Aerial work platform intelligent supervision method and system based on data analysis

Abstract

The invention discloses an aerial work platform intelligent supervision method and system based on data analysis. The method comprises the following steps: S1, collecting operation state data of an aerial work platform and behavior video stream data of an operator; s2, inputting a dynamic time warping network to carry out cross-modal time alignment, and outputting a key state feature sequence; s3, inputting the behavior video stream into a double-path Transform model, and generating an operation behavior representation sequence; s4, jointly modeling key state features and behavior characterization, and constructing a risk scoring function; s5, the risk scoring result is compared with a grade threshold value, an early warning mechanism is triggered, and a control instruction is generated; s6, executing platform control actions, such as speed limiting, descent, locking, pause and recording feedback; and S7, updating the model by taking the key state characteristics, the behavior characterization, the scoring result and the feedback as samples. According to the invention, intelligent identification and early warning of high-altitude operation states and behaviors are realized, and operation safety and supervision efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial intelligent supervision, and in particular to a method and system for intelligent supervision of aerial work platforms based on data analysis. Background Art

[0002] In the practical application of aerial work platforms, platform operational safety and operator behavior compliance remain key factors influencing operational safety. Traditional monitoring methods primarily rely on manual inspections, timed recording, and limited video surveillance, which suffer from issues such as poor real-time performance, insufficient coverage, and significant errors in subjective judgment. To improve monitoring effectiveness, sensors have been gradually introduced in recent years to collect platform operational data, and cameras have been deployed to record operator behavior for later analysis. However, most current systems only analyze single-modal data and lack the ability to deeply integrate multi-source information, making it difficult to fully reflect the risk level of the operating status.

[0003] On a technical level, some methods have attempted to introduce deep learning models for state prediction or anomaly detection. However, these models often only process a single data type, such as platform status or image behavior, and ignore cross-modal correlation issues such as temporal misalignment and semantic inconsistency between platform operational characteristics and human behavior. Especially in real-world scenarios, due to differences in sampling frequency, data latency, and operational habits, data from different modalities often exhibit significant temporal distortions, making it difficult to ensure accurate risk assessments with simple synchronization. Furthermore, existing models typically employ static fusion strategies that are unable to dynamically adjust the importance of each modal feature based on changes in the operating environment, resulting in a lack of contextual sensitivity in the fusion results.

[0004] Furthermore, there is currently a lack of a complete system solution that can combine aerial work platform status data with worker behavior videos for cross-modal alignment, joint modeling, and risk scoring. Existing technologies, particularly in the closed-loop execution of feature alignment, risk identification, and control responses, suffer from simple modeling approaches, fragmented response mechanisms, and limited learning capabilities, making them incapable of meeting the requirements for operational safety supervision in complex environments.

[0005] Therefore, how to provide an intelligent monitoring method and system for aerial work platforms based on data analysis is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of the present invention is to propose a data-analysis-based intelligent monitoring method and system for aerial work platforms. This method fully integrates a dynamic time warping network with a dual-path Transformer architecture. By constructing a cross-modal alignment and joint risk modeling mechanism, it achieves deep collaborative perception of platform operating status data and operator behavior video streams. The method also describes in detail the intelligent processing flow for risk score generation, early warning triggering, and control linkage. This method boasts high accuracy in identifying operational risks, rapid response, and strong adaptive learning capabilities, effectively improving safety monitoring and platform operation intelligence in aerial work scenarios.

[0007] According to an embodiment of the present invention, a method for intelligently monitoring an aerial work platform based on data analysis includes the following steps:

[0008] S1. Collect the operation status data of the aerial work platform and the video stream data of the operator's behavior, perform timestamp marking and cache storage;

[0009] S2. Input the operation platform operation status data into the dynamic time warping network, perform cross-modal time alignment, identify the time warping relationship between different data sources through the dynamic warping mechanism, and output the aligned key state feature sequence;

[0010] S3. Construct a dual-path Transformer structure model, input the operator behavior video stream data into the dual-path Transformer structure model, and generate a work behavior representation sequence;

[0011] S4. Jointly model the key state feature sequence and the operation behavior representation sequence, construct a joint risk scoring function, calculate the risk score of the current operation state based on the coupling analysis of multi-dimensional features, and output the risk scoring result;

[0012] S5. Compare the risk score result with the risk level threshold. If the risk score exceeds the risk level threshold, trigger the early warning mechanism and output an audible and visual alarm signal, remote notification information or control command according to the risk level;

[0013] S6. Linking the platform execution controller in the aerial work platform according to the control command to perform automatic speed limit, platform return, lock control or work pause operation, and storing the operation feedback together with the risk scoring result;

[0014] S7. Input the key state feature sequence, operation behavior representation sequence, risk score results and operation feedback as training samples into the dynamic time warping network and dual-path Transformer structure model to perform parameter update and structure adjustment.

[0015] Optionally, the S2 specifically includes:

[0016] S21. Obtain the operation status data of the operation platform and construct a multi-channel time series input. Let the status data of the cth channel be Where c=1,2,…,C, all channel data are combined to form the input sequence set x c (t i ) represents the time point t i At , the state value collected by the cth channel, N is the number of time steps on a single channel, and C is the total number of operating state data channels that can be collected in the operation platform;

[0017] S22, the status data of each channel data X c Input modal projection encoder f c (), map different modal features to a unified aligned embedding space to obtain the state embedding sequence Z c =f c (X c ), merge all channel embeddings into the initial state embedding sequence

[0018] S23. Construct the operation platform operation status sensing channel graph structure G = (V, E), where V is the node set of the graph, E is the edge set of the graph, and each node v i ∈V represents a state channel, edge e ij ∈E represents the physical connection relationship or control coupling relationship between modules;

[0019] S24. Embed the initial state into sequence Z s Input graph neural network module, integrate graph structure adjacency information, and generate graph enhanced state embedding feature H s ;

[0020] S25, based on the graph enhanced state embedding feature H s Construct a derivable path attention matrix A, where A ij for:

[0021]

[0022] Among them, τ is the temperature adjustment coefficient, A ij is the attention coefficient of the i-th row and j-th column of the derivable path attention matrix, H s,i is the feature vector of the i-th time step in the state embedding sequence after graph enhancement, H s,j is the feature vector of the jth time step in the state embedding sequence after graph enhancement, H s,k is the feature vector of the kth time step in the state embedding sequence after graph enhancement, exp is the exponential function, and T is the length of the time series;

[0023] S26. Based on the derivable path attention matrix A, calculate the weighted feature representation of each time step and fuse the original features, build a residual gating mechanism, and calculate the final aligned feature sequence for:

[0024]

[0025] Among them, α t is the gate weight of the t-th time step, which is dynamically generated by the position information, A tj is the attention coefficient of the t-th row and j-th column of the derivable path attention matrix, Embed the state path vector for the jth time step;

[0026] S27, based on the final alignment feature sequence Z final Embedded with the initial state sequence Z s Based on the feature distribution deviation between them, an alignment consistency driven feedback mechanism is constructed to generate an alignment consistency score vector C, which dynamically adjusts the regional weight distribution of the path attention matrix in the inference stage.

