Optimization method and system of AI fool-proof algorithm and digital model

Abstract

The invention relates to the technical field of data processing, and provides an AI fool-proof algorithm and digital model optimization method and system, and the method comprises the steps: collecting multi-modal process data, carrying out the time alignment and event segmentation, obtaining a standard behavior sequence and an abnormal behavior sequence, inputting a preset deep learning behavior prediction model according to the standard behavior sequence, and carrying out the prediction of an abnormal behavior. And carrying out structure matching on the behavior prediction result and the abnormal behavior sequence to obtain an optimized instruction set and semantic prompt information, and then optimizing a preset deep learning behavior prediction model to obtain an optimized fool-proof intelligent prediction model. By optimizing an original deep learning behavior prediction model, the adaptive recognition capability and risk avoidance capability of the model for abnormal behaviors are improved, and the problem that in a complex practical application environment, high-frequency operation environment changes cannot be flexibly processed, so that part of high-risk behaviors cannot be recognized or corrected in time, and the risk of abnormal behaviors is reduced is solved. And the problems of low accuracy and poor adaptability exist.

Description

Technical Field

[0001] The present application relates to the technical field of data processing, and in particular to an optimization method and system for AI fool-proofing algorithms and digital models. Background Art

[0002] An increasing number of industries are leveraging AI algorithms and digital models to improve efficiency, optimize decision-making processes, and predict behavioral trends. However, in many industries, particularly in manufacturing, finance, and healthcare, abnormal operational processes still frequently occur, posing risks to production efficiency and safety. These "foolproofing" issues often arise from operator error or environmental uncertainty. Therefore, researching and implementing AI foolproofing algorithms, combined with efficient digital model optimization methods, is a key technology for improving automation and reducing human error.

[0003] Among the relevant technical means, the current mainstream error-proofing algorithms use traditional supervised learning or unsupervised learning methods to label, cluster, and classify historical data. For the detection of abnormal behavior, existing technologies mostly use rule-based judgment methods or clustering algorithms, which can effectively identify abnormal behavior and issue alarms in most cases, thereby reducing human errors.

[0004] Regarding the above technical solutions, although the existing technology can effectively realize the identification and early warning of abnormal behaviors in some standard scenarios through clustering, rule judgment and other means, in complex actual application environments, the existing error-proofing algorithms are difficult to cope with complex data with diverse data types, high dimensions and time series, and cannot flexibly handle high-frequency changes in the operating environment. This has led to some high-risk behaviors not being identified or corrected in a timely manner, resulting in problems of low accuracy and poor adaptability. Summary of the Invention

[0005] In order to improve the problem of being unable to flexibly handle high-frequency operating environment changes in complex actual application environments, resulting in some high-risk behaviors not being identified or corrected in a timely manner, and the existence of low accuracy and poor adaptability, this application provides an optimization method and system for AI anti-foolproofing algorithms and digital models.

[0006] The present invention provides an optimization method for an AI fool-proofing algorithm and a digital model, comprising: collecting multimodal process data, performing time alignment and event segmentation on the multimodal process data to obtain a standard behavior sequence and an abnormal behavior sequence; inputting the standard behavior sequence into a preset deep learning behavior prediction model to obtain a behavior prediction result, performing structural matching on the behavior prediction result and the abnormal behavior sequence to obtain fused feature data; performing high-dimensional clustering and deviation correction on the fused feature data to obtain a corrected clustering result and boundary residual information, generating a fool-proofing judgment rule set based on the corrected clustering result, using the fool-proofing judgment rule set to map and match the boundary residual information to obtain a risk label and optimization suggestion; mapping the risk label to a behavior segment corresponding to the fused feature data to generate a converged optimization feature representation, using the converged optimization feature representation to perform association annotation and behavior semantic mapping on the optimization suggestion to obtain an optimization instruction set and semantic prompt information; and optimizing the preset deep learning behavior prediction model using the optimization instruction set and the semantic prompt information to obtain an optimized fool-proofing intelligent prediction model.

[0007] As a preferred solution, the steps of collecting multimodal process data, time-aligning and event-segmenting the multimodal process data, and obtaining standard behavior sequences and abnormal behavior sequences include: synchronously collecting action signals, environmental parameters, and user interaction logs during the operation process through a multi-channel sensing device to construct multimodal process data, time-aligning and event-segmenting the multimodal process data to obtain event fragment sequences and context window information; inputting the event fragment sequence into a preset graph convolutional network model to extract spatial correlation features and temporal relationship features of the operation action, performing abnormal pattern discrimination on the context window information to obtain an abnormal score matrix and a behavior credibility vector; fusing the behavior credibility vector with the temporal relationship features to generate a standard behavior sequence; jointly analyzing the abnormal score matrix and the spatial correlation features, and constructing an abnormal behavior sequence based on the analysis results.

[0008] As a preferred solution, the step of inputting the standard behavior sequence into a preset deep learning behavior prediction model to obtain a behavior prediction result, performing structural matching on the behavior prediction result and the abnormal behavior sequence to obtain fused feature data includes: performing sequence embedding encoding on the standard behavior sequence to obtain an encoded standard behavior sequence, inputting the encoded standard behavior sequence into a deep learning behavior prediction model based on a Transformer structure, and outputting a multi-dimensional behavior prediction result; performing hierarchical attention analysis on the behavior prediction result to extract key behavior nodes, constructing a structural difference vector using the key behavior nodes and the abnormal behavior sequence, and extracting trigger offset features based on the structural difference vector; performing residual feedback on the structural difference vector and the intermediate features of the deep learning behavior prediction model to obtain a feature compensation vector and an error inverse map; constructing a behavior error fusion feature tensor based on the trigger offset feature and the feature compensation vector, and performing tensor-level fusion on the behavior error fusion feature tensor and the error inverse map to obtain fused feature data.

[0009] As a preferred solution, the steps of performing hierarchical attention analysis on the behavior prediction results to extract key behavior nodes, and constructing a structural difference vector using the key behavior nodes and the abnormal behavior sequence include: weighting the outputs of each layer of the behavior prediction results through a hierarchical attention mechanism, extracting key behavior nodes based on the weighted results, and applying a dynamic time warping algorithm to calculate the similarity between the key behavior nodes; identifying the deviation part in the abnormal behavior sequence, and constructing a structural difference vector using the similarity and the deviation part.

[0010] As a preferred solution, the steps of performing high-dimensional clustering and deviation correction on the fused feature data to obtain corrected clustering results and boundary residual information, generating a fool-proof judgment rule set based on the corrected clustering results, mapping and matching the boundary residual information using the fool-proof judgment rule set, and obtaining risk labels and optimization suggestions include: performing dimensionality reduction preprocessing on the fused feature data according to principal component analysis to obtain dimensionality reduction features, performing high-dimensional clustering on the dimensionality reduction features to obtain clustering results and a behavior clustering map; performing deviation modeling on the clustering results and the behavior clustering map to calculate the behavior offset, dynamically adjusting the boundary conditions in the clustering results according to the behavior offset, correcting the behavior category division range, and obtaining corrected clustering results and boundary residual information; constructing a fool-proof judgment rule set based on the corrected clustering results, mapping and matching the boundary residual information with the fool-proof judgment rule set, and extracting rule conflict groups and risk level factors based on the matching results; comprehensively analyzing the risk level factors and the historical frequency weights of the rule conflict groups, and generating corresponding risk labels and optimization suggestions based on the analysis results.

[0011] As a preferred solution, the step of mapping the risk label to the behavior segment corresponding to the fused feature data to generate a converged optimization feature representation, and using the converged optimization feature representation to associate and annotate the optimization suggestion with behavioral semantics to obtain an optimization instruction set and semantic prompt information includes: mapping the risk label to the behavior segment corresponding to the fused feature data to extract high-risk behavior area features, using a convolutional coding network to perform context modeling and semantic normalization on the high-risk behavior area features to generate a candidate optimization parameter set; performing a weighted combination of the candidate optimization parameter set and the original fused feature to construct a modified input tensor, and performing a weighted combination of the candidate optimization parameter set and the original fused feature to generate a candidate optimization parameter set; The corrected input tensor undergoes multiple rounds of residual regression and error inversion processing to generate a converged optimization feature representation; wherein the original fused feature refers to a high-dimensional feature set after the fusion of user interaction data, sensor data, and historical behavior data; the converged optimization feature representation and the optimization suggestion are associated and labeled and behavioral semantically mapped to construct an optimization instruction set; the task target items in the optimization instruction set are converted into prompt statement templates, and the prompt statement templates are combined with the user behavior context to generate semantic prompt information; wherein the user behavior context refers to the timestamp, operation sequence, previous state and environmental conditions of the user's current operation.

