A semantic-level collaborative reasoning method, system, and storage medium for industrial AI models in intelligent manufacturing production lines.

By constructing a unified semantic space and semantic mapping function, the raw observation data of heterogeneous devices are mapped to the unified semantic space, which solves the accuracy problem of cross-layer collaborative reasoning of industrial AI models on intelligent manufacturing production lines, realizes the plug-and-play and adaptive evolution of models, and improves the collaborative reasoning ability of the system in complex environments.

CN123117366APending Publication Date: 2026-11-17QILU INST OF TECH
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611627996.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, industrial AI models struggle to achieve accurate cross-layer collaborative reasoning on smart manufacturing production lines, primarily due to the lack of a unified semantic benchmark. This results in a lack of consistent constraints across model adaptation, semantic transmission, and reasoning decision-making, making it impossible to maintain efficient collaboration under conditions of heterogeneous equipment and changing operating conditions.

Method used

By constructing a unified semantic space, the raw observation data of heterogeneous devices are mapped to the unified semantic space using a semantic mapping function. Joint training is performed through cross-device semantic alignment loss to generate a workstation data feature dictionary. Combined with the workstation semantic prototype, a comprehensive adaptation index is calculated to achieve plug-and-play functionality and semantic communication of the model, and to dynamically adjust the division of inference tasks between the edge and the cloud.

Benefits of technology

It improves the accuracy of cross-layer collaborative reasoning under conditions of heterogeneous equipment and changing operating conditions, realizes plug-and-play and adaptive evolution of the model, and enhances the collaborative reasoning capability of the system in complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117366A_ABST
    Figure CN123117366A_ABST
Patent Text Reader

Abstract

This application relates to the field of intelligent manufacturing technology, specifically to a semantic-level collaborative reasoning method, system, and storage medium for industrial AI models in intelligent manufacturing production lines. The method includes: taking raw observation data from heterogeneous equipment as input, mapping it to a unified semantic space through a semantic mapping function, and jointly training it using cross-device semantic alignment loss to obtain a unified semantic feature vector; generating a workstation data feature dictionary based on the unified semantic space; mapping the output of the model to be deployed to the unified semantic space, calculating a comprehensive fit index by combining the workstation semantic prototype and the workstation data feature dictionary, and determining whether to deploy based on the comprehensive fit index; for the deployed model, performing semantic communication using the unified semantic vector as the transmission object, and entering edge reasoning after semantic consistency verification; calculating the multi-dimensional fusion confidence of edge reasoning, and triggering cloud reasoning if it is lower than a first threshold. This application improves the accuracy of cross-layer collaborative reasoning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing technology, specifically to a semantic-level collaborative reasoning method, system, and storage medium for industrial AI models for intelligent manufacturing production lines. Background Technology

[0002] With the deepening development of Industry 4.0 and intelligent manufacturing, industrial AI models have been widely applied in key areas such as equipment health monitoring, tool wear prediction, and quality status identification. However, intelligent manufacturing production lines typically consist of multiple layers, including equipment, control, and manufacturing execution layers. These layers have complex data types, heterogeneous communication protocols, and dynamically changing operating environments, making it difficult for industrial AI models to be deeply integrated with the production line's operating system. Their engineering deployment capabilities and cross-layer collaboration capabilities remain significantly insufficient.

[0003] To address the deployment and operation challenges of edge AI models, existing technologies have proposed cloud-edge collaborative inference solutions. For example, invention patent CN121919915A discloses a large-scale model edge-cloud collaborative inference method and system that balances efficiency and privacy. This method deploys the model at both the edge and the cloud, with the edge device performing part of the inference and then uploading the feature tensor to the cloud to complete the remaining computation. Furthermore, academic research has proposed a cloud-edge collaborative inference framework based on uncertainty calibration, deploying a lightweight model at the edge and a high-performance model at the cloud. When the edge model's inference confidence is insufficient, it automatically initiates a collaborative inference request to the cloud.

[0004] The aforementioned existing technical solutions have the following drawbacks: There is a lack of unified semantic constraints across the model adaptation, semantic transmission, and inference decision-making stages. Model adaptation is based on label or feature matching, semantic transmission on independent encoders, and inference decision-making on internal model confidence. Each of these three operates independently based on different data representations, resulting in a lack of a unified semantic benchmark for cross-layer collaboration. Furthermore, the edge and cloud rely on additional manual adaptation or preprocessing to achieve semantic alignment under heterogeneous data conditions. Once equipment is added or removed, or processes change, the existing alignment relationship becomes invalid, requiring manual reconfiguration. Due to the lack of a unified semantic benchmark, when operational conditions drift and data distribution shifts, model adaptation evaluation, semantic transmission quality, and inference confidence cannot be dynamically calibrated based on the same semantic reference system. Deviations in each stage cannot be identified and corrected in a timely manner, leading to a continuous decline in the overall accuracy of cross-layer collaborative inference under heterogeneous equipment and changing operational conditions. Summary of the Invention

[0005] In view of this, this application provides a semantic-level collaborative reasoning method, system and storage medium for industrial AI models for intelligent manufacturing production lines, in order to solve the problem of the continuous decline in the accuracy of cross-layer collaborative reasoning in the prior art.

[0006] In a first aspect, embodiments of this application provide a semantic-level collaborative reasoning method for industrial AI models in intelligent manufacturing production lines, including: Using raw observation data from heterogeneous devices as input, the data is mapped to a unified semantic space through a semantic mapping function, and then jointly trained with cross-device semantic alignment loss to obtain a unified semantic feature vector. A workstation data feature dictionary is generated based on the unified semantic space. The workstation data feature dictionary records the input and output semantic format and workstation semantic prototype of each workstation. The output of the model to be deployed is mapped to the unified semantic space, and the comprehensive adaptation index is calculated by combining the workstation semantic prototype and the workstation data feature dictionary. The deployment is then determined based on the comprehensive adaptation index. After deployment, semantic communication is performed using a unified semantic vector as the transmission object, and the model enters edge inference after semantic consistency verification. Calculate the multi-dimensional fusion confidence score of edge inference. If it is lower than the first threshold, cloud inference is triggered.

