Multi-dimensional dynamic monitoring and early warning system and method for hazardous environments of explosion-proof units

Abstract

This invention discloses a multi-dimensional dynamic monitoring and early warning system and method for hazardous environments in explosion-proof units. It collects node detection data and historical data in real time, generates node feature vectors, and completes initialization. In each early warning cycle, the node feature vectors and periodic feedback feature vectors are fused to form a periodic feature vector, which is input into an improved MobileNetV3 regression model, outputting fitting parameters for the evolution trend of hazardous deviations. Based on the fitting parameters, a family of multi-scale risk deviation curves is generated and fused into a comprehensive risk deviation curve, dynamically adjusting the sliding windows at each scale. Multiple evaluation and prediction paths are scheduled according to the node state differences, and early warning status is output. After the cycle ends, the path response contribution, resource consumption, and fitting error are statistically analyzed to construct a path response evaluation tensor, calculate path update weights, hierarchically optimize model weights, and generate a periodic feedback feature vector to feed back into the next cycle, achieving real-time global risk prediction and adaptive closed-loop optimization of explosion-proof units in hazardous environments.

Description

Technical Field

[0001] This invention relates to the field of industrial safety technology, specifically to a multi-dimensional dynamic monitoring and early warning system and method for hazardous environments of explosion-proof units. Background Technology

[0002] During the dynamic evolution of hazardous behaviors in explosion-proof units, monitoring nodes experience performance degradation and limited resource redundancy (inability to add more nodes or equipment), making it difficult for the system to ensure long-term stable operation through traditional physical redundancy methods. Simultaneously, the randomness and complexity of environmental disturbances gradually render traditional prediction and feedback mechanisms based on single models ineffective. Existing hazard warning systems typically rely on state assessment models based on node performance weight adjustments and a single prediction model feedback correction mechanism. When system resources are sufficient and environmental disturbances are relatively stable, a certain degree of warning deviation correction and system stability assurance can be achieved through continuous optimization of weight allocation and model parameter adjustments. However, the following technical shortcomings remain:

[0003] 1. The monitoring data has limited dimensions and lacks integration capabilities, making it difficult to identify chain-like risk transmission processes across nodes. Risk judgment is generally based on fixed periods and static thresholds, and the analysis window or weight allocation is not dynamically adjusted according to changes in the operating status, resulting in insufficient accuracy and timeliness of risk assessment, and easy to cause delayed warnings or false alarms.

[0004] 2. The model structure and parameter configuration are usually static and fixed, mostly staying at the stage of data collection and simple alarm. They lack the dynamic closed-loop management capability to drive model optimization and strategy adjustment based on prediction results. They are unable to cope with multi-dimensional risk changes in complex environments and cannot dynamically adjust the model structure and parameters in combination with path contribution, resource consumption level and prediction error, resulting in a decline in prediction accuracy in long-term operation.

[0005] In view of this, the present invention provides a multi-dimensional dynamic monitoring and early warning system and method for hazardous environments of explosion-proof units, thereby solving the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-dimensional dynamic monitoring and early warning system and method for explosion-proof units in hazardous environments, which solves the problem that multi-dimensional monitoring data of explosion-proof units in hazardous environments cannot be integrated and analyzed in real time, dynamically predicted for risks, and optimized for adaptive early warning strategies.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] The first aspect is a multi-dimensional dynamic monitoring and early warning method for hazardous environments in explosion-proof units, including the following steps:

[0009] Real-time collection of detection data and historical data from each monitoring node; analysis of operating status and resource consumption levels; generation of node feature vectors; and completion of initial configuration.

[0010] At the beginning of each warning period, the node feature vector is fused with the periodic feedback feature vector generated by the model optimization module in the previous warning period to form a periodic feature vector, which is then input into the improved MobileNetV3 regression model and outputs the fitting parameters of the risk deviation evolution trend.

[0011] A family of multi-scale risk deviation curves is generated based on the fitted parameters and merged into a comprehensive risk deviation curve. The sliding window of each scale is dynamically adjusted according to the rate of change of the comprehensive risk deviation curve.

[0012] Based on the performance status characteristics and the comprehensive risk deviation curve, the node status difference is extracted. Based on the node status difference, the path scheduling priority of multiple evaluation and prediction paths is generated. The warning status of each path is output for the current warning period.

[0013] After the cycle ends, the contribution of each path response, resource consumption level and fitting error are statistically analyzed, a path response evaluation tensor is constructed, the path update weights are calculated, the weights of the fully connected layer and bottleneck layer of the improved MobileNetV3 regression model are optimized hierarchically, and the periodic feedback feature vector for the next early warning cycle is extracted.

[0014] As a preferred embodiment of the first aspect of the present invention, the improved MobileNetV3 regression model includes:

[0015] Based on the MobileNetV3 network architecture, convolutional layers and bottleneck layers are retained. The convolutional layers are used for temporal feature extraction, and the bottleneck layers are used for resource state modeling. A nested fusion mechanism of residual connections and graph convolution is introduced.

[0016] The output layer has a regression structure and uses L1 loss optimization to fit the deviation evolution trend. The output is a fitting parameter that characterizes the evolution trend of the system's dangerous deviation. The fitting parameter includes the deviation trend slope, deviation fluctuation amplitude, and deviation convergence parameter.

[0017] As a preferred embodiment of the first aspect of the present invention, the generation logic of the risk deviation curve family is as follows:

[0018] A family of nested risk deviation curves with multiple scales is constructed based on a preset time scale. The time scale is constructed using a sliding window and is divided into a short-term scale to capture instantaneous fluctuations, a medium-term scale to reflect trend changes, and a long-term scale to accumulate risk.

[0019] Deviation trend time series are constructed on short-term, medium-term and long-term scales, and corresponding risk deviation curves are generated at each scale. The risk deviation curves are accumulated according to the time sequence of the early warning cycle to generate deviation trend time series. Based on the deviation trend time series, the fitting parameter vector of each cycle is mapped to the same time axis to realize the connection of risk trends and the recording of change trajectories between consecutive cycles.

[0020] Based on the aforementioned deviation trend time series, a risk deviation curve is generated using a trend fitting function to characterize the changing trend of dangerous status within the current warning period. The horizontal axis of the risk deviation curve represents time, and the vertical axis represents the risk deviation value. The curve shape can simultaneously reflect short-term fluctuation characteristics and long-term evolution trends.

[0021] The risk deviation curves corresponding to each time scale are aligned along the time axis and combined in a nested manner to form a multi-scale nested risk deviation curve family.

[0022] As a preferred embodiment of the first aspect of the present invention, the fusion logic of the comprehensive risk deviation curve is as follows:

[0023] Calculate the short-term volatility change, medium-term trend slope change rate, and long-term baseline drift of the risk deviation curve at short-term, medium-term, and long-term scales, respectively.