[0027] S28. Through the back-propagation algorithm, the structural parameters of the modal projection encoder, graph neural network, path attention mechanism and residual fusion channel are jointly optimized to achieve the precise alignment of multi-channel operation status data by the dynamic time warping network and output the key state feature sequence Z final .

[0028] Optionally, the S3 specifically includes:

[0029] S31, collect the operator behavior video stream data, decompose it into continuous frame-level image sequence X v ={x1,x2,…,x T}, where each frame image x t ∈R H×W×C , perform unified time regularization to form a time series length T that is consistent with the status data of the operation platform, where R is a real number set, H is the image height, W is the image width, and C is the number of image channels;

[0030] S32, perform image block segmentation and linear projection on the image sequence, and map each frame into a visual embedding vector z t , combined to obtain the visual embedding sequence Z v , and introduce position encoding P to obtain input representation E;

[0031] S33. Construct a dual-path Transformer structure model, which includes a behavior evolution perception path and a global supervision fusion path. The behavior evolution perception path uses a local window attention mechanism and a temporal factor encoder to model behavioral micro-changes, while the global supervision fusion path captures global trends through a cross-temporal fully connected attention mechanism.

[0032] S34, the behavior evolution perception path introduces a frame-level change detection module to calculate the feature difference vector Δz between each frame image and the previous frame t , used for time evolution guidance, and performs a time-weighted self-attention mechanism on each window to output a local representation sequence H evo ;

[0033] S35, the global supervision fusion path uses a multi-head attention module to model the full sequence execution behavior context and introduces a cross-modal reference guidance weight G t The cross-modal reference guided weight is modeled by associating the visual embedding features of each frame image with the aligned state features of the corresponding time step, and using the weight coefficient reflecting the correlation between the two modalities generated by the learnable mapping function to generate an enhanced global representation H by aligning the attention with the state features. global ;

[0034] S36, fuse the behavior evolution perception path and the global supervision fusion path output, and calculate the fusion representation H through the residual attention gating mechanism fuse (t):

[0035] H fuse (t) = γ t ·H evo (t)+(1-γ t )·H global (t);

[0036] Among them, γ t =σ(W r ·[H evo (t),H global (t)]) is the gating coefficient, σ is the activation function, W r is a learnable parameter;

[0037] S37, introduce the behavior dynamic fuzzy enhancement module to fusion representation H fuse Perform fuzzy perception filtering, and obtain the enhanced sequence H by structural noise suppression and background mask filtering. enhanced ;

[0038] S38, the enhanced sequence H enhanced Input a two-layer feedforward neural network to extract deep abstract semantic features and generate a task behavior representation sequence R v ;

[0039] S39. Execute the feature orthogonality constraint strategy on the job behavior representation sequence, calculate the cosine similarity matrix for each time step feature, and perform diagonal enhancement operation to suppress feature redundancy.

[0040] Optionally, the S4 specifically includes:

[0041] S41. Obtain the key state feature sequence Z output by the dynamic time warping network final ={z1,z2,…,z T}, where z t represents the platform operation status characteristics after alignment, and T is the length of the time series;

[0042] S42. Obtain the task behavior representation sequence R output by the dual-path Transformer structure v ={r1,r2,…,r T}, where r t represents the behavioral characteristics of the operator at the tth time step;

[0043] S43, key state feature sequence Z final and task behavior representation sequence R v Perform feature concatenation according to the time step and construct a joint feature sequence U={u1,u2,…,u T};

[0044] S44, input the joint feature sequence U into the joint risk scoring sub-network, use the multi-layer feedforward neural network to extract the coupling feature representation, and generate the risk representation sequence S = {s1, s2, ..., s T}, where s t represents the candidate risk score value at the t-th time step;

[0045] S45. Introduce a modal perception-guided aggregation mechanism to interactively map behavioral features with joint state features, guide the weighting of risk scores at each time step, and generate a fusion score value.

[0046]

[0047] Among them, ρ t is the modal guided gating coefficient, which is determined by the behavioral characteristics r t and platform operating status characteristics t The matching relationship is generated; φ(r t ,u t ) is a nonlinear mapping function used to generate risk compensation items;

[0048] S46, fusion score value Compare with the set risk level threshold value set Θ = {θ1, θ2, θ3}, which corresponds to θ1 mild, θ2 moderate, and θ3 severe risk levels respectively;

[0049] S47. Determine a risk level label L∈{0,1,2,3} based on the comparison result, where 0 represents no risk, 1 represents mild risk, 2 represents moderate risk, and 3 represents severe risk;

[0050] S48: Output the fusion score and the corresponding risk level label L, which serves as the input for triggering the early warning mechanism, generating control instructions, and executing the control logic.

[0051] Optionally, the S5 specifically includes:

[0052] S51. Receive the output risk score result and the corresponding risk level label as input information for risk assessment of the current operation status;

[0053] S52. Set a multi-level risk level threshold standard, compare and analyze the input risk score results, and determine whether they exceed the set mild, moderate or severe risk level thresholds;

[0054] S53. When the risk score exceeds any level threshold, the early warning mechanism is automatically activated, and the risk level trigger level is determined based on the comparison results;

[0055] S54. Select a corresponding warning output method based on the triggered risk level, including: activation of an audible and visual alarm device, remote notification from the platform control center, or a pop-up reminder on a mobile terminal;

[0056] S55. If the risk level is moderate or severe, then while outputting the notification, a corresponding control instruction is generated for the platform controller to execute a linkage response operation;

[0057] S56. Record the current risk score value, risk level label and the executed early warning response type together as a control log and risk feedback data archive.