[0007] In one possible implementation, the process of taking raw observation data from heterogeneous devices as input, mapping it to a unified semantic space through a semantic mapping function, and jointly training it with cross-device semantic alignment loss to obtain a unified semantic feature vector specifically includes: Raw multi-source industrial data output by the i-th device As input, through the semantic mapping function corresponding to the device. Mapping to a unified semantic space yields semantic feature vectors: ,in For the learnable parameters of the mapping function, Unified semantic space dimension; The semantic mapping function is jointly trained using cross-device semantic alignment loss and physical constraint terms; The cross-device semantic alignment loss is: ,in, The total number of devices participating in semantic alignment; i and j are the distinct device numbers, and Each pair of devices is counted only once; These are the sequence numbers of semantic samples of the same type. This represents the total number of semantic samples of the same type. For equipment The original samples belonging to the k-th semantic class; For equipment The original samples belonging to the k-th semantic class; For equipment The corresponding semantic mapping function; For equipment The corresponding semantic mapping function; It is the square of the L2 norm; The physical constraint term is the physical consistency loss corresponding to the physical laws of the equipment. ; Calculate the total loss: ,in, This refers to the loss of physical consistency corresponding to the physical laws of the equipment. These are the physical constraint weights.

[0008] In one possible implementation, the step of generating a workstation data feature dictionary based on the unified semantic space, wherein the workstation data feature dictionary records the input and output semantic format and workstation semantic prototype of each workstation, specifically includes: The set of historical semantic samples for the j-th workstation Calculate the mean to obtain the semantic prototype of this workstation: ; The semantic prototype The data is stored in association with workstation identifier, data modality, feature dimension, and protocol type to form the workstation data feature dictionary.

[0009] In one possible implementation, mapping the output of the model to be deployed to the unified semantic space, calculating a comprehensive adaptation index by combining the workstation semantic prototype and the workstation data feature dictionary, and determining whether to deploy based on the comprehensive adaptation index specifically includes: The output of the model to be deployed is mapped to the unified semantic space via a semantic mapping function to obtain the model output semantic vector. ; calculate Workstation semantic prototype semantic fit between , where σ is the distance scale parameter; The semantic fit A comprehensive compatibility index is obtained by weighting and integrating factors such as functional category matching, data modality matching, performance compliance, resource adaptability, and protocol compatibility. ; Plug and play deployment is executed when Ω is greater than or equal to the first threshold. Adaptive fine-tuning is performed when the second threshold ≤ Ω < the first threshold; Deployment is rejected when Ω is less than the second threshold.

[0010] In one possible implementation, the step of performing semantic communication on the deployed model using a unified semantic vector as the transmission object, and then entering edge inference after semantic consistency verification, specifically includes: The edge terminal takes the unified semantic feature vector as input, encodes it into compact semantic features by a semantic encoder, and dynamically adjusts the coding rate according to the channel signal-to-noise ratio: Where SNR is the signal-to-noise ratio of the channel between the edge and the cloud; Dynamic coding rate for semantic transmission; This represents the upper limit of the encoding bitrate; Based on the baseline bit rate; This is the signal-to-noise ratio adjustment coefficient; According to the semantic dimensions preserved after compression The compact semantic features are semantically compressed to obtain compressed semantic features, wherein... To unify the dimensions of the semantic space; The compressed semantic features are transmitted to the cloud via a channel; The semantic features are recovered by the semantic decoder in the cloud. The cosine similarity between edge-encoded features and cloud-decoded recovered features is used as a semantic consistency verification index. When the cosine similarity is lower than a preset similarity threshold, semantic transmission is determined to be distorted, and cloud inference is directly triggered. When the cosine similarity is not lower than the preset similarity threshold, edge inference is initiated.

[0011] In one possible implementation, if the multi-dimensional fusion confidence score of the calculated edge inference is lower than a first threshold, cloud inference is triggered, specifically including: The deployed model is decomposed into a lightweight front-end layer at the edge and a heavyweight back-end layer in the cloud. At the edge, preliminary inference is performed to output edge prediction results, and multi-dimensional fusion confidence is calculated: in, Maximize the output of edge reasoning probability; It is the difference between the highest and second-highest probabilities; The variance of the feature vector is output for the edge. To integrate weights, satisfy ; when When, the edge directly outputs the inference result; when At that time, the cloud-based weighted back layer is triggered to execute the complete inference; The Initialize using the set quantiles of the historical marginal confidence distribution, and dynamically update according to the following formula: in, To update the learning step size for the threshold, To determine the actual accuracy of edge inference based on real-time statistics from the production line. Set target accuracy for business operations.

[0012] One possible implementation also includes: Record the sequence of working conditions and the frequency of cloud inference triggered by the edge. When the proportion of low-confidence inference at the edge in the sliding window exceeds the set threshold, use the cloud model as the teacher model, generate a lightweight student model for edge devices through knowledge distillation and channel pruning, and migrate and deploy the student model to the edge. The semantic mapping function parameters of the unified semantic space and the workstation semantic prototype are updated synchronously, and the teacher training set is reconstructed based on the updated unified semantic space.

[0013] In one possible implementation, the method of using a cloud-based model as the teacher model and generating a lightweight student model for edge devices through knowledge distillation and channel pruning specifically includes: Channels to be pruned are selected based on channel importance assessment, where the channel importance is the absolute value of the L1 norm of the convolution kernel weights or the scaling factor of the BN layer. A layer-by-layer differentiated pruning strategy is adopted, with the pruning rate of shallow feature extraction layers being lower than that of deep task-specific layers, resulting in an overall pruning rate of... ; Construct a composite loss function: Train the student model, where For classification cross-entropy loss, For the KL divergence distillation loss of the output layer, This is due to the distillation loss of the intermediate characteristic layer. To output the distillation loss weight, Weights for intermediate feature distillation loss; When the accuracy of the student model is not lower than a set percentage of the accuracy of the teacher model, the student model will be migrated and deployed to the edge.

[0014] Secondly, embodiments of this application provide an industrial AI model semantic-level collaborative reasoning system for intelligent manufacturing production lines, including: The semantic modeling module is used to take the original observation data from heterogeneous devices as input, map it to a unified semantic space through a semantic mapping function, and jointly train it with cross-device semantic alignment loss to obtain a unified semantic feature vector. The data feature dictionary generation module is used to generate a workstation data feature dictionary based on the unified semantic space. The dictionary records the input and output semantic format and workstation semantic prototype of each workstation. The adaptability assessment and instantiation module is used to map the output of the model to be deployed to the unified semantic space, combine the workstation semantic prototype and the workstation data feature dictionary to perform adaptability assessment, obtain a comprehensive adaptability index, and determine whether to deploy based on the comprehensive adaptability index. The semantic communication module is used to perform semantic communication on the deployed model, using a unified semantic vector as the transmission object, and then enter the edge inference after semantic consistency verification. The hierarchical inference engine module is used to calculate the multi-dimensional fusion confidence of edge inference. If it is lower than the first threshold, cloud inference is triggered.

[0015] Thirdly, embodiments of this application provide an electronic device, including: processor; Memory; And a computer program, wherein the computer program is stored in the memory, the computer program including instructions that, when executed by the processor, cause the electronic device to perform the method described in any one of the first aspects.

[0016] Fourthly, embodiments of this application provide a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the method described in any one of the first aspects.

