), shortening the model adaptation cycle by more than 12 times and significantly reducing the dependence on large-scale labeled data and the cost of manual parameter tuning.
[0087] Furthermore, the dynamic confidence filter and task-aware parameter isolation mechanism designed in this application enable the model to automatically identify and downweight anomalous samples in high-noise industrial environments with a signal-to-noise ratio below 20dB. Simultaneously, a hybrid strategy of Elastic Weight Consolidation (EWC) and replay buffer protects key parameters of historical knowledge. This mechanism effectively suppresses model performance fluctuations, controlling performance fluctuations under high noise levels to within 3% (compared to 5%–10% for traditional methods), and ensuring that the accuracy of identifying historical working conditions remains above 95%. This achieves continuous evolution capabilities that are forget-free, highly robust, and require minimal intervention, providing a solid foundation for the long-term autonomous operation of industrial intelligent agents.
[0088] In one overall embodiment, the online model evolution of an energy management system is described as follows.
[0089] 1. System Composition
[0090] Data sources include:
[0091] Text: Daily enterprise energy consumption report, energy-saving renovation plan, carbon emission accounting guide; Visual: Infrared thermal imaging image of power distribution room equipment, OCR image of electricity meter readings; Time series: Hourly data such as total enterprise electricity consumption, steam consumption, compressed air flow, etc. (signal-to-noise ratio ≈ 18dB); Structured knowledge: Enterprise-industry-energy efficiency benchmark value knowledge base provided by the enterprise.
[0092] System Module: Integrated into the "Energy Management Service Platform", serving as its core continuous learning engine, it includes all functional modules such as dual-channel learning, noise filtering, and parameter isolation.
[0093] 2. Implementation Steps
[0094] Step 1: The initial model was trained based on three months of historical data from the enterprise;
[0095] Step 2: When a new chemical company (providing only 32 energy consumption samples) joins the platform, the online learning process is triggered;
[0096] Step 3: The system automatically constructs a subset of rules for "abnormal energy consumption - equipment overheating - policy violation" and encodes it as an industrial adapter;
[0097] Step 4: The dynamic confidence filter identifies 8 high-noise samples (such as instantaneous sensor jumps) and assigns them low weights;
[0098] Step 5: Fine-tune only the adapter and cue vectors, freeze the backbone network, and complete the model update within 22 hours;
[0099] Step 6: After the new model goes online, it will automatically monitor the company's energy consumption anomalies and generate energy-saving suggestions.
[0100] 3. Effects
[0101] The model completed effective transfer with only 32 samples, shortening the adaptation period by 12 times (traditional methods require more than 10 days); in a high-noise environment with a signal-to-noise ratio of 18dB, the model performance fluctuation was only 2.7% (traditional continuous learning methods fluctuated by 8.4%); the energy consumption diagnosis accuracy for 100 historical enterprises remained at 96.3%, with no significant forgetting; manual intervention was reduced by 68%, supporting the platform to achieve "daily" dynamic monitoring of energy consumption and helping enterprises save electricity.
[0102] In another embodiment, the rapid generalization of the virtual simulation training model is described as follows.
[0103] 1. System Composition
[0104] Data source:
[0105] Text: Emergency plan for crude oil leaks containing hydrogen sulfide, operating procedures for refining equipment; Visual: AR inspection video, high-definition images of equipment appearance, screenshots of DCS operation interface; Timing: Data from temperature, pressure, and level sensors of the reactor (sampling frequency 1Hz, including sudden noise); Structured knowledge: Equipment knowledge graph (including safety rules such as "pumps must not run dry" and "reactor temperature ≤300℃").
[0106] System Module: Embedded platform AI training engine, serving as its core for small sample scenario adaptation.
[0107] 2. Implementation Steps
[0108] Step 1: The basic model has covered 6 process scenarios including refining, storage and transportation;
[0109] Step 2: When a new "catalytic cracking" unit training requirement is added, only 45 sets of multimodal samples (including 15 sets of data with noise) are collected.
[0110] Step 3: The system constructs a rule subgraph of "catalytic cracking-reaction temperature-catalyst activity" and injects dual-channel structural knowledge channels;
[0111] Step 4: The semantic cognition optimization channel loads the instruction template "When [the device] experiences [overheating], the feed rate should be reduced";
[0112] Step 5: Perform fine-tuning of the composite loss with EWC regularization and replay buffer, and complete model evolution within 20 hours;
[0113] Step 6: The new model supports trainees in conducting emergency response training for catalytic cracking in an AR environment.
