A method and system for intelligent safety management and control of boiler walkway grating.

Abstract

This invention provides an intelligent safety management method and system for boiler walkway gratings, relating to the field of industrial safety technology. The method includes: S1, acquiring in real-time the obstruction signal sequence generated by an infrared beam sensor deployed at the grating inlet and the weight signal sequence generated by a pressure sensor deployed below the grating, and synchronizing the obstruction and weight signal sequences with timestamps; S2, inputting the synchronized obstruction and weight signal sequences into a pre-constructed lightweight spatiotemporal feature extraction network to extract multi-dimensional temporal features representing personnel flow, object presence, and weight distribution; and inputting these multi-dimensional temporal features into a risk fusion inference engine built based on Dempster-Shafer evidence theory. This system can eliminate perceptual ambiguity, overcome decision-making rigidity, and mine data value in the harsh edge environment of boiler walkways, achieving progressive intelligent safety management from precise risk identification to dynamic and flexible assessment and early hazard insight.

Description

Technical Field

[0001] This invention relates to the field of industrial safety technology, and in particular to an intelligent safety management method and system for boiler walkway gratings. Background Technology

[0002] With the continuous improvement of industrial safety standards and the in-depth application of intelligent technologies, real-time and reliable safety monitoring of critical infrastructure in core areas such as thermal power plants and chemical plants has become increasingly important. Boiler walkway gratings, as essential passageways for personnel inspections and equipment maintenance, directly impact personnel safety and the safety of main equipment through their structural safety and load-bearing capacity. These areas typically experience extreme and harsh environments such as high temperatures, dust, high humidity, and continuous vibration, posing a significant challenge to the environmental tolerance, reliability, and real-time performance of any monitoring system.

[0003] To address the aforementioned issues, existing technologies primarily employ two approaches. The first approach utilizes computer vision. For example, patent CN121030671A, entitled "A Multimodal Collaborative Security Monitoring Method, Device, Equipment, and Storage Medium," proposes fusing data from multiple sensors, including images and infrared thermal imaging, and utilizing visual recognition algorithms for security event assessment. However, its drawback lies in insufficient environmental robustness: dust and moisture permeating boiler walkways can severely obscure and contaminate lenses, leading to a sharp decline in image quality; environmental vibrations can cause camera shifts, creating blind spots; and complex background and lighting changes can increase the algorithm's false alarm rate. More critically, to achieve complex visual analysis, such solutions typically rely on high computing power, making real-time processing difficult on cost- and power-constrained edge devices. They often require transmitting video data back to the cloud or server, thus becoming limited by network bandwidth and latency, failing to meet the immediacy requirements of security alerts, and rendering the system completely inoperable during network outages. Therefore, while intelligent, this type of solution is fragile and expensive in harsh industrial environments, making large-scale reliable deployment difficult.

[0004] The second technical approach is based on simple and reliable non-visual physical sensors such as infrared beam detectors and pressure sensors. For example, infrared beam obstruction can be used for personnel counting, combined with pressure sensors to measure total weight, and an alarm can be triggered at the local edge computing unit based on a preset fixed threshold. This solution successfully solves the problems of environmental tolerance, real-time performance, and cost. However, its core flaw lies in its rudimentary level of intelligence, which fails to achieve effective risk identification and decision support. Specifically: First, the perception dimension is low, leading to perceptual ambiguity: the system can only acquire low-dimensional information such as "obstruction / non-obstruction" and "total weight value," unable to distinguish between real load risks and signal artifacts caused by sensor temperature drift, environmental vibration and shock, or momentary external interference such as the passage of heavy vehicles in the vicinity. This results in frequent false alarms at night or when operating conditions change; for example, after triggering a "weight over-limit" warning, on-site personnel may find no abnormalities. After several repetitions, alarm fatigue can easily occur among on-duty personnel, severely weakening their trust in the system and potentially leading to manual shutdown of the alarm function, rendering the system ineffective. The contradiction lies in the fact that the system can only mechanically report what exceeds the threshold without explaining why it does so, lacking the ability to assess the reliability of the data. Secondly, the decision-making logic is rigid and lacks dynamic flexibility. The system's early warning response and safety interlocks are entirely based on preset, static threshold logic, unable to dynamically adjust according to actual working conditions, event confidence levels, or historical patterns, such as the need to temporarily increase personnel during emergency maintenance. For example, when the system locks a passage due to a temporary over-limit of people at the entrance, it may be during an emergency maintenance, urgently requiring support personnel and equipment. This rigid "one-size-fits-all" control hinders actual production safety operations, exposing the drawbacks of its "black box" decision-making. This easily leads to over-intervention or alarm fatigue, ultimately being ignored by on-site personnel. Thirdly, the system only uses data to judge whether there is an instantaneous "exceeding the standard," treating massive amounts of continuous process data, such as subtle fluctuations in pressure or slow changes in sensor baselines, as "compliant" information, simply storing or overwriting them, creating a "data graveyard." This prevents the system from uncovering early potential hazards hidden within the data. For example, the load-bearing capacity of a grating plate may slowly decline due to fatigue, which could be reflected in a slight drift of the unloaded baseline on the pressure sensor. However, this may not be detected until deformation occurs during a load test (which may still not exceed the threshold). Because the system only cares about "whether the line is crossed," it wastes the opportunity to perform trend analysis and predictive maintenance, and loses the ability to "prevent problems before they occur."

[0005] In summary, existing technologies have formed two fragmented paradigms: solutions represented by video surveillance, while committed to "intelligent assessment," suffer from "poor survivability"; solutions represented by simple threshold alarms, while "robust and reliable," have "low levels of intelligence." Neither has solved the problem of building a progressive intelligent safety monitoring system in harsh edge environments such as boiler walkways, which possesses both extremely high environmental robustness and the ability to achieve "precise risk identification," "dynamic and flexible assessment," and "early hazard detection." Summary of the Invention

[0006] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is to provide a method and system for intelligent safety management and control of boiler walkway gratings. This system can balance high robustness and real-time performance in the harsh edge environment of boiler walkways, eliminate perceptual ambiguity, overcome decision-making rigidity, and extract data value, thereby achieving progressive intelligent safety management and control from accurate risk identification to dynamic and flexible assessment and early detection of hidden dangers.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The present invention provides an intelligent safety management and control method and system for boiler walkway gratings, comprising the following steps: S1. Multi-source sensor data acquisition and synchronization steps: acquire the occlusion signal sequence generated by the infrared beam sensor deployed at the grid plate inlet and the weight signal sequence generated by the pressure sensor deployed below the grid plate in real time, and synchronize the occlusion signal sequence and the weight signal sequence with timestamps. S2, Uncertainty Risk Perception Step: The synchronized occlusion signal sequence and weight signal sequence are input into a pre-constructed lightweight spatiotemporal feature extraction network to extract multi-dimensional temporal features representing personnel flow, object presence, and weight distribution; the multi-dimensional temporal features are input into a risk fusion inference engine built based on Dempster-Shafer evidence theory, and the risk fusion inference engine outputs at least one risk judgment result with confidence rating, including personnel exceeding limits, object stacking, and sensor anomalies; S3. Dynamic cognition and prediction steps: Input the risk assessment result with confidence rating and the weight signal sequence into a pre-trained long short-term memory network prediction model; use the attention mechanism integrated in the long short-term memory network prediction model to dynamically allocate the weight of the risk assessment result and the weight signal sequence in the current prediction; execute the forward inference algorithm of the long short-term memory network prediction model to output the prediction of the change trend of the load-bearing state of the grating plate in the future period of time. S4. Closed-loop decision generation steps: Store the current risk assessment result with confidence rating and the predicted trend as security status information in the local knowledge base; call the pre-stored policy rules in the local knowledge base to compare and analyze the security status information with similar information stored in history; based on the comparison and analysis results, generate and output a closed-loop decision report containing the current status evaluation, potential risk diagnosis and operation and maintenance suggestions.

[0008] In the preferred embodiment, step S1, the specific process of acquiring and synchronizing multi-source sensor data includes: S11. Sensor Deployment and Signal Acquisition: Based on the width of the grating inlet channel and the location of the load-bearing beam nodes, infrared through-beam sensors are deployed on both sides of the central axis of the inlet channel, and pressure sensors are deployed at the load-bearing beam nodes below the grating. A sampling period is set, and each sensor is driven according to the sampling period to acquire the original occlusion signal sequence in real time. and the original weight signal sequence ; The sampling period includes the sampling period of the infrared beam sensor. and pressure sensor sampling period ; S12. Timestamp synchronization processing: Extract the original occlusion signal sequence and the original weight signal sequence The sampling timestamp, through a hardware-triggered synchronization mechanism, is used to generate the original occlusion signal sequence. and the original weight signal sequence Each data sample in the process is appended with a high-precision timestamp uniformly distributed by the master clock of the edge computing module. This yields a timestamped signal sequence. Alignment verification is performed on timestamped signal sequences to identify signal data segments with inconsistent time nodes; a linear interpolation synchronization algorithm is used, employing the formula... Perform signal value completion calculations on inconsistent data segments, where and The times that need to be filled in are respectively The two most recent valid sampling timestamps are used to obtain a synchronized occlusion signal sequence with perfectly aligned timestamps. and synchronous weight signal sequence .