[0024] The magnitude of change is obtained by calculating the difference between the changes in short-term volatility intensity, the rate of change in medium-term trend slope, and the amount of long-term baseline drift and the corresponding changes in short-term volatility intensity, the rate of change in medium-term trend slope, and the amount of long-term baseline drift in the previous warning period.

[0025] Based on preset risk assessment weights, the magnitude of changes at each scale is fused to generate a comprehensive risk deviation curve. When the short-term fluctuation intensity change in the comprehensive risk deviation curve exceeds the first threshold, the short-term sliding window is shortened and the sliding step size is reduced. When the long-term baseline drift is lower than the second threshold and the rate of change of the medium-term trend slope tends to stabilize, the long-term sliding window is extended and the sliding step size is increased.

[0026] As a preferred embodiment of the first aspect of the present invention, the logic for extracting the node state difference is as follows:

[0027] Arrange the performance status characteristics of the current early warning period into a performance characteristic matrix according to the monitoring node number;

[0028] Map the slope of the deviation trend, volatility, and convergence parameters of the comprehensive risk deviation curve in the current period to the deviation feature matrix by node;

[0029] The difference vector of the node is obtained by performing a difference operation on the corresponding dimensions of the performance feature matrix and the deviation feature matrix for the same monitoring node;

[0030] The node difference vectors of all monitoring nodes are compared with the safety baseline state to obtain the baseline deviation measure. The baseline deviation measures are stacked row by row to form a baseline deviation matrix. The norm value of the baseline deviation measure of each node in the baseline deviation matrix is ​​marked as the node state difference quantity.

[0031] As a preferred embodiment of the first aspect of the present invention, the output logic of the warning status of the current warning period is as follows:

[0032] Based on the topological mapping relationship between the path and the monitoring nodes, the average difference of the nodes covered by each path is calculated; the average difference is used as the path risk impact factor, and combined with the historical performance fluctuation index of the path, the path correlation score is calculated.

[0033] Paths are sorted from highest to lowest based on their path relevance scores. Scheduling priority values ​​are then assigned to the sorted paths in sequence, and a priority list is generated.

[0034] The risk threshold conditions for the corresponding paths are activated in order of priority; the risk warning status output by each path is collected, and the warning status output for the current warning period is synthesized by a comprehensive voting method.

[0035] As a preferred embodiment of the first aspect of the present invention, the logic for obtaining the path response evaluation tensor is as follows:

[0036] For each path during the current cycle, the weighted voting ratio of the warning status output for the corresponding path in the current warning cycle is calculated, and the weighted voting ratio is used as the response contribution.

[0037] Record the CPU usage, memory usage, and execution time of the path during the runtime cycle, and synthesize the resource consumption rate after normalization according to a unified unit.

[0038] Calculate the mean squared error between the predicted output of the path in the current period and the actual risk status, and use it as a fitting error index.

[0039] The response contribution, resource consumption rate, and fitting error of each path are arranged in order of path number; the three types of indicators are stacked into a three-dimensional tensor along the path dimension, with the dimensions corresponding to the path number, indicator type, and indicator value, respectively; min-max normalization is performed on the three-dimensional tensor along the indicator type dimension to obtain the path response evaluation tensor.

[0040] As a preferred embodiment of the first aspect of the present invention, the driving logic for performing hierarchical optimization of the weights of the fully connected layer and bottleneck layer of the improved MobileNetV3 regression model is as follows:

[0041] Based on the path combination parameters, read the resource consumption limit and parallel scheduling limit of each path, and calculate the resource gating factor of the path;

[0042] The importance score of the path is obtained by weighting the path's response contribution, fitting error, and node state difference according to a preset ratio.

[0043] The importance scores of all paths are sorted numerically, and a set of paths allowed to participate in model updates is selected in combination with parallel scheduling constraints. The scores of the paths in the set are multiplied by the corresponding resource gating factor and normalized to obtain the path update weights.

[0044] The update coefficients of the fully connected layer are obtained by summing the path update weights along the path dimension. The path update weights are then allocated to the corresponding channels according to the mapping relationship between the path and the bottleneck layer channels and summed to obtain the update coefficients of each channel of the bottleneck layer. The update coefficients are then used to scale the weight update magnitude of the corresponding layer to perform hierarchical optimization.

[0045] As a preferred embodiment of the first aspect of the present invention, the logic for obtaining the periodic feedback feature vector is as follows:

[0046] For each detection node in the set of paths allowed to participate in model updates, calculate the residual between the model prediction and the actual observation of the detection node in the current warning period; perform a weighted average according to the distribution of the detection nodes in the set of paths, where the weight is the product of the path update weight and the node state difference, to obtain the path residual value of each path.

[0047] All path residual values ​​are combined in order of path number to form residual feedback factors.

[0048] The short-term scale portion is extracted from the comprehensive risk deviation curve, and the rate of change between the current period and the previous warning period is calculated. The rate of change sequence is smoothed and filtered to remove high-frequency noise, and the short-term trend disturbance value of each node is obtained.

[0049] The path mean of the node trend disturbance values ​​is calculated according to the path mapping relationship, and combined into a trend disturbance factor;

[0050] The residual feedback factor and the trend disturbance factor are concatenated according to the corresponding positions of the path numbers to obtain the periodic feedback feature vector as the next early warning period.

[0051] Secondly, the present invention provides a multi-dimensional dynamic monitoring and early warning system for hazardous environments of explosion-proof units, based on the implementation of the first aspect, including a data preprocessing module, a fitting parameter generation module, a risk situation analysis module, an early warning status output module, and a model optimization module, with each module connected by wired and / or wireless means.

[0052] The data preprocessing module is used to collect the node detection data and historical node detection data of each monitoring node in real time after the explosion-proof unit is equipped with monitoring nodes. Based on the node detection data and historical node detection data, the module performs operation status and resource consumption analysis to generate node feature vectors and completes the initial configuration of each monitoring node.

[0053] The fitting parameter generation module merges the node feature vector with the periodic feedback feature vector generated by the model optimization module in the previous warning period at the beginning of each warning period to form a periodic feature vector, which is then input into the improved MobileNetV3 regression model and outputs fitting parameters of the risk deviation evolution trend.

[0054] The risk situation analysis module generates a family of multi-scale risk deviation curves based on the fitted parameters and merges them into a comprehensive risk deviation curve. It dynamically adjusts the sliding window of each scale according to the rate of change of the comprehensive risk deviation curve.

[0055] The early warning status output module extracts the node status difference based on performance status characteristics and comprehensive risk deviation curve, generates path scheduling priorities for multiple evaluation and prediction paths based on the node status difference, and schedules each path to output the early warning status for the current early warning period.

[0056] The model optimization module calculates the contribution of each path response, resource consumption level, and fitting error after the cycle ends, constructs a path response evaluation tensor, calculates path update weights, performs hierarchical optimization on the weights of the fully connected layer and bottleneck layer of the improved MobileNetV3 regression model, and extracts residual feedback factors and trend disturbance factors to generate the periodic feedback feature vector for the next early warning cycle, which is then fed back into the fitting parameter generation module.