[0012] As a preferred solution, the step of using the optimization instruction set and the semantic hint information to optimize the preset deep learning behavior prediction model to obtain an optimized fool-proof intelligent prediction model includes: using the semantic hint information to perform label enhancement on the intermediate layer behavior representation of the preset deep learning behavior prediction model to obtain an enhanced deep learning behavior prediction model; modularly decomposing the optimization instruction set, constructing a corresponding parameter tuning path based on the decomposition result, and using the parameter tuning path to optimize the training hyperparameters, activation strategy and inter-layer connection method of the enhanced deep learning behavior prediction model to obtain an optimized fool-proof intelligent prediction model.

[0013] The present application also provides an optimization system for AI fool-proofing algorithm and digital model, including: an acquisition module for acquiring multimodal process data, performing time alignment and event segmentation on the multimodal process data, and obtaining standard behavior sequences and abnormal behavior sequences; a matching module for inputting the standard behavior sequences into a preset deep learning behavior prediction model to obtain behavior prediction results, performing structural matching on the behavior prediction results and the abnormal behavior sequences to obtain fused feature data; a correction module for performing high-dimensional clustering and deviation correction on the fused feature data to obtain corrected clustering results and boundary residual information, and based on the correction module, The positive clustering results generate a fool-proof judgment rule set, which is used to map and match the boundary residual information to obtain risk labels and optimization suggestions; a mapping module is used to map the risk labels to the behavior segments corresponding to the fused feature data to generate a convergent optimization feature representation, and the convergent optimization feature representation is used to associate and label the optimization suggestions with behavioral semantics to obtain an optimization instruction set and semantic prompt information; an optimization module is used to optimize the preset deep learning behavior prediction model using the optimization instruction set and the semantic prompt information to obtain an optimized fool-proof intelligent prediction model.

[0014] Compared with the existing technology, this application has the following beneficial effects: high accuracy and strong adaptability. By synchronously collecting multimodal process data through multiple channels and performing time alignment and event segmentation, a comprehensive perception of user behavior and operation scenarios is ensured; by inputting standard behavior sequences into a preset deep learning behavior prediction model, combined with hierarchical attention analysis, dynamic time warping algorithm and structural matching of abnormal behavior sequences, structural difference vectors and trigger offset features are effectively identified and extracted, and high-dimensional clustering and boundary residual information correction technology are used to construct a fool-proof judgment rule set and output accurate risk labels and personalized optimization suggestions; based on the mapping relationship between risk labels and fused feature data, high-risk behavior area features are extracted and converged optimization feature representations are generated, and optimization instruction sets and semantic prompt information are generated through the semantic understanding mechanism to achieve a closed-loop connection between behavior control and model feedback; the original deep learning behavior prediction model is optimized by combining semantic prompt information and optimization instruction sets to improve the model's adaptive recognition ability and risk avoidance ability for abnormal behaviors, and obtain an optimized fool-proof intelligent prediction model, which improves the problem of being unable to flexibly handle high-frequency operating environment changes in complex actual application environments, resulting in some high-risk behaviors not being identified or corrected in a timely manner, and having low accuracy and poor adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] The structures, proportions, sizes, etc. depicted in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not intended to limit the conditions under which the present invention can be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportional relationships, or adjustments in size should still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and objectives that can be achieved by the present invention.

[0017] Figure 1 1 is a flow chart of an AI fool-proofing algorithm and a digital model optimization method provided by an embodiment of the present invention; Figure 2 It is a schematic block diagram of the structure of the AI ​​fool-proofing algorithm and digital model optimization system provided by an embodiment of the present invention.

[0018] Description of reference numerals: 10. AI anti-mistake algorithm and digital model optimization system; 11. Acquisition module; 12. Matching module; 13. Correction module; 14. Mapping module; 15. Optimization module. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined, or partially merged, so the actual execution order may vary depending on the actual situation.

[0021] It should also be understood that the terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit the present application. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0022] It should be further understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and through specific implementation methods.

[0024] Example 1: like Figure 1 As shown, the AI ​​fool-proofing algorithm and digital model optimization method and system provided in the embodiment of the present application include steps S100 to S300.

[0025] Example 1: This application provides an optimization method for an AI foolproof algorithm and a digital model, which specifically includes the following steps: Step S100: Collect multimodal process data, perform time alignment and event segmentation on the multimodal process data, and obtain standard behavior sequences and abnormal behavior sequences.

[0026] In this step, multi-channel sensing devices, including visual cameras, motion capture sensors, voice recognition modules, and environmental monitoring sensors, are deployed in the operating environment to synchronously collect multimodal process data in real time. This data covers a variety of information, including motion signals, voice commands, environmental conditions (such as temperature and humidity), and device response status, that occur during the operation. Specifically, the multimodal process data is time-aligned, and the sliding window technology is used to divide the data into event segments. The event segments are then segmented according to preset rules to extract standard behavior sequences and abnormal behavior sequences. The standard behavior sequence represents the normal operating process, and the abnormal behavior sequence represents abnormal operations that do not conform to the standard behavior.

[0027] For example, on a production line, operators collect hand movements (via motion sensors), voice commands (via voice recognition), and environmental data (via temperature and humidity sensors) as they perform tasks. This information is synchronized using a time alignment algorithm and divided into 10-second event segments. This allows the normal operating process to be identified and a standard behavior sequence to be constructed. Furthermore, abnormal operations that deviate from the standard process are identified and aberrant behavior sequences are formed.

[0028] Step S200: Input the standard behavior sequence into a preset deep learning behavior prediction model to obtain a behavior prediction result, perform structural matching on the behavior prediction result and the abnormal behavior sequence to obtain fused feature data.

[0029] In this step, the standard behavior sequence is input into a preset deep learning-based behavior prediction model, specifically using a deep learning behavior prediction model based on the Transformer structure to obtain the behavior prediction results; specifically, the behavior prediction results include future behavior patterns predicted based on the standard behavior sequence. The model uses a hierarchical attention mechanism to weight the key behavior nodes in the standard behavior sequence, and then performs structural matching with the abnormal behavior sequence to extract the structural differences and behavior deviation information, and finally fuses these features into fused feature data.

[0030] For example, if the standard behavior sequence is "Action A - B - C," the deep learning model will predict "A' - B' - C'." If the abnormal behavior sequence is "A - X - C," the inconsistency between "B" and "X" is detected through structural matching and fused with the behavior deviation features to generate fused feature data.

[0031] Step S300: Perform high-dimensional clustering and deviation correction on the fused feature data to obtain the corrected clustering results and boundary residual information, generate a fool-proof judgment rule set based on the corrected clustering results, use the fool-proof judgment rule set to map and match the boundary residual information, and obtain risk labels and optimization suggestions.

[0032] In this step, principal component analysis (PCA) is performed on the fused feature data to reduce its dimensionality, retaining the main abnormal feature subspace and obtaining the feature data after dimensionality reduction. Specifically, a density-based adaptive clustering algorithm (such as DBSCAN) is used to perform high-dimensional clustering on the reduced feature data to obtain preliminary clustering results and behavioral clustering maps. Then, the boundary offset between the behavioral category and the clustering result is calculated through deviation modeling, and the clustering result is dynamically adjusted according to the boundary residual information to generate a corrected clustering result. Based on the corrected result, a fool-proof judgment rule set is constructed, and the boundary residual information is mapped and matched using this rule set to obtain risk labels and optimization suggestions.

[0033] For example, suppose that after dimensionality reduction, behavioral features are represented as 10-dimensional data. After DBSCAN clustering, multiple behavioral categories are obtained. During the clustering process, it is discovered that some behavioral categories overlap. Through deviation modeling, the boundaries are adjusted, and based on this adjustment, a foolproof rule is generated, such as "Operation sequences that do not comply with regulations should be marked as high risk." This generates a corresponding risk label and provides an optimization suggestion: "Recommend adjusting the operation sequence."