[0009] In the preferred embodiment, step S2, the construction and feature extraction process of the lightweight spatiotemporal feature extraction network, includes: S21. Network Architecture Search: Obtain the computing power and memory constraint parameters of the ARM processor in the edge computing module; adopt hardware-aware neural architecture search technology to automatically search and evaluate network structures based on the temporal characteristics of infrared occlusion signals and pressure change curves in a preset search space containing various convolutional kernel sizes and network depths, and obtain evaluation results; based on the evaluation results, automatically search, generate and determine an ultra-lightweight spatiotemporal feature extraction network architecture containing truncated convolutional layers, dimension reduction networks and local Transformer layers; S22. Temporal Feature Extraction: Extracting the synchronized occlusion signal sequence... and synchronous weight signal sequence Data is spliced ​​together according to the data channel dimension to form a unified network input data; the unified network input data is then input into an ultra-lightweight spatiotemporal feature extraction network. The input data is processed by a truncated convolutional layer to extract local spatial features, resulting in an initial feature map. The initial feature map is input into a dimension reduction network, which performs linear combination and compression transformation on all feature channels through its 1×1 convolutional kernel. Based on the computation... Reduce the original number of feature channels from Compress to This yields a low-dimensional feature map; S23. Multidimensional feature output: The low-dimensional feature map is sliced ​​along the time dimension to divide it into continuous local temporal segments; the local temporal segments are sequentially input into the local Transformer layer, and the feature dependencies of different time steps within the segment are modeled through the self-attention mechanism to obtain high-level features that integrate spatiotemporal context information. A global average pooling layer is used to perform dimensionality reduction and normalization operations on high-level features in the spatiotemporal dimension, and the output includes short-term occlusion count features. Pressure change rate characteristics Weight distribution variance characteristics Multidimensional temporal feature set ; In step S2, the risk fusion reasoning process based on the Dempster-Shafer evidence theory includes: S24. Risk Identification Framework Definition: Define a risk identification framework based on the objectives of safety monitoring of boiler walkway grating. ,in Characteristic of personnel exceeding limits, Propositions that characterize the stacking of objects Characterizing sensor anomalies Propositions that characterize the state of security; S25. Basic Probability Assignment (BPA) Transformation: Transforming Multidimensional Temporal Feature Sets For each feature in the recognition framework, calculate its relationship to the recognition framework. Each proposition support Based on formula The support of each feature for each proposition is converted into basic probability assignment values, where The minimum constant is set to prevent the denominator from being zero; thus, the basic probability allocation functions derived from the characteristics of infrared blocking signals are obtained. and the fundamental probability assignment function derived from the weight signal characteristics ; S26. Evidence Fusion: Integrating the Basic Probability Assignment Function and Evidence fusion is performed using the Dempster combination rule, and the synthesis formula is as follows: ; The integrated basic probability distribution after fusion is obtained This reflects the degree of joint support for each proposition after fusing evidence from both types of sensors; S27. Confidence rating output: Based on the integrated basic probability allocation. Calculate each proposition Trust function According to the trust function Determine the confidence level of the corresponding risk proposition and output the risk assessment result with the confidence level.

[0010] In the preferred embodiment, step S3, the weight allocation process of the attention mechanism includes: S31. Attention Module Construction: Acquire spatial location information and time series information of multi-sensor data streams, and construct a lightweight spatiotemporal attention module based on the spatial location information and time series information. The spatiotemporal attention module includes parallel spatial attention branches and temporal attention branches. S32. Spatial Attention Calculation: Obtain the physical deployment location coordinates of each sensor, encode the physical deployment location coordinates of each sensor into a spatial location vector, perform a linear transformation on the spatial location vector according to the spatial attention branch, and generate a spatial query vector through the linear transformation. and key vector This yields spatial vector pairs; based on the formula Similarity calculation and normalization are performed on spatial vector pairs to obtain the spatial attention weight matrix; where, For vector dimensions; S33. Temporal Attention Calculation: Calculating Multidimensional Temporal Feature Sets The rate of change of time is obtained by the difference between adjacent time steps. ; the rate of change over time The input to the multilayer perceptron in the temporal attention branch performs a nonlinear mapping to obtain a temporal feature vector; the temporal feature vector is then compressed by the sigmoid activation function to map it to the (0,1) interval, thus obtaining the temporal attention weights. ; S34. Dynamic Weight Fusion: Set or adaptively adjust the balance coefficient between spatial attention and temporal attention based on the dynamic characteristics of the current monitoring scene. Based on formula Spatial attention weight matrix And time attention weight Perform a weighted summation operation to generate the final dynamically assigned weights. ;Utilize dynamic weight allocation The features input to the Long Short-Term Memory Network prediction model are weighted.

[0011] In the preferred embodiment, step S3, the training and inference process of the Long Short-Term Memory (LSTM) prediction model, includes: S35. Model Initialization: Based on the time-series prediction requirements of the load-bearing state of the grating, initialize the input layer dimension and the number of hidden layer units of the LSTM network. In the output layer dimension, ReLU is used as the activation function to obtain a pre-trained LSTM model; S36. Incremental learning optimization: Obtain the newly acquired weight signal sequence within the most recent update cycle. The corresponding actual load-bearing state is then input into the pre-trained LSTM model for forward inference, and the loss function between the pre-trained LSTM model's predicted value and the actual value is calculated. The loss value is obtained; based on the loss value, the network parameters are updated using the gradient descent algorithm, with the update formula being: ,in For learning rate, The gradient of the loss function with respect to the old parameters is used to obtain the incrementally updated LSTM model; S37. Forward Inference Execution: This involves linking the risk assessment result with confidence rating to the synchronous weight signal sequence. Perform feature concatenation to form the model input vector at the current time step. ; input vector Input the incrementally updated LSTM model, and sequentially perform the calculations for the forget gate, input gate, cell state update, and output gate, as shown in the following formulas: ; ; ; ; ; in, This is the output of the hidden layer from the previous time step. This represents the cell state at the previous moment. The function is the Sigmoid function; through the above gating operation, the current hidden layer output state is obtained. .

[0012] In the preferred embodiment, step S3, the output process for predicting the trend of load-bearing state changes, includes: S38. Forecast Window Setting: Based on the actual needs of the industrial site for advance warning, set the length of the future forecast time window. This yields the time window parameters; S39. Trend Assessment: Analyze the hidden layer state sequence output by the LSTM model. The input is fed into a fully connected layer, and the hidden layer state sequence is transformed through a linear transformation. Mapping to the future The sequence of predicted load-bearing states at each time step ,in Based on the predicted value sequence Calculate from the present moment to the future Average rate of change at time To obtain trend indicators; based on the average rate of change The positive and negative signs and the absolute value of the trend indicator are used to classify the future trend of the load-bearing state, and the output is a prediction of increasing, decreasing or stable trend.

[0013] In the preferred embodiment, step S4, the local knowledge base storage procedure includes: S41, DIK architecture instantiation: Obtain the hierarchical definition of data, information, and knowledge, and instantiate the three-layer data-information-knowledge storage architecture according to the hierarchical definition of data, information, and knowledge. The data layer is used to store raw sensor data, the information layer is used to store processed risk and prediction information, and the knowledge layer is used to store operation and maintenance strategies and historical patterns. S42. Data storage optimization: Synchronization blockage signal sequence and synchronous weight signal sequence Data is categorized and stored in the data layer according to its collection timestamp, and a cyclic overwrite mechanism is configured for the data layer, with a set data retention period. To manage storage capacity; to encapsulate risk assessment results with confidence ratings and trend predictions through the spatiotemporal attributes of event occurrences, forming structured security status information, which is then stored in the information layer; and to organize preset operation and maintenance strategy rules and state correlation patterns mined from historical information into triples. The structure is stored in the knowledge layer, where As a triggering condition, For the corresponding operation, For rule confidence; In step S4, the comparison and correlation analysis process of security status information includes: S43. Similarity Calculation: Extracting current security status information from the local knowledge base. Similar security status information stored historically The current security status information With each historical security status information Each feature vector is converted into a standardized feature vector with the same dimension to obtain a comparable vector, and the current security status information is then calculated. With each historical security status information The cosine similarity between them is calculated using the following formula: ; Obtain a similarity score; S44. Association Pattern Mining: Setting Similarity Thresholds From all historical security status information Select those with similarity scores higher than By analyzing historical information, a highly correlated historical dataset is obtained; then, an association rule mining algorithm is used to analyze the current state characteristics of the highly correlated historical dataset. and subsequent state evolution The strength of the causal relationship between them, and the degree of correlation are increased by the degree of correlation. The measurement and calculation formula is as follows: By analyzing the causal relationship between the current state and subsequent historical states, the laws governing state evolution are obtained, among which... Represents probability. This is the current state. This is the state that follows history.

[0014] In the preferred embodiment, step S4, the process of generating the closed-loop decision report, includes: S45. Current Status Evaluation: Obtain the confidence rating of the risk assessment result and the highest similarity score obtained from the similarity calculation, set the status evaluation threshold, perform level classification on the current safety status of the grating, and output a status evaluation of safe, low risk, medium risk or high risk. S46. Potential Risk Diagnosis: Input the predicted trend and state evolution law into the grey prediction model, and then, according to the formula... By performing risk outbreak time prediction, the expected occurrence time of potential risks is obtained, whereby... For development coefficient, The original data sequence is used as the basis for analysis. Combining the risk type and state evolution pattern in the risk assessment results, the causes leading to the current risk state are analyzed, and the risk diagnosis results are output. S47. Operation and Maintenance Suggestions: Call the operation and maintenance strategy rule triplet stored in the knowledge layer of the local knowledge base, match the risk type, diagnosis results and strategy rule execution to obtain a preliminary operation and maintenance plan; adjust the operation and maintenance priority according to the risk level in the current status assessment, and generate operation and maintenance suggestions for targeted operations; integrate the current status assessment, potential risk diagnosis results and operation and maintenance suggestions to form a closed-loop decision report and output it.

[0015] In a preferred embodiment, the present invention also provides an intelligent safety management and control system for boiler walkway gratings, comprising a multimodal sensor module, an edge computing module, a local storage module, and a hierarchical early warning module connected in sequence via communication. The system is used to execute the intelligent safety management and control method for boiler walkway gratings as described above, wherein: The multimodal sensor module is used to acquire multi-source sensing data related to the grid panel and to achieve timestamp synchronization; The edge computing module is used to perform uncertainty risk perception, dynamic cognition and prediction, and closed-loop decision generation on synchronized sensor data; The local storage module is used to store security status information, historical data, and a dynamic policy knowledge base; The tiered early warning module is used to output targeted early warning signals and operation and maintenance suggestions based on the closed-loop decision results.