[0058] According to an embodiment of the present invention, an intelligent monitoring system for aerial work platforms based on data analysis includes the following modules:

[0059] The data acquisition module is used to collect platform operation status data and operator behavior video stream data, and perform timestamp marking and cache storage;

[0060] The time warping module is used to perform cross-modal time alignment on the platform operation status data and output the aligned key state feature sequence;

[0061] The behavior extraction module is used to input the operator behavior video stream data into the dual-path Transformer structure model to extract the work behavior representation sequence;

[0062] Joint assessment module, used to fuse key state feature sequences and operational behavior representation sequences, construct joint risk functions, and output risk scoring results;

[0063] The risk comparison module is used to compare the risk score results with the risk level threshold to determine the risk level;

[0064] Early warning output module, used to output sound and light alarms, remote notifications or terminal reminders according to risk levels;

[0065] The control execution module is used to link the platform controller to execute speed limit, return, lock or pause operations according to the control instructions;

[0066] The continuous optimization module is used to use key state feature sequences, operation behavior representation sequences, risk scoring results and operation feedback as training samples to perform joint parameter updates and structural adjustments.

[0067] The beneficial effects of the present invention are:

[0068] The present invention solves the problems of cross-modal data time misalignment and insufficient feature fusion capabilities in the existing aerial work platform supervision process by introducing a dynamic time warping network and a dual-path Transformer structure. Compared with traditional monitoring systems that rely on a single modality or static rules, the present invention can uniformly time-align and semantically model the operating status data of the work platform and the behavioral video stream of the operator, thereby achieving accurate identification and dynamic assessment of potential risks in the aerial work process. Especially in the case of asynchronous acquisition and inconsistent sampling frequency of multi-source data, the dynamic time warping mechanism can effectively correct the time distortion relationship, extract synchronous key feature sequences, and improve the accuracy of data fusion.

[0069] At the same time, the dual-path Transformer structure possesses powerful feature modeling and behavior recognition capabilities. It can extract multi-dimensional, context-sensitive operational behavior features from visual information and deeply couple them with platform status features to form a more comprehensive and accurate risk scoring result. By setting multi-level risk level thresholds and linking the control system, this invention not only achieves early warning of risks but also automatically triggers response actions such as speed limiting, reversing, and locking, forming a data-driven closed-loop control mechanism, greatly improving the automation and intelligence level of the system.

[0070] The technical solution of the present invention has significant advantages such as high risk identification accuracy, fast system response efficiency, and sustainable model updating. It can be widely applied to real-time intelligent supervision in complex and high-risk operation scenarios, significantly improving the operation safety level and management efficiency, and has good engineering practical value and industrial promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0072] Figure 1This is a flow chart of an intelligent monitoring method for aerial work platforms based on data analysis proposed by the present invention;

[0073] Figure 2 This is a structural diagram of an intelligent monitoring system for aerial work platforms based on data analysis proposed by the present invention. DETAILED DESCRIPTION

[0074] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0075] refer to Figure 1 , an intelligent monitoring method for aerial work platforms based on data analysis, comprising the following steps:

[0076] S1. Collect the operation status data of the aerial work platform and the video stream data of the operator's behavior, perform timestamp marking and cache storage;

[0077] S2. Input the operation platform operation status data into the dynamic time warping network, perform cross-modal time alignment, identify the time warping relationship between different data sources through the dynamic warping mechanism, and output the aligned key state feature sequence;

[0078] S3. Construct a dual-path Transformer structure model, input the operator behavior video stream data into the dual-path Transformer structure model, and generate a work behavior representation sequence;

[0079] S4. Jointly model the key state feature sequence and the operation behavior representation sequence, construct a joint risk scoring function, calculate the risk score of the current operation state based on the coupling analysis of multi-dimensional features, and output the risk scoring result;

[0080] S5. Compare the risk score result with the risk level threshold. If the risk score exceeds the risk level threshold, trigger the early warning mechanism and output an audible and visual alarm signal, remote notification information or control command according to the risk level;

[0081] S6. Linking the platform execution controller in the aerial work platform according to the control command to perform automatic speed limit, platform return, lock control or work pause operation, and storing the operation feedback together with the risk scoring result;

[0082] S7. Input the key state feature sequence, operation behavior representation sequence, risk score results and operation feedback as training samples into the dynamic time warping network and dual-path Transformer structure model to perform parameter update and structure adjustment.

[0083] The present invention achieves in-depth perception and risk assessment of the operating status and personnel behavior of aerial work platforms by constructing an intelligent supervision method that integrates a dynamic time warping network and a dual-path Transformer structure, significantly improving the safety management level of the work scene. This method can effectively perform cross-modal time alignment of platform operating status data and behavioral video data, solving problems such as inconsistent sampling frequencies and information mismatches of different data sources, and extracting synchronized and critical state features. By extracting the multi-dimensional behavioral representation of the operating personnel through the dual-path Transformer model and jointly modeling it with the platform status features, a risk scoring mechanism is constructed to achieve accurate judgment of operational risks. Furthermore, the system can automatically trigger early warning and control responses based on the risk level, and the platform can execute operations such as speed limit, return or lock to form a response closed loop. At the same time, the model is continuously optimized and updated in combination with feedback data, and has good adaptive and learning capabilities, ensuring the stability and accuracy of the system in a dynamic environment, with high security, high reliability and wide application and promotion value.