[0034] Step S400: Map the risk label to the behavior segment corresponding to the fusion feature data to generate a converged optimization feature representation, use the converged optimization feature representation to associate and label the optimization suggestions with behavioral semantic mapping, and obtain the optimization instruction set and semantic prompt information.

[0035] In this step, risk labels are mapped to behavioral segments corresponding to the fused feature data, identifying high-risk behavioral regions. Specifically, a convolutional coding network is used to contextually model these high-risk behavioral regions and generate a set of candidate optimization parameters. This set of optimization parameters is then weighted and combined with the original fused feature data to construct a modified input tensor. Multiple rounds of residual regression and error inversion are performed to obtain a converged optimized feature representation. The semantic understanding module then associates this optimized feature representation with optimization suggestions, annotating them and mapping them to behavioral semantics. This extracts action targets, triggering conditions, and risk avoidance paths, ultimately yielding an optimized instruction set and semantically informed information.

[0036] For example, if "Operation B" is detected to be executed too quickly, a risk label of "Operation sequence too fast" is generated. The convolutional coding network analyzes the context of the segment and derives the optimization parameters for "increasing operation delay." After regression processing, a converged optimization feature representation is obtained. The semantic understanding module generates an optimization instruction: "Please increase the completion time of Operation A," and displays a semantic prompt on the user interface: "Please check the task sequence."

[0037] Step S500: Optimize the preset deep learning behavior prediction model using the optimization instruction set and semantic prompt information to obtain an optimized fool-proof intelligent prediction model.

[0038] In this step, semantic cues are used to label and enhance the intermediate-layer features of the deep learning behavior prediction model. The model's learning objectives are then adjusted based on the optimization instruction set. Specifically, the optimization instruction set is modularized to extract the model's optimization factors, such as training hyperparameters, activation strategies, and inter-layer connectivity, and a tuning path is constructed. This tuning path optimizes the deep learning behavior prediction model, resulting in an optimized, foolproof intelligent prediction model.

[0039] For example, by optimizing the instruction set, adjusting the weights of the Transformer layer 3 attention heads and reducing the training learning rate (for example, from 0.001 to 0.0005), the model's sensitivity to abnormal behavior was further enhanced. After training, the optimized model improved its prediction accuracy for abnormal behavior in complex environments by 15%, successfully reducing false alarms.

[0040] In this embodiment, multimodal process data is synchronously collected using a multi-channel sensing device, encompassing motion signals, environmental parameters, and user interaction logs during the operation. Next, the multimodal process data is time-aligned and event-segmented to extract standard and abnormal behavior sequences. The standard behavior sequences are input into a pre-defined deep learning behavior prediction model, which generates behavior prediction results. Next, the behavior prediction results are structurally matched with the abnormal behavior sequences to generate fused feature data. High-dimensional clustering and bias correction are then performed on the fused feature data to generate corrected clustering results and boundary residual information. A set of foolproofing judgment rules is generated based on the corrected clustering results. The boundary residual information is mapped and matched using the foolproofing judgment rule set to generate risk labels and optimization recommendations. The risk labels are then mapped to the corresponding behavior segments in the fused feature data, generating a converged and optimized feature representation. Based on the converged and optimized feature representation, the optimization recommendations are associated and annotated with behavior semantics, resulting in an optimized instruction set and semantic hint information. Finally, the optimized instruction set and semantic hint information are used to optimize the pre-defined deep learning behavior prediction model, resulting in an optimized foolproofing intelligent prediction model.

[0041] Through the efficient collection and analysis of multimodal process data, accurate identification and anomaly detection of complex behavior patterns are achieved. On the basis of the existing fool-proofing algorithm, through deep learning and high-dimensional data processing technology, the behavior prediction results are integrated with the abnormal behavior sequence, which effectively improves the prediction accuracy. Through high-dimensional clustering and deviation correction, the present invention can accurately identify and correct abnormal behaviors, and generate a set of fool-proofing judgment rules, which greatly enhances the ability to cope with changing environments and improves the inability to flexibly handle high-frequency operating environment changes in complex actual application environments, resulting in some high-risk behaviors not being identified or corrected in time, and there are problems of low accuracy and poor adaptability. In addition, optimizing the generation method of instruction sets and semantic prompt information can not only timely feedback and correct potential risks, but also effectively improve the accuracy and safety of user operations, optimize the overall performance of the fool-proofing intelligent prediction model, and ensure a more stable and secure operating process.

[0042] Example 2: In step S100, the steps of collecting multimodal process data, performing time alignment and event segmentation on the multimodal process data, and obtaining standard behavior sequences and abnormal behavior sequences specifically include: The motion signals, environmental parameters and user interaction logs during the operation are synchronously collected through multi-channel sensing devices to construct multimodal process data. The multimodal process data is time-aligned and event-segmented to obtain event fragment sequences and context window information.

[0043] Multimodal process data is constructed by deploying multi-channel sensing devices in the operating environment to collect data from heterogeneous signal sources. Specifically, the multi-channel sensing devices include gesture capture devices (for collecting motion signals), environmental monitoring sensors (for collecting environmental parameters such as temperature, humidity, and illumination), console log interfaces (for recording user interactions and device responses), and voice recognition modules (for capturing voice control commands). Multimodal process data consists of the combined behavioral data stream formed by arranging the above sensor data by timestamp.

[0044] Time alignment uses a unified time base (e.g., millisecond-level operation timestamps) to synchronize processing of various signals. A window sliding algorithm is used to segment continuous multimodal process data into sequences of fixed-duration (e.g., 10-second) event segments. Furthermore, combined with a pre- and post-event context feature extraction module, a time window (e.g., ±5 seconds) is introduced around each segment to extract contextual information for subsequent causal modeling and state determination in behavioral assessment.

[0045] For example, on an automobile assembly line, a worker issues a "lock" command via voice while simultaneously performing manual operations to secure a part. The "lock" command is captured by the voice recognition module, gesture capture captures the hand's grip and rotation angle, environmental sensors record the current temperature and illumination, and the operation log records the device's response status. This information is synchronously integrated into a multimodal process data record, segmented into a sequence of event segments, such as "tightening a bolt." The context window includes information such as "previous operation: positioning the part," "current temperature: 28°C," and "normal lighting."

[0046] The event fragment sequence is input into the preset graph convolutional network model to extract the spatial correlation features and temporal relationship features of the operation action, perform abnormal pattern discrimination on the context window information, and perform comparative learning with historical label samples to obtain the anomaly score matrix and behavior credibility vector.

[0047] By constructing a graph-based behavioral node network, the event segment sequence is mapped into a graph data structure and fed into a graph convolutional network (GCN) model to extract spatial and temporal features. Specifically, each event segment is modeled as a node in the graph, and the collaborative relationships and sequence between different operation events are connected through edges. The GCN uses a multi-layer aggregation structure to perform aggregation calculations on each node and its neighbors, obtaining a representation of the spatial dependencies and temporal behavior of each segment.

[0048] Context window information is incorporated into node attributes as auxiliary features to enhance the ability to determine the current behavior's state. The extracted features are then compared with historically annotated sequences of normal and abnormal behaviors, using contrastive loss functions such as Triplet Loss or NT-Xent to train feature discrimination. The output includes an anomaly score matrix representing the degree of abnormality for each segment (higher scores indicate greater abnormality), and a behavior credibility vector composed of corresponding credibility scores (indicating the confidence that the current behavior is "standard behavior").

[0049] For example, in a certain process, the normal operation flow is "grasping - positioning - tightening." The graph model constructs a directed edge graph of these three steps. After input into the GCN, node 1 (grasping) and node 2 (positioning) are strongly connected. If the "grasping-tightening" skip behavior occurs in the current data, the GCN detects that this sequence violates the normal structure. Contextual information, such as "missing device response logs," is encoded into node attributes to further enhance recognition. This anomalous sequence is assigned an anomaly score of 0.87 (between 0 and 1), and the behavior credibility vector is [0.3, 0.1, 0.9], indicating that the confidence of the first two steps is low, while the confidence of the tightening operation is high.

[0050] The behavior credibility vector is fused with the temporal relationship features, and a standard behavior sequence is generated through confidence reconstruction.