[0016] In a preferred embodiment, the multimodal sensor module includes an infrared beam sensor submodule, a pressure sensor submodule, a vibration sensor submodule, a vision sensor submodule, and a temperature sensor submodule, wherein: The infrared beam sensor submodule includes at least two pairs of infrared beam sensors deployed in the grid panel entrance channel and on both sides, configured to perform deployment and signal acquisition, and output a sequence of occlusion signals; The pressure sensor submodule includes at least four resistance strain gauge pressure sensors deployed at the load-bearing beam nodes below the grating, configured to perform deployment and data acquisition, and output a sequence of weight signals; The vibration sensor submodule is deployed at the load-bearing structure of the grating plate to collect dynamic mechanical vibration signals of the grating plate; The visual sensor submodule is deployed on the bracket above the grid plate and integrates a dust compensation hardware unit for acquiring visual images of the grid plate surface. The temperature sensor submodule is deployed in the environment surrounding the grating to collect temperature data in the industrial field; Each submodule communicates with the edge computing module through an industrial Ethernet interface and achieves timestamp alignment through a hardware-triggered synchronization mechanism.

[0017] In the preferred embodiment, the edge computing module is an embedded hardware computing unit, including an ARM series processor, a multi-protocol communication interface, an FPGA coprocessor unit, and a real-time operating system, wherein: The processor is used to perform the lightweight spatiotemporal feature extraction network and DS evidence theory fusion reasoning in step S2 above, and the spatiotemporal attention and incremental LSTM model operation in step S3; the processor integrates an incremental learning processor and a spatiotemporal attention controller; Incremental learning processor is an integrated circuit that integrates a long short-term memory network computing unit. The long short-term memory network computing unit dynamically and locally updates its internal weight parameters through an online gradient descent algorithm to achieve incremental learning of the model. The spatiotemporal attention controller is a logic chip that works in conjunction with the incremental learning processor. It includes separate spatial attention computing units and temporal attention computing units, integrates real-time computing and generates dynamic weights to control the weighted fusion of feature vectors input to the incremental learning processor. FPGA coprocessor units are used to accelerate the parallel preprocessing of multimodal data; The real-time operating system adopts a multi-threaded scheduling mechanism, with the thread priority as follows: pressure resistance testing thread > weight monitoring thread > personnel / object quantity monitoring thread > entrance flow monitoring thread; The edge computing module also integrates a hardware-aware neural architecture search optimization engine for customizing ultra-lightweight feature extraction networks for multimodal time-series data.

[0018] In the preferred embodiment, the local storage module is an industrial-grade non-volatile storage unit, and the hierarchical early warning module includes a local audible and visual alarm unit and a host computer communication unit, wherein: The local storage module adopts a DIK three-layer storage architecture. The data layer stores the raw sensor data, the information layer stores the risk assessment results and trend prediction data, and the knowledge layer stores the dynamic strategy knowledge base and incremental learning and updating model parameters. It supports cyclic overwrite and incremental backup and is used to execute step S4 as described above. The local audible and visual alarm unit includes alarm lights of different colors and frequencies, and a voice broadcaster with adjustable volume. It outputs differentiated warnings according to the risk level and is used to perform the warning outputs in steps S15, S25, S35, and S43 as described above. The host computer communication unit uses the TCP / IP protocol to upload early warning information, cancellation signals and detection reports to the host computer system. The edge computing module also integrates an incremental learning hardware acceleration unit to realize online parameter updates of the LSTM model.

[0019] This invention provides an intelligent safety management method and system for boiler walkway grating plates. Through the coordination of the above structures, compared with existing methods, it has the following advantages: First, by adopting a multimodal sensor module and relying on the lightweight model of the edge computing module and the parallel acceleration capability of FPGA, the monitoring failure problem caused by dust obstruction and vibration displacement in the pure vision solution is avoided. It also gets rid of its dependence on high computing power and cloud transmission. Real-time analysis can be completed at the edge without data transmission back. It can operate stably in extreme industrial environments and meet the requirements of immediacy for safety early warning. Secondly, it solves the problem of perception ambiguity and improves the accuracy of risk identification: by extracting the spatiotemporal features of multimodal data and fusing reasoning with DS evidence theory, it can obtain multi-dimensional perception information and assess the credibility of data, effectively distinguish between real load risks and signal illusions such as sensor temperature drift and external vibration interference, avoid the frequent false alarms of traditional threshold schemes, eliminate the "alarm fatigue" of on-duty personnel, and clarify the causes of risks, thus solving the perception defect of traditional systems that "only report phenomena but do not understand the causes"; Third, through the dynamic weight allocation of the spatiotemporal attention controller and the dynamic strategy rules of the local knowledge base, the monitoring threshold and response logic can be adjusted according to different working conditions such as emergency maintenance and normal operation, avoiding the rigid control of the traditional "one-size-fits-all" approach. This ensures the effectiveness of safety management and control without hindering the flexible implementation of actual production operations, thus overcoming the drawbacks of the "black box" of rigid decision-making. Fourth, by relying on the incremental LSTM prediction model to analyze the trend of process data, and combined with the DIK three-layer storage architecture to perform layered accumulation and correlation mining of data, information and knowledge, it is possible to capture early hidden dangers such as the decline in the load-bearing capacity of the grating, and transform the "compliant" process data that is simply stored or covered in the traditional solution into knowledge assets, thus realizing the upgrade of predictive maintenance capabilities from "post-event handling" to "prevention before problems occur". Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a main view diagram of the process structure of this invention; Figure 2 This is a diagram showing the overall appearance and structure of the present invention. Detailed Implementation

[0021] To better understand the purpose, system architecture, and functional implementation of this embodiment, the embodiments and features in the embodiments of this application can be combined with each other without conflict. The exemplary embodiments disclosed in this application will be described below with reference to the accompanying drawings, which include specific technical details disclosed in this embodiment to aid understanding; however, these details should be considered exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the invention. Similarly, for clarity, detailed descriptions of well-known technologies, functions, and structures (such as standard image processing algorithms and common communication protocols) are omitted in the following description.

[0022] In the field of artificial intelligence, pre-trained language models are widely used in natural language processing. Pre-training allows for the learning of general semantic representations from large-scale data through self-supervised learning. In many tasks, using pre-trained language models can significantly reduce the amount of training data required and improve the accuracy of the language model.

[0023] With the development of artificial intelligence technology, pre-trained language models have gradually expanded from monolingual to multilingual and multimodal tasks for application in cognitive advanced AI tasks. Multimodal pre-trained language models can learn representations of multimodal information such as text, images, and videos within a unified semantic space, typically achieving better results than traditional monomodal pre-trained language models. In cognitive advanced AI tasks, multimodal pre-trained language models often require external knowledge to make correct inferences and decisions. However, common multimodal information does not explicitly contain this knowledge, making it difficult for multimodal pre-trained models to directly learn it.

[0024] When related technologies train pre-trained models using external knowledge, the pre-trained language models are limited to text-based pre-trained language models in terms of information modality processing, and are limited to text fragments in terms of knowledge type utilization. Therefore, the method of training pre-trained models using external knowledge in related technologies has certain limitations, which makes it difficult to meet the application requirements of multimodal information in advanced artificial intelligence tasks.

[0025] Example 1 This embodiment details a method for intelligent safety management and control of boiler walkway gratings, as well as the training steps, parameter configuration, and implementation process of the model in the system.

[0026] In the preferred embodiment, both the training and testing data are derived from a multimodal sensor network deployed on-site at the boiler walkway grating of a large chemical plant. The grating specifications are: a single piece 6m in length and 1.5m in width, with a designed maximum load capacity of 1000kg. The environmental parameters of the deployment area are: temperature -10℃ to 70℃, relative humidity 30% to 95%, and dust concentration ≤5mg / m³. The sensor deployment strictly adheres to the following configuration: Infrared beam sensors: One pair is deployed on both sides of the central axis of the entrance channel, at a height of 1.3m, with a sampling period of 100ms, an emission wavelength of 940nm, and a detection distance of 1.5m; Three pairs (spaced 2m apart) are deployed along the length of both sides of the grid plate, at a height of 0.9m, with the monitoring range of adjacent sensors overlapping by 0.25m, and the output signal is a high level of 3.3V (unobstructed) and a low level of 0V (obstructed). Pressure sensors: Four sensors are deployed at the load-bearing beam nodes at the four corners of the grating plate, with a range of 0~500kg, an accuracy of 0.5 grade, a sampling period of 200ms, and an output of 4-20mA current signal; Auxiliary sensors: Simultaneously deploy vibration sensors (sampling period 50ms, range 0-10g) and temperature sensors (sampling period 1s, range -20℃ to 80℃) to supplement multimodal characteristics.

[0027] In this embodiment, the data collection period was 30 days, with 24-hour continuous data collection per day, totaling 1.2TB of data. This included five main categories of data: normal operating conditions, personnel exceeding limits, object stacking, grating corrosion / deformation / loosening damage, and sensor anomaly data. Before data preprocessing, a quality screening was performed to remove invalid data such as sensor disconnections and signal interruptions, resulting in 98.7% valid data.

[0028] Specifically, the linear interpolation synchronization algorithm performs the following process for aligning the timestamps of multi-source sensor signals: During the data acquisition phase, a hardware-triggered synchronization mechanism is used to add a high-precision timestamp to each sensor signal sample. (Accurate to 1ms), where the infrared beam sensor timestamp sequence is: The pressure sensor timestamp sequence is ; Identifying asynchronous time points: Traverse the two sequences and calculate the timestamp difference. ,when The nodes were identified as asynchronous nodes, and the percentage of asynchronous nodes was 3.2%. Linear interpolation calculation: for asynchronous times Select the closest synchronization time before and after. and ,in Through formula Complete the signal value. For example, an infrared sensor in... No sampling points The signal is high (3.3V). When the signal is low (0V), the interpolation is obtained. ; Post-synchronization verification: The timestamp deviation of the synchronized signal is calculated to be ≤2ms, which meets the time alignment requirements of multimodal data fusion.

[0029] In this embodiment, the data cleaning algorithm is used to remove redundancy and noise. The specific steps are as follows: Redundancy removal: A sliding window deduplication method is used, with the window size set to 5 sampling periods. When the signal value change within the window is ≤0.01V (infrared sensor) or ≤0.1kg (pressure sensor), the first sampled value within the window is retained, and the remaining redundant data is removed. After redundancy removal, the amount of data is reduced by 28%. Noise filtering: The infrared sensor signal is filtered using median filtering (window size 3) to remove instantaneous fluctuations caused by dust interference; the pressure sensor signal is filtered using Gaussian filtering (standard deviation). The vibration interference was smoothed out, and the signal-to-noise ratio of the filtered signal was improved from 25dB to 42dB.