[0017] In this embodiment, the solution constructs a unified semantic space that directly takes the original observation data of heterogeneous devices as input, enabling all stages of model adaptation evaluation, semantic transmission, and edge inference to operate on the same semantic benchmark, thus avoiding semantic deviations caused by data heterogeneity and stage fragmentation. On this basis, the semantic distance between the model output and the workstation requirements is quantified using the workstation semantic prototype, and the division of inference between the edge and the cloud is dynamically controlled through semantic consistency verification and multi-dimensional fusion confidence, so that the system can maintain semantic alignment and linkage correction of each stage when the operating conditions drift, ultimately effectively improving the accuracy of cross-layer collaborative inference under conditions of equipment heterogeneity and changing operating conditions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of the reasoning method provided in the embodiments of this application; Figure 2 A system core architecture diagram of the inference method provided in the embodiments of this application; Figure 3 A flowchart of hierarchical progressive edge-cloud collaborative inference and model adaptive migration provided in the embodiments of this application; Figure 4This is a diagram illustrating the construction of an industrial AI agent as provided in an embodiment of this application. Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] To better understand the technical solution of this application, the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0021] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0022] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0023] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0024] Reference Figures 1 to 3 In this application embodiment, an industrial AI model semantic-level collaborative reasoning method for intelligent manufacturing production lines includes: S1 takes the original observation data from heterogeneous devices as input, maps it to a unified semantic space through a semantic mapping function, and jointly trains it with cross-device semantic alignment loss to obtain a unified semantic feature vector; In this embodiment, the intelligent manufacturing production line is equipped with various heterogeneous devices, including vibration sensors, visual inspection cameras, current acquisition devices, acoustic emission sensors, and PLC controllers. The raw observation data output by these devices differs significantly in data type, sampling frequency, dimensions, and communication protocols. This step does not rely on pre-constructed knowledge ontologies or manually designed structured features; instead, it directly uses the raw observation data output by each device as input and maps it to a unified semantic space through a learnable semantic mapping function.

[0025] Specifically, the raw multi-source industrial data output by the i-th device As input, through the semantic mapping function corresponding to the device. Mapping to a unified semantic space yields semantic feature vectors: in, For the first The raw, multi-source industrial data output by the equipment; For the first The semantic mapping function corresponding to the device is provided. The semantic mapping function adopts a neural network structure composed of a multilayer perceptron (MLP) and a modal encoder. The image data adopts a convolutional encoder, the time-series signals such as vibration, acoustic emission and current adopt a one-dimensional convolutional and LSTM joint encoder, the discrete state of PLC adopts an embedded encoder, and each modality is finally mapped to a semantic vector of a unified dimension d.

[0026] These are the learnable parameters of the mapping function; The mapped unified semantic vector, To unify the dimensions of the semantic space.

[0027] To ensure that data with the same physical meaning in different devices are mapped to similar semantic locations, a cross-device semantic alignment loss is constructed: in, The total number of devices participating in semantic alignment; i and j are the distinct device numbers, and Each pair of devices is counted only once; These are the sequence numbers of semantic samples of the same type. This represents the total number of semantic samples of the same type. For equipment The original samples belonging to the k-th semantic class are as follows. For equipment The original samples belonging to the k-th semantic class; For equipment The corresponding semantic mapping function; For equipment The corresponding semantic mapping function; The squared norm 2 measures the semantic vector distance across devices.

[0028] Furthermore, to ensure the physical consistency of the unified semantic space, a physical constraint term is introduced based on the semantic alignment loss: in, The total loss for semantic mapping, For physical constraint weights; The physical consistency loss, representing the physical laws governing equipment such as wear evolution and bearing dynamics, is defined using physical residual constraints: in, For the physical law model of the equipment, the wear evolution empirical model is used for the tool scenario, and the bearing dynamics degradation model is used for the bearing scenario. This is the predicted physical state after semantic mapping, thereby ensuring that the unified semantic space satisfies the physical consistency of the device.

[0029] This step outputs a unified semantic representation space S.

[0030] In existing technologies, most solutions are industrial semantic ontology and knowledge graphs, which process structured features and ontological entities that have already been manually extracted. They rely on pre-built knowledge ontology and cannot directly process raw sensor sampling data.

[0031] This invention directly addresses the raw, multi-source observation data of devices, designing a learnable semantic mapping function. By superimposing cross-device alignment loss with device physical mechanism constraint loss, it achieves cross-device semantic normalization at the raw data level, eliminating the need for manual pre-design of ontology and features. Simultaneously, this semantic space serves as a common underlying input for subsequent model adaptation evaluation, semantic communication transmission, and hierarchical collaborative reasoning, forming a tight coupling with subsequent steps.

[0032] S2, Generate a workstation data feature dictionary based on the unified semantic space, wherein the workstation data feature dictionary records the input and output semantic format and workstation semantic prototype of each workstation; In this embodiment, based on the unified semantic expression space constructed by S1, semantic requirements of each workstation on the production line are parsed and standardized and encapsulated to construct a globally unified industrial data feature dictionary.

[0033] Unlike traditional input / output recording formats, this step further establishes a standard semantic prototype for the workstation based on a unified semantic vector. Let the first... The historical semantic sample set of each workstation is as follows Its semantic prototype is defined as: in, Let be the semantic prototype of the j-th workstation, representing the standard expression of the input and output semantics of that workstation. Indicates the number of samples; For the first The unified semantic vector of the nth sample within the historical semantic sample set of each workstation.

[0034] semantic prototype The data is stored in association with workstation identifier, data modality, feature dimension, and protocol type, forming a workstation data feature dictionary. Each workstation record in the dictionary includes the following fields as shown in Table 1: Table 1 Workstation Record Field Contents S2 Output Workstation Semantic Prototype This provides a direct basis for the semantic adaptation of the S3 model. When new equipment is added to the production line or the sensor configuration changes, it is only necessary to perform semantic space projection on the mapping results of the new equipment and match the similarity with the existing dictionary entries. If the similarity is lower than the preset threshold, a new entry is added to the dictionary; if it is higher than the preset threshold, it is merged into the existing entries, realizing the incremental update of the dictionary.

[0035] S3, map the output of the model to be deployed to the unified semantic space, calculate the comprehensive adaptation index by combining the workstation semantic prototype and the workstation data feature dictionary, and determine whether to deploy based on the comprehensive adaptation index; In this embodiment, the inputs include S1 outputting a unified semantic space, S2 outputting a unified data feature dictionary for workstations, and the industrial AI model to be deployed.