[0051] By fusing temporal relationship features extracted from graph convolutional networks with behavior credibility vectors, the behavior path is reconstructed to generate a more accurate standard behavior sequence. Specifically, the trusted operation nodes represented by the behavior credibility vectors are screened, retaining high-confidence event segments while discarding or filling in low-confidence segments. The fused behavior path is then optimized through temporal alignment and sequence rearrangement to form a coherent and representative standard path.

[0052] This process is equivalent to extracting credible subgraphs from the original graph structure and then outputting them as sequences according to the order in which behaviors are executed, forming a standard behavior sequence. This sequence is then used for subsequent comparative analysis with abnormal behavior sequences, model training, and foolproof rule construction.

[0053] For example, for the behavioral sequence "grab - position - tighten - place," if the confidence level of "position" is 0.2 and the confidence levels of the others are all above 0.8, the "position" step can be removed, and the reconstructed standard behavioral sequence "grab - tighten - place" can be used. If the confidence level is low but the structure should be retained, a representative sample derived from historical frequency matching can be inserted as a replacement to ensure sequence integrity and rationality.

[0054] The anomaly score matrix and spatial correlation features are jointly analyzed to screen out segments matching abnormal patterns and their behavior labels, and abnormal behavior sequences are constructed based on the analysis results.

[0055] By jointly calculating the anomaly score matrix obtained above with the spatial correlation features extracted by GCN, event segments that deviate from the normal path in spatial structure and have higher scores are identified; specifically, a set threshold (such as anomaly score > 0.7) is used as the initial screening condition, and then node structure anomalies such as skipping, loops, and missing nodes are identified through topological variations in the behavior graph.

[0056] At the same time, combined with historical label samples in the abnormal behavior database, the current abnormal fragments are labeled and matched, and behavioral labels (such as "wrong operation sequence", "device unresponsive", "unauthorized operation", etc.) are automatically assigned. The filtered fragments are arranged in the original chronological order to form an abnormal behavior sequence with semantic information.

[0057] For example, during a robot operation, a sequence of "pick up an object - move it - place it down - pick up an object" was detected. GCN spatial analysis revealed repeated "pick up an object" without any intermediate tasks, thus identifying it as an anomaly. Combined with the anomaly score matrix, the score for this "repeated pick up" behavior was 0.91, exceeding the set threshold of 0.7. The anomaly label library was used to label this behavior as "repeated execution," and this sequence was constructed as an abnormal behavior sequence for subsequent behavioral model analysis and error-proofing strategy learning.

[0058] In step S200, the standard behavior sequence is input into a preset deep learning behavior prediction model to obtain a behavior prediction result, and the behavior prediction result and the abnormal behavior sequence are structurally matched to obtain fusion feature data, which specifically includes: The standard behavior sequence is sequence-embedded and encoded to obtain the encoded standard behavior sequence. The encoded standard behavior sequence is input into the deep learning behavior prediction model based on the Transformer structure to output multi-dimensional behavior prediction results.

[0059] By performing sequence embedding encoding on standard behavior sequences to fully extract the spatial and temporal features of the behavior sequences, the embedding results are input into a preset deep learning behavior prediction model based on the Transformer structure, and multi-dimensional behavior prediction results are output from the model. Specifically, an independent vector representation is generated for each behavior node in the standard behavior sequence, and the attributes of the behavior node (such as duration, device response status, and environmental context) are compressed into a continuous vector of fixed dimension through embedding. The embedding sequences corresponding to multiple nodes are then input into the embedding layer of the Transformer model. The Transformer model uses a multi-head self-attention mechanism to encode the features of the node sequence, extracting node-level features (the time dependency and state association of a single behavior node) and sequence-level features (the overall time evolution trend of the entire behavior sequence) respectively, and obtaining multi-dimensional behavior prediction results including time dimension, spatial dimension, and behavior feature dimension at the output layer of the model.

[0060] For example, suppose a standard behavior sequence is "Action A - Action B - Action C," where each action node contains corresponding attributes such as "Duration: 5 seconds," "Device status: Successful response," and "Environment: Temperature: 35°C." The embedding encoding process generates a vector for each node, such as "Action A: [0.2, 0.5, 0.3]," which forms the sequence input to the Transformer. The model's multi-head attention mechanism detects that "Action A and Action C are highly correlated" and that "device status anomalies occur during Action B." It then outputs a multi-dimensional prediction result, including features such as the predicted next action node, time span, and device response time, providing a basis for further analysis.

[0061] A hierarchical attention analysis is performed on the behavior prediction results to extract key behavior nodes. The structural difference vector is constructed using the key behavior nodes and abnormal behavior sequences. The trigger offset features are extracted based on the structural difference vector. By comparing and analyzing the changes in behavior patterns in different situations, the key offset features that can trigger abnormal behaviors are obtained, which are defined as trigger offset features.

[0062] By performing hierarchical attention analysis on the output of each layer of the behavior prediction results, the importance order of the behavior nodes is calculated to extract the key behavior nodes. Subsequently, the extracted key behavior nodes are compared and analyzed with the abnormal behavior sequences through a structural matching algorithm, and a structural difference vector is generated using this structural matching. Specifically, a multi-layer attention weight reduction mechanism is set up (such as doubling the encoding weights of the last two layers), and the time deviation between the key behavior nodes and the abnormal behavior nodes is quantified through the Dynamic Time Warping (DTW) algorithm to extract the temporal offset characteristics of the behavior. Then, by statistically analyzing and comparing the frequencies of associated behavior sequences in different scenarios, the contribution of behavior offset to risk triggering is further analyzed, and the key behavior variables defined as trigger offset characteristics are obtained.

[0063] For example, for the predicted result "Operation A - Operation B - Operation C" and the abnormal behavior sequence "Operation A - Operation X - Operation C," the key behavior node corresponding to "Operation B" is identified. Using a hierarchical attention mechanism, the "Operation B weight = 0.9" and "Operation X weight = 0.2" are calculated, determining that the trigger risk for deviation behavior in Operation X is high. Combined with contextual frequency analysis, it is found that "Operation X often occurs in emergency stop environments." "Skipping steps when the action is not completed" is defined as a trigger deviation feature for subsequent error-proofing rule optimization.

[0064] The structural difference vector is subjected to residual feedback with the intermediate features of the deep learning behavior prediction model to obtain the feature compensation vector and the error reverse map.

[0065] By inputting the structural difference vector into the middle layer of a deep learning behavior prediction model, a residual feedback comparison is performed with the model's multi-head attention output. Specifically, during the prediction process of each behavior node, a negative gradient correction is applied to the feature output of low-weight nodes (those with abnormal behaviors), making low-confidence abnormal behaviors more dominant in the model's features. Simultaneously, a residual mechanism is used to generate an error inversion map by calculating the modulus of the difference vector and the partial derivative of the model's activation function. The purpose of the error inversion map is to impose "negative correlation weights" on low-impact nodes. Using a compensation mechanism, the model is retrained to make the behavior predictions more consistent with actual abnormal patterns, generating a feature compensation vector for model optimization.

[0066] For example, if "Operation B" is determined to be a key node but its predicted weight is too low, at 0.4, the residual feedback mechanism calculates the characteristic deviation of this node and corrects the weight using the function f(x) = αx + β (where α represents the model correction coefficient and β represents the compensation offset), PB [compensated weight = 0.8]. The error inversion graph shows that the model is less sensitive to the trigger mechanism of "Operation X," requiring an increase in the activation partial derivative as a correction path.

[0067] A behavioral error fusion feature tensor is constructed according to the trigger offset feature and the feature compensation vector. The behavioral error fusion feature tensor is fused with the error inverse map at the tensor level to obtain the fused feature data.

[0068] By utilizing the obtained trigger offset features to perform weighted adjustments on the feature compensation vector at the feature level, a behavioral error fusion feature tensor is formed. This tensor is then gradually fused with the error inverse map at the tensor level. Specifically, the fusion process involves using matrix operations based on tensor products to extract salient information points from the feature compensation vectors corresponding to multiple behavioral categories using the trigger offset features. The weight deviations and anomaly locations are then redistributed to generate a stable global behavioral feature tensor, or fused feature data. This process not only uniformly describes the differences between standard and abnormal behavior sequences but also effectively reflects the temporal offset and behavioral anomaly patterns across the entire graph structure feature space.