[0057] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0058] This invention collects real-time data on the operating status and resource consumption of monitoring nodes and performs multi-dimensional feature fusion to ensure the integrity and diversity of input information. Then, it utilizes the advantages of the improved MobileNetV3 regression model in multi-scale fusion and long-range dependency modeling to achieve high-precision fitting of the evolution trend of risk deviation, thereby obtaining a family of multi-scale risk deviation curves for short-term, medium-term and long-term risks, which are then fused into a comprehensive risk deviation curve to reflect the overall situation changes.

[0059] The sliding window at each time scale is dynamically adjusted based on the rate of change of the composite curve, enabling the system to respond sensitively to sudden events while maintaining stable tracking of long-term trends; path scheduling priorities are generated using node state differences to ensure that critical paths are executed first when resources are limited, thereby improving the timeliness of early warnings and the efficiency of resource utilization.

[0060] Furthermore, by constructing a path response evaluation tensor and calculating path update weights, the model's fully connected layer and bottleneck layer are optimized in a hierarchical manner. The residual feedback factor and trend disturbance factor are then fed back into the next cycle to achieve cross-cycle adaptive optimization, thus maintaining the accuracy of monitoring, the timeliness of early warning, and the stability of system operation in complex and dangerous environments over a long period of time. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0062] Figure 1 This is a structural block diagram of the explosion-proof unit hazardous environment multi-dimensional dynamic monitoring and early warning system of the present invention;

[0063] Figure 2 This is a schematic diagram illustrating the application logic of the improved MobileNetV3 regression model of this invention.

[0064] Figure 3 This is a schematic diagram of the multi-scale risk deviation curve family and the comprehensive risk deviation curve of the present invention;

[0065] Figure 4 This is a schematic diagram illustrating the change of the multi-scale dynamic weights over time in this invention;

[0066] Figure 5 This is a schematic diagram of the multi-dimensional dynamic monitoring and early warning method for hazardous environments of explosion-proof units according to the present invention. Detailed Implementation

[0067] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art. The drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0068] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more exemplary embodiments. Numerous specific details are provided in the following description to give a full understanding of exemplary embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more specific details omitted, or other methods, components, steps, etc., can be employed. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0069] Example 1

[0070] like Figure 1 As shown, this embodiment provides a multi-dimensional dynamic monitoring and early warning system for hazardous environments of explosion-proof units, including a data preprocessing module, a fitting parameter generation module, a risk situation analysis module, an early warning status output module, and a model optimization module. The modules are connected to each other via wired and / or wireless means.

[0071] The data preprocessing module is used to collect the node detection data and historical node detection data of each monitoring node in real time after the explosion-proof unit is equipped with monitoring nodes. Based on the node detection data and historical node detection data, the module performs operation status and resource consumption analysis to generate node feature vectors and completes the initial configuration of each monitoring node.

[0072] It should be noted that the explosion-proof unit is equipped with multiple monitoring nodes, including temperature, pressure, and vibration nodes. Each node collects real-time detection data and historical detection data; among which:

[0073] Node detection data includes real-time collected operational status data and node resource redundancy data; among which:

[0074] Operational status data includes performance indicators such as temperature, pressure, current, voltage, and vibration amplitude. Based on the operational status data, the performance status characteristics of each monitoring node over the past multiple warning periods are extracted to characterize the node's operational trend and performance fluctuation characteristics.

[0075] Node resource redundancy data includes the resource consumption rate and remaining capacity of the currently monitored nodes during the early warning period. The resource consumption rate and remaining capacity are used as resource status characteristics to input into the system.

[0076] Specifically, the logic for obtaining the node feature vector is as follows:

[0077] The node feature vector includes performance status features and resource status features;

[0078] The sampling data from multiple monitoring nodes are synchronized according to a unified timestamp to eliminate feature misalignment caused by sampling delays at different nodes; interpolation or imputation strategies are used to process missing data to ensure the integrity of the time series.

[0079] Based on historical reference intervals and sliding window statistics, outliers caused by single-point mutations, over-range sampling, and sensor malfunctions are identified; outliers are removed or replaced (such as replacement with the nearest mean or trend extrapolation replacement) to prevent them from affecting trend modeling and threshold determination.

[0080] Standard normalization is applied to various performance indicators to map different physical quantities to a comparable scale; resource consumption rate and remaining capacity characteristics are normalized to form resource status characteristics, maintaining the consistency of numerical distribution of cross-node characteristics.

[0081] Calculate the statistical indicators corresponding to the running status data within a sliding window; the statistical indicators include the performance status mean, performance status standard deviation, fluctuation range, trend slope and anomaly frequency. Combine the statistical results with the real-time collected values ​​to form performance status features. The performance status features also include adding a unique node identifier and feature category label to each monitoring node and its features, which facilitates global information propagation and differential modeling in a multi-node environment.

[0082] To further explain, the initialization settings include configuring the safety baseline status, path combination parameters, and risk threshold conditions for each monitoring node; among which:

[0083] The safety baseline state is dynamically constructed using a sliding window statistical method, by extracting the closest values ​​for each monitoring node. The performance status characteristics of each historical early warning period are used to calculate the mean and standard deviation of the performance status, forming a reference baseline interval: ;

[0084] in: Indicates the first The average performance status of each monitoring node within the sliding window. Indicates the first The standard deviation of the performance status of each monitoring node within the sliding window This is the sensitivity adjustment coefficient. Indicates the first The reference range of each monitoring node within the sliding window This represents the maximum value within the reference range. This represents the minimum value within the reference range;

[0085] The path combination parameters include initial path weight, upper limit of resource consumption, and parallel scheduling limit, which are used to guide path selection and resource scheduling during the cold start phase of the model.

[0086] ;

[0087] in: For path numbering, Indicates the total number of paths. Indicates the first The initial selection weights of each path (set during the model cold start phase) represent the weights of each path among all available paths. The initial probability or priority of being selected; This indicates that the sum of the initial weights of all paths is 1, forming a path selection probability distribution; Indicates the first The current actual resource consumption of this path measures the system resources used during path evaluation or prediction. Depending on the definition of system resources, this could be CPU %, memory (MB), bandwidth (Mbps), execution time, etc. Indicates the first The maximum resource consumption allowed for each path is used to prevent a path from consuming too many resources due to complex logic. In other words, the current resource consumption of each path cannot exceed its maximum allowed consumption limit. This indicates the maximum number of paths that can be executed in parallel simultaneously. It controls the number of paths that are started at the same time in each warning cycle to prevent system overload.