[0114] 3. Effects
[0115] The new scenario model migration requires only 45 samples, with an adaptation period of less than 24 hours, reducing the workload of manual annotation by 75%; under conditions containing sudden noise (such as the instantaneous zeroing of pressure sensors), the model performance fluctuation is controlled within 2.9%; the recognition accuracy of the original 6 major process scenarios is maintained above 95.8%, achieving continuous learning without forgetting; the efficiency of training resource generation is improved by 15 times, supporting the digital transformation of simulation training in the upstream, midstream and downstream of Sinopec.
[0116] The two embodiments described above cover two core industrial scenarios: dual-carbon energy consumption management and high-risk process training. They fully demonstrate the significant technical advantages of this application in terms of efficient transfer in small samples, high noise robustness, and continuous evolution without forgetting, and have high feasibility and wide promotion value.
[0117] In summary, this application includes the following beneficial effects.
[0118] 1. Dual-channel adaptive continuous learning architecture for industrial scenarios
[0119] The first parallel architecture combining "structural knowledge injection channel" and "semantic cognition optimization channel" is proposed.
[0120] Structured knowledge injection channel: Encode structured knowledge such as equipment operation and maintenance rules, safety specifications, and process constraints into a lightweight industrial adapter, embed it between the Transformer layers of the pre-trained large model, and only fine-tune the adapter parameters;
[0121] Semantic cognitive optimization channel: Construct an industrial instruction template library (such as "When [equipment] exhibits [phenomenon], [operation] should be performed"), and combine it with the Knowledge-Enhanced Prompt Engineering framework to perform semantic correction on the model output.
[0122] This dual-channel design allows the model to still follow industrial logic even under small sample conditions, avoiding blind migration driven purely by data.
[0123] 2. Dynamic confidence filtering and task-aware parameter isolation mechanism
[0124] Design a dynamic confidence filter: automatically identify and deweight high-noise samples (such as sensor instantaneous changes and abnormal operating conditions) based on predicted confidence to prevent noise from polluting model updates;
[0125] A task-aware parameter isolation strategy is introduced: only sub-networks related to the new task (such as adapters, cue vectors, and classification heads) are fine-tuned, while the remaining parameters are frozen. This is combined with a hybrid mechanism of Elastic Weight Consolidation (EWC) and Replay Buffer to effectively prevent catastrophic forgetting.
[0126] 3. Efficient transfer learning with small sample sizes and online self-optimization process
[0127] Define a standardized method for constructing small-sample industrial training data: based on multimodal data (text, image, time series) aligned with device ID and timestamp and structured rules, form a high-quality training set of <50 examples;
[0128] Construct a semantic drift detection trigger mechanism: when the semantic distance between a new sample and the historical knowledge base exceeds a threshold (θ=0.35), automatically initiate incremental fine-tuning;
[0129] Achieve an online continuous learning closed loop with a model self-optimization cycle of ≤24 hours and a reduction of ≥60% in manual intervention.
[0130] Corresponding to the above method embodiments, this specification also provides an embodiment of a dual-channel adaptive continuous learning device for small-sample industrial scenarios. Figure 2 This specification illustrates a schematic diagram of a dual-channel adaptive continuous learning device for small-sample industrial scenarios, based on one embodiment of this specification. Figure 2 As shown, the device includes:
[0131] The data acquisition module 201 is configured to acquire a small sample industrial training dataset, which includes multimodal industrial data and corresponding annotations as well as a subset of associated rules.
[0132] Engine building module 202 is configured to build a dual-channel adaptive continuous learning engine. The dual-channel adaptive continuous learning engine includes a structural knowledge injection channel and a semantic cognitive optimization channel. The structural knowledge injection channel is used to encode a subset of rules into an industrial domain adapter and embed it into a pre-trained model. The semantic cognitive optimization channel is used to perform knowledge enhancement prompting engineering on multimodal inputs based on industrial instruction templates.
[0133] The model fine-tuning module 203 is configured to perform incremental fine-tuning with noise filtering, including using a dynamic confidence filter to identify and deweight samples in a small industrial training dataset, employing a task-aware parameter isolation mechanism to fine-tune only the sub-networks relevant to the new task, and combining elastic weight solidification and replay buffers to prevent catastrophic forgetting.
[0134] The model deployment module 204 is configured to deploy an online continuous learning and model update mechanism, including incremental fine-tuning triggered based on semantic drift detection, and to manage model versions.
[0135] In one possible implementation, obtaining a small-sample industrial training dataset includes:
[0136] Collect multimodal industrial data, including text modality, visual modality, time-series modality, and structured knowledge modality;
[0137] Alignment of multimodal industrial data based on equipment identifiers and timestamps;
[0138] Industrial terms in the text are linked to rule base nodes using named entity recognition and entity linking technologies.