[0036] First, the output of the industrial AI model to be deployed (including classification probabilities, regression predictions, or intermediate feature vectors) is input into the shared semantic mapping network and projected onto S1 using the same semantic mapping function with the same parameters to obtain a unified semantic vector output by the model: in: The raw output of the industrial AI model to be deployed. Output the corresponding semantic mapping function for the model; Output semantic vectors for the model after mapping to a unified semantic space.

[0037] Then, the distance between the model semantics and the workstation semantic prototype is calculated, and a unified semantic fit is defined: , in, This is the semantic prototype of the workstation target. The distance scale parameter is 0.5 in this example.

[0038] Further establish six-dimensional adaptation indicators: in, To ensure consistent semantic adaptability; For functional category matching degree; Data modality matching degree is used to measure the consistency between the workstation data modality and the input modality of the model to be deployed; To assess performance compliance, the model's inference accuracy and inference latency metrics are measured to determine whether they meet the real-time requirements of the workstation and the accuracy constraints of the business. For the adaptation of computing resources; For communication protocol compatibility.

[0039] in It is used to measure the distance between the model output semantics and the target semantics of the workstation in the S1 unified semantic space, rather than traditional label matching.

[0040] Data modality matching degree is defined as: in, A set of data modes (images, vibrations, currents, acoustic emissions, etc.) supported by the workstation. The set of input modes required by the model; A higher value indicates higher modal compatibility.

[0041] Functional category matching degree is defined as: in, To support a set of tasks for the model, This is a collection of tasks for each workstation.

[0042] Performance compliance is defined as: in: The inference accuracy compliance score measures the degree to which the actual inference accuracy of the model meets the accuracy requirements of the workstation, and its value range is [range missing]. When the model's actual inference accuracy Meets or exceeds the required accuracy of the workstation hour, When the actual inference accuracy is lower than the required accuracy, .

[0043] The inference latency compliance metric measures how well the model's actual inference latency meets the upper limit of the allowable latency for a workstation. Its value range is [range missing]. When the model's actual inference latency Not exceeding the workstation's required delay limit hour, When the actual inference latency exceeds the required upper limit, .

[0044] in: For the actual inference accuracy of the model, Precision is required for the workstation; The maximum delay is required for the workstation. The actual inference latency of the model is represented by the above indicators. The accuracy and latency are normalized respectively, so that both are converted into evaluation indicators with larger values ​​and higher degree of compliance.

[0045] Resource adaptability is defined as: In the formula, The computing resources available at the workstation level. The computing resources required for model inference.

[0046] Communication protocol compatibility This indicates the degree of compatibility between the model interface and the workstation communication protocol.

[0047] The weights are not manually set, but rather obtained through offline learning based on historical deployment samples. Assume there are already... The group deployment sample, its actual deployment effect is The weights are then solved using constrained least squares optimization: ; Constraints: .

[0048] in, For the predicted deployment effect of the i-th sample, This represents the actual deployment effect for the i-th sample. When the number of historical deployment samples is less than 30 and the constraint-based least squares stable solution cannot be satisfied, uniform weight initialization is used. .

[0049] As production line deployment samples continue to accumulate, when the number of samples reaches more than 30 sets, the background offline re-executes constrained least squares optimization, iteratively updates the weights of the six-dimensional indicators, and achieves continuous optimization of the comprehensive adaptation index.

[0050] Based on the comprehensive compatibility index For deployment decisions, the first threshold is set at 0.85, and the second threshold is set at 0.60. Plug and play deployment at any time; Automatic fine-tuning Deployment was rejected at that time. The values ​​of 0.85 and 0.60 are derived from the optimal ROC split points of historical deployment samples.

[0051] Current technologies for matching models with production line workstations mostly rely on superficial matching based on tags, task names, computing resources, and communication protocols. They fail to measure the intrinsic consistency between the model's output semantics and the semantics of the workstation's business requirements. Even if computing power and tags are fully matched, the model's semantic representation and the actual business semantics of the workstation may still be misaligned, leading to inference failure after deployment.

[0052] This invention adds a semantic adaptability index, which measures the distance between the model output and the workstation semantic prototype in a unified semantic vector space. At the same time, it constructs a six-dimensional comprehensive adaptability index, with the weights automatically optimized by least squares constraints from historical deployment samples. Furthermore, it determines the three decision thresholds of deployment / fine-tuning / rejection based on the ROC curve, truly realizing plug-and-play AI models and solving the problem of a large amount of manual code rewriting on heterogeneous production lines.

[0053] S4 performs semantic communication on the deployed model using a unified semantic vector as the transmission object, and enters edge inference after semantic consistency verification; In this embodiment, S4 directly transmits the unified semantic vector S output by S1, replacing traditional bit-level data transmission with semantic encoding, semantic compression, and semantic consistency verification.

[0054] The edge first encodes the unified semantic vector: in, The unified semantic vector output in step S1 Edge semantic encoder, Compact semantic features after edge encoding These are the parameters for the edge semantic encoder.

[0055] The coding rate is dynamically adjusted based on the real-time channel signal-to-noise ratio (SNR). SNR: Signal-to-noise ratio of wireless / industrial network channels between the edge and the cloud; Semantic transmission dynamic coding bitrate; The upper limit of the encoding bitrate is used to prevent excessive bitrate from causing bandwidth overload. Baseline bitrate; Signal-to-noise ratio adjustment coefficient; The parameter values ​​in the example are as follows: The SNR in the formula uses a linear signal-to-noise ratio. The signal-to-noise ratio described in the examples is typically expressed in decibels (dB). Therefore, before substituting the signal-to-noise ratio into the formula, the decibel value needs to be converted to a linear value. The conversion relationship is as follows: ;in, Signal-to-noise ratio expressed in decibels. This represents the linear signal-to-noise ratio after conversion.

[0056] When SNR≥25dB, the coding rate is approximately 0.54; when SNR=10dB, the coding rate is approximately 0.49.

[0057] When SNR ≤ 3dB, the industrial network is determined to be in a severely degraded state, the system stops semantic compression transmission, and directly triggers full inference in the cloud; when SNR ≥ 30dB, the encoding rate is limited to [value missing]. Avoid exceeding network bandwidth constraints.

[0058] The semantic dimensions retained after compression are: After receiving the semantic data packet in the cloud, the recovery process is executed: Where d is the dimension of the unified semantic space. It is a cloud-based semantic decoder. For transmitting semantic packets in the channel, These are parameters for the cloud-based semantic decoder.

[0059] Define semantic consistency metrics: Indicates semantic consistency metrics; This represents the semantic feature vector after encoding at the edge and before transmission. This represents the semantic feature vector received and decoded in the cloud. This metric uses cosine similarity to measure the directional consistency between the semantic features before transmission and the semantic features recovered after transmission. For non-zero vectors, its value ranges from [value missing]. The closer the value is to 1, the higher the consistency of the directions between the two.