[0088] The risk threshold conditions include risk warning states determined based on path correlation scoring. These risk warning states include normal operation, fluctuating operation, and abnormal operation, and will be dynamically optimized in conjunction with a feedback mechanism. The determination logic is as follows:

[0089] When the path correlation score is less than the minimum threshold for judging abnormal fluctuations, the current risk warning status is judged as normal operation status.

[0090] When the path correlation score is between the minimum and maximum values ​​of the fluctuation anomaly judgment threshold, the current risk warning status is determined to be a fluctuation operation status.

[0091] When the path correlation score is greater than the maximum value of the fluctuation anomaly judgment threshold, the current risk warning status is judged as an abnormal operation status;

[0092] Among them, the minimum and maximum values ​​of the fluctuation anomaly detection threshold can be adaptively adjusted through a feedback mechanism.

[0093] The fitting parameter generation module is used to fuse the node feature vectors of each monitoring node and the periodic feedback feature vector of the previous warning cycle at the beginning of each warning cycle to form a periodic feature vector, and input it into the improved MobileNetV3 regression model to output fitting parameters that characterize the evolution trend of the system's risk deviation.

[0094] Specifically, the periodic feedback feature vector includes a residual feedback factor and a trend disturbance factor. The residual feedback factor is used to quantify the difference between the predicted value and the actual observed value of each monitoring node in the previous warning period. In the case of a multidimensional feature vector, the residual feedback factor adopts the mean square error between the predicted value and the actual observed value in the previous warning period. The trend disturbance factor is used to characterize the trend change magnitude of each monitoring node in adjacent warning periods, and is defined as the difference between the smoothed trend function in adjacent periods.

[0095] The periodic feature vector is composed of the performance and resource status characteristics of each monitoring node in the current warning period, as well as the residual feedback factor and trend disturbance factor from the previous warning period, concatenated in a preset order. In other words:

[0096] For the node numbered The performance status characteristics and resource status characteristics of the monitoring nodes are concatenated by channel to obtain the node feature vector, which is: ;in: There are a total of monitoring nodes indivual, It is a positive integer. This indicates the node number corresponding to the monitoring node, the first... The performance status characteristics of each monitoring node are as follows: and resource status characteristics ; Periodic feedback feature vector of the previous early warning period ;

[0097] ;

[0098] ;

[0099] ;

[0100] The feedback feature vector from the previous warning period Node feature vectors concatenated in the current early warning period The ends of the nodes form the node extension vector. The node extension vectors of all monitoring nodes are concatenated sequentially according to the node number, and then flattened row by row into a periodic feature vector of one dimension. ;

[0101] .

[0102] It should also be noted that the improved MobileNetV3 classification network model is used to achieve global joint modeling of node performance status and resource redundancy characteristics, as well as fitting of the evolution trend of danger deviation. For example... Figure 2 As shown, the improved MobileNetV3 regression model includes:

[0103] Using the MobileNetV3 network architecture as the basic framework, convolutional layers and bottleneck layers are retained. By sliding the convolutional layers in the time dimension, feature extraction of node feature vectors of each monitoring node is achieved, capturing local transient fluctuations and long-term degradation trends. In the bottleneck layer, performance status features and resource status features are concatenated in a preset channel order, and multi-scale fusion expression is achieved through linear bottleneck compression.

[0104] The nested fusion mechanism introduces channel-level attention weights into the fused feature tensor to enhance the expression of the coupling relationship between performance state features and resource state features, highlighting the dynamic impact of resource changes on performance degradation.

[0105] By introducing node feature vectors and periodic feedback feature vectors into the residual path through different levels, the long-range transmission of node feature vectors and periodic feedback feature vectors in the network is maintained, realizing cross-node information propagation and global feature modeling, and improving the model's ability to perceive the chain evolution trend of systemic risks.

[0106] The original fully connected classification layer of MobileNetV3 is replaced with a regression output layer, which outputs fitting parameters characterizing the evolution trend of the system's risk deviation. These fitting parameters include the deviation trend slope, deviation fluctuation amplitude, and deviation convergence parameters; wherein:

[0107] Deviation trend slope: measures the direction and speed of change in system risk indicators;

[0108] Deviation fluctuation amplitude: quantifies the intensity of short-term fluctuations and is used to reflect operational stability;

[0109] Deviation convergence parameter: assesses whether risk changes tend to stabilize or continue to worsen.

[0110] By introducing the L1 loss function, the model's ability to fit subtle performance changes and outlier samples is enhanced. The L1 loss function is used to optimize the regression output, suppress the excessive interference of outliers on the model's prediction results, and improve the accuracy and stability of the deviation trend fitting.

[0111] It should be noted that the improved MobileNetV3 classification network model integrates feature extraction and fusion coding mechanisms to enhance the model's ability to express complex performance-resource coupling relationships. Global joint modeling not only improves the accuracy of identifying local node deviations but also enables dynamic prediction of the chain evolution trend of systemic deviations, providing highly reliable model support for subsequent risk deviation curve generation and early warning status determination. The improved MobileNetV3 regression model can achieve global joint modeling of the performance status and resource redundancy characteristics of multiple nodes in explosion-proof units while ensuring low computational resource consumption. It outputs highly accurate and robust prediction results of dangerous deviation evolution trends, effectively supporting dynamic early warning and resource adaptive optimization decisions of explosion-proof systems.

[0112] The risk situation analysis module generates a family of nested risk deviation curves at multiple scales based on fitted parameters and integrates the magnitude of change to generate a comprehensive risk deviation curve. It then dynamically adjusts the sliding windows for short-term, medium-term, and long-term scales based on the comprehensive risk deviation curve.

[0113] Specifically, the generation logic of the family of risk deviation curves is as follows:

[0114] A family of nested risk deviation curves with multiple scales is constructed based on a preset time scale. The time scale is constructed using a sliding window and is divided into a short-term scale to capture instantaneous fluctuations, a medium-term scale to reflect trend changes, and a long-term scale to accumulate risk.

[0115] We construct time series of deviation trends on short-term, medium-term, and long-term scales, and generate corresponding risk deviation curves at each scale.

[0116] More specifically, the logic for obtaining the risk deviation curve is as follows:

[0117] Accumulate data according to the time sequence of the early warning cycle to construct a deviation trend time series that reflects changes in system risk. Based on the deviation trend time series, map the fitted parameter vector of each cycle onto the same time axis to realize the connection of risk trends and record the trajectory of changes between consecutive cycles.

[0118] Based on the aforementioned deviation trend time series, a risk deviation curve representing the changing trend of dangerous status within the current warning period is generated using a trend fitting function (such as polynomial fitting, exponential smoothing, or spline interpolation). The horizontal axis of the risk deviation curve represents time, and the vertical axis represents the risk deviation value. The curve shape can simultaneously reflect short-term fluctuation characteristics and long-term evolution trends.

[0119] The risk deviation curves corresponding to each time scale are aligned along the time axis and combined in a nested manner to form a multi-scale nested risk deviation curve family.