[0139] Construct a training sample set, where each sample includes a multimodal input, a label, and a subset of associated rules.
[0140] In one possible implementation, the structural knowledge injection channel includes encoding a subset of rules into an industry domain adapter, which is a lightweight fully connected network; embedding the industry domain adapter between Transformer layers of a pre-trained industrial large model; fine-tuning only the parameters of the industry domain adapter while freezing the backbone network parameters of the pre-trained industrial large model.
[0141] In one possible implementation, the semantic cognition optimization channel includes: constructing an industrial instruction template library, which includes fillable templates for equipment, phenomena, and operations; concatenating the industrial instruction templates with multimodal inputs to form contextualized prompts; and introducing a semantic correction algorithm to perform rule consistency verification on the model output and to provide feedback correction signals when rules are violated.
[0142] In one possible implementation, a dynamic confidence filter is used to identify and reduce noise in samples from a small industrial training dataset, including:
[0143] For each sample, calculate the confidence level of the model prediction;
[0144] If the confidence level is below 0.3 or above 0.95, it is judged as a noise sample;
[0145] Noisy samples are assigned a low first weight, and normal samples are assigned a second weight; wherein the first weight is less than the second weight.
[0146] In one possible implementation, the task-aware parameter isolation mechanism includes:
[0147] An importance matrix of key parameters for historical tasks is solidified using flexible weights.
[0148] Construct a replay buffer to store historical high-confidence samples;
[0149] Define a composite loss function, which includes the new task loss calculated based on sample weights, the replay loss calculated based on samples in the replay buffer, and the elastic weight solidification regularization term calculated based on parameter importance and original parameter values.
[0150] Only the subnetworks relevant to the new task are fine-tuned, including the industry domain adapter, cue vectors, and classification head.
[0151] One possible implementation involves deploying an online continuous learning and model update mechanism, including:
[0152] Calculate the average semantic distance between the new sample and the historical knowledge base; if the average semantic distance exceeds 0.35, trigger incremental fine-tuning; complete the fine-tuning within 24 hours and generate a new version of the model; use semantic version numbers to manage model versions, and support A / B testing and rollback.
[0153] The above is an illustrative scheme of a dual-channel adaptive continuous learning device for small-sample industrial scenarios according to this embodiment. It should be noted that the technical solution of this dual-channel adaptive continuous learning device for small-sample industrial scenarios and the technical solution of the dual-channel adaptive continuous learning method for small-sample industrial scenarios described above belong to the same concept. Details not described in detail in the technical solution of the dual-channel adaptive continuous learning device for small-sample industrial scenarios can be found in the description of the technical solution of the dual-channel adaptive continuous learning method for small-sample industrial scenarios described above.
[0154] Figure 3 A structural block diagram of a computing device 300 according to one embodiment of this specification is shown. The components of the computing device 300 include, but are not limited to, a memory 310 and a processor 320. The processor 320 is connected to the memory 310 via a bus 330, and a database 350 is used to store data.
[0155] The computing device 300 also includes an access device 340, which enables the computing device 300 to communicate via one or more networks 360. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 340 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.
[0156] In one embodiment of this specification, the aforementioned components of the computing device 300 and Figure 3 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 3 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.
[0157] The computing device 300 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 300 can also be a mobile or stationary server.
[0158] The processor 320 executes the following computer-executable instructions, which, when executed by the processor, implement the steps of the aforementioned dual-channel adaptive continuous learning method for small-sample industrial scenarios. The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the aforementioned dual-channel adaptive continuous learning method for small-sample industrial scenarios belong to the same concept. Details not described in detail in the technical solution of the computing device can be found in the description of the technical solution of the aforementioned dual-channel adaptive continuous learning method for small-sample industrial scenarios.
[0159] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the aforementioned dual-channel adaptive continuous learning method for small-sample industrial scenarios.
[0160] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solution of the dual-channel adaptive continuous learning method for small-sample industrial scenarios described above. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the dual-channel adaptive continuous learning method for small-sample industrial scenarios described above.
[0161] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described dual-channel adaptive continuous learning method for small-sample industrial scenarios.
[0162] The above is an illustrative example of a computer program in this embodiment. It should be noted that the technical solution of this computer program belongs to the same concept as the technical solution of the dual-channel adaptive continuous learning method for small-sample industrial scenarios described above. Details not described in detail in the computer program's technical solution can be found in the description of the technical solution of the dual-channel adaptive continuous learning method for small-sample industrial scenarios described above.
[0163] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.
[0164] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.
[0165] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.
[0166] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0167] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.