[0030] In practice, the data normalization algorithm is used to unify the dimensions of resources from different modalities. The steps are as follows: Statistical analysis of extreme values ​​for each sensor: Maximum value of infrared sensor signal Minimum value Maximum weight of pressure sensor Minimum value Maximum value of temperature sensor Minimum value ; Using linear normalization formula Map all data to the (0,1) interval. For example, a pressure sensor sample value of 300 kg would be normalized to... A temperature sensor's sampled value is 30℃, which, after normalization, is... ; Validation after normalization: The mean of each modality data is 0.48-0.52, and the variance is 0.18-0.22, which meets the data distribution requirements for model training.

[0031] In the preferred scheme, the resource frequency statistics algorithm is used to calculate the frequency value of sensed resources, and the resources are processed in three categories: Single-variable numerical resources (such as pressure sensor weight data): Statistical analysis of a certain value range Frequency within Total frequency Through formula Calculate the frequency value. For example, if the frequency of a weight in the 800-1000kg range is 120, and the total frequency is 10000, then the frequency value is... ; Multivariate numerical resources (such as infrared + pressure fusion data): K-means clustering (k=4 clusters) is used to analyze resource distribution and calculate the probability of each category. The maximum value is taken as the resource value. For example, if the probability of a sample belonging to the "weight overload" category after clustering is 0.92, then the resource value is 0.92. Non-numerical resources (such as sensor abnormal state indicators): Count the frequency of occurrence of a certain state. Through formula Calculate the frequency value. For example, if the infrared sensor's "occlusion anomaly" state occurs 35 times out of a total of 10,000 times, then the frequency value is 0.0035.

[0032] In this embodiment, the hardware-aware neural architecture search technique is used to generate a spatiotemporal feature extraction network adapted to the ARM Cortex-A53 edge processor. The training steps are as follows: Search space settings: Define the search range of the network structure, including convolutional layer type (normal convolution, depthwise separable convolution), convolutional kernel size (3×3, 5×5), number of channels (16, 32, 64), pooling method (max pooling, average pooling), number of Transformer heads (2, 4, 8), and number of Transformer layers (1, 2, 3). Hardware constraints configuration: Input the upper limit of ARM Cortex-A53 computing power (1.2GHz main frequency, 16GB storage), latency constraint (≤100ms), and power consumption constraint (≤5W); Search algorithm implementation: The Progressive Search Pruning (PSP) strategy is adopted, and the search is carried out in three stages: Phase 1: Randomly sample 1000 candidate architectures, filter architectures with latency ≤ 150ms and parameter count ≤ 2M, and retain 200; Phase 2: Validate the inference latency and accuracy of candidate architectures on edge devices, filter out architectures with latency ≤ 100ms and accuracy ≥ 85%, and retain 50; The third stage: reinforcement learning is used to optimize the architecture performance, with "accuracy - 0.1 × latency" as the reward function. Finally, the optimal architecture is searched and obtained, which includes a truncated ResNet50 front end, a dimension reduction network, and two local Transformer layers. Search results verification: The optimal architecture has 1.8M parameters, an inference latency of 86ms, and a power consumption of 4.2W, which meets the requirements for edge deployment.

[0033] Specifically, the truncated ResNet50 is used for spatial feature extraction, and the training steps are as follows: Network architecture configuration: Extract the first four convolutional blocks of ResNet50, remove the fully connected layers and the final pooling layer, and set the parameters for each layer as follows: First layer: 7×7 convolution, 64 filters, stride 2, padding 3, output feature map size is 112×112×64; The second layer consists of three residual blocks, each containing a 1×1 convolution (64 filters), a 3×3 convolution (64 filters), and a 1×1 convolution (256 filters), with an output feature map size of 56×56×256. The third layer consists of four residual blocks, each containing a 1×1 convolution (128 filters), a 3×3 convolution (128 filters), and a 1×1 convolution (512 filters), with an output feature map size of 28×28×512. The fourth layer consists of 6 residual blocks, each containing a 1×1 convolution (256 filters), a 3×3 convolution (256 filters), and a 1×1 convolution (1024 filters), with an output feature map size of 14×14×1024. Pre-training and fine-tuning: The network was initialized with ImageNet pre-trained weights, the first two convolutional blocks were frozen, and the last two convolutional blocks were fine-tuned. The training parameters were: batch size 32, learning rate 0.001, momentum 0.9, weight decay 1e-4, and number of iterations 100 rounds. Training effect verification: On the grid plate surface feature dataset, the spatial features extracted by the fine-tuned network have a similarity of 89.7% with the features of the original ResNet50, and the computational cost is reduced by 68%, which meets the edge computing power requirements.

[0034] In the preferred scheme, the dimensionality reduction network is used for feature channel compression, and the training steps are as follows: Network architecture design: It adopts 3 layers of 1×1 convolutional layers, with 1024 input channels (truncated ResNet50 output), 512 and 256 intermediate channels respectively, and 64 output channels. Batch Normalization and ReLU activation functions are added to each layer. Training data preparation: Spatial feature maps (14×14×1024) of 100,000 grid surface images were selected as training data, and the labels were the category identifiers of the original high-dimensional features (normal, erosion, deformation, loosening). Training parameter configuration: batch size 64, learning rate 0.0005, optimizer Adam, loss function cross-entropy loss, number of iterations 50; Feature Preservation Validation: The feature retention rate (FRR) of the network after training is 89.7% ± 2.3%, which means that the average cosine similarity between the compressed 64-channel features and the original 1024-channel features is 0.897, ensuring that key spatial features are not lost.

[0035] In this embodiment, the local Transformer layer is used for temporal dependency modeling, and the training steps are as follows: Network structure configuration: Set up a 2-layer Transformer encoder, each layer contains 4 attention heads with a head dimension of 16, the feedforward network hidden layer dimension is 256, and the local attention window size is 16 (corresponding to 16 time steps). Time series data processing: The normalized sensor time series data is divided into sequences of length 32 according to time steps. Each sequence contains data of "current time step + previous 31 time steps", generating a total of 500,000 training sequences. Training parameter configuration: batch size 128, learning rate 0.0001, optimizer AdamW, loss function Temporal Prediction Loss (MSE), number of iterations 80, dropout (probability 0.1) is used during training to prevent overfitting; Training effect verification: After training, the network achieved a modeling accuracy of 92.3% for temporal dependencies, effectively capturing the correlation between temporal features such as pressure change rate and infrared occlusion duration.

[0036] In practice, global average pooling is used for feature dimension regularization, and the steps are as follows: Input features: The feature map size output by the local Transformer layer is 32×14×14×64 (time step × height × width × number of channels). Pooling operation: Perform global average pooling on the spatial feature map (14×14×64) at each time step, calculate the mean of each channel, and obtain a 32×64 temporal feature vector; Dimensional normalization: The 64-dimensional feature vector is mapped to 32 dimensions through 1×1 convolution, resulting in a 32×32 multi-dimensional temporal feature set. ,in This is a short-term occlusion counting feature. The characteristic of pressure change rate, This represents the variance characteristic of the weight distribution.

[0037] In the preferred scheme, the DS evidence theory fusion engine is used for uncertainty risk reasoning, and the training steps are as follows: Risk identification framework definition: Define the risk identification framework ,in Because the number of personnel exceeded the limit, For stacking objects The sensor is malfunctioning. For safety; Training data preparation: Select 100,000 labeled multidimensional time-series features. As training data, each data point is labeled with its corresponding real risk category ( ~ ); Basic Probability Assignment (BPA) Training: Construct a BPA transformation model with multidimensional time-series features as input. The output is the BPA value for each proposition. ; The Softmax activation function is used to map the features to a probability distribution. The training parameters are: batch size 64, learning rate 0.001, optimizer SGD, loss function KL divergence loss, and number of iterations 60. The accuracy of the BPA conversion model after training is 91.5%. For example, for a certain feature set... Corresponding to "object stacking" ( The BPA value is 0.85, which is considered "safe". The BPA value is 0.12, indicating "uncertainty". ; Evidence combination rule calibration: Select 50,000 data points containing conflicting evidence to train the conflict coefficient of the evidence combination rules; Using formula Evidence is combined, and the conflict coefficient is optimized through gradient descent to maximize the matching degree between the combined BPA value and the true label. Trust function and likelihood function calculation: The trained model is calculated using the formula... Calculate the trust function using the formula Calculate the likelihood function, in order to As a confidence level output.

[0038] In this embodiment, the lightweight spatiotemporal attention module is used for dynamic weight allocation of multimodal data, and the training steps are as follows: Network structure construction: Construct a dual-branch attention module, with the spatial attention branch processing sensor location information and the temporal attention branch processing temporal data changes; Spatial attention training: The input is a spatial vector (16 dimensions) transformed from the physical location coordinates of the sensor, and a query vector is generated through linear transformation. and key vector Spatial attention weights are calculated using the Softmax activation function. ,in The dimension is vector; the training data consists of 50,000 samples containing location information and true weights, and the training parameters are: batch size 32, learning rate 0.0008, loss function MSE, and number of iterations 40 rounds. Temporal attention training: Input is temporal feature difference (32-dimensional), a non-linear mapping is performed using a multilayer perceptron (MLP), with the MLP structure being 32→64→32; the value range is compressed using a sigmoid activation function to obtain the temporal attention weights. The training data consists of 80,000 samples containing temporal variations and true weights. The training parameters are consistent with spatial attention, and the number of iterations is 50 rounds. Dual-path fusion training: Introducing the AlphaBlender module approach and setting balance coefficients. (Initial value 0.5), using the formula Weight fusion; dynamically adjusted during training This minimizes the error between the merged weights and the actual contributions, ultimately determining... The dynamic adjustment range is 0.3~0.7 (static scene). Dynamic scenes Training effect verification: After training, the weight allocation accuracy of the attention module was 93.1%. For example, when the data of a pressure sensor drifted, its weight automatically dropped from 0.2 to 0.05; when the system was in "emergency maintenance" mode, the weight of the entrance flow sensor dropped from 0.3 to 0.1.