[0120] like Figure 3 As shown, during the risk situation analysis process, short-term, medium-term, and long-term risk deviation curves are generated based on the fitted parameters, each corresponding to a different time window length. Specifically, the short-term risk deviation curve reflects transient fluctuations and sudden anomalies within a short period; the medium-term risk deviation curve describes trend changes over a medium time span, considering both fluctuations and trends; and the long-term risk deviation curve depicts the risk baseline changes over a long period, reflecting the health evolution trend of the equipment. A comprehensive risk deviation curve is obtained by weighting and fusing the three curves according to risk assessment weights. This comprehensive curve serves as the overall output of the risk situation analysis, driving the decision logic of the early warning status output module and the model optimization module.

[0121] The fusion logic of the comprehensive risk deviation curve is as follows:

[0122] Calculate the short-term volatility change, medium-term trend slope change rate, and long-term baseline drift of the risk deviation curve at short-term, medium-term, and long-term scales, respectively.

[0123] The magnitude of change is obtained by calculating the difference between the changes in short-term volatility intensity, the rate of change in medium-term trend slope, and the amount of long-term baseline drift and the corresponding changes in short-term volatility intensity, the rate of change in medium-term trend slope, and the amount of long-term baseline drift in the previous warning period.

[0124] Based on preset risk assessment weights, the magnitude of changes at each scale is fused to generate a comprehensive risk deviation curve. When the short-term fluctuation intensity change in the comprehensive risk deviation curve exceeds the first threshold, the short-term sliding window is shortened and the sliding step size is reduced. When the long-term baseline drift is lower than the second threshold and the rate of change of the medium-term trend slope tends to stabilize, the long-term sliding window is extended and the sliding step size is increased.

[0125] For example, the comprehensive risk deviation curve for:

[0126] ;

[0127] in: , and These are the changes in short-term volatility intensity, the rate of change in medium-term trend slope, and the long-term baseline drift during the current warning period, respectively. , and These are the changes in short-term volatility intensity, the rate of change in medium-term trend slope, and the long-term baseline drift in the previous warning period, respectively. , and These are the risk assessment weights for short-term, medium-term, and long-term timeframes, respectively. .

[0128] like Figure 4 As shown, the risk assessment weights of the short-term, medium-term, and long-term risk deviation curves change over time. The rate of change and trend characteristics of the combined risk deviation curves are used to trigger adaptive adjustments to the sliding window length: when short-term fluctuations exceed a threshold, the short-term window is shortened to improve sensitivity to sudden changes; when the medium-term curve changes tend to stabilize, the medium-term window is extended to enhance trend smoothness; when the long-term baseline shifts rapidly, the long-term window is shortened to accelerate the response to long-term trend changes; and when short-term fluctuations remain below a threshold for an extended period, the short-term window is extended to avoid overreacting to minor fluctuations.

[0129] The risk assessment weights are determined by a combination of system operating status, risk trend characteristics, and resource allocation strategies, and are updated in each early warning cycle. They are used to allocate the importance of each time scale in the calculation of the comprehensive curve, thereby realizing closed-loop control of multi-scale risk perception and adaptive weight adjustment.

[0130] The early warning status output module extracts the node status difference based on performance status characteristics and comprehensive risk deviation curve, generates path scheduling priorities for multiple evaluation and prediction paths based on the node status difference, and schedules each path to output the early warning status of the current early warning period.

[0131] Specifically, the logic for extracting the node state difference is as follows:

[0132] Arrange the performance status characteristics of the current early warning period into a performance characteristic matrix according to the monitoring node number;

[0133] Map the slope of the deviation trend, volatility, and convergence parameters of the comprehensive risk deviation curve in the current period to the deviation feature matrix by node;

[0134] The difference vector of the node is obtained by performing a difference operation on the corresponding dimensions of the performance feature matrix and the deviation feature matrix for the same monitoring node;

[0135] The node difference vectors of all monitoring nodes are compared with the safety baseline state to obtain the baseline deviation measure. The baseline deviation measures are stacked row by row to form a baseline deviation matrix. The norm value of the baseline deviation measure of each node in the baseline deviation matrix is ​​marked as the node state difference quantity.

[0136] To further explain, the output logic for the warning status of the current warning period is as follows:

[0137] Based on the topological mapping relationship between the path and the monitoring nodes, the average difference of the nodes covered by each path is calculated; the average difference is used as the path risk impact factor, and combined with the historical performance fluctuation index of the path, the path correlation score is calculated.

[0138] Sort the paths by path relevance score from high to low, assign scheduling priority values ​​(e.g., 1 is the highest priority) to the sorted paths, and generate a priority list.

[0139] The risk threshold conditions for the corresponding paths are activated in order of priority; the risk warning status output by each path is collected, and the warning status output for the current warning period is synthesized by a comprehensive voting method.

[0140] The model optimization module calculates the response contribution, resource consumption rate, and fitting error of each path after the warning period ends, constructs a path response evaluation tensor, and performs hierarchical optimization on the weights of the fully connected layer and bottleneck layer of the improved MobileNetV3 regression model based on the path response evaluation tensor. At the same time, it extracts the residual feedback factor and trend disturbance factor as feature fusion inputs for the next warning period.

[0141] Specifically, the logic for obtaining the path response evaluation tensor is as follows:

[0142] For each path during the current cycle, the weighted voting ratio of the warning status output for the corresponding path in the current warning cycle is calculated, and the weighted voting ratio is used as the response contribution.

[0143] Record the CPU usage, memory usage, and execution time of the path during the runtime cycle, and synthesize the resource consumption rate after normalization according to a unified unit.

[0144] Calculate the mean squared error between the predicted output of the path in the current period and the actual risk status, and use it as a fitting error index.

[0145] The response contribution, resource consumption rate, and fitting error of each path are arranged in order of path number; the three types of indicators are stacked into a three-dimensional tensor along the path dimension, with the dimensions corresponding to the path number, indicator type, and indicator value, respectively; min-max normalization is performed on the three-dimensional tensor along the indicator type dimension to obtain the path response evaluation tensor.

[0146] To further explain, the driving logic for performing hierarchical optimization of the weights of the fully connected layers and bottleneck layers in the improved MobileNetV3 regression model is as follows:

[0147] Based on the path combination parameters, read the resource consumption limit and parallel scheduling limit of each path, and calculate the resource gating factor of the path;

[0148] The importance score of the path is obtained by weighting the path's response contribution, fitting error, and node state difference according to a preset ratio.

[0149] The importance scores of all paths are sorted numerically, and a set of paths allowed to participate in model updates is selected in combination with parallel scheduling constraints. The scores of the paths in the set are multiplied by the corresponding resource gating factor and normalized to obtain the path update weights.