[0069] For example, for the standard behavior sequence "Operation A - Operation B - Operation C" and the abnormal behavior sequence "Operation A - Operation X - Operation C", the fused feature data will trigger the offset feature to focus on the "time offset value of Operation X" of 5 seconds. By fusing the feature tensor with the behavior error, the "representative weight of Operation X in the abnormal environment" is enhanced. The abnormal behavior characteristics are visualized as risk propagation paths through tensor graphs for subsequent optimization module analysis.

[0070] The steps of performing hierarchical attention analysis on the behavior prediction results to extract key behavior nodes and constructing a structural difference vector using the key behavior nodes and abnormal behavior sequences include: The output of each layer of the behavior prediction results is weighted through the hierarchical attention mechanism, and the behavior nodes with high attention are extracted based on the weighted results to obtain the key behavior nodes. The key behavior nodes are the nodes that best represent the specific operation mode and behavior type in the prediction model. The dynamic time warping algorithm is applied to calculate the similarity between the key behavior nodes.

[0071] By utilizing a hierarchical attention mechanism to perform significance analysis on the outputs of each layer of a deep learning behavior prediction model and dynamically weighting them based on the attention weights, the model extracts behavior nodes with high attention as key behavior nodes. Specifically, the hierarchical attention mechanism divides the multi-layer output of the behavior prediction model into different levels, such as the feature encoding layer, the attention mechanism extraction layer, and the behavior prediction output layer. The output weights of each layer are based on a preset attention evaluation function (such as activation value strength or convergence information during model training). Highly weighted behavior nodes (e.g., those assigned a weight greater than 0.7) are defined as "key behavior nodes" and are considered to best represent the current operating mode and behavior type in the model.

[0072] Subsequently, the Dynamic Time Warping (DTW) algorithm was used to calculate the temporal similarity between key behavior nodes. This algorithm aims to measure the temporal variation between standard and abnormal behavior sequences, as well as the regularity of behavioral connections. DTW calculates the temporal offset of each behavior node point by point and establishes the optimal matching path, thereby generating a temporal similarity distribution between behavior nodes.

[0073] For example, when predicting the standard behavior sequence "Action A - B - C," the hierarchical attention mechanism detected an attention weight of 0.85 for "Action B," thus defining it as a key behavior node. Subsequently, when comparing it to the abnormal sequence "Action A - X - C," a DTW calculation revealed that the matching score between "Action X" and the key node "Action B" was 0.68, indicating a significant time offset and weak sequence association for "Action X." Therefore, it was considered a potential abnormal operation for further analysis.

[0074] Identify the deviation part in the abnormal behavior sequence and construct a structural difference vector using the similarity and deviation part.

[0075] Through hierarchical attention expansion analysis based on key behavior nodes and combined with the temporal similarity calculation results of the dynamic time warping algorithm, abnormal operations in abnormal behavior sequences are screened and accurately identified, and the deviation features are extracted to construct a structural difference vector. Specifically, for behavior nodes in abnormal behavior sequences that have significant time offset, sequence disorder, or semantic inconsistency with key behavior nodes, the time offset detection module is first used to calculate the offset value (such as delay, advance, and skipping between operations), and then the sequence matching module is used to determine the correlation distance between nodes (such as the degree of order disorder between two behaviors). The features of the deviant behavior nodes are quantified to generate a structural difference vector containing multi-dimensional features such as time offset, sequence difference, and semantic deviation.

[0076] The structural difference vector is used to describe the deviation between the standard behavior sequence and the abnormal behavior sequence in terms of spatial position, temporal connection and behavior logic chain, providing information basis for subsequent trigger feature extraction and error-proofing judgment rule generation.

[0077] For example, suppose the standard behavior sequence is "Operation A - Operation B - Operation C," while the abnormal behavior sequence is "Operation A - Operation X - Operation C." The key behavior node "Operation B" is replaced by "Operation X" in the abnormal sequence. The time offset detection module calculates that "Operation X" completes 3 seconds later than "Operation B." The sequence analysis module determines that the correlation between "Operation X" and "Operation A" is only 0.45. Combining this deviation information, a structural difference vector is generated with a time offset of 3 seconds, a sequence matching degree of 0.45, and a semantic difference degree of 0.7, which represents the degree of temporal and logical deviation of this abnormal behavior from the standard behavior.

[0078] In step S300, high-dimensional clustering and deviation correction are performed on the fused feature data to obtain corrected clustering results and boundary residual information. A set of error-proofing judgment rules is generated based on the corrected clustering results. The boundary residual information is mapped and matched using the set of error-proofing judgment rules to obtain risk labels and optimization suggestions. The steps specifically include: The fusion feature data is preprocessed for dimensionality reduction based on principal component analysis, and the key abnormal feature subspace is retained to obtain the dimensionality reduction features. The dimensionality reduction features are then clustered in high dimensions using a density-based adaptive clustering algorithm to obtain clustering results and behavioral clustering maps.

[0079] Principal Component Analysis (PCA), a statistical analysis technique, is used to reduce the dimensionality of the fused feature data to preserve the key anomaly feature subspace. Specifically, the variance of each dimension in the fused feature data is first calculated, and the top N principal components with the highest variance contributions are selected (N is determined based on a set threshold, such as a cumulative variance contribution of 95% or higher) to generate a reduced-dimensional feature representation. This reduced-dimensionality feature not only reduces computational complexity but also focuses on the significant characteristics of abnormal behavior, facilitating subsequent cluster analysis.

[0080] The density-based adaptive clustering algorithm (DBSCAN) then performs high-dimensional clustering on the reduced feature data. By defining a "density threshold" and a "radius threshold," it automatically clusters the behavior into categories. The clustering results can be divided into multiple behavior categories. The features of each category are annotated and mapped to a behavior cluster map, forming a visualization of the behavior distribution in the high-dimensional feature space.

[0081] For example, assuming the fused feature data is a high-dimensional matrix containing 100 features, PCA analysis extracts the top 10 principal components, which collectively account for over 97% of the variance. The reduced 10-dimensional data is then fed into the DBSCAN algorithm, which automatically clusters the data using a radius threshold of 1.5 and a density threshold of 5. The results reveal three behavioral categories: "Normal Action A," "Normal Action B," and "Abnormal Action C." The relationships between these categories and their spatial distribution are then mapped onto a visualization to create a behavioral cluster map, which supports subsequent differential analysis.

[0082] Deviation modeling is performed between the clustering results and the behavior clustering map to calculate the behavior offset. The boundary conditions in the clustering results are dynamically adjusted according to the behavior offset, and the behavior category division range is corrected to obtain the corrected clustering results and boundary residual information.

[0083] By designing a deviation modeling mechanism, we analyze boundary samples within behavioral categories and calculate behavioral offsets at category boundaries. Specifically, we perform statistical analysis on samples within each category in the clustering results, calculating the distribution of abnormal feature values ​​(such as feature mean and feature variance) at category boundary points. We also screen out boundary samples in category intersection areas and dynamically adjust category boundary conditions by calculating the degree of feature offset (behavioral offset) between boundary samples and center samples. By adjusting category boundaries, we can reclassify overlapping samples into their corresponding categories, generating more accurate corrected clustering results. We also output boundary residual information to annotate the characteristics of the samples after boundary adjustment.

[0084] For example, suppose the original clustering results are divided into behavioral categories "normal action A" and "abnormal action C." Some samples are found in the boundary area close to both categories A and C. By calculating the sample feature offset in this area, for example, if a sample is 0.7 from the center of category A and 0.3 from the center of category C, it is determined to be closer to category A and dynamically classified as "normal action A," forming a revised clustering result. Boundary residual information is also generated, recording the adjusted sample feature values ​​and deviations to support subsequent rule generation and analysis.

[0085] Based on the modified clustering results, a set of error-proofing judgment rules is constructed. The rule content includes behavior pattern thresholds, timing conditions and action logic chains; the boundary residual information is mapped and matched with the error-proofing judgment rule set, and the rule conflict group and risk level factor are extracted based on the matching results.

[0086] By analyzing the corrected clustering results and boundary residual information, a standardized set of error-proofing judgment rules is constructed. Specifically, this set of error-proofing judgment rules is based on the boundary conditions of the sample category, the characteristic value range, and the behavioral temporal characteristics. For example, behavioral pattern threshold limits can be set, such as requiring that the execution time of an action must not exceed a specified range, or specifying a sequential logic chain (e.g., action A must be executed after action B). Boundary residual information is used to perform item-by-item mapping and matching with the rules to identify whether sample features violate the preset conditions of the rules, extract rule conflict groups (i.e., behavior items that violate multiple rules), and further define risk level factors (e.g., a weighted score of behavior frequency and conflict intensity).