[0060] During semantic consistency verification, Similarity threshold Comparison: When When the semantic consistency check passes, it is determined that the semantic consistency check has passed; when If the semantic consistency check fails, the priority of cloud-based reasoning is increased according to the collaborative reasoning decision-making mechanism in step S5.

[0061] Task performance retention rate For semantic transmission quality evaluation metrics: in, Indicates the task performance retention rate; This indicates the inference accuracy of the model on the test dataset after employing semantic encoding, compression, transmission, and decoding mechanisms. This represents the model's inference accuracy on the same test dataset when inference is performed directly using the corresponding raw input data without semantic communication processing. The same test samples, task definitions, and accuracy calculation methods are used under both conditions to ensure consistency in the comparison results.

[0062] when hour, This represents the ratio of inference accuracy under two communication conditions. The closer this metric is to 1, the smaller the impact of semantic communication on model inference accuracy; when it is below a preset performance retention rate threshold... If the performance loss caused by semantic communication exceeds the allowable range, the corresponding semantic transmission parameter adjustment or cloud inference guarantee mechanism will be triggered.

[0063] The preset similarity threshold is obtained based on statistics from the training set: in The average cosine similarity; This represents the standard deviation. In the example... ; ; .

[0064] when If the semantic recovery is successful, proceed to S5; otherwise, the semantic transmission is deemed distorted, and cloud-based inference is triggered directly.

[0065] This invention transmits a unified semantic vector obtained by mapping raw industrial data. The semantic coding rate is dynamically adjusted based on the industrial channel signal-to-noise ratio, semantic similarity verification is added, and a linkage between communication distortion and the inference process is established: if semantic verification fails, edge inference is skipped and cloud inference is triggered. The semantic communication module is not an independent communication component, but rather provides input for hierarchical progressive collaborative inference and is deeply coupled with the preceding and following model adaptation and confidence decision stages.

[0066] S5 calculates the multi-dimensional fusion confidence score of edge inference. If it is lower than the first threshold, cloud inference is triggered.

[0067] In this embodiment, the industrial AI model deployed by S3 is divided into an edge lightweight front-end layer. With cloud weight back layer The overall model is represented as: The edge terminal first completes the initial reasoning: Simultaneously calculate the multi-dimensional fusion confidence score. : in Maximize the output of edge reasoning Probability reflects class confidence; The difference between the maximum and second-highest probabilities represents the class discrimination margin; The variance of the feature vector is output for the edge. The fusion weights are initialized as follows: ,satisfy .

[0068] The initial weights are derived from the feature importance analysis of the historical validation set. The contribution of the three indicators to the correctness of the reasoning is calculated by logistic regression and normalized to obtain the initial values. After the system completes every 5000 edge inferences, the fusion weights are retrained based on the latest validation data to achieve adaptive updates.

[0069] With S3 They are independent of each other; the former is used for confidence fusion in marginal inference, while the latter is used for model fit evaluation.

[0070] The marginal confidence decision threshold is initialized using historical statistics: in, This represents the 20th percentile of the historical marginal confidence distribution.

[0071] Dynamically updated during operation: in, , : Update the confidence decision thresholds before and after; Threshold update learning step size; Real-time statistics from the production line provide the actual accuracy of edge inference. The business sets the target accuracy. (Example: In the embodiment...) It is approximately 0.7.

[0072] Threshold update learning step size Defined as: in, This is the maximum update step size, with a default value of 0.05. This represents the cumulative number of running windows. To stabilize the window length. As the cumulative number of running windows increases, Gradually increase and approach This allows the update step size to be smoothly adjusted from small to large.

[0073] The reasoning and decision-making rules are as follows: Output directly from the edge; Real-time cloud-based fine-grained inference, through dynamic thresholds, enables a collaborative mechanism for rapid edge processing of simple samples and fine-grained cloud-based inference of complex samples.

[0074] Most existing cloud-edge collaborative inference solutions rely solely on the model's output Softmax maximum probability as a single confidence criterion, with the inference threshold being a manually preset fixed value. This fails to adapt to changes in data distribution after shifts in production line conditions. A single maximum probability cannot fully characterize the reliability of the inference results: there are unreliable inference samples with high probabilities but low class distinction margins and divergent features.

[0075] This invention integrates the maximum probability of the category, the difference between the maximum and second-highest probabilities, and the variance of the output features to obtain the fused inference confidence. At the same time, the confidence decision threshold is dynamically and adaptively updated according to the actual inference accuracy of the production line, enabling fast edge processing of simple samples and fine inference in the cloud for fuzzy and complex samples.

[0076] In another specific embodiment, it also includes: S6 records the sequence of changes in operating conditions and the frequency of cloud inference triggered by the edge side. When the proportion of low-confidence inference at the edge within the sliding window exceeds the set threshold, the model is triggered to adapt and evolve.

[0077] In this embodiment, the system continuously records the sequence of operating condition changes and the frequency of edge-side triggering cloud inference. When the proportion of low-confidence inference within the sliding window exceeds a set threshold, a persistent operating condition drift is determined to have occurred. The triggering condition is: in: Model adaptive evolution trigger flag, This represents the initiation of cloud-based adaptive evolution; Total number of sliding observation windows for edge inference; Edge model The confidence level of the next inference output; : Indicator function, the value is 1 if the condition inside the parentheses is true, and 0 otherwise; The edge reasoning confidence alarm threshold is recommended to be 0.7. If the confidence of a single reasoning is lower than this threshold, the reasoning result is considered unreliable. The threshold for the proportion of low-confidence inference within the sliding window is recommended to be 0.4. When the proportion of unreliable inference within the window is greater than or equal to this threshold, it is determined that a continuous operating condition drift has occurred, triggering adaptive evolution of the cloud model.

[0078] Upon triggering, the unified semantic space is updated first: in, Let be the learnable parameters of the unified semantic space mapping function at iteration t. To provide learnable parameters for the unified semantic space mapping function at iteration t+1 after the update, The learning step size is used to update the semantic mapping network parameters, with a value range of [0.005, 0.05] and a default value of 0.02. When the edge precision of three consecutive statistical windows is consistently higher than the target precision, Automatic decay to a minimum value of 0.005 to improve system stability; when edge accuracy continuously decreases, It can be increased to 0.05 to speed up the threshold convergence.

[0079] Simultaneously update the workstation semantic prototype: in, For the j-th workstation, the semantic prototype vector of the workstation after the update at time t+1 is given. The unified semantic vector is obtained by semantic mapping for the newly collected nth sample.

[0080] Based on the updated unified semantic space, the teacher training set is reconstructed, and knowledge distillation and channel pruning are used to generate a new lightweight edge model, so as to realize the synchronous evolution of the semantic space and the edge model.