[0150] The update coefficients of the fully connected layer are obtained by summing the path update weights along the path dimension. The path update weights are then allocated to the corresponding channels according to the mapping relationship between the path and the bottleneck layer channels and summed to obtain the update coefficients of each channel of the bottleneck layer. The update coefficients are then used to scale the weight update magnitude of the corresponding layer to perform hierarchical optimization.

[0151] To further explain, the logic for obtaining the periodic feedback feature vector is as follows:

[0152] For each detection node in the set of paths allowed to participate in model updates, calculate the residual between the model prediction and the actual observation of the detection node in the current warning period; perform a weighted average according to the distribution of the detection nodes in the set of paths, where the weight is the product of the path update weight and the node state difference, to obtain the path residual value of each path.

[0153] All path residual values ​​are combined in order of path number to form residual feedback factors.

[0154] The short-term scale portion is extracted from the comprehensive risk deviation curve, and the rate of change between the current period and the previous warning period is calculated. The rate of change sequence is smoothed and filtered to remove high-frequency noise, and the short-term trend disturbance value of each node is obtained.

[0155] The path mean of the node trend disturbance values ​​is calculated according to the path mapping relationship, and combined into a trend disturbance factor;

[0156] The residual feedback factor and the trend disturbance factor are concatenated according to the corresponding positions of the path numbers to obtain the periodic feedback feature vector as the next early warning period.

[0157] Example 2

[0158] like Figure 5 As shown, the parts not detailed in this embodiment are as described in Embodiment 1. This embodiment provides a method for multi-dimensional dynamic monitoring and early warning of hazardous environments in explosion-proof units, including the following steps:

[0159] Real-time collection of detection data and historical data from each monitoring node; analysis of operating status and resource consumption levels; generation of node feature vectors; and completion of initial configuration.

[0160] At the beginning of each warning period, the node feature vector is fused with the periodic feedback feature vector generated by the model optimization module in the previous warning period to form a periodic feature vector, which is then input into the improved MobileNetV3 regression model and outputs the fitting parameters of the risk deviation evolution trend.

[0161] A family of multi-scale risk deviation curves is generated based on the fitted parameters and merged into a comprehensive risk deviation curve. The sliding window of each scale is dynamically adjusted according to the rate of change of the comprehensive risk deviation curve.

[0162] Based on the performance status characteristics and the comprehensive risk deviation curve, the node status difference is extracted. Based on the node status difference, the path scheduling priority of multiple evaluation and prediction paths is generated. The warning status of each path is output for the current warning period.

[0163] After the cycle ends, the contribution of each path response, resource consumption level and fitting error are statistically analyzed, a path response evaluation tensor is constructed, the path update weights are calculated, the weights of the fully connected layer and bottleneck layer of the improved MobileNetV3 regression model are optimized hierarchically, and the periodic feedback feature vector for the next early warning cycle is extracted.

[0164] The improved MobileNetV3 regression model includes:

[0165] Based on the MobileNetV3 network architecture, convolutional layers and bottleneck layers are retained. The convolutional layers are used for temporal feature extraction, and the bottleneck layers are used for resource state modeling. A nested fusion mechanism of residual connections and graph convolution is introduced.

[0166] The output layer has a regression structure and uses L1 loss optimization to fit the deviation evolution trend. The output is a fitting parameter that characterizes the evolution trend of the system's dangerous deviation. The fitting parameter includes the deviation trend slope, deviation fluctuation amplitude, and deviation convergence parameter.

[0167] The generation logic of the risk deviation curve family is as follows:

[0168] A family of nested risk deviation curves with multiple scales is constructed based on a preset time scale. The time scale is constructed using a sliding window and is divided into a short-term scale to capture instantaneous fluctuations, a medium-term scale to reflect trend changes, and a long-term scale to accumulate risk.

[0169] Deviation trend time series are constructed on short-term, medium-term and long-term scales, and corresponding risk deviation curves are generated at each scale. The risk deviation curves are accumulated according to the time sequence of the early warning cycle to generate deviation trend time series. Based on the deviation trend time series, the fitting parameter vector of each cycle is mapped to the same time axis to realize the connection of risk trends and the recording of change trajectories between consecutive cycles.

[0170] Based on the aforementioned deviation trend time series, a risk deviation curve is generated using a trend fitting function to characterize the changing trend of dangerous status within the current warning period. The horizontal axis of the risk deviation curve represents time, and the vertical axis represents the risk deviation value. The curve shape can simultaneously reflect short-term fluctuation characteristics and long-term evolution trends.

[0171] The risk deviation curves corresponding to each time scale are aligned along the time axis and combined in a nested manner to form a multi-scale nested risk deviation curve family.

[0172] The fusion logic of the comprehensive risk deviation curve is as follows:

[0173] Calculate the short-term volatility change, medium-term trend slope change rate, and long-term baseline drift of the risk deviation curve at short-term, medium-term, and long-term scales, respectively.

[0174] The magnitude of change is obtained by calculating the difference between the changes in short-term volatility intensity, the rate of change in medium-term trend slope, and the amount of long-term baseline drift and the corresponding changes in short-term volatility intensity, the rate of change in medium-term trend slope, and the amount of long-term baseline drift in the previous warning period.

[0175] Based on preset risk assessment weights, the magnitude of changes at each scale is fused to generate a comprehensive risk deviation curve. When the short-term fluctuation intensity change in the comprehensive risk deviation curve exceeds the first threshold, the short-term sliding window is shortened and the sliding step size is reduced. When the long-term baseline drift is lower than the second threshold and the rate of change of the medium-term trend slope tends to stabilize, the long-term sliding window is extended and the sliding step size is increased.

[0176] The logic for extracting the node state difference is as follows:

[0177] Arrange the performance status characteristics of the current early warning period into a performance characteristic matrix according to the monitoring node number;

[0178] Map the slope of the deviation trend, volatility, and convergence parameters of the comprehensive risk deviation curve in the current period to the deviation feature matrix by node;

[0179] The difference vector of the node is obtained by performing a difference operation on the corresponding dimensions of the performance feature matrix and the deviation feature matrix for the same monitoring node;

[0180] The node difference vectors of all monitoring nodes are compared with the safety baseline state to obtain the baseline deviation measure. The baseline deviation measures are stacked row by row to form a baseline deviation matrix. The norm value of the baseline deviation measure of each node in the baseline deviation matrix is ​​marked as the node state difference quantity.

[0181] The output logic for the warning status of the current warning period is as follows:

[0182] Based on the topological mapping relationship between the path and the monitoring nodes, the average difference of the nodes covered by each path is calculated; the average difference is used as the path risk impact factor, and combined with the historical performance fluctuation index of the path, the path correlation score is calculated.

[0183] Paths are sorted from highest to lowest based on their path relevance scores. Scheduling priority values ​​are then assigned to the sorted paths in sequence, and a priority list is generated.

[0184] The risk threshold conditions for the corresponding paths are activated in order of priority; the risk warning status output by each path is collected, and the warning status output for the current warning period is synthesized by a comprehensive voting method.