[0084] In this embodiment, S2 specifically includes:

[0085] S21. Obtain the operation status data of the operation platform and construct a multi-channel time series input. Let the status data of the cth channel be Where c=1,2,…,C, all channel data are combined to form the input sequence set x c (t i ) represents the time point t i At , the state value collected by the cth channel, N is the number of time steps on a single channel, and C is the total number of operating state data channels that can be collected in the operation platform;

[0086] S22, the status data of each channel data X c Input modal projection encoder f c (), map different modal features to a unified aligned embedding space to obtain the state embedding sequence Z c =f c (X c ), merge all channel embeddings into the initial state embedding sequence

[0087] S23. Construct the operation platform operation status sensing channel graph structure G = (V, E), where V is the node set of the graph, E is the edge set of the graph, and each node v i ∈V represents a state channel, edge e ij ∈E represents the physical connection relationship or control coupling relationship between modules;

[0088] S24. Embed the initial state into sequence Zs Input graph neural network module, integrate graph structure adjacency information, and generate graph enhanced state embedding feature H s ;

[0089] S25, based on the graph enhanced state embedding feature H s Construct a derivable path attention matrix A, where A ij for:

[0090]

[0091] Among them, τ is the temperature adjustment coefficient, A ij is the attention coefficient of the i-th row and j-th column of the derivable path attention matrix, H s,i is the feature vector of the i-th time step in the state embedding sequence after graph enhancement, H s,j is the feature vector of the jth time step in the state embedding sequence after graph enhancement, H s,k is the feature vector of the kth time step in the state embedding sequence after graph enhancement, exp is the exponential function, and T is the length of the time series;

[0092] A ij The practical significance of this formula lies in its ability to model the dynamic relationships between multi-channel state data through a "guided path attention mechanism." The core idea of ​​this mechanism is to automatically calculate the degree of attention each time point pays to other time points based on the feature differences and correlation strengths exhibited by different state channels at different times, thereby forming a set of attention weights that are physically structure-aware and time-sensitive. This attention mechanism not only incorporates the graph-structured relationships between channels in terms of device connectivity and control coupling, but also dynamically adjusts the attention region based on the similarity between state features. This enables the network to selectively capture the most representative temporal path information even in complex situations such as time warping and sampling asynchrony in sensor data. Through this mechanism, the state features at each moment are determined not only by the current channel data itself but also by the importance of other relevant time points, forming a fused representation with a stronger sense of overall trends and local criticality. This path-aware attention modeling approach effectively improves the accuracy and robustness of dynamic time warping networks in state feature alignment and is a key technical foundation for achieving high-precision cross-modal alignment and risk identification.

[0093] S26. Based on the derivable path attention matrix A, calculate the weighted feature representation of each time step and fuse the original features, build a residual gating mechanism, and calculate the final aligned feature sequence for:

[0094]

[0095] Among them, α tis the gate weight of the t-th time step, which is dynamically generated by the position information, A tj is the attention coefficient of the t-th row and j-th column of the derivable path attention matrix, Embed the state path vector for the jth time step;

[0096] The practical significance of this formula lies in the construction of a residual gating mechanism that fuses contextual information at the current time point with original state features to generate the final aligned key state feature sequence. During dynamic time warping, raw multi-channel state data often suffers from issues such as varying sampling frequencies and temporal misalignment. Relying solely on local features makes it difficult to accurately capture the full picture of the true operational state. To this end, this mechanism uses a path attention mechanism to calculate a weighted representation of each time step relative to other time steps, extracting feature information with temporal and structural dependencies. During the fusion process, gating weights are dynamically generated based on the positional information at each time point, adaptively adjusting the fusion ratio between historical information and current features. This ensures the complete representation of key state information while preserving intuitive features in the original data relevant to the actual platform operation, effectively avoiding overfitting and feature degradation. The resulting aligned feature sequence exhibits global perception capabilities while maintaining temporal continuity and operational stability, providing a reliable state input foundation for subsequent joint behavioral modeling and risk scoring analysis. This method is highly innovative and has excellent engineering adaptability in the field of multi-source asynchronous state data processing.

[0097] S27, based on the final alignment feature sequence Z final Embedded with the initial state sequence Z s Based on the feature distribution deviation between them, an alignment consistency driven feedback mechanism is constructed to generate an alignment consistency score vector C, which dynamically adjusts the regional weight distribution of the path attention matrix in the inference stage.

[0098] S28. Through the back-propagation algorithm, the structural parameters of the modal projection encoder, graph neural network, path attention mechanism and residual fusion channel are jointly optimized to achieve the precise alignment of multi-channel operation status data by the dynamic time warping network and output the key state feature sequence Z final .

[0099] The present invention realizes high-precision alignment of multi-channel operating status data of aerial work platforms by improving the dynamic time warping network, which significantly improves the effectiveness and stability of cross-modal data fusion. The introduction of modal projection encoder can map different physical channel data to a unified embedding space, overcoming the dimensional difference and representation mismatch problems between channels; by constructing a graph structure and fusing graph neural networks, the physical connection relationship and control coupling characteristics between state channels are fully utilized to enhance the structural perception ability of embedded features; the further proposed directional path attention mechanism combines the adjacency graph structure to generate a path attention matrix, which is trainable and interpretable; the residual gated fusion mechanism realizes dynamic weight adjustment while maintaining the original feature information, effectively improving the alignment robustness; the alignment consistency feedback mechanism can dynamically perceive the embedding distribution difference and optimize the distribution response of path attention in the time series dimension. The overall solution can adapt to the time series distortion of state data in different scenarios, improve the quality of cross-channel information alignment, and provide accurate and interpretable input features for subsequent risk assessment, with significant accuracy, stability and engineering practical value.

[0100] In this embodiment, S3 specifically includes:

[0101] S31, collect the operator behavior video stream data, decompose it into continuous frame-level image sequence X v ={x1,x2,…,x T}, where each frame image x t ∈R H×W×C , perform unified time regularization to form a time series length T that is consistent with the status data of the operation platform, where R is a real number set, H is the image height, W is the image width, and C is the number of image channels;

[0102] S32, perform image block segmentation and linear projection on the image sequence, and map each frame into a visual embedding vector z t , combined to obtain the visual embedding sequence Z v , and introduce position encoding P to obtain input representation E;

[0103] S33. Construct a dual-path Transformer structure model, which includes a behavior evolution perception path and a global supervision fusion path. The behavior evolution perception path uses a local window attention mechanism and a temporal factor encoder to model behavioral micro-changes, while the global supervision fusion path captures global trends through a cross-temporal fully connected attention mechanism.