[0087] For example, suppose the error prevention rule set includes: "(1) Step time ≤ 5 seconds; (2) Action A and Action B must be executed continuously; (3) Abnormal feature deviation ≤ 20%." The sample characteristics in the modified clustering results are matched with the rules one by one. It is found that a certain sample does not meet the requirements in both Rule 1 and Rule 3, forming a rule conflict group. The risk level factor of this sample is calculated based on the number of execution conflicts and is 80 (where 100 is the maximum risk). This sample is output to the high-risk behavior list for error prevention analysis and optimization suggestion generation.

[0088] A comprehensive analysis is conducted on the risk level factors and the historical frequency weights of the rule conflict groups, and corresponding risk labels and optimization suggestions are generated based on the analysis results.

[0089] By combining the historical records of rule conflict groups and sample risk level factors to comprehensively analyze behavioral characteristics, refined risk labels and optimization suggestions are generated. Specifically, the historical frequency weights of rule conflict groups in the database (such as the number of times a certain behavior item violates a certain rule) are weighed together with the sample risk level factors to generate risk labels for the behaviors (such as "high-risk operations" or "low-risk behaviors"), and the content of optimization suggestions is analyzed (such as increasing time buffers, adjusting the order of operations, or rectifying the range of abnormal characteristic values).

[0090] For example, for a sample containing the rule violation groups "Step time mismatch" and "Logical sequence error," the frequency of this behavior item in the historical records was found to be 50. When assessing the risk weight, combined with the risk level factor of 80, the risk label "High-risk operation" was generated. Furthermore, the recommended optimization content was to "adjust the operation time limit to ≤ 3 seconds and prohibit operations from skipping logical chain links" to prompt users to correct or adjust their operation methods.

[0091] In step S400, the risk labels are mapped to the behavior segments corresponding to the fused feature data to generate a converged optimization feature representation. The converged optimization feature representation is used to associate and annotate the optimization suggestions with behavioral semantics to obtain the optimization instruction set and semantic prompt information. Specifically, the steps include: The risk labels are mapped to the behavior segments corresponding to the fused feature data to extract the features of high-risk behavior areas. The convolutional coding network is used to perform context modeling and semantic normalization on the features of high-risk behavior areas to generate a set of candidate optimization parameters.

[0092] By leveraging the mapping relationship between risk labels and fused feature data, we locate characteristic regions of high-risk behavior segments and construct a contextual semantic model to understand the causal relationships between risky behaviors. Specifically, using risk labels as indexes, we compare them point-by-point with the segment sequences in the fused feature data using a graph structure matching method to extract the relevant temporal, spatial, and logical features of the risky behavior segments. A convolutional coding network (CNN) is used to model the identified high-risk behavior regions and extract contextual features of the segmented behaviors, including information about the time window before and after the behavior, associated logic, and device response status. Subsequently, a semantic normalization module is used to normalize the contextual features and generate a set of candidate optimization parameters for subsequent optimization.

[0093] For example, suppose the risk label "Operation Timeout" is detected for the action segment "Action A - B," with a device response time of 15 seconds (exceeding the 10-second threshold). CNN modeling reveals a voice command delay before "Action A," while there's no significant interaction between "Action B" and the device state. Contextual information is extracted, such as "Action A voice response time of 5 seconds" and "Action B device inactive." After normalization, candidate optimization parameters are generated, such as "Adjust voice response time to 3 seconds" and "Set minimum device activity time to 1 second," for further optimization analysis.

[0094] The candidate optimization parameter set is weightedly combined with the original fusion features to construct a corrected input tensor, and the corrected input tensor is subjected to multiple rounds of residual regression and error inversion to generate a converged optimization feature representation; among them, the original fusion features refer to the high-dimensional feature set after the fusion of user interaction data, sensor data, and historical behavior data, covering information such as behavior patterns, temporal relationships, and environmental context.

[0095] A modified input tensor is constructed by weightedly combining the normalized set of candidate optimization parameters with the original fused features. This modified input tensor is then iteratively optimized using residual regression and error inversion mechanisms to generate a converged optimized feature representation. Specifically, the original fused features include user interaction data (such as mouse clicks and voice commands), sensor data (such as temperature and pressure signals), and historical behavioral data (such as operation paths and time series). These data are formed into a feature set through time alignment and context embedding. The weighted combination process amplifies or reduces each feature dimension based on the weight of the optimization parameter set (which depends on the priority and historical frequency of the risk label), for example, by increasing the weight of risk feature data and decreasing the weight of redundant feature data. Residual regression techniques are used to dynamically adjust the weights from error inversion to ensure that the converged optimized feature representation meets the fit requirements of the behavioral model.

[0096] For example, consider the original fused features containing "Duration Data = [5, 7, 15]" and "Device Response Time = [3, 6, 12]," along with the adjusted values ​​of the optimized parameter set: "Duration reduction weight = 0.2" and "Device item increase weight = 0.3." After weighted combination, the modified input tensor is constructed as 6.4, 8.1, and 12.2. Error inversion further adjusts the device response time weight by a factor of 4 to obtain feature representations of 6.8, 9.0, and 14.8, ensuring that the optimized feature representations converge quickly and are used for model feedback.

[0097] The converged optimization feature representation is input into the semantic understanding module, associated with the optimization suggestions and labeled with behavioral semantic mapping, and the action goals, triggering conditions and risk avoidance paths in the behavioral semantic mapping are extracted to construct an optimization instruction set.

[0098] By leveraging the semantic understanding module's label reinforcement and behavioral pattern matching mechanisms, the module extracts semantic information from the convergence optimization feature representation, associates it with optimization suggestions, and generates actionable behavioral semantic mappings. Specifically, the semantic understanding module, based on a feature parser, breaks down the convergence optimization feature representation into action goals (e.g., behavior end state), trigger conditions (e.g., behavior start time), and risk avoidance paths (e.g., alternatives to high-risk nodes). By querying a pre-set semantic rule database, each goal, condition, and path is annotated and its logical chain reconstructed to form a set of optimization instructions, ensuring that user-actionable behavioral instructions are clearly expressed in real-world scenarios.

[0099] For example, if the convergence optimization feature representation includes the goal of "device status recovery", the condition of "operation interval adjustment to 5 seconds", and the path of "skipping abnormal devices", combined with the semantic rule database, an optimization instruction set is generated, such as "please ensure that the device recovers normally before executing the next step", "adjust the operation time interval to more than 5 seconds", and "bypass the device abnormal status node to complete the task", and these instruction sets are provided for user interface operations.

[0100] The task target items in the optimization instruction set are converted into prompt statement templates, and the prompt statement templates are combined with the user behavior context to generate semantic prompt information; among them, the user behavior context refers to the timestamp, operation sequence, previous state and environmental conditions of the user's current operation, which serves as the contextual basis for model reasoning.

[0101] By integrating user behavior context into a templated set of optimization instructions, the task objectives are transformed into concise and readable semantic prompts. Specifically, the user behavior context extraction module collects operation timestamps, operation sequences, previous states, and current environmental data, and embeds this information into the optimization instruction template (e.g., "Action goal: Complete current task" and "Current state: Device idle"), thereby generating semantic prompts with contextual explanations. The prompts not only provide users with the exact content of the optimization instructions but also intuitively present the optimization action path through semantic expression.

[0102] For example, in a certain process, the user operation timestamp is "12:00:15," the device status is "Idle," and the previous status is "Data upload completed." Combined with the optimization instruction set, prompts are generated, such as "Please pause for 5 seconds after 12:00:15 to complete the upload" and "The current device idle state allows continued operation; it is recommended to start the next process." These semantic prompts not only guide user operations but also provide detailed explanations of optimization suggestions through context.

[0103] In step S500, the steps of optimizing the preset deep learning behavior prediction model using the optimization instruction set and semantic prompt information to obtain the optimized foolproof intelligent prediction model include: Semantic hint information is used to enhance the labels of the intermediate-layer behavior representation of the preset deep learning behavior prediction model, and the learning objectives are semantically aligned to obtain the enhanced deep learning behavior prediction model.