[0081] Upon triggering, a lightweight student model for edge devices is generated using a large cloud-based model as the teacher model, combined with knowledge distillation and channel pruning. Channel importance is assessed using L1 norm or BN scaling factor to filter channels for pruning. The channel importance calculation formula is as follows: in: : No. The importance of each convolutional channel; The L1 norm of the convolution kernel weights for this channel; : The absolute value of the scaling factor corresponding to the BN layer.

[0082] Set channel pruning rate Prioritize retaining the channels with the highest importance One channel. The pruning rate represents the proportion of channels that have been pruned out of the total number of channels. This represents the total number of channels in this layer of the network. The value range constraint is based on the following: a pruning rate greater than 0.3 can easily cause an irreversible decrease in model accuracy, while a pruning rate less than 0.1 results in negligible compression benefits; values ​​within this range represent light to moderate pruning, and subsequent knowledge distillation can compensate for the accuracy loss caused by pruning. This is the preferred value in engineering practice. .

[0083] Constructing a knowledge distillation composite loss function: in: Cross-entropy loss for student model classification: To determine the output layer distillation loss, KL divergence is used to align the probability distributions of teacher and student softening categories: The intermediate feature layer distillation loss constrains the student model to mimic the intermediate feature output of the teacher model: in, Number of training samples; Number of categories; :sample Authentic one-hot tags; Student model on samples Output Category The predicted probability; These are the category probability distributions of students and teachers after temperature smoothing, respectively. Distillation temperature, recommended value 6, is used to soften the probability distribution and uncover implicit correlations between categories. Used for loss scale normalization; Samples for student and teacher models respectively Intermediate features of the output; Output distillation loss weight, recommended value is 0.7, responsible for guiding students to learn the knowledge of soft labels output by teachers; : Weight of intermediate feature distillation loss, recommended value is 0.3, to compensate for the loss of intermediate representation information caused by channel pruning.

[0084] When the pruning rate is taken as the upper limit of 0.3, it can be... Increase the weight to 0.4~0.5 to strengthen feature constraints; the weight should not be too large to avoid suppressing cross-entropy loss and causing the model to deviate from the true label.

[0085] After the lightweight model is trained, an accuracy check is performed. The check conditions are as follows: in, To validate the accuracy of the student model on the validation set. To verify the accuracy of the large model validation set for teachers in the cloud; only when the accuracy of the student model is not lower than 95% of that of the teacher model will the new model be allowed to be automatically migrated and deployed to edge devices, thereby completing the "self-maintenance" of the edge AI model; this threshold serves as a fallback constraint for the project, avoiding the generation of low-quality models by pruning-distillation and their distribution to edge nodes.

[0086] This invention uses the edge confidence sliding window statistical results output by hierarchical collaborative inference as the trigger condition for work condition drift. Once adaptive evolution is triggered, not only is a lightweight AI model trained, but the unified semantic mapping parameters and workstation semantic prototypes of the previous steps are also updated in reverse, realizing the synchronous evolution of the semantic space and the edge AI model. Furthermore, engineering safety fallback conditions are set. Only when the accuracy of the lightweight student model reaches more than 95% of that of the cloud teacher model is it allowed to migrate and be deployed to edge devices, avoiding production accidents caused by the deployment of low-quality models after pruning and distillation on the production line.

[0087] In another specific embodiment, it also includes: S7 builds an industrial AI intelligent agent to achieve human-machine collaborative semantic interaction and intelligent scheduling.

[0088] In this embodiment, an industrial AI agent is constructed for intelligent manufacturing production lines. This AI agent possesses three core capabilities: natural language semantic parsing, industrial knowledge base retrieval, and intelligent model task scheduling. For natural language operation and maintenance commands, process adjustment commands, and model scheduling commands input by operators, the agent first performs semantic parsing specific to the industrial scenario, identifying the corresponding workstation, task, and model requirements. Then, it searches the production line knowledge graph and model resource library to match the optimal execution strategy. Finally, it automatically completes operations such as model start / stop, model switching, inference-level scheduling, and parameter updates.

[0089] For example, when an operator asks, "Does the current tool need to be replaced?", the industrial AI agent analyzes the instruction, queries historical data, and calls up the wear recognition model to output replacement suggestions and confidence information.

[0090] This step enables an intelligent upgrade from natural language processing to human-machine collaboration and autonomous model scheduling, improving the convenience of production line operation and maintenance and intelligent decision-making capabilities.

[0091] Furthermore, it also includes step S8, a cross-scenario zero-shot validation and model safety rollback mechanism based on meta-learning.

[0092] In this embodiment, a cross-scenario zero-shot validation and rollback strategy based on meta-learning is implemented. This embodiment uses the First-Order Model-Independent Meta-Learning (FOMAML) algorithm as the meta-learning framework. Compared to standard MAML, FOMAML omits second-order gradient calculations, making it more suitable for rapid small-sample adaptation of industrial edge devices, reducing training computational overhead while maintaining cross-scenario generalization capabilities. Small-sample meta-tasks simulating production line configuration changes, process parameter shifts, and equipment replacements are constructed in batches to build a meta-model with cross-scenario generalization capabilities. When the trained AI model needs to be embedded into a new production line or faces on-site process changes, the system only needs to collect a very small amount of production line idle data to quickly output the current model embedding scheme's configuration reliability through the meta-model.

[0093] If the cross-scenario adaptation reliability obtained from meta-learning inference is lower than a preset safety threshold, the system determines that the current model weights cannot adapt to the new working conditions. The system automatically performs a version rollback, restoring the previous stable version of the model to prevent production line inference failure caused by the deployment of an abnormal model. Simultaneously, it automatically triggers offline re-adaptation training in the simulation environment to re-complete small-sample fine-tuning and cross-scenario adaptation. The model is then deployed online only after successful verification, achieving safety, zero-sample adaptability, and controllable rollback capability for model migration across production lines.

[0094] Reference Figure 4 Application Scenario Example 1: This example uses the method described above to achieve online monitoring of tool wear status.

[0095] Perception and semantic modeling: Collect spindle vibration, cutting force, temperature signals and MES tool number; the semantic modeling module constructs a semantic expression space for tool operation status.

[0096] Adaptive embedding: Generate a data feature dictionary; the adaptability evaluation algorithm matches and instantiates the tool wear prediction model for the tool monitoring station.

[0097] Semantic Communication and Edge Inference: The edge controller receives semantic state information, performs lightweight layer inference, and outputs the wear stage in real time.

[0098] Intelligent agent interaction: When an operator asks, "Does the current tool need to be replaced?" the industrial AI agent interprets the command, queries historical data, calls up the wear recognition model, and outputs a replacement suggestion.