[0185] The logic for obtaining the path response evaluation tensor is as follows:

[0186] For each path during the current cycle, the weighted voting ratio of the warning status output for the corresponding path in the current warning cycle is calculated, and the weighted voting ratio is used as the response contribution.

[0187] Record the CPU usage, memory usage, and execution time of the path during the runtime cycle, and synthesize the resource consumption rate after normalization according to a unified unit.

[0188] Calculate the mean squared error between the predicted output of the path in the current period and the actual risk status, and use it as a fitting error index.

[0189] The response contribution, resource consumption rate, and fitting error of each path are arranged in order of path number; the three types of indicators are stacked into a three-dimensional tensor along the path dimension, with the dimensions corresponding to the path number, indicator type, and indicator value, respectively; min-max normalization is performed on the three-dimensional tensor along the indicator type dimension to obtain the path response evaluation tensor.

[0190] The driving logic for performing hierarchical optimization of the weights of the fully connected layers and bottleneck layers in the improved MobileNetV3 regression model is as follows:

[0191] Based on the path combination parameters, read the resource consumption limit and parallel scheduling limit of each path, and calculate the resource gating factor of the path;

[0192] The importance score of the path is obtained by weighting the path's response contribution, fitting error, and node state difference according to a preset ratio.

[0193] The importance scores of all paths are sorted numerically, and a set of paths allowed to participate in model updates is selected in combination with parallel scheduling constraints. The scores of the paths in the set are multiplied by the corresponding resource gating factor and normalized to obtain the path update weights.

[0194] The update coefficients of the fully connected layer are obtained by summing the path update weights along the path dimension. The path update weights are then allocated to the corresponding channels according to the mapping relationship between the path and the bottleneck layer channels and summed to obtain the update coefficients of each channel of the bottleneck layer. The update coefficients are then used to scale the weight update magnitude of the corresponding layer to perform hierarchical optimization.

[0195] The logic for obtaining the periodic feedback feature vector is as follows:

[0196] For each detection node in the set of paths allowed to participate in model updates, calculate the residual between the model prediction and the actual observation of the detection node in the current warning period; perform a weighted average according to the distribution of the detection nodes in the set of paths, where the weight is the product of the path update weight and the node state difference, to obtain the path residual value of each path.

[0197] All path residual values ​​are combined in order of path number to form residual feedback factors.

[0198] The short-term scale portion is extracted from the comprehensive risk deviation curve, and the rate of change between the current period and the previous warning period is calculated. The rate of change sequence is smoothed and filtered to remove high-frequency noise, and the short-term trend disturbance value of each node is obtained.

[0199] The path mean of the node trend disturbance values ​​is calculated according to the path mapping relationship, and combined into a trend disturbance factor;

[0200] The residual feedback factor and the trend disturbance factor are concatenated according to the corresponding positions of the path numbers to obtain the periodic feedback feature vector as the next early warning period.

[0201] The multi-dimensional dynamic monitoring and early warning method for hazardous environments of explosion-proof units provided in this embodiment is used to execute the multi-dimensional dynamic monitoring and early warning system for hazardous environments of explosion-proof units provided in the above embodiments of the present invention. For details of the specific methods and processes for realizing the corresponding functions of each structure included in the multi-dimensional dynamic monitoring and early warning method for hazardous environments of explosion-proof units, please refer to the embodiments of the multi-dimensional dynamic monitoring and early warning system for hazardous environments of explosion-proof units mentioned above, which will not be repeated here.

[0202] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for multi-dimensional dynamic monitoring and early warning of hazardous environments in explosion-proof units, characterized in that, Includes the following steps; Real-time collection of detection data and historical data from each monitoring node; analysis of operating status and resource consumption levels; generation of node feature vectors; and completion of initial configuration. At the beginning of each warning period, the node feature vector is fused with the periodic feedback feature vector generated by the model optimization module in the previous warning period to form a periodic feature vector, which is then input into the improved MobileNetV3 regression model and outputs the fitting parameters of the risk deviation evolution trend. A family of multi-scale risk deviation curves is generated based on the fitted parameters and merged into a comprehensive risk deviation curve. The sliding window of each scale is dynamically adjusted according to the rate of change of the comprehensive risk deviation curve. Based on the performance status characteristics and the comprehensive risk deviation curve, the node status difference is extracted. Based on the node status difference, the path scheduling priority of multiple evaluation and prediction paths is generated. The warning status of each path is output for the current warning period. After the cycle ends, the contribution of each path response, resource consumption level and fitting error are statistically analyzed, a path response evaluation tensor is constructed, the path update weights are calculated, the weights of the fully connected layer and bottleneck layer of the improved MobileNetV3 regression model are optimized hierarchically, and the periodic feedback feature vector for the next early warning cycle is extracted.

2. The method for multi-dimensional dynamic monitoring and early warning of hazardous environments in explosion-proof units according to claim 1, characterized in that, The improved MobileNetV3 regression model includes: Based on the MobileNetV3 network architecture, convolutional layers and bottleneck layers are retained. The convolutional layers are used for temporal feature extraction, and the bottleneck layers are used for resource state modeling. A nested fusion mechanism of residual connections and graph convolution is introduced. The output layer has a regression structure and uses L1 loss optimization to fit the deviation evolution trend. The output is a fitting parameter that characterizes the evolution trend of the system's dangerous deviation. The fitting parameter includes the deviation trend slope, deviation fluctuation amplitude, and deviation convergence parameter.

3. The method for multi-dimensional dynamic monitoring and early warning of hazardous environments in explosion-proof units according to claim 2, characterized in that, The generation logic of the risk deviation curve family is as follows: A family of nested risk deviation curves with multiple scales is constructed based on a preset time scale. The time scale is constructed using a sliding window and is divided into a short-term scale to capture instantaneous fluctuations, a medium-term scale to reflect trend changes, and a long-term scale to accumulate risk. Deviation trend time series are constructed on short-term, medium-term and long-term scales, and corresponding risk deviation curves are generated at each scale. The risk deviation curves are accumulated according to the time sequence of the early warning cycle to generate deviation trend time series. Based on the deviation trend time series, the fitting parameter vector of each cycle is mapped to the same time axis to realize the connection of risk trends and the recording of change trajectories between consecutive cycles. Based on the aforementioned deviation trend time series, a risk deviation curve is generated using a trend fitting function to characterize the changing trend of dangerous status within the current warning period. The horizontal axis of the risk deviation curve represents time, and the vertical axis represents the risk deviation value. The curve shape can simultaneously reflect short-term fluctuation characteristics and long-term evolution trends. The risk deviation curves corresponding to each time scale are aligned along the time axis and combined in a nested manner to form a multi-scale nested risk deviation curve family.