[0104] By inputting semantic cue information into the middle layer of a pre-set deep learning behavior prediction model, the labels of the behavior representations are strengthened, and semantic alignment is performed during the training process. Specifically, the middle layer feature representations in the deep learning behavior prediction model are first extracted. These middle layer features contain the temporal characteristics and semantic patterns of the operational behavior. Then, the labels of these feature representations are strengthened using semantic cue information. By increasing the labeling of high-risk and low-confidence behaviors, the model can pay more attention to and correct these behaviors. In this way, the model not only optimizes standard behaviors but also increases its sensitivity to abnormal or borderline behaviors, thereby enhancing the model's predictive capabilities.

[0105] During this process, the model's learning objectives are also semantically aligned, that is, the goals, conditions, and semantic elements in the optimization instructions are matched with the model's training objectives to ensure that the model's output behavior sequence can be better aligned with the semantic hierarchy, thereby improving the model's understanding and execution of task semantics.

[0106] For example, if the semantic hint information in the optimized instruction set includes "Please perform operation B immediately after operation A is completed", then this goal is converted into the training goal of the model through semantic alignment, so that the model pays more attention to the temporal relationship between operation A and operation B during the prediction process, thereby improving the accuracy and timeliness of task completion.

[0107] The optimization instruction set is modularly decomposed, and the corresponding parameter tuning path is constructed based on the decomposition results. The parameter tuning path is used to optimize the training hyperparameters, activation strategy and inter-layer connection method of the enhanced deep learning behavior prediction model to obtain the optimized fool-proof intelligent prediction model.

[0108] By modularizing the optimization instruction set, complex task objectives are converted into small, easily optimized modules. Parameter tuning paths are then constructed based on the functional requirements of each module. Specifically, each task objective in the optimization instruction set (such as "operation sequence adjustment" and "time delay requirement") is first broken down, and core parameters (such as time threshold, action sequence, and delay) are extracted. An optimization path is then constructed for each parameter. This optimization path is generated based on the adjustments required during model training, including hyperparameters (such as learning rate and regularization factor), activation function selection (such as ReLU and Sigmoid), and inter-layer connectivity (such as fully connected layers and convolutional layers). This results in an optimized, fool-proof intelligent prediction model.

[0109] Through these tuning paths, the model training process can be dynamically adjusted to ensure that each module can effectively improve the model's predictive ability for error-proofing tasks, while ensuring that the model's structure can cope with changing operating environments.

[0110] For example, suppose a task in the optimized instruction set is "Operation B must be completed within 2 seconds of Operation A." This goal is broken down into a time delay requirement ("complete within 2 seconds of Operation A") and a task sequence requirement ("Operation B follows Operation A"). Based on this, the model's tuning path will focus on adjusting the model's time series processing capabilities, including increasing the weight of the attention mechanism on the timing of operations, while optimizing the learning rate and activation function to ensure that the model can accurately predict the relationship between Operation A and Operation B.

[0111] In this embodiment, a fully designed AI fool-proofing algorithm and digital model optimization method achieve a closed-loop process from multimodal data acquisition to deep model optimization, addressing the technical challenges of behavior prediction and exception handling in complex operating environments. Specifically, multi-channel sensing devices are first used to synchronously collect multimodal process data such as user behavior, device response, and environmental attributes. The data is then time-aligned and event-segmented to extract standard and abnormal behavior sequences. A graph convolutional network and contrastive learning algorithm are then combined to further analyze the spatial correlation features and temporal relationships in the behavior sequences. An anomaly score matrix and behavior credibility vector are calculated, and a standard behavior sequence is reconstructed based on the confidence level. Abnormal behavior sequences are also constructed to locate high-risk operations. In the behavior prediction stage, a deep learning model is used to embed the standard behavior sequence, and a hierarchical attention mechanism is used to extract key behavior nodes. These are then structurally matched with the abnormal behavior sequence to generate a structural difference vector. Subsequently, a feature compensation mechanism and an error inversion map are combined to generate a fused feature tensor for the behavior error. The fused feature data is then constructed to provide data support for high-dimensional clustering and bias correction. In the high-dimensional clustering phase, principal component analysis (PCA) dimensionality reduction and density-adaptive clustering algorithms are used to categorize behavior. Deviation modeling is used to dynamically adjust cluster boundaries to generate corrected clustering results and boundary residual information. Based on this information, a set of error-proofing judgment rules is constructed, mapping rule conflicts to risk level factors, and generating targeted risk labels and optimization recommendations. In the optimization instruction generation phase, convergent optimization feature representation is used to perform contextual modeling and semantic mapping of high-risk areas, forming an optimized instruction set and semantic prompt information. Through label enhancement and parameter tuning path optimization of the deep learning model, predictive capabilities and exception handling response capabilities are enhanced, and the optimized error-proofing intelligent prediction model is output.

[0112] This embodiment scheme has achieved remarkable results in the accuracy, timeliness and model adaptability of standard behavior prediction and abnormal behavior processing, further ensuring the security, stability and efficiency of complex scenario operation processes.

[0113] Example 3: like Figure 2 As shown, the present application also provides an AI foolproof algorithm and digital model optimization system 10, including the following modules: The acquisition module 11 is used to acquire multimodal process data, perform time alignment and event segmentation on the multimodal process data, and obtain standard behavior sequences and abnormal behavior sequences.

[0114] The acquisition module 11 uses a multi-channel sensing device and a unified timestamp to time-align and segment the data, analyze user behavior, device status and environmental parameters, generate standard behavior sequences and abnormal behavior sequences, and provide basic data support for subsequent analysis.

[0115] The matching module 12 is used to input the standard behavior sequence into a preset deep learning behavior prediction model to obtain the behavior prediction result, perform structural matching on the behavior prediction result and the abnormal behavior sequence to obtain fused feature data.

[0116] The matching module 12 inputs the standard behavior sequence into the deep learning behavior prediction model, combines the behavior prediction results with the abnormal behavior sequence for structural matching, further extracts the behavior pattern deviation, and generates fusion feature data.

[0117] The correction module 13 is used to perform high-dimensional clustering and deviation correction on the fused feature data to obtain the corrected clustering results and boundary residual information, generate a set of error-proofing judgment rules based on the corrected clustering results, and use the set of error-proofing judgment rules to map and match the boundary residual information to obtain risk labels and optimization suggestions.

[0118] Correction Module 13 uses principal component analysis to reduce the dimensionality of high-dimensional fused feature data, retaining key feature subspaces. It then uses a density-adaptive clustering algorithm to perform high-dimensional cluster analysis on behavioral categories. It also calculates deviations based on boundary residual information and dynamically adjusts clustering ranges to achieve precise classification of behavioral categories and generate corrected clustering results. Correction Module 13 also generates a set of foolproof judgment rules, extracts rule conflicts through a mapping and matching mechanism, and outputs risk labels and optimization recommendations.

[0119] The mapping module 14 is used to map the risk label to the behavior segment corresponding to the fusion feature data to generate a convergent optimization feature representation, and use the convergent optimization feature representation to associate the optimization suggestions with the behavior semantic mapping to obtain the optimization instruction set and semantic prompt information.

[0120] The mapping module 14 further maps the risk labels with the high-risk segments of the fused feature data to construct a converged optimization feature representation, and realizes the semantic normalization and context association of the optimization suggestions through convolutional network modeling, thereby generating an optimization instruction set and semantic prompt information.

[0121] The optimization module 15 is used to optimize the preset deep learning behavior prediction model using the optimization instruction set and semantic prompt information to obtain an optimized fool-proof intelligent prediction model.

[0122] On this basis, the optimization module 15 comprehensively adjusts the training objectives, training parameters and inter-layer connection strategies of the deep learning behavior prediction model. The optimized fool-proof intelligent prediction model has higher prediction accuracy and adaptability to abnormal behaviors.

[0123] In this embodiment, the systematic design consisting of the acquisition module 11, the matching module 12, the correction module 13, the mapping module 14 and the optimization module 15 realizes the full-chain error-proofing algorithm and digital model optimization from multimodal process data acquisition to deep learning model optimization, significantly improving the system's recognition accuracy, risk avoidance ability and error-proofing effect for complex behavior patterns in actual operation, and effectively supporting intelligent operation management under changing scenarios.