[0099] Adaptive Evolution: When changes in the processing technology cause low confidence on the edge side to trigger an excessive number of cloud inference attempts, the system automatically performs knowledge distillation to generate a new lightweight layer and distributes it to restore low-latency inference.

[0100] Cross-scenario verification: When a new type of tool is replaced, a small amount of no-load data is collected. The meta-model quickly verifies the confidence of the embedding scheme. If it is higher than the safety threshold, the embedding is successful.

[0101] Example 2: This example demonstrates the robustness of the system under production line reorganization.

[0102] When a new quality inspection station was added to the production line, the system collected four idle samples. The meta-model inference yielded an optimal embedding parameter confidence level of 0.65 (below the first safety threshold of 0.70). The system automatically rolled back to the previous stable model version, triggered the simulation offline re-adaptation module, and re-executed adaptation and knowledge distillation in the digital twin environment. After generating a model version that met the conditions of the new station, it was formally embedded, without affecting production throughout the process.

[0103] Corresponding to the above embodiments, this application also provides an industrial AI model semantic-level collaborative reasoning system for intelligent manufacturing production lines, including the following modules: Multi-source industrial sensing module: used to collect production line operating status parameters.

[0104] Semantic modeling module: used to construct a unified semantic representation space for industrial states.

[0105] Data Feature Dictionary Generation Module: Used to generate a unified data feature dictionary describing the semantic format of input and output at each workstation.

[0106] Model repository module: Used to store registered industrial AI models and their metadata.

[0107] Adaptability assessment and instantiation module: Implements the model-production line adaptability assessment algorithm and dynamically instantiates model services.

[0108] Semantic communication module: Enables semantic-level information transmission between cloud, edge, and device.

[0109] Layered inference engine module: Deployed on the edge controller, it performs lightweight front-end inference and confidence judgment.

[0110] Cloud-based collaborative reasoning module: Deployed on a cloud server, it performs heavyweight back-layer reasoning, knowledge distillation, and channel pruning.

[0111] Industrial AI Intelligent Agent Module: Enables semantic parsing, knowledge retrieval, model scheduling, and policy execution.

[0112] Meta-learning and rollback control module: Enables meta-learning training, cross-scenario zero-shot inference, security verification, and version rollback.

[0113] The offline simulation re-adaptation module provides a simulation environment and re-executes model adaptation offline after a rollback.

[0114] Furthermore, the meta-learning and rollback control module is equipped with multiple levels of safety thresholds, including a first safety threshold. The recommended value is 0.70, the second safety threshold. The recommended value is 0.85. When the confidence level is lower than the first safety threshold, a rollback is triggered, immediately rolling back to the previous stable version and starting offline re-adaptive training; when the confidence level is between the first and second safety thresholds, the current model continues to run, while performing digital twin simulation verification and issuing an early warning; when the confidence level is greater than 0.85, the new model can be deployed directly.

[0115] It should be noted that the specific content involved in the embodiments of this application can be found in the description of the above method embodiments, and will not be repeated here for the sake of brevity.

[0116] Corresponding to the above embodiments, this application also provides an electronic device.

[0117] See Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device 30 may include a processor 301, a memory 302, and a communication unit 303. These components communicate via one or more buses. Those skilled in the art will understand that the electronic device structure shown in the figures does not constitute a limitation on the embodiments of this application. It may be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0118] The communication unit 303 is used to establish a communication channel, thereby enabling the electronic device to communicate with other devices.

[0119] The processor 301 serves as the control center of the electronic device, connecting various parts of the device via various interfaces and lines. It executes software programs and / or modules stored in the memory 302, and calls data stored in the memory to perform various functions and / or process data. The processor can be composed of integrated circuits (ICs), such as a single packaged IC or multiple packaged ICs with the same or different functions connected together. For example, the processor 301 may consist only of a central processing unit (CPU). In this embodiment, the CPU may have a single processing core or include multiple processing cores.

[0120] Memory 302 is used to store the execution instructions of processor 301. Memory 302 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0121] When the execution instructions in memory 302 are executed by processor 301, the electronic device 30 is able to perform some or all of the steps in the above method embodiments.

[0122] Corresponding to the above embodiments, this application also provides a computer-readable storage medium, wherein the computer-readable storage medium may store a program, wherein when the program runs, it can control the device where the computer-readable storage medium is located to execute some or all of the steps in the above method embodiments. In specific implementation, the computer-readable storage medium may be a magnetic disk, an optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0123] Corresponding to the above embodiments, this application also provides a computer program product containing executable instructions that, when executed on a computer, cause the computer to perform some or all of the steps in the above method embodiments.

[0124] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0125] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0126] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0127] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0128] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A semantic-level collaborative reasoning method for industrial AI models in intelligent manufacturing production lines, characterized in that, include: Using raw observation data from heterogeneous devices as input, the data is mapped to a unified semantic space through a semantic mapping function, and then jointly trained with cross-device semantic alignment loss to obtain a unified semantic feature vector. A workstation data feature dictionary is generated based on the unified semantic space. The workstation data feature dictionary records the input and output semantic format and workstation semantic prototype of each workstation. The output of the model to be deployed is mapped to the unified semantic space, and the comprehensive adaptation index is calculated by combining the workstation semantic prototype and the workstation data feature dictionary. The deployment is then determined based on the comprehensive adaptation index. After deployment, semantic communication is performed using a unified semantic vector as the transmission object, and the model enters edge inference after semantic consistency verification. Calculate the multi-dimensional fusion confidence score of edge inference. If it is lower than the first threshold, cloud inference is triggered.

2. The semantic-level collaborative reasoning method for industrial AI models for intelligent manufacturing production lines according to claim 1, characterized in that, The process involves taking raw observation data from heterogeneous devices as input, mapping it to a unified semantic space through a semantic mapping function, and jointly training it with cross-device semantic alignment loss to obtain a unified semantic feature vector. Specifically, this includes: Raw multi-source industrial data output by the i-th device As input, through the semantic mapping function corresponding to the device. Mapping to a unified semantic space yields semantic feature vectors: ,in For the learnable parameters of the mapping function, To unify the dimensions of the semantic space; The semantic mapping function is jointly trained using cross-device semantic alignment loss and physical constraint terms; The cross-device semantic alignment loss is: ,in, The total number of devices participating in semantic alignment; i and j are the distinct device numbers, and Each pair of devices is counted only once; These are the sequence numbers of semantic samples of the same type. This represents the total number of semantic samples of the same type. For equipment The original samples belonging to the k-th semantic class; For equipment The original samples belonging to the k-th semantic class; For equipment The corresponding semantic mapping function; For equipment The corresponding semantic mapping function; It is the square of the L2 norm; The physical constraint term is the physical consistency loss corresponding to the physical laws of the equipment. ; Calculate the total loss: ,in, This refers to the loss of physical consistency corresponding to the physical laws of the equipment. These are the physical constraint weights.