Claims
1. A dual-channel adaptive continuous learning method for small-sample industrial scenarios, characterized in that, include: Obtain a small sample industrial training dataset, which includes multimodal industrial data, corresponding annotations, and associated rule subsets; A dual-channel adaptive continuous learning engine is constructed, comprising a structural knowledge injection channel and a semantic cognitive optimization channel. The structural knowledge injection channel is used to encode the rule subset into an industrial domain adapter and embed it into a pre-trained model. The semantic cognitive optimization channel is used to perform knowledge enhancement prompting engineering on multimodal inputs based on industrial instruction templates. Incremental fine-tuning with noise filtering is performed, including using a dynamic confidence filter to identify and deweight the samples in the small industrial training dataset, employing a task-aware parameter isolation mechanism to fine-tune only the sub-networks relevant to the new task, and combining elastic weight solidification and replay buffers to prevent catastrophic forgetting. Deploy online continuous learning and model update mechanisms, including incremental fine-tuning triggered by semantic drift detection, and manage model versions; The structural knowledge injection channel includes encoding the rule subset into an industrial domain adapter, wherein the industrial domain adapter is a lightweight fully connected network; The industrial domain adapter is embedded between the Transformer layers of the pre-trained industrial large model; the parameters of the industrial domain adapter are fine-tuned only, and the backbone network parameters of the pre-trained industrial large model are frozen. The semantic cognition optimization channel includes: constructing an industrial instruction template library, which includes fillable templates for equipment, phenomena, and operations; concatenating the industrial instruction templates with multimodal inputs to form contextualized prompts; and introducing a semantic correction algorithm to perform rule consistency verification on the model output and to provide feedback correction signals when rules are violated. The step of using a dynamic confidence filter to identify and reduce noise in the samples of the small industrial training dataset includes: For each sample, calculate the confidence level of the model prediction; If the confidence level is lower than 0.3 or higher than 0.95, it is determined to be a noise sample; The noisy samples are assigned a low first weight, and the normal samples are assigned a second weight; wherein the first weight is less than the second weight. The task-aware parameter isolation mechanism includes: An importance matrix of key parameters for historical tasks is solidified using flexible weights. Construct a replay buffer to store historical high-confidence samples; Define a composite loss function, which includes a new task loss calculated based on sample weights, a replay loss calculated based on samples in the replay buffer, and an elastic weight solidification regularization term calculated based on parameter importance and original parameter values. Only the sub-networks relevant to the new task are fine-tuned, including the industry domain adapter, cue vectors, and classification heads; The deployment of the online continuous learning and model update mechanism includes: Calculate the average semantic distance between the new sample and the historical knowledge base; if the average semantic distance exceeds 0.35, trigger incremental fine-tuning; complete the fine-tuning within 24 hours and generate a new version of the model; manage model versions using semantic version numbers, supporting A / B testing and rollback.
2. The method according to claim 1, characterized in that, The acquisition of the small sample industrial training dataset includes: Collect multimodal industrial data, including text modality, visual modality, time-series modality, and structured knowledge modality; The multimodal industrial data is aligned based on device identifiers and timestamps; Industrial terms in the text are linked to rule base nodes using named entity recognition and entity linking technologies. Construct a training sample set, where each sample includes a multimodal input, a label, and a subset of associated rules.
3. A dual-channel adaptive continuous learning device for small-sample industrial scenarios, characterized in that, To implement the dual-channel adaptive continuous learning method for small-sample industrial scenarios as described in claim 2, the method includes: The data acquisition module is configured to acquire a small sample industrial training dataset, which includes multimodal industrial data and corresponding annotations and associated rule subsets. The engine building module is configured to build a dual-channel adaptive continuous learning engine, which includes a structural knowledge injection channel and a semantic cognitive optimization channel. The structural knowledge injection channel is used to encode the rule subset into an industrial domain adapter and embed it into a pre-trained model. The semantic cognitive optimization channel is used to perform knowledge enhancement prompting engineering on multimodal inputs based on industrial instruction templates. The model fine-tuning module is configured to perform incremental fine-tuning with noise filtering, including using a dynamic confidence filter to identify and deweight the samples in the small industrial training dataset for noise, employing a task-aware parameter isolation mechanism to fine-tune only the sub-networks that are relevant to the new task, and combining elastic weight solidification and replay buffers to prevent catastrophic forgetting. The model deployment module is configured to deploy online continuous learning and model update mechanisms, including incremental fine-tuning triggered based on semantic drift detection, and managing model versions.
4. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the dual-channel adaptive continuous learning method for small-sample industrial scenarios as described in claim 2.
5. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the dual-channel adaptive continuous learning method for small-sample industrial scenarios as described in claim 2.