[0104] S34, the behavior evolution perception path introduces a frame-level change detection module to calculate the feature difference vector Δz between each frame image and the previous frame t , used for time evolution guidance, and performs a time-weighted self-attention mechanism on each window to output a local representation sequence H evo ;

[0105] S35, the global supervision fusion path uses a multi-head attention module to model the full sequence execution behavior context and introduces a cross-modal reference guidance weight G t The cross-modal reference guided weight is modeled by associating the visual embedding features of each frame image with the aligned state features of the corresponding time step, and using the weight coefficient reflecting the correlation between the two modalities generated by the learnable mapping function to generate an enhanced global representation H by aligning the attention with the state features. global ;

[0106] S36, fuse the behavior evolution perception path and the global supervision fusion path output, and calculate the fusion representation H through the residual attention gating mechanism fuse (t):

[0107] H fuse (t) = γ t ·H evo (t)+(1-γ t )·H global (t);

[0108] Among them, γ t =σ(W r ·[H evo (t),H global (t)]) is the gating coefficient, σ is the activation function, W r is a learnable parameter;

[0109] H fuse The practical significance of the (t) formula lies in its dynamic fusion of behavioral features output by the dual-path Transformer architecture to construct a more discriminative and robust comprehensive representation of task behavior. Specifically, this formula introduces a residual attention gating mechanism to weightedly integrate features from the behavior evolution perception path and the global supervision fusion path. The gating coefficient is a learnable, dynamic adjustment factor generated based on feature content, which determines the contribution ratio of the two types of features to the final fused representation. This mechanism effectively addresses issues such as redundant information and representation offset that may exist between different paths. This allows the fused features to preserve subtle behavioral changes while also capturing global contextual trends, improving the ability to represent complex behavioral patterns. Furthermore, the introduction of a residual structure enhances training stability and the consistency of feature representation, preventing feature degradation or distortion during deep fusion. The resulting fused behavioral representation is not only sensitive to the dynamic evolution of behavioral actions but also accurately captures the location and intensity of abnormal behaviors in the global sequence. This provides high-quality behavioral feature input with clear semantics and distinct hierarchies for risk scoring models, demonstrating strong engineering practicality and modeling innovation value.

[0110] S37, introduce the behavior dynamic fuzzy enhancement module to fusion representation H fusePerform fuzzy perception filtering, and obtain the enhanced sequence H by structural noise suppression and background mask filtering. enhanced ;

[0111] S38, the enhanced sequence H enhanced Input a two-layer feedforward neural network to extract deep abstract semantic features and generate a task behavior representation sequence R v ;

[0112] S39. Execute the feature orthogonality constraint strategy on the job behavior representation sequence, calculate the cosine similarity matrix for each time step feature, and perform diagonal enhancement operation to suppress feature redundancy.

[0113] The present invention achieves high-dimensional semantic modeling and fine feature extraction of operator behavior video streams by constructing an improved dual-path Transformer structure, significantly improving the accuracy and robustness of behavior recognition. By decomposing the video stream into frame-level images and unifying the time regularization, the time dimension of the visual sequence and the platform state data are effectively aligned to ensure the consistency of cross-modal analysis; image block segmentation and position encoding methods are used to form a stable visual embedding representation; in terms of structural design, the behavior evolution perception path and the global supervision fusion path are innovatively introduced. The former uses local window attention and time factor encoder to capture short-term micro-behavior changes, and the latter combines cross-modal reference weights and multi-head attention mechanism to extract global context trends and effectively fuse state data guidance information; the residual attention gating mechanism strengthens the collaborative modeling of the two paths to ensure feature transfer efficiency and expression diversity; the introduction of the behavior dynamic fuzzy enhancement module can enhance the target for blur, occlusion and noise in the image and improve feature quality; finally, the redundancy of the time step features is constrained by the feature orthogonality constraint strategy, making the generated operation behavior representation clearer and more independent. The overall solution has extremely strong behavior recognition accuracy, temporal robustness and semantic representation capabilities, providing a solid visual feature foundation for subsequent risk modeling.

[0114] In this embodiment, the S4 specifically includes:

[0115] S41. Obtain the key state feature sequence Z output by the dynamic time warping network final ={z1,z2,…,z T}, where z t represents the platform operation status characteristics after alignment, and T is the length of the time series;

[0116] S42. Obtain the task behavior representation sequence R output by the dual-path Transformer structure v ={r1,r2,…,r T}, where r t represents the behavioral characteristics of the operator at the tth time step;

[0117] S43, key state feature sequence Z final and task behavior representation sequence R v Perform feature concatenation according to the time step and construct a joint feature sequence U={u1,u2,…,u T};

[0118] S44, input the joint feature sequence U into the joint risk scoring sub-network, use the multi-layer feedforward neural network to extract the coupling feature representation, and generate the risk representation sequence S = {s1, s2, ..., s T}, where s t represents the candidate risk score value at the t-th time step;

[0119] S45. Introduce a modal perception-guided aggregation mechanism to interactively map behavioral features with joint state features, guide the weighting of risk scores at each time step, and generate a fusion score value.

[0120]

[0121] Among them, ρ t is the modal guided gating coefficient, which is determined by the behavioral characteristics r t and platform operating status characteristics t The matching relationship is generated; φ(r t ,u t ) is a nonlinear mapping function used to generate risk compensation items;

[0122] The practical significance of the formula lies in the dynamic weighting and optimization of risk scoring results by introducing a modality-aware guided aggregation mechanism, thereby improving the accuracy and robustness of job status assessment. This mechanism comprehensively considers the correlation between job behavior characteristics and platform status characteristics, generating a modality-guided gating coefficient for each time step. The final risk score weight is adaptively adjusted based on the behavioral characteristics' ability to explain the state characteristics. In other words, if the behavioral characteristics at a certain time point are highly coupled with the platform state, the gating coefficient will be given a higher weight, emphasizing their importance in risk assessment; otherwise, their influence will be suppressed. Furthermore, a risk compensation term generated by nonlinear mapping is introduced to emphasize abnormal states or potential risk trends, effectively addressing the lack of discriminative power of single-modality scores in critical conditions. This entire aggregation process not only strengthens the cross-modal collaborative analysis capabilities but also enhances the model's sensitivity to minor anomalies and potential risks, thereby enabling more reliable intelligent early warnings in multi-source heterogeneous data fusion scenarios. This is a key step in building an efficient and precise risk identification mechanism.

[0123] S46, fusion score value Compare with the set risk level threshold value set Θ = {θ1, θ2, θ3}, which corresponds to θ1 mild, θ2 moderate, and θ3 severe risk levels respectively;

[0124] S47. Determine a risk level label L∈{0,1,2,3} based on the comparison result, where 0 represents no risk, 1 represents mild risk, 2 represents moderate risk, and 3 represents severe risk;

[0125] S48: Output the fusion score and the corresponding risk level label L, which serves as the input for triggering the early warning mechanism, generating control instructions, and executing the control logic.