[0039] In the preferred scheme, the LSTM prediction model is used to predict the trend of the load-bearing state of the grating. The training steps are as follows: Network architecture configuration: Input layer dimension is 64 (32-dimensional temporal features + 32-dimensional attention-weighted features); Hidden layers are set to 2 layers, each with 128 units, using the ReLU activation function; Output layer dimension is 1 (predicting the future). The trend value of load-bearing status change within seconds); the network includes input gate, forget gate, cell state, and output gate, and the activation function of the gated unit is Sigmoid; Training data preparation: Select 200,000 load-bearing time series data of grating plates. Each data point contains "historical 30-second data + future 10-second data". Historical data is used as input, and the average rate of change of future data is used as label. Data augmentation: Time flipping and noise injection (Gaussian noise) are employed. The data was expanded, increasing the total number of records to 400,000. Offline training process: Training parameters: batch size 64, learning rate 0.001, optimizer Adam, loss function MSE, number of iterations 100, early stopping strategy (patience=10) is used to prevent overfitting; Training process monitoring: The training set loss decreased from the initial 0.12 to 0.02, the validation set loss decreased from 0.11 to 0.03, and the validation set loss reached its minimum value after 58 iterations, at which point training stopped; Incremental learning mechanism training: Set the incremental update period to 30 seconds (3000 samples), freeze the first hidden layer of the LSTM network, and only update the parameters of the second hidden layer and the output layer; prepare 50,000 new scene data (such as load-bearing data after slight deformation of the grating plate) to simulate online incremental updates: First incremental update: Input 1000 new samples, adjust the learning rate to 0.0001, iterate 10 times, and the model prediction accuracy improves from 92.3% to 93.1%; Fifth incremental update: A total of 5,000 new samples were input, and the model prediction accuracy remained stable at 94.7%, with no "catastrophic forgetting" (the original scene prediction accuracy remained above 91.8%). Gated unit training results: Input gate It can effectively filter important input features, with features having an activation value ≥ 0.8 accounting for 32%; Forgotten Gate It can forget outdated information, with a forgetting weight of 0.75 for historical data from 10 seconds ago; Cell state It can accumulate key timing information with a state update error ≤ 0.01. Output gate and hidden layer output The prediction error RMSE is 0.025.

[0040] In practical implementation, the grey prediction model is used to predict the timing of potential risk outbreaks. The training steps are as follows: Training data preparation: 50,000 damage evolution data of grid plates were selected. Each data point contains a complete time series from "initial damage state to outbreak state", and the outbreak time is marked. ; Model parameter training: using a grey prediction model ,in For development coefficient, The original data sequence is used; the development coefficients are fitted using the least squares method. The training data was divided into a training set (80%) and a validation set (20%); the evolution coefficient after training... The mean value is 0.12, and the average error of the model in predicting the outbreak time is 2.3 hours, which meets the requirements for potential risk early warning; Prediction effect verification: For corrosion damage data of a certain grating plate, the model input initial corrosion degree (0.1mm) predicted the outbreak time to be 72 hours, the actual outbreak time was 75 hours, and the error was 4%.

[0041] In the preferred scheme, the association rule algorithm is used to mine state evolution patterns, and the training steps are as follows: Training data preparation: Select 100,000 historical security status information entries, each entry containing a mapping relationship of "current state → subsequent state"; Similarity calculation training: using the cosine similarity algorithm Calculate state similarity; the training data consists of 50,000 labeled similarity state pairs. The weight parameters for similarity calculation are optimized using gradient descent, achieving a similarity determination accuracy of 90.8%. Lift calculation training: using association rule algorithm Discovering causal relationships, among which This is the current state. For subsequent historical states; during training, the lift thresholds for various state associations are statistically analyzed. When a strong association is identified, the accuracy rate of strong association mining after training is 89.3%.

[0042] In this embodiment, the effects of each activation function are as follows: Softmax activation function: used for BPA transformation and attention weight normalization, normalizing the output probability to the (0,1) interval. For example, after BPA transformation... , , , The sum is 1.0; The Sigmoid activation function is used for calculating temporal attention weights and activating LSTM gate units, compressing the value range to (0,1). For example, the output after calculating temporal attention weights is 0.32, 0.45, etc., and the activation values ​​of the gate units are 0.21 (input gate), 0.78 (forget gate), etc. ReLU activation function: used for non-linear activation of LSTM hidden layers to alleviate the gradient vanishing problem. During training, the mean of the hidden layer output is 0.42, the variance is 0.18, and the proportion of non-zero activation values ​​is 68%.

[0043] Specifically, 1×1 convolutions are used for feature channel compression. In the dimension reduction network, 1×1 convolutions compress 1024 channels to 64 channels, reducing the computational cost by 94% while maintaining a feature retention rate of 89.7%. Cross-attention layers are used for multimodal feature association analysis. Drawing on the idea of ​​text context injection, infrared features are used as query vectors and pressure features are used as key / value vectors. The feature fusion accuracy after association analysis is improved by 5.2%.

[0044] Example 2 The intelligent safety management and control method for boiler walkway grating provided by the present invention will be further optimized and described below. The implementation process of the method described below can be referred to in correspondence with the lightweight spatiotemporal feature extraction network, DS evidence theory fusion engine and incremental LSTM prediction model pre-trained in Example 1 above.

[0045] like Figure 1 , 2 As shown, the intelligent safety management method for boiler walkway grating provided in Embodiment 2 of this application specifically includes the following: In this embodiment, the application scenario is the walkway of boiler No. 3 in a large chemical enterprise. The grating specifications are 8m in length and 1.8m in width per piece, with a maximum designed load-bearing capacity of 1200kg. The environmental parameters of the deployment area are: temperature -20℃ to 80℃, relative humidity 40% to 98%, dust concentration ≤8mg / m³, and interference factors such as equipment vibration (vibration frequency 50Hz to 200Hz) and sudden changes in light intensity (500lux to 1500lux). The system deployment strictly follows industrial safety regulations, and all equipment has passed explosion-proof and high-temperature resistance certifications.

[0046] In this embodiment, step S1 focuses on optimizing environmental interference suppression and multimodal timing alignment accuracy based on data synchronization in embodiment 1, solving the problem of data distortion caused by sensor signals being susceptible to vibration and dust interference in the prior art, and ensuring the consistency and reliability of multi-source data.

[0047] In the preferred scheme, sensor selection and deployment are strictly based on industrial environment adaptation requirements, and parameter settings are based on scenario characteristics and pre-trained model input requirements. The specific configuration is shown in Table 1 below: Table 1 Multimodal Sensor Deployment and Parameter Configuration Table Sensor type Deployment location and method Core parameter settings Adaptation and Optimization Design Infrared beam sensor At the entrance passage, sensors are deployed on both sides of the central axis (1.4m high); a pair of sensors (0.95m high) are placed every 1.8m along the edges of the grid panel, with adjacent sensors overlapping by 0.3m. Sampling period T_1=80ms, emission wavelength 940nm, detection distance adjustable from 0.5-8m, response time ≤3ms, protection rating IP67. The addition of a metal dust cover and vibration damping bracket, along with a hydrophobic anti-fog coating on the lens, reduces the impact of dust adhesion and vibration shift. pressure sensor At the load-bearing beam nodes below the four corners of the grating, two additional sensors (a total of six) are added in the middle, spaced 2m apart, with a bolt fixing torque of 5N·m. Measuring range 0-600kg, accuracy 0.3 grade, sampling period T_2=150ms, output 4-20mA current signal, vibration resistance ≥8g High-temperature compensated resistance strain gauges with built-in temperature self-calibration chips are used to eliminate temperature drift within the range of -20℃ to 80℃. Vibration sensor There are three pressure sensors located at both ends and in the middle of the load-bearing beam of the grating, each corresponding to a different pressure sensor position. Sampling period T_3=50ms, measurement range 0-20g, frequency response 5Hz-500Hz, output ±5V voltage signal A piezoelectric sensor is used, with damping adhesive applied to the mounting surface. The filter bandwidth focuses on the 50Hz-200Hz frequency range of the equipment's vibration. Temperature sensor There are four evenly distributed below the grating and at the top of the entrance passage. Sampling period T_4=1s, measurement range -40℃~1000℃, accuracy ±0.5℃, digital signal output. It adopts a type K thermocouple, and the probe is encapsulated in a high-temperature resistant ceramic sheath, with a response time ≤100ms. Specifically, before deploying the sensors, a site survey is required. A portable vibration tester (model: VM-63A) is used to measure the vibration frequency distribution of the walkway to determine the filtering bandwidth of the vibration sensors. A dust concentration detector (model: LD-5C) is used to record 24-hour dust concentration changes to optimize the dust cover structure of the infrared sensors. All sensors are connected to the edge computing module (model: NVIDIA Jetson Nano 4GB version) via an RS485 bus. The bus uses shielded twisted-pair cable, and wiring should be kept at least 30cm away from power cables to reduce electromagnetic interference.

[0048] In this embodiment, data synchronization adopts a dual mechanism of "hardware triggering + software calibration" to solve the fusion error problem caused by the timing misalignment of multimodal data in the prior art. The specific steps are as follows: Hardware Synchronization Trigger: A GPIO synchronization trigger circuit is integrated into the edge computing module. All sensors obtain a synchronization clock signal through this circuit, with the clock frequency set to 1MHz to ensure that the sampling start time deviation of each sensor is ≤1μs. The trigger signal uses differential signal transmission to enhance anti-interference capability, and the synchronization circuit is powered by an isolated power supply module to avoid the influence of power supply noise.

[0049] Timing calibration algorithm optimization: Based on the linear interpolation synchronization algorithm of Example 1, a dynamic time warping (DTW) calibration step is added to handle the timing offset caused by differences in sensor sampling periods. The specific implementation is as follows: First, extract the timestamp sequence of each modality data. , , ; Using the highest sampling frequency (50ms) of the vibration sensor as a benchmark, the DTW algorithm is used to calculate the optimal matching path between other modal data and the benchmark time series. The formula is as follows: ; in The signal difference between the i-th reference time and the j-th time to be calibrated is calculated using Euclidean distance; Resampling of infrared and pressure signals based on the optimal matching path yields synchronized data with perfectly aligned timing. , , .