3. The semantic-level collaborative reasoning method for industrial AI models for intelligent manufacturing production lines according to claim 1, characterized in that, The step of generating a workstation data feature dictionary based on the unified semantic space, wherein the workstation data feature dictionary records the input and output semantic format and workstation semantic prototype of each workstation, specifically includes: The set of historical semantic samples for the j-th workstation Calculate the mean to obtain the semantic prototype of this workstation: ; The semantic prototype The data is stored in association with workstation identifier, data modality, feature dimension, and protocol type to form the workstation data feature dictionary.

4. The semantic-level collaborative reasoning method for industrial AI models for intelligent manufacturing production lines according to claim 1, characterized in that, The process of mapping the output of the model to be deployed to the unified semantic space, calculating a comprehensive adaptation index by combining the workstation semantic prototype and the workstation data feature dictionary, and determining whether to deploy based on the comprehensive adaptation index specifically includes: The output of the model to be deployed is mapped to the unified semantic space via a semantic mapping function to obtain the model output semantic vector. ; calculate Workstation semantic prototype semantic fit between , where σ is the distance scale parameter; The semantic fit A comprehensive compatibility index is obtained by weighting and integrating factors such as functional category matching, data modality matching, performance compliance, resource adaptability, and protocol compatibility. ; Plug and play deployment is executed when Ω is greater than or equal to the first threshold. Adaptive fine-tuning is performed when the second threshold ≤ Ω < the first threshold; Deployment is rejected when Ω is less than the second threshold.

5. The semantic-level collaborative reasoning method for industrial AI models for intelligent manufacturing production lines according to claim 1, characterized in that, The process of deploying the model, using a unified semantic vector as the transmission object, performs semantic communication, and after semantic consistency verification, proceeds to edge inference. Specifically, this includes: The edge terminal takes the unified semantic feature vector as input, encodes it into compact semantic features by a semantic encoder, and dynamically adjusts the coding rate according to the channel signal-to-noise ratio: Where SNR is the signal-to-noise ratio of the channel between the edge and the cloud; Dynamic coding rate for semantic transmission; This represents the upper limit of the encoding bitrate; Based on the baseline bit rate; This is the signal-to-noise ratio adjustment coefficient; According to the semantic dimensions preserved after compression The compact semantic features are semantically compressed to obtain compressed semantic features, wherein... To unify the dimensions of the semantic space; The compressed semantic features are transmitted to the cloud via a channel; The semantic features are recovered by the semantic decoder in the cloud. The cosine similarity between edge-encoded features and cloud-decoded recovered features is used as a semantic consistency verification index. When the cosine similarity is lower than a preset similarity threshold, semantic transmission is determined to be distorted, and cloud inference is directly triggered. When the cosine similarity is not lower than the preset similarity threshold, edge inference is initiated.

6. The semantic-level collaborative reasoning method for industrial AI models for intelligent manufacturing production lines according to claim 1, characterized in that, If the multi-dimensional fusion confidence score of the calculated edge inference is lower than a first threshold, cloud inference is triggered, specifically including: The deployed model is decomposed into a lightweight front-end layer at the edge and a heavyweight back-end layer in the cloud. At the edge, preliminary inference is performed to output edge prediction results, and multi-dimensional fusion confidence is calculated: in, Maximize the output of edge reasoning probability; It is the difference between the highest and second-highest probabilities; The variance of the feature vector is output for the edge. To integrate weights, satisfy ; when When, the edge directly outputs the inference result; when At that time, the cloud-based weighted back layer is triggered to execute the complete inference; The Initialize using the set quantiles of the historical marginal confidence distribution, and dynamically update according to the following formula: in, To update the learning step size for the threshold, To determine the actual accuracy of edge inference based on real-time statistics from the production line. Set target accuracy for business operations.

7. The semantic-level collaborative reasoning method for industrial AI models for intelligent manufacturing production lines according to claim 1, characterized in that, Also includes: Record the sequence of working conditions and the frequency of cloud inference triggered by the edge. When the proportion of low-confidence inference at the edge in the sliding window exceeds the set threshold, use the cloud model as the teacher model, generate a lightweight student model for edge devices through knowledge distillation and channel pruning, and migrate and deploy the student model to the edge. The semantic mapping function parameters of the unified semantic space and the workstation semantic prototype are updated synchronously, and the teacher training set is reconstructed based on the updated unified semantic space.

8. The semantic-level collaborative reasoning method for industrial AI models for intelligent manufacturing production lines according to claim 7, characterized in that, The method of using a cloud-based model as the teacher model and generating a lightweight student model for edge devices through knowledge distillation and channel pruning specifically includes: Channels to be pruned are selected based on channel importance assessment, where the channel importance is the absolute value of the L1 norm of the convolution kernel weights or the scaling factor of the BN layer. A layer-by-layer differentiated pruning strategy is adopted, with the pruning rate of shallow feature extraction layers being lower than that of deep task-specific layers, resulting in an overall pruning rate of... ; Construct a composite loss function: Train the student model, where For classification cross-entropy loss, For the KL divergence distillation loss of the output layer, This is due to the distillation loss of the intermediate characteristic layer. To output the distillation loss weight, Weights for intermediate feature distillation loss; When the accuracy of the student model is not lower than a set percentage of the accuracy of the teacher model, the student model will be migrated and deployed to the edge.

9. A semantic-level collaborative reasoning system for industrial AI models in intelligent manufacturing production lines, used to implement the collaborative reasoning method as described in any one of claims 1-8, characterized in that, include: The semantic modeling module is used to take the original observation data from heterogeneous devices as input, map it to a unified semantic space through a semantic mapping function, and jointly train it with cross-device semantic alignment loss to obtain a unified semantic feature vector. The data feature dictionary generation module is used to generate a workstation data feature dictionary based on the unified semantic space. The dictionary records the input and output semantic format and workstation semantic prototype of each workstation. The adaptability assessment and instantiation module is used to map the output of the model to be deployed to the unified semantic space, combine the workstation semantic prototype and the workstation data feature dictionary to perform adaptability assessment, obtain a comprehensive adaptability index, and determine whether to deploy based on the comprehensive adaptability index. The semantic communication module is used to perform semantic communication on the deployed model, using a unified semantic vector as the transmission object, and then enter the edge inference after semantic consistency verification. The hierarchical inference engine module is used to calculate the multi-dimensional fusion confidence of edge inference. If it is lower than the first threshold, cloud inference is triggered.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the collaborative reasoning method as described in any one of claims 1-8.

Citation Information

Patent Citations

  • Large-model end-cloud collaborative reasoning method and system giving consideration to efficiency and privacy

    CN121919915A