[0126] This invention, through the design of a joint state-behavior modeling and risk scoring mechanism, achieves collaborative risk perception of the status of personnel and equipment in aerial work scenarios, significantly improving the accuracy and reliability of early warning identification. By extracting aligned state features and work behavior features from a dynamic time warping network and a dual-path Transformer architecture, and then concatenating them after temporal alignment, a joint feature sequence is constructed, effectively integrating cross-modal information. Furthermore, a joint risk scoring subnetwork is introduced, mining multimodal coupling features through a multi-layer neural architecture to accurately generate a risk representation reflecting the work status. The innovative modality-aware guided aggregation mechanism adaptively adjusts the weight of the risk score at each time step based on the matching relationship between behavioral and state features, achieving behavior-driven dynamic risk compensation and enhancing the sensitivity and interpretability of the score. Multi-level risk level thresholds are also set and level labels are output, providing highly reliable input for early warning mechanisms and platform control command generation and execution. This solution offers advantages such as fine recognition granularity, strong cross-modal fusion, and precise risk assessment. It provides key support for intelligent and automated supervision of high-risk work scenarios, demonstrating outstanding safety control effectiveness and engineering application value.

[0127] In this embodiment, the S5 specifically includes:

[0128] S51. Receive the output risk score result and the corresponding risk level label as input information for risk assessment of the current operation status;

[0129] S52. Set a multi-level risk level threshold standard, compare and analyze the input risk score results, and determine whether they exceed the set mild, moderate or severe risk level thresholds;

[0130] S53. When the risk score exceeds any level threshold, the early warning mechanism is automatically activated, and the risk level trigger level is determined based on the comparison results;

[0131] S54. Select a corresponding warning output method based on the triggered risk level, including: activation of an audible and visual alarm device, remote notification from the platform control center, or a pop-up reminder on a mobile terminal;

[0132] S55. If the risk level is moderate or severe, then while outputting the notification, a corresponding control instruction is generated for the platform controller to execute a linkage response operation;

[0133] S56. Record the current risk score value, risk level label and the executed early warning response type together as a control log and risk feedback data archive.

[0134] By designing an intelligent early warning response mechanism based on multi-level risk scoring, this invention enables automatic identification and graded intervention of risk states for aerial work platforms, significantly improving the efficiency of safety control and intelligent response at work sites. The system receives and analyzes the risk score and corresponding risk level label output by the preceding risk modeling module, and dynamically determines the risk level of the current work state based on pre-set multi-level risk level thresholds. Based on different risk levels, the graded early warning mechanism is automatically activated, flexibly incorporating multiple early warning methods such as audible and visual alarms, notifications from a remote control center, and pop-up notifications on mobile terminals, ensuring accurate and rapid delivery of early warning information in different scenarios. In particular, in moderate or severe risk states, the system can interact with the platform controller to automatically generate control instructions to execute safety actions such as speed limiting and locking, forming a closed-loop response chain from early warning to control. This method also systematically archives risk scoring results, level labels, and early warning execution records, providing data support for subsequent risk analysis and strategy optimization. This integrated mechanism offers advantages such as timely identification, rapid response, efficient linkage, and complete records, and possesses significant application value in improving platform safety management and accident prevention capabilities.

[0135] refer to Figure 2 , an intelligent monitoring system for aerial work platforms based on data analysis, including the following modules:

[0136] The data acquisition module is used to collect platform operation status data and operator behavior video stream data, and perform timestamp marking and cache storage;

[0137] The time warping module is used to perform cross-modal time alignment on the platform operation status data and output the aligned key state feature sequence;

[0138] The behavior extraction module is used to input the operator behavior video stream data into the dual-path Transformer structure model to extract the work behavior representation sequence;

[0139] Joint assessment module, used to fuse key state feature sequences and operational behavior representation sequences, construct joint risk functions, and output risk scoring results;

[0140] The risk comparison module is used to compare the risk score results with the risk level threshold to determine the risk level;

[0141] Early warning output module, used to output sound and light alarms, remote notifications or terminal reminders according to risk levels;

[0142] The control execution module is used to link the platform controller to execute speed limit, return, lock or pause operations according to the control instructions;

[0143] The continuous optimization module is used to use key state feature sequences, operation behavior representation sequences, risk scoring results and operation feedback as training samples to perform joint parameter updates and structural adjustments.

[0144] Example 1:

[0145] In order to verify the feasibility of the present invention in implementation, the present invention was applied to a certain aerial work platform operation base. Three aerial work platforms of model "GTZ24-Ⅱ" were selected as test equipment. A two-hour test verification was carried out from 8:00 to 10:00 on the morning of May 10, 2025, and the intelligent supervision method for aerial work platforms based on data analysis proposed in the present invention was fully applied.

[0146] The test scenario simulates a real-world construction environment, encompassing typical activities such as multi-person collaborative work, material handling, and overhead control. Platform operating status data covers 12 channels, including platform height, platform inclination, hydraulic system pressure, drive motor speed, and vibration data, sampled at a 10Hz per second. Video streams of operator activity are captured using a high-definition fisheye camera and uploaded in real time to an edge processing unit.

[0147] First, all state data and video streams are timestamp-aligned and cached. The platform state data is then fed into an improved dynamic time warping network, which integrates a modality projection encoder, graph neural architecture modeling, a path attention mechanism, and a residual alignment mechanism. This accurately identifies the temporal warping relationships between different modalities and outputs a sequence of key state features.

[0148] Subsequently, the worker behavior video stream is fed into a dual-path Transformer architecture model. The model extracts behavioral representation sequences from two dimensions: behavioral evolutionary trends and global action structure. This is then combined with state features to create a risk score. The system incorporates a modal perception guidance mechanism to fuse risk values ​​and outputs the current risk level based on a set threshold level.

[0149] Once the system identifies a moderate risk or higher, it immediately triggers an audible and visual alarm and sends a control command to the platform control center via the edge module. For example, at 8:36:21, a platform operator was not wearing a seatbelt and performed an illegal lifting maneuver. The system risk score reached 2.78 (exceeding the moderate threshold), automatically triggering a moderate warning. The system then slowed down the platform and restricted its height, averting a potential accident.

[0150] Ultimately, all the characteristic data, risk levels, and control feedback collected by the system are combined as dynamic training samples and fed into the model, automatically completing parameter updates. Throughout this process, the system maintains a real-time response within 0.3 seconds, with a significantly lower false alarm rate than traditional systems.