[0050] Enhanced data preprocessing: Multi-stage filtering is added to address the high dust and vibration interference in the boiler environment. The infrared signal is filtered using median filtering (window size 5) to remove momentary dust obstruction interference, and then filtered using adaptive threshold filtering to eliminate false obstruction signals caused by vibration. The threshold calculation formula is as follows: ; in The mean signal value is the signal value within a sliding window (window size 100ms). Standard deviation; The pressure signal is filtered using a Kalman filter to suppress vibration noise. The state equation and observation equation are set as follows: ; ; in This represents the true pressure value. The sensor measurement value, , , Process noise The variance was set to 0.01, and the observation noise was... The variance is dynamically adjusted by calculating the signal variance in real time; All preprocessed data are normalized to the [0,1] interval using the formula: ; in and This is the limit value of the sensor's measurement range, ensuring that the data distribution meets the input requirements of the pre-trained model.

[0051] In practice, the optimization effect was verified by comparing the data consistency before and after synchronization. Ten minutes of continuous running data were selected, and the timestamp deviation and signal correlation of each modality were calculated. The results are shown in Table 2 below. Table 2 Comparison of Data Synchronization Results Evaluation indicators Before synchronization After synchronization Optimization and improvement Maximum timestamp deviation 8.3 0.9 89.2% Infrared-pressure signal correlation 0.62 0.87 40.3% Vibration-pressure signal correlation 0.58 0.83 43.1% Signal-to-noise ratio after data preprocessing 38.5 52.7 36.9% As shown in Table 2, the synchronization optimization scheme in this embodiment significantly improves the temporal consistency and signal quality of multi-source data, providing a reliable data foundation for subsequent risk perception steps and effectively solving the perception ambiguity problem caused by data distortion in the prior art.

[0052] In this embodiment, step S2 optimizes the feature extraction targeting and the robustness of uncertainty reasoning based on the lightweight spatiotemporal feature extraction network and DS evidence theory fusion engine pre-trained in embodiment 1, and solves the perceptual ambiguity defect in the prior art that "can only determine whether it exceeds the standard, but cannot explain the reason for exceeding the standard".

[0053] In the preferred embodiment, the network pre-trained in Example 1 is fine-tuned and optimized to enhance its ability to identify damage features such as corrosion, deformation, and loosening, taking into account the multimodal data characteristics of the boiler environment. Network structure fine-tuning: The truncated ResNet50 front end and local Transformer layers are retained, and the channel compression ratio of the dimensionality reduction network is adjusted. The infrared signal feature channels are compressed from 64 to 32, and the pressure and vibration signal feature channels are compressed from 64 to 48. By adding a dedicated convolutional layer for damage features (3×3 convolution, 64 filters), the extraction of damage-related features of the grid plate is enhanced.

[0054] Feature set expansion: In Example 1 Based on this, new damage-related features are added: Vibration frequency characteristics The vibration signal is converted to the frequency domain using Fast Fourier Transform (FFT), and the energy proportion in the 50Hz-200Hz frequency band is extracted using the following formula: ;in The amplitude in the frequency domain of the vibration signal; Pressure distribution uniformity characteristics The formula for calculating the ratio of the standard deviation to the mean of the six pressure sensors is as follows: ;in The standard deviation of the calibrated pressure value. The mean; Temperature gradient characteristics The formula for calculating the ratio of temperature difference at different locations on the grating to distance is: ;in , Temperature values ​​at different locations, The distance between the two points.

[0055] Network fine-tuning training: A transfer learning strategy was adopted, freezing the first 80% of the parameters of the pre-trained network and training only the newly added convolutional and fully connected layers. The training data consisted of 100,000 labeled data points collected on-site (including 7 scenarios: normal working conditions, personnel exceeding limits, object stacking, corrosion damage, deformation hazards, loosening risks, and sensor failure). The training parameters were set as follows: batch size 16, learning rate 0.0003, optimizer AdamW, loss function cross-entropy loss, 50 iterations, and an early stopping strategy (patience=8) was used to prevent overfitting.

[0056] In this embodiment, the Basic Probability Assignment (BPA) transformation and conflict evidence handling mechanism of the DS evidence theory are optimized to improve the accuracy of uncertainty risk perception: Recognition framework extension: In Example 1 Based on, expand to ,in Because the number of personnel exceeded the limit, For stacking objects The sensor is malfunctioning. For safety, For corrosion damage, To prevent potential deformation, To mitigate risks.

[0057] BPA Transformation Optimization: Based on the expanded feature set, a multi-feature fusion BPA transformation model is constructed, and the support of each proposition is calculated using a weighted summation method. : Overstaffing : ( , ); Object stacking : ( , ); Corrosion damage : ( , ); Deformation risks : ( , ); risk of loosening : ( , ); After support normalization, the formula is used. ( Convert to BPA value.

[0058] Conflict Evidence Handling: Introducing a Conflict Coefficient The formula for determining the degree of conflict between pieces of evidence is: ; when In cases of high conflict, a weighted average method is used to correct the BPA value. The formula is as follows: ;in , , The conflict coefficient between single-modal evidence and other evidence; when At that time, the traditional Dempster combination rule was used for fusion.

[0059] In practice, five typical scenarios were selected for testing, and the perception effects of the scheme in Implementation Example 1 and the optimized scheme in this embodiment were compared. The results are shown in Table 3 below: Table 3 Comparison of Uncertainty Risk Perception Effects Test Scenario True risk type Traditional sensor judgment results Optimization scheme determination result in this embodiment Accuracy of judgment Five people entered the grating at the same time Excessive personnel (A1) Personnel exceeded limit (0.82) Personnel exceeded limit (0.91) 100% Tools are stacked in the middle of the grating. Object stacking (A2) Safety (0.65) / Object stacking (0.32) Object stacking (0.89) 100% Loose bolts at the corner of the grating plate Risk of loosening (A7) Safety (0.71) Risk of loosening (0.83) 100% Pressure sensor temperature drift Sensor malfunction (A3) Weight exceeded limit (0.68) Sensor malfunction (0.87) 100% Localized corrosion of grating and personnel passage Corrosion damage (A5) Safety (0.53) Corrosion damage (0.79) and personnel exceeding limits (0.18) 100% As shown in Table 3, the optimized scheme of this embodiment can accurately identify multiple risk types, effectively distinguish between real risks and sensor anomalies, solve the perceptual ambiguity problem of the prior art, and improve the confidence level by more than 15% on average compared with Embodiment 1, providing an accurate risk basis for subsequent dynamic cognitive prediction.

[0060] In this embodiment, step S3 optimizes the attention mechanism and incremental learning strategy based on the incremental LSTM prediction model pre-trained in embodiment 1, thereby solving the defects of rigid decision-making logic and inability to adapt to changes in working conditions in the prior art, and realizing dynamic and flexible judgment.

[0061] In the preferred solution, the dynamic adjustment logic of spatial attention and temporal attention is integrated, and the weight allocation strategy is optimized in combination with the working conditions of industrial scenarios: Spatial attention weight refinement: A position weight matrix is ​​constructed based on the differences in the importance of sensor deployment locations. The corners (pressure sensors 1 and 4) and the middle (pressure sensors 3 and 6) of the grating are critical load-bearing areas, with a spatial attention weight baseline value set to 0.8; the edge areas (pressure sensors 2 and 5) have a weight baseline value set to 0.6. Simultaneously, based on the real-time pressure distribution uniformity characteristics... Dynamically adjust weights: ; in The position reference weight for the i-th sensor is determined when the pressure distribution is uneven ( (Increase), the weight of key areas is further enhanced, and the focus is on high-risk locations.

[0062] Time attention weight optimization: Introducing working condition recognition factor It distinguishes between three operating conditions: normal operation, maintenance, and emergency work, and dynamically adjusts the response sensitivity of time attention. Operating condition identification is based on personnel flow characteristics. Determined by operation instructions (input via host computer): Normal operating conditions ( The formula for calculating time attention weights remains unchanged. ; Maintenance conditions ( The time attention weight for pedestrian flow characteristics is reduced by 30%, and the formula is adjusted as follows: ; Emergency operation conditions ( The time attention weight for the weight change rate feature is increased by 20%, and the formula is adjusted as follows: .

[0063] Dual-path fusion strategy upgrade: Adopting an improved version of the AlphaBlender module, introducing a risk confidence factor. ( (To maximize the risk confidence level output by step S2) The fusion ratio of spatial and temporal weights is dynamically adjusted. ; When the risk confidence level is high ( When the risk confidence level is low, strengthen spatial weighting and focus on the location where the risk occurs; when the risk confidence level is low ( When doing so, strengthen the time weighting to capture potential trend changes.

[0064] In this embodiment, the incremental learning mechanism and prediction output strategy are optimized to improve the model's adaptability to changes in operating conditions and its prediction accuracy. Dynamic adjustment of incremental learning cycle: Abandoning the fixed 30-second update cycle of Example 1, a data drift detection mechanism is used to dynamically set the update cycle. This is achieved by calculating the distribution difference between real-time data and model training data (measured using KL divergence). ;in For real-time data distribution, For the training data distribution. When When there is a significant drift, the update cycle is set to 10 seconds; when When there is a slight drift, the update cycle is set to 30 seconds; when (Without drift) The update cycle is set to 60 seconds.

[0065] Adaptive Prediction Window Setting: Adjusts the prediction time window based on the operating condition type. : Normal operating conditions: Seconds, focusing on short-term load-bearing changes; Maintenance conditions: Seconds, adaptable to scenarios with frequent personnel and equipment movement; Emergency operation conditions: Seconds are all it takes to respond quickly to sudden risks.

[0066] The prediction results have been refined: in addition to outputting the trend of load-bearing status changes, risk level prediction and development rate prediction have been added. Risk levels are based on the ratio of the predicted load-bearing value to a threshold value. Low risk: ; Medium risk: ; High risk: .

[0067] The growth rate is calculated using the slope of the predicted sequence: It provides richer predictive information for operation and maintenance decisions.

[0068] In this embodiment, three typical operating condition change scenarios were selected to verify the model's dynamic adaptability and prediction accuracy. The results are shown in Table 4 below: Table 4 Validation Table of Dynamic Cognitive Prediction Effect

[0069] As shown in Table 4, the optimized incremental LSTM prediction model can dynamically adjust the prediction strategy according to the working conditions, control the prediction error within 2%, and accurately output the risk level and development rate. It solves the problem of rigid decision-making logic in existing technologies and provides strong support for flexible decision-making.