[0124] It should be noted that, those skilled in the art will clearly understand that, for the sake of convenience and brevity of description, the specific working processes of the above-described system and each module can refer to the corresponding processes in the aforementioned embodiment 1 and will not be repeated here.

[0125] The structures, proportions, sizes, etc. depicted in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not intended to limit the conditions under which the present invention can be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportional relationships, or adjustments in size should still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and objectives that can be achieved by the present invention.

[0126] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI foolproof algorithm and digital model optimization method, characterized in that: include: Collecting multimodal process data, performing time alignment and event segmentation on the multimodal process data, and obtaining standard behavior sequences and abnormal behavior sequences; Inputting the standard behavior sequence into a preset deep learning behavior prediction model to obtain a behavior prediction result, performing structural matching on the behavior prediction result and the abnormal behavior sequence to obtain fused feature data; Performing high-dimensional clustering and deviation correction on the fused feature data to obtain a corrected clustering result and boundary residual information, generating a fool-proofing judgment rule set based on the corrected clustering result, and using the fool-proofing judgment rule set to perform mapping and matching on the boundary residual information to obtain risk labels and optimization suggestions; Mapping the risk labels to the behavior segments corresponding to the fused feature data to generate a converged optimization feature representation, and using the converged optimization feature representation to associate and annotate the optimization suggestions with behavioral semantics to obtain an optimization instruction set and semantic prompt information; The preset deep learning behavior prediction model is optimized using the optimization instruction set and the semantic prompt information to obtain an optimized fool-proof intelligent prediction model.

2. The AI ​​foolproof algorithm and digital model optimization method according to claim 1, characterized in that: The step of collecting multimodal process data, performing time alignment and event segmentation on the multimodal process data, and obtaining standard behavior sequences and abnormal behavior sequences includes: The multi-channel sensing device synchronously collects the action signals, environmental parameters and user interaction logs during the operation to construct multimodal process data, and performs time alignment and event segmentation on the multimodal process data to obtain event segment sequences and context window information; Input the event fragment sequence into a preset graph convolutional network model to extract the spatial correlation features and temporal relationship features of the operation action, perform abnormal pattern discrimination on the context window information, and obtain an abnormality score matrix and behavior credibility vector; fusing the behavior credibility vector with the temporal relationship feature to generate a standard behavior sequence; The anomaly score matrix and the spatial correlation features are jointly analyzed, and an abnormal behavior sequence is constructed based on the analysis results.

3. The AI ​​foolproof algorithm and digital model optimization method according to claim 1, characterized in that: The step of inputting the standard behavior sequence into a preset deep learning behavior prediction model to obtain a behavior prediction result, and performing structural matching on the behavior prediction result and the abnormal behavior sequence to obtain fused feature data includes: Performing sequence embedding encoding on the standard behavior sequence to obtain an encoded standard behavior sequence, inputting the encoded standard behavior sequence into a deep learning behavior prediction model based on a Transformer structure, and outputting a multi-dimensional behavior prediction result; performing hierarchical attention analysis on the behavior prediction results to extract key behavior nodes, constructing a structural difference vector using the key behavior nodes and the abnormal behavior sequence, and extracting trigger offset features based on the structural difference vector; Performing residual feedback on the structural difference vector and the intermediate features of the deep learning behavior prediction model to obtain a feature compensation vector and an error reverse map; A behavior error fusion feature tensor is constructed according to the trigger offset feature and the feature compensation vector, and the behavior error fusion feature tensor is fused with the error inverse map at a tensor level to obtain fusion feature data.

4. The AI ​​fool-proofing algorithm and digital model optimization method according to claim 3, characterized in that: The step of performing hierarchical attention analysis on the behavior prediction results to extract key behavior nodes, and constructing a structural difference vector using the key behavior nodes and the abnormal behavior sequence includes: The output of each layer of the behavior prediction results is weighted through a hierarchical attention mechanism, key behavior nodes are extracted based on the weighted results, and the similarity between the key behavior nodes is calculated using a dynamic time warping algorithm; A deviation portion in the abnormal behavior sequence is identified, and a structural difference vector is constructed using the similarity and the deviation portion.

5. The AI ​​fool-proofing algorithm and digital model optimization method according to claim 1, characterized in that: The steps of performing high-dimensional clustering and deviation correction on the fused feature data to obtain a corrected clustering result and boundary residual information, generating a foolproofing judgment rule set based on the corrected clustering result, and mapping and matching the boundary residual information using the foolproofing judgment rule set to obtain a risk label and optimization suggestion include: Performing dimensionality reduction preprocessing on the fused feature data according to principal component analysis to obtain dimensionality reduction features, performing high-dimensional clustering on the dimensionality reduction features to obtain clustering results and a behavior clustering map; Deviation modeling is performed between the clustering result and the behavior clustering map to calculate the behavior offset, and boundary conditions in the clustering result are dynamically adjusted according to the behavior offset to correct the behavior category division range to obtain a corrected clustering result and boundary residual information; Constructing a foolproof determination rule set based on the modified clustering result, mapping and matching the boundary residual information with the foolproof determination rule set, and extracting a rule conflict group and a risk level factor based on the matching result; The risk level factor and the historical frequency weight of the rule conflict group are comprehensively analyzed, and corresponding risk labels and optimization suggestions are generated based on the analysis results.

6. The AI ​​foolproof algorithm and digital model optimization method according to claim 1, characterized in that: The step of mapping the risk label to the behavior segment corresponding to the fused feature data to generate a converged optimization feature representation, and using the converged optimization feature representation to associate and annotate the optimization suggestion with behavior semantic mapping to obtain an optimization instruction set and semantic prompt information includes: Mapping the risk labels to the behavior segments corresponding to the fused feature data to extract high-risk behavior area features, performing context modeling and semantic normalization on the high-risk behavior area features using a convolutional coding network, and generating a set of candidate optimization parameters; Performing a weighted combination of the candidate optimization parameter set and the original fused features to construct a modified input tensor, and performing multiple rounds of residual regression and error inversion on the modified input tensor to generate a converged optimized feature representation; wherein the original fused features refer to a high-dimensional feature set obtained by fusing user interaction data, sensor data, and historical behavior data; Associating and annotating the convergence optimization feature representation with the optimization suggestion and mapping the behavior semantics to construct an optimization instruction set; The task target items in the optimization instruction set are converted into prompt statement templates, and the prompt statement templates are combined with the user behavior context to generate semantic prompt information; wherein, the user behavior context refers to the timestamp, operation sequence, previous state and environmental conditions of the user's current operation.

7. The AI ​​foolproof algorithm and digital model optimization method according to claim 1, characterized in that: The step of optimizing the preset deep learning behavior prediction model using the optimization instruction set and the semantic prompt information to obtain an optimized fool-proof intelligent prediction model includes: Using the semantic hint information, label enhancement is performed on the intermediate layer behavior representation of the preset deep learning behavior prediction model to obtain an enhanced deep learning behavior prediction model; The optimization instruction set is modularly decomposed, and a corresponding parameter tuning path is constructed based on the decomposition results. The parameter tuning path is used to optimize the training hyperparameters, activation strategy and inter-layer connection method of the enhanced deep learning behavior prediction model to obtain an optimized fool-proof intelligent prediction model.

8. An AI foolproof algorithm and digital model optimization system, characterized in that: include: An acquisition module is used to acquire multimodal process data, perform time alignment and event segmentation on the multimodal process data, and obtain standard behavior sequences and abnormal behavior sequences; A matching module is used to input the standard behavior sequence into a preset deep learning behavior prediction model to obtain a behavior prediction result, and perform structural matching on the behavior prediction result and the abnormal behavior sequence to obtain fused feature data; A correction module is used to perform high-dimensional clustering and deviation correction on the fused feature data to obtain a corrected clustering result and boundary residual information, generate a fool-proofing judgment rule set based on the corrected clustering result, and use the fool-proofing judgment rule set to map and match the boundary residual information to obtain risk labels and optimization suggestions; A mapping module is used to map the risk label to the behavior segment corresponding to the fused feature data to generate a converged optimization feature representation, and use the converged optimization feature representation to associate and annotate the optimization suggestions with behavioral semantics to obtain an optimization instruction set and semantic prompt information; An optimization module is used to optimize the preset deep learning behavior prediction model using the optimization instruction set and the semantic prompt information to obtain an optimized fool-proof intelligent prediction model.

Images