[0151] Table 1 Comparative effect data of the present invention and traditional method

[0152] index Traditional methods Method of the present invention Behavior recognition accuracy 84.7％ 96.3％ State recognition accuracy 81.3％ 94.1％ Risk score and threshold matching rate 78.9％ 92.8％ Warning response time (seconds) 12.4 3.2 Alarm trigger success rate 85.2％ 95.6％

[0153] According to the comparative analysis of the key indicators in Table 1, it can be seen that the method of the present invention is significantly superior to the traditional method in many aspects, reflecting strong practicality and technical advantages.

[0154] First, in terms of behavior recognition accuracy, the proposed method achieved 96.3%, an increase of nearly 12 percentage points compared to the 84.7% of traditional methods. This is mainly due to the deep integration of the behavior evolution perception path and the global supervision fusion path in the dual-path Transformer structure, which effectively improves the modeling and recognition capabilities of complex behavior sequences.

[0155] Secondly, in terms of state recognition accuracy, this invention, by introducing a dynamic time warping network and supplementing it with graph-enhanced state modeling technology, achieves a recognition accuracy of 94.1%, far exceeding the 81.3% of traditional methods. This demonstrates that the system can more accurately perceive the platform's operating status, significantly reducing the possibility of monitoring errors.

[0156] The matching rate between risk score and threshold reached 92.8% for the present invention, while the traditional method only achieved 78.9%. This shows that the present invention is more accurate in characterizing risks, triggering early warning mechanisms that are more closely aligned with actual risk conditions, and avoiding false triggering or missed reports.

[0157] Regarding the key safety metric of early warning response time, the proposed method achieves a response time of 3.2 seconds, significantly faster than the 12.4 seconds of the traditional method. This substantial improvement is crucial for high-altitude operations, enabling risk assessment and response to be completed in a fraction of the time, significantly improving operational safety.

[0158] The proposed method achieved a 95.6% alarm trigger success rate, compared to 85.2% for traditional methods, demonstrating a more stable early warning capability across different risk levels. This capability relies on the deep integration of multimodal coupling analysis and risk scoring mechanisms, ensuring the system strikes a good balance between real-time performance and accuracy.

[0159] In summary, the tabular data fully demonstrates that the method proposed in this invention has significant advantages in behavior understanding, state perception, risk identification and response efficiency, and provides a safer, more reliable and efficient solution for the intelligent supervision of aerial work platforms.

[0160] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for intelligent monitoring of aerial work platforms based on data analysis, characterized in that: The steps include: S1. Collect the operation status data of the aerial work platform and the video stream data of the operator's behavior, perform timestamp marking and cache storage; S2. Input the operation platform operation status data into the dynamic time warping network, perform cross-modal time alignment, identify the time warping relationship between different data sources through the dynamic warping mechanism, and output the aligned key state feature sequence; S3. Construct a dual-path Transformer structure model, input the operator behavior video stream data into the dual-path Transformer structure model, and generate a work behavior representation sequence; S4. Jointly model the key state feature sequence and the operation behavior representation sequence, construct a joint risk scoring function, calculate the risk score of the current operation state based on the coupling analysis of multi-dimensional features, and output the risk scoring result; S5. Compare the risk score result with the risk level threshold. If the risk score exceeds the risk level threshold, trigger the early warning mechanism and output an audible and visual alarm signal, remote notification information or control command according to the risk level; S6. Linking the platform execution controller in the aerial work platform according to the control command to perform automatic speed limit, platform return, lock control or work pause operation, and storing the operation feedback together with the risk scoring result; S7. Input the key state feature sequence, operation behavior representation sequence, risk score results and operation feedback as training samples into the dynamic time warping network and dual-path Transformer structure model to perform parameter update and structure adjustment.

2. The intelligent monitoring method for aerial work platforms based on data analysis according to claim 1 is characterized in that: The S2 specifically includes: S21. Obtain the operation status data of the operation platform and construct a multi-channel time series input. Let the status data of the cth channel be Where c=1,2,…,C, all channel data are combined to form the input sequence set x c (t i ) represents the time point t i At , the state value collected by the cth channel, N is the number of time steps on a single channel, and C is the total number of operating state data channels that can be collected in the operation platform; S22, the status data of each channel data X c Input modal projection encoder f c (), map different modal features to a unified aligned embedding space to obtain the state embedding sequence Z c =f c (X c ), merge all channel embeddings into the initial state embedding sequence S23. Construct the operation platform operation status sensing channel graph structure G = (V, E), where V is the node set of the graph, E is the edge set of the graph, and each node v i ∈V represents a state channel, edge e ij ∈E represents the physical connection relationship or control coupling relationship between modules; S24. Embed the initial state into sequence Z s Input graph neural network module, integrate graph structure adjacency information, and generate graph enhanced state embedding feature H s ; S25, based on the graph enhanced state embedding feature H s Construct a derivable path attention matrix A, where A ij for: Among them, τ is the temperature adjustment coefficient, A ij is the attention coefficient of the i-th row and j-th column of the derivable path attention matrix, H s,i is the feature vector of the i-th time step in the state embedding sequence after graph enhancement, H s,j is the feature vector of the jth time step in the state embedding sequence after graph enhancement, H s,k is the feature vector of the kth time step in the state embedding sequence after graph enhancement, exp is the exponential function, and T is the length of the time series; S26. Based on the derivable path attention matrix A, calculate the weighted feature representation of each time step and fuse the original features, build a residual gating mechanism, and calculate the final aligned feature sequence for: Among them, α t is the gate weight of the t-th time step, which is dynamically generated by the position information, A tj is the attention coefficient of the t-th row and j-th column of the derivable path attention matrix, Embed the state path vector for the jth time step; S27, based on the final alignment feature sequence Z final Embedded with the initial state sequence Z s Based on the feature distribution deviation between them, an alignment consistency driven feedback mechanism is constructed to generate an alignment consistency score vector C, which dynamically adjusts the regional weight distribution of the path attention matrix in the inference stage. S28. Through the back-propagation algorithm, the structural parameters of the modal projection encoder, graph neural network, path attention mechanism and residual fusion channel are jointly optimized to achieve the precise alignment of multi-channel operation status data by the dynamic time warping network and output the key state feature sequence Z final .