[0070] In this embodiment, step S4, based on the closed-loop decision-making in embodiment 1, introduces a DIK (Data-Information-Knowledge) architecture and a dynamic strategy library to solve the defects of data dormancy and inability to accumulate knowledge in the prior art, and realizes the upgrade from "data collection" to "knowledge reuse".

[0071] In this embodiment, the knowledge base adopts a DIK three-layer architecture to achieve layered storage and association of data, information, and knowledge: Data Layer: Stores preprocessed raw sensor data and synchronization data using a time-series database (InfluxDB). Data retention is 90 days, indexed by "Device ID-Timestamp" for fast querying. Stored fields include: infrared occlusion signal, pressure calibration value, vibration signal, temperature value, and preprocessed data for each modality.

[0072] Information Layer: Stores the risk assessment results from step S2 and the prediction results from step S3, using a relational database (SQLite) with a retention period of 180 days. Stored fields include: risk type, confidence level, occurrence time, occurrence location, predicted load factor, risk level, and development rate.

[0073] Knowledge Layer: Construct a dynamic strategy knowledge base to store patterns and operational rules mined from the information layer, using triples. Structured storage, specifically including: Environmental adaptation rules: such as "when the ambient temperature is >60℃ and the humidity is >90%, the zero-point drift probability of the pressure sensor increases by 30%, and the calibration cycle needs to be shortened to 12 hours"; Risk evolution rules: such as "when a loosening risk (A7) is detected and the development rate is >0.05kg / s, the probability of the risk level rising to high risk within 72 hours is 85%"; Operation and maintenance optimization rules: such as "when the pressure distribution in the middle of the grating plate is uniform". "At that time, priority should be given to inspecting the central support structure"; Dynamic threshold rules: such as "under maintenance conditions, the personnel over-limit threshold can be increased by 20%, and the weight over-limit threshold can be increased by 15%".

[0074] In practice, the closed-loop decision generation process is divided into three stages: information association, knowledge matching, and strategy output, to ensure the accuracy and flexibility of decision-making. Information association phase: The current security status information (risk assessment result + prediction result) is correlated with historical information, and the similarity is calculated using the cosine similarity algorithm. Historical cases with a similarity greater than 0.8 were selected, and their handling experience and operational effectiveness were extracted as a reference for current decision-making.

[0075] Knowledge matching phase: The dynamic strategy library of the knowledge layer is invoked to match the current operating conditions, risk types, prediction results, and rule conditions. A fuzzy matching algorithm is used to handle differences caused by changes in operating conditions. The matching priority is: emergency operation rules > maintenance rules > normal operation rules; high-risk rules > medium-risk rules > low-risk rules.

[0076] Strategy Output Phase: Based on the information association results and knowledge matching results, a closed-loop decision report is generated, which includes three parts: Current status assessment: Identify the risk type, location, confidence level, and risk grade; Potential risk diagnosis: Based on prediction results and evolution rules, analyze the risk development trend and potential secondary risks; Operation and maintenance recommendations: including emergency response measures, priorities, execution time limits, and suggestions for adjusting calibration / maintenance cycles.

[0077] In this embodiment, an automatic knowledge mining and strategy iteration mechanism is established to periodically mine new patterns from information layer data and update the strategy base. Knowledge mining cycle: set at 7 days, using association rule algorithms to mine causal relationships in information layer data and calculate lift. : ; when When a strong association is identified, new knowledge rules are generated.

[0078] Strategy iteration process: Newly generated knowledge rules are verified on-site (the validity of which is confirmed by operations and maintenance personnel) and then entered into the strategy library. At the same time, the validity of the rules in the strategy library is evaluated regularly (every 30 days), rules with an accuracy rate of less than 70% are removed, and the confidence parameters are optimized.

[0079] Taking "Risk of loosening in the middle of the grating (A7) and maintenance conditions" as an example, the closed-loop decision report output is as follows: Current status assessment: There is a risk of loosening in the middle of the grating (sensor 3 location), with a confidence level of 0.83 and a risk level of medium risk; the current load-bearing capacity is 820kg (design threshold 1200kg), and the predicted load-bearing capacity is 850kg in 30 seconds, with a development rate of 0.1kg / s.

[0080] Potential risk diagnosis: Based on knowledge rules, the probability of this risk escalating to high risk within 72 hours under maintenance conditions is 82%, which may lead to deformation of the central support structure and consequently cause uneven overall load-bearing.

[0081] Operation and maintenance recommendations: Priority Level 1 (to be handled within 72 hours); Emergency response measures are to "limit the load in the central area of ​​the grating, with a maximum of 2 people passing through at a time"; Execution time limit is to "complete bolt tightening within 3 days"; Maintenance cycle is adjusted to "shorten the vibration sensor calibration cycle in this area from 7 days to 3 days".

[0082] In this embodiment, the overall performance of the system under harsh industrial conditions was verified through a 30-day field trial run. The key indicators of the existing technical solutions and the solution of this invention were compared, and the results are shown in Table 5 below: Table 5 Overall Performance Comparison Table Evaluation indicators Existing technology 1 (threshold alarm) Existing technology 2 (visual fusion) Examples of this application Increase Risk identification accuracy 72.3 85.6 (Good environment) / 62.1 (High dust) 95.8 32.5% False alarm rate 18.7 12.3 (Good environment) / 25.8 (High dust) 3.2 82.9% Response time 350 800 (Cloud Transmission) 45 87.1% Accuracy of operation and maintenance suggestions - - 91.5 - Data utilization 15.2 38.5 89.7 490.1% Environmental adaptability good Poor excellent - As shown in Table 5, the solution of Embodiment 2 of the present invention is significantly superior to the prior art in terms of core indicators such as risk identification accuracy, false alarm rate, and response time. At the same time, it realizes the generation of operation and maintenance suggestions and efficient data utilization, and completely solves the three major problems of perception ambiguity, decision rigidity, and data dormancy in the prior art. It provides a reliable engineering solution for the intelligent safety management and control of boiler walkway gratings.

[0083] The optimization process of this embodiment strictly follows the environmental characteristics and safety regulations of the industrial site. All technical details are based on the transfer adaptation and scenario-based optimization of the pre-trained model, ensuring that those skilled in the art can reproduce the deployment and operation of the entire system through the above description, thus meeting the patent law's requirement of "sufficient disclosure".

[0084] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0085] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for intelligent safety management and control of boiler walkway grating, characterized in that, Includes the following steps: S1. Real-time acquisition of the occlusion signal sequence generated by the infrared beam sensor deployed at the entrance of the grid plate and the weight signal sequence generated by the pressure sensor deployed below the grid plate, and time-stamp synchronization of the occlusion signal sequence and the weight signal sequence. S2. Input the synchronized occlusion signal sequence and weight signal sequence into a pre-constructed lightweight spatiotemporal feature extraction network to extract multi-dimensional temporal features representing personnel flow, object presence and weight distribution; input the multi-dimensional temporal features into a risk fusion inference engine based on Dempster-Shafer evidence theory; the risk fusion inference engine outputs at least one risk judgment result with confidence rating, including personnel exceeding limits, object stacking and sensor anomalies. S3. Input the risk assessment result with confidence rating and the weight signal sequence into the pre-trained long short-term memory network prediction model; use the attention mechanism integrated in the long short-term memory network prediction model to dynamically allocate the weights of the risk assessment result and the weight signal sequence in the current prediction; execute the forward inference algorithm of the long short-term memory network prediction model to output the prediction of the change trend of the load-bearing state of the grating plate in the future period. S4. Store the current risk assessment result with confidence rating and the predicted trend as security status information in the local knowledge base; call the pre-stored policy rules in the local knowledge base to compare and analyze the security status information with similar information stored in the past; based on the comparison and analysis results, generate and output a closed-loop decision report containing the current status evaluation, potential risk diagnosis and operation and maintenance suggestions.

2. The intelligent safety management and control method for boiler walkway gratings according to claim 1, characterized in that, In step S1, the specific process of acquiring and synchronizing multi-source sensor data includes: S11. Based on the width of the grating inlet channel and the position of the load-bearing beam node, deploy infrared through-beam sensors on both sides of the central axis of the inlet channel, and deploy pressure sensors at the load-bearing beam node below the grating. Set the sampling period, drive each sensor according to the sampling period, and collect the original occlusion signal sequence in real time. and the original weight signal sequence ; The sampling period includes the sampling period of the infrared beam sensor. and pressure sensor sampling period ; S12. Extract the original occlusion signal sequence and the original weight signal sequence The sampling timestamp, through a hardware-triggered synchronization mechanism, is used to generate the original occlusion signal sequence. and the original weight signal sequence Each data sample in the process is appended with a high-precision timestamp uniformly distributed by the master clock of the edge computing module. This yields a timestamped signal sequence. Alignment verification is performed on timestamped signal sequences to identify signal data segments with inconsistent time nodes; a linear interpolation synchronization algorithm is used, employing the formula... Perform signal value completion calculations on inconsistent data segments, where and The times that need to be filled in are respectively The two most recent valid sampling timestamps are used to obtain a synchronized occlusion signal sequence with perfectly aligned timestamps. and synchronous weight signal sequence .

3. The intelligent safety management and control method for boiler walkway gratings according to claim 1, characterized in that, In step S2, the construction and feature extraction process of the lightweight spatiotemporal feature extraction network includes: S21. Obtain the computing power and memory constraint parameters of the ARM processor in the edge computing module; adopt hardware-aware neural architecture search technology to automatically search and evaluate the network structure based on the temporal characteristics of infrared occlusion signals and pressure change curves in a preset search space containing various convolutional kernel sizes and network depths, and obtain the evaluation results; automatically search, generate and determine an ultra-lightweight spatiotemporal feature extraction network architecture containing truncated convolutional layers, dimension reduction networks and local Transformer layers based on the evaluation results. S22, synchronize the synchronization blocking signal sequence and synchronous weight signal sequence Data is spliced ​​together according to the data channel dimension to form a unified network input data; the unified network input data is then input into an ultra-lightweight spatiotemporal feature extraction network. The input data is processed by a truncated convolutional layer to extract local spatial features, resulting in an initial feature map. The initial feature map is input into a dimension reduction network, which performs linear combination and compression transformation on all feature channels through its 1×1 convolutional kernel. Based on the computation... Reduce the original number of feature channels from Compress to This yields a low-dimensional feature map; S23. Slice the low-dimensional feature map along the time dimension to divide it into continuous local temporal segments; input the local temporal segments into the local Transformer layer in sequence, and model the feature dependencies of different time steps within the segment through the self-attention mechanism to obtain high-level features that integrate spatiotemporal context information. A global average pooling layer is used to perform dimensionality reduction and normalization operations on high-level features in the spatiotemporal dimension, and the output includes short-term occlusion count features. Pressure change rate characteristics Weight distribution variance characteristics Multidimensional temporal feature set ; In step S2, the risk fusion reasoning process based on the Dempster-Shafer evidence theory includes: S24. Define a risk identification framework based on the objectives of safety monitoring of boiler walkway grating. ,in Characteristic of personnel exceeding limits, Propositions that characterize the stacking of objects Characterizing sensor anomalies Propositions that characterize the state of security; S25. Multidimensional temporal feature sets For each feature in the recognition framework, calculate its relationship to the recognition framework. Each proposition support Based on formula The support of each feature for each proposition is converted into basic probability assignment values, where The minimum constant is set to prevent the denominator from being zero; thus, the basic probability allocation functions derived from the characteristics of infrared blocking signals are obtained. and the fundamental probability assignment function derived from the weight signal characteristics ; S26. Assign the basic probability function and Evidence fusion was performed using Dempster's combination rules to obtain the fused comprehensive basic probability allocation. This reflects the degree of joint support for each proposition after fusing evidence from both types of sensors; S27. Assigning based on the integrated basic probability after fusion. Calculate each proposition Trust function According to the trust function Determine the confidence level of the corresponding risk proposition and output the risk assessment result with the confidence level.

4. The intelligent safety management and control method for boiler walkway grating plates according to any one of claims 1 to 3, characterized in that, In step S3, the weight allocation process of the attention mechanism includes: S31. Acquire spatial location information and time series information of multi-sensor data streams, and construct a lightweight spatiotemporal attention module based on the spatial location information and time series information. The spatiotemporal attention module includes parallel spatial attention branches and temporal attention branches. S32. Obtain the physical deployment location coordinates of each sensor, encode the physical deployment location coordinates of each sensor into a spatial location vector, perform a linear transformation on the spatial location vector according to the spatial attention branch, and generate a spatial query vector through the linear transformation. and key vector This yields spatial vector pairs; based on the formula Similarity calculation and normalization are performed on spatial vector pairs to obtain the spatial attention weight matrix; where, For vector dimensions; S33. Calculate the multidimensional time series feature set The rate of change of time is obtained by the difference between adjacent time steps. ; the rate of change over time The input to the multilayer perceptron in the temporal attention branch performs a nonlinear mapping to obtain a temporal feature vector; the temporal feature vector is then compressed by the sigmoid activation function to map it to the (0,1) interval, thus obtaining the temporal attention weights. ; S34. Set or adaptively adjust the balance coefficient between spatial attention and temporal attention according to the dynamic characteristics of the current monitoring scene. Based on formula Spatial attention weight matrix And time attention weight Perform a weighted summation operation to generate the final dynamically assigned weights. ;Utilize dynamic weight allocation The features input to the Long Short-Term Memory Network prediction model are weighted.

5. The intelligent safety management and control method for boiler walkway gratings according to claim 1, characterized in that, In step S3, the training and inference process of the Long Short-Term Memory Network prediction model includes: S35. Based on the time-series prediction requirements of the load-bearing state of the grating, initialize the input layer dimension and the number of hidden layer units of the LSTM network. In the output layer dimension, ReLU is used as the activation function to obtain a pre-trained LSTM model; S36. Obtain the newly acquired weight signal sequence within the most recent update cycle. The corresponding actual load-bearing state is then input into the pre-trained LSTM model for forward inference, and the loss function between the pre-trained LSTM model's predicted value and the actual value is calculated. The loss value is obtained; based on the loss value, the network parameters are updated using the gradient descent algorithm, with the update formula being: ,in For learning rate, The gradient of the loss function with respect to the old parameters is used to obtain the incrementally updated LSTM model; S37. The risk assessment results with confidence ratings are compared with the synchronous weight signal sequence. Perform feature concatenation to form the model input vector at the current time step. ; input vector Input the incrementally updated LSTM model, and sequentially perform calculations for the forget gate, input gate, cell state update, and output gate to obtain the current hidden layer output state. .

6. The intelligent safety management and control method for boiler walkway gratings according to claim 1, characterized in that, In step S3, the output process for predicting the trend of load-bearing state changes includes: S38. Based on the actual needs of industrial sites for advance warning, set the length of the future forecast time window. This yields the time window parameters; S39. Output the hidden layer state sequence of the LSTM model. The input is fed into a fully connected layer, and the hidden layer state sequence is transformed through a linear transformation. Mapping to the future The sequence of predicted load-bearing states at each time step ,in Based on the predicted value sequence Calculate from the present moment to the future Average rate of change at time To obtain trend indicators; based on the average rate of change The positive and negative signs and the absolute value of the trend indicator are used to classify the future trend of the load-bearing state, and the output is a prediction of increasing, decreasing or stable trend.

7. The intelligent safety management and control method for boiler walkway gratings according to claim 1, characterized in that, In step S4, the stored procedure for the local knowledge base includes: S41. Obtain the hierarchical definition of data, information, and knowledge, and instantiate a three-layer data-information-knowledge storage architecture based on the hierarchical definition of data, information, and knowledge. The data layer is used to store raw sensor data, the information layer is used to store processed risk and prediction information, and the knowledge layer is used to store operation and maintenance strategies and historical patterns. S42, Synchronization blocking signal sequence and synchronous weight signal sequence Data is categorized and stored in the data layer according to its collection timestamp, and a cyclic overwrite mechanism is configured for the data layer, with a set data retention period. To manage storage capacity; to encapsulate risk assessment results with confidence ratings and trend predictions through the spatiotemporal attributes of event occurrences to form structured security status information, which is then stored in the information layer; and to organize preset operation and maintenance strategy rules and status correlation patterns mined from historical information into a triple structure, which is then stored in the knowledge layer. In step S4, the comparison and correlation analysis process of security status information includes: S43. Extract current security status information from the local knowledge base. Similar security status information stored historically The current security status information With each historical security status information Each feature vector is converted into a standardized feature vector with the same dimension to obtain a comparable vector, and the current security status information is then calculated. With each historical security status information The cosine similarity between them is used to obtain a similarity score; S44. Set a similarity threshold. From all historical security status information Select those with similarity scores higher than By analyzing historical information, a highly correlated historical dataset is obtained; then, an association rule mining algorithm is used to analyze the current state characteristics of the highly correlated historical dataset. and subsequent state evolution The strength of the causal relationship between them, and the degree of correlation are increased by the degree of correlation. The measurement and calculation formula is as follows: By analyzing the causal relationship between the current state and subsequent historical states, the laws governing state evolution are obtained, among which... Represents probability. This is the current state. This is the state that follows history.

8. The intelligent safety management and control method for boiler walkway gratings according to claim 1, characterized in that, In step S4, the process of generating the closed-loop decision report includes: S45. Obtain the confidence rating of the risk assessment result and the highest similarity score obtained from the similarity calculation, set the status evaluation threshold, perform level classification on the current safety status of the grating, and output the status evaluation of safe, low risk, medium risk or high risk. S46. Input the trend prediction and state evolution law into the grey prediction model, and then apply the formula... By performing risk outbreak time prediction, the expected occurrence time of potential risks is obtained, whereby... For development coefficient, The original data sequence is used as the basis for analysis. Combining the risk type and state evolution pattern in the risk assessment results, the causes leading to the current risk state are analyzed, and the risk diagnosis results are output. S47. Call the triplet of operation and maintenance strategy rules stored in the knowledge layer of the local knowledge base, match the risk type, diagnosis result and strategy rule execution to obtain a preliminary operation and maintenance plan; adjust the operation and maintenance priority according to the risk level in the current status assessment, and generate operation and maintenance operation suggestions for targeted operations; integrate the current status assessment, potential risk diagnosis results and operation and maintenance operation suggestions to form a closed-loop decision report and output it.

9. A smart safety management and control system for boiler walkway grating, characterized in that, The system comprises a multimodal sensor module, an edge computing module, a local storage module, and a hierarchical early warning module, which are sequentially connected in communication. The system is used to execute the intelligent safety management method for boiler walkway gratings as described in any one of claims 1 to 8, wherein: The multimodal sensor module is used to acquire multi-source sensing data related to the grid panel and to achieve timestamp synchronization; The edge computing module is used to perform uncertainty risk perception, dynamic cognition and prediction, and closed-loop decision generation on synchronized sensor data; The local storage module is used to store security status information, historical data, and a dynamic policy knowledge base; The tiered early warning module is used to output targeted early warning signals and operation and maintenance suggestions based on the closed-loop decision results.

10. The intelligent safety management and control system for boiler walkway gratings according to claim 9, characterized in that, The multimodal sensor module includes an infrared beam sensor submodule, a pressure sensor submodule, a vibration sensor submodule, a vision sensor submodule, and a temperature sensor submodule, wherein: The infrared beam sensor submodule includes at least two pairs of infrared beam sensors deployed in the grid panel entrance channel and on both sides, configured to perform deployment and signal acquisition, and output a sequence of occlusion signals; The pressure sensor submodule includes at least four resistance strain gauge pressure sensors deployed at the load-bearing beam nodes below the grating, configured to perform deployment and data acquisition, and output a sequence of weight signals; The vibration sensor submodule is deployed at the load-bearing structure of the grating plate to collect dynamic mechanical vibration signals of the grating plate; The visual sensor submodule is deployed on the bracket above the grid plate and integrates a dust compensation hardware unit for acquiring visual images of the grid plate surface. The temperature sensor submodule is deployed in the environment surrounding the grating to collect temperature data in the industrial field; Each submodule communicates with the edge computing module through an industrial Ethernet interface and achieves timestamp alignment through a hardware-triggered synchronization mechanism.

Images