3. The method and system for intelligent monitoring of aerial work platforms based on data analysis according to claim 1, characterized in that: The S3 specifically includes: S31, collect the operator behavior video stream data, decompose it into continuous frame-level image sequence X v ={x1,x2,…,x T }, where each frame image x t ∈R H×W×C , perform unified time regularization to form a time series length T that is consistent with the status data of the operation platform, where R is a real number set, H is the image height, W is the image width, and C is the number of image channels; S32, perform image block segmentation and linear projection on the image sequence, and map each frame into a visual embedding vector z t , combined to obtain the visual embedding sequence Z v , and introduce position encoding P to obtain input representation E; S33. Construct a dual-path Transformer structure model, which includes a behavior evolution perception path and a global supervision fusion path. The behavior evolution perception path uses a local window attention mechanism and a temporal factor encoder to model behavioral micro-changes, while the global supervision fusion path captures global trends through a cross-temporal fully connected attention mechanism. S34, the behavior evolution perception path introduces a frame-level change detection module to calculate the feature difference vector Δz between each frame image and the previous frame t , used for time evolution guidance, and performs a time-weighted self-attention mechanism on each window to output a local representation sequence H evo ; S35, the global supervision fusion path uses a multi-head attention module to model the full sequence execution behavior context and introduces a cross-modal reference guidance weight G t The cross-modal reference guided weight is modeled by associating the visual embedding features of each frame image with the aligned state features of the corresponding time step, and using the weight coefficient reflecting the correlation between the two modalities generated by the learnable mapping function to generate an enhanced global representation H by aligning the attention with the state features. global ; S36, fuse the behavior evolution perception path and the global supervision fusion path output, and calculate the fusion representation H through the residual attention gating mechanism fuse (t): H fuse (t)=γ t ·H evo (t)+(1-γ t )·H global (t); Among them, γ t =σ(W r ·[H evo (t),H global (t)]) is the gating coefficient, σ is the activation function, W r is a learnable parameter; S37, introduce the behavior dynamic fuzzy enhancement module to fusion representation H fuse Perform fuzzy perception filtering, and obtain the enhanced sequence H by structural noise suppression and background mask filtering. enhanced ; S38, the enhanced sequence H enhanced Input a two-layer feedforward neural network to extract deep abstract semantic features and generate a task behavior representation sequence R v ; S39. Execute the feature orthogonality constraint strategy on the job behavior representation sequence, calculate the cosine similarity matrix for each time step feature, and perform diagonal enhancement operation to suppress feature redundancy.

4. The intelligent monitoring method for aerial work platforms based on data analysis according to claim 1 is characterized in that: The S4 specifically includes: S41. Obtain the key state feature sequence Z output by the dynamic time warping network final ={z1,z2,…,z T }, where z t represents the platform operation status characteristics after alignment, and T is the length of the time series; S42. Obtain the task behavior representation sequence R output by the dual-path Transformer structure v ={r1,r2,…,r T }, where r t represents the behavioral characteristics of the operator at the tth time step; S43, key state feature sequence Z final and task behavior representation sequence R v Perform feature concatenation according to the time step and construct a joint feature sequence U={u1,u2,…,u T }; S44, input the joint feature sequence U into the joint risk scoring sub-network, use the multi-layer feedforward neural network to extract the coupling feature representation, and generate the risk representation sequence S = {s1, s2, ..., s T }, where s t represents the candidate risk score value at the t-th time step; S45. Introduce a modal perception-guided aggregation mechanism to interactively map behavioral features with joint state features, guide the weighting of risk scores at each time step, and generate a fusion score value. Among them, ρ t is the modal guided gating coefficient, which is determined by the behavioral characteristics r t and platform operating status characteristics t The matching relationship is generated; φ(r t ,u t ) is a nonlinear mapping function used to generate risk compensation items; S46, fusion score value Compare with the set risk level threshold value set Θ = {θ1, θ2, θ3}, which corresponds to θ1 mild, θ2 moderate, and θ3 severe risk levels respectively; S47. Determine a risk level label L∈{0,1,2,3} based on the comparison result, where 0 represents no risk, 1 represents mild risk, 2 represents moderate risk, and 3 represents severe risk; S48: Output the fusion score and the corresponding risk level label L, which serves as the input for triggering the early warning mechanism, generating control instructions, and executing the control logic.

5. The intelligent monitoring method for aerial work platforms based on data analysis according to claim 1 is characterized in that: The S5 specifically includes: S51. Receive the output risk score result and the corresponding risk level label as input information for risk assessment of the current operation status; S52. Set a multi-level risk level threshold standard, compare and analyze the input risk score results, and determine whether they exceed the set mild, moderate or severe risk level thresholds; S53. When the risk score exceeds any level threshold, the early warning mechanism is automatically activated, and the risk level trigger level is determined based on the comparison results; S54. Select a corresponding warning output method based on the triggered risk level, including: activation of an audible and visual alarm device, remote notification from the platform control center, or a pop-up reminder on a mobile terminal; S55. If the risk level is moderate or severe, then while outputting the notification, a corresponding control instruction is generated for the platform controller to execute a linkage response operation; S56. Record the current risk score value, risk level label and the executed early warning response type together as a control log and risk feedback data archive.

6. An intelligent monitoring system for aerial work platforms based on data analysis, an intelligent monitoring method for aerial work platforms based on data analysis according to any one of claims 1 to 5, characterized in that: Includes the following modules: The data acquisition module is used to collect platform operation status data and operator behavior video stream data, and perform timestamp marking and cache storage; The time warping module is used to perform cross-modal time alignment on the platform operation status data and output the aligned key state feature sequence; The behavior extraction module is used to input the operator behavior video stream data into the dual-path Transformer structure model to extract the work behavior representation sequence; Joint assessment module, used to fuse key state feature sequences and operational behavior representation sequences, construct joint risk functions, and output risk scoring results; The risk comparison module is used to compare the risk score results with the risk level threshold to determine the risk level; Early warning output module, used to output sound and light alarms, remote notifications or terminal reminders according to risk levels; The control execution module is used to link the platform controller to execute speed limit, return, lock or pause operations according to the control instructions; The continuous optimization module is used to use key state feature sequences, operation behavior representation sequences, risk scoring results and operation feedback as training samples to perform joint parameter updates and structural adjustments.

Images