4. The method for multi-dimensional dynamic monitoring and early warning of hazardous environments in explosion-proof units according to claim 3, characterized in that, The fusion logic of the comprehensive risk deviation curve is as follows: Calculate the short-term volatility change, medium-term trend slope change rate, and long-term baseline drift of the risk deviation curve at short-term, medium-term, and long-term scales, respectively. The magnitude of change is obtained by calculating the difference between the changes in short-term volatility intensity, the rate of change in medium-term trend slope, and the amount of long-term baseline drift and the corresponding changes in short-term volatility intensity, the rate of change in medium-term trend slope, and the amount of long-term baseline drift in the previous warning period. Based on preset risk assessment weights, the magnitude of changes at each scale is fused to generate a comprehensive risk deviation curve. When the short-term fluctuation intensity change in the comprehensive risk deviation curve exceeds the first threshold, the short-term sliding window is shortened and the sliding step size is reduced. When the long-term baseline drift is lower than the second threshold and the rate of change of the medium-term trend slope tends to stabilize, the long-term sliding window is extended and the sliding step size is increased.

5. The method for multi-dimensional dynamic monitoring and early warning of hazardous environments in explosion-proof units according to claim 4, characterized in that, The logic for extracting the node state difference is as follows: Arrange the performance status characteristics of the current early warning period into a performance characteristic matrix according to the monitoring node number; Map the slope of the deviation trend, volatility, and convergence parameters of the comprehensive risk deviation curve in the current period to the deviation feature matrix by node; The difference vector of the node is obtained by performing a difference operation on the corresponding dimensions of the performance feature matrix and the deviation feature matrix for the same monitoring node; The node difference vectors of all monitoring nodes are compared with the safety baseline state to obtain the baseline deviation measure. The baseline deviation measures are stacked row by row to form a baseline deviation matrix. The norm value of the baseline deviation measure of each node in the baseline deviation matrix is ​​marked as the node state difference quantity.

6. The method for multi-dimensional dynamic monitoring and early warning of hazardous environments in explosion-proof units according to claim 5, characterized in that, The output logic for the warning status of the current warning period is as follows: Based on the topological mapping relationship between the path and the monitoring nodes, the average difference of the nodes covered by each path is calculated; the average difference is used as the path risk impact factor, and combined with the historical performance fluctuation index of the path, the path correlation score is calculated. Paths are sorted from highest to lowest based on their path relevance scores. Scheduling priority values ​​are then assigned to the sorted paths in sequence, and a priority list is generated. Activate the risk threshold conditions for the corresponding paths in priority order; Collect the risk warning status output from each path, and synthesize the warning status output for the current warning period using a comprehensive voting method.

7. The method for multi-dimensional dynamic monitoring and early warning of hazardous environments in explosion-proof units according to claim 6, characterized in that, The logic for obtaining the path response evaluation tensor is as follows: For each path during the current cycle, the weighted voting ratio of the warning status output for the corresponding path in the current warning cycle is calculated, and the weighted voting ratio is used as the response contribution. Record the CPU usage, memory usage, and execution time of the path during the runtime cycle, and synthesize the resource consumption rate after normalization according to a unified unit. Calculate the mean squared error between the predicted output of the path in the current period and the actual risk status, and use it as a fitting error index. The response contribution, resource consumption rate, and fitting error of each path are arranged in order of path number; the three types of indicators are stacked into a three-dimensional tensor along the path dimension, with the dimensions corresponding to the path number, indicator type, and indicator value, respectively; min-max normalization is performed on the three-dimensional tensor along the indicator type dimension to obtain the path response evaluation tensor.

8. The method for multi-dimensional dynamic monitoring and early warning of hazardous environments in explosion-proof units according to claim 7, characterized in that, The driving logic for performing hierarchical optimization of the weights of the fully connected layers and bottleneck layers in the improved MobileNetV3 regression model is as follows: Based on the path combination parameters, read the resource consumption limit and parallel scheduling limit of each path, and calculate the resource gating factor of the path; The importance score of the path is obtained by weighting the path's response contribution, fitting error, and node state difference according to a preset ratio. The importance scores of all paths are sorted numerically, and a set of paths allowed to participate in model updates is selected in combination with parallel scheduling constraints. The scores of the paths in the set are multiplied by the corresponding resource gating factor and normalized to obtain the path update weights. The update coefficients of the fully connected layer are obtained by summing the path update weights along the path dimension. The path update weights are then allocated to the corresponding channels according to the mapping relationship between the path and the bottleneck layer channels and summed to obtain the update coefficients of each channel of the bottleneck layer. The update coefficients are then used to scale the weight update magnitude of the corresponding layer to perform hierarchical optimization.

9. The method for multi-dimensional dynamic monitoring and early warning of hazardous environments in explosion-proof units according to claim 8, characterized in that, The logic for obtaining the periodic feedback feature vector is as follows: For each detection node in the set of paths allowed to participate in model updates, calculate the residual between the model prediction and the actual observation of the detection node in the current warning period; The path residual value of each path is obtained by weighting the distribution of the detected nodes in the path set, where the weight is the product of the path update weight and the node state difference. All path residual values ​​are combined into residual feedback factors in order of path number. The short-term scale portion is extracted from the comprehensive risk deviation curve, and the rate of change between the current period and the previous warning period is calculated. The rate of change sequence is smoothed and filtered to remove high-frequency noise, and the short-term trend disturbance value of each node is obtained. The path mean of the node trend disturbance values ​​is calculated according to the path mapping relationship, and combined into a trend disturbance factor; The residual feedback factor and the trend disturbance factor are concatenated according to the corresponding positions of the path numbers to obtain the periodic feedback feature vector as the next early warning period.

10. A multi-dimensional dynamic monitoring and early warning system for hazardous environments of explosion-proof units, based on the implementation of any one of the multi-dimensional dynamic monitoring and early warning methods for hazardous environments of explosion-proof units according to claims 1-9, characterized in that, It includes a data preprocessing module, a fitting parameter generation module, a risk situation analysis module, an early warning status output module, and a model optimization module, with each module connected via wired and / or wireless means. The data preprocessing module is used to collect the node detection data and historical node detection data of each monitoring node in real time after the explosion-proof unit is equipped with monitoring nodes. Based on the node detection data and historical node detection data, the module performs operation status and resource consumption analysis to generate node feature vectors and completes the initial configuration of each monitoring node. The fitting parameter generation module merges the node feature vector with the periodic feedback feature vector generated by the model optimization module in the previous warning period at the beginning of each warning period to form a periodic feature vector, which is then input into the improved MobileNetV3 regression model and outputs fitting parameters of the risk deviation evolution trend. The risk situation analysis module generates a family of multi-scale risk deviation curves based on the fitted parameters and merges them into a comprehensive risk deviation curve. It dynamically adjusts the sliding window of each scale according to the rate of change of the comprehensive risk deviation curve. The early warning status output module extracts the node status difference based on performance status characteristics and comprehensive risk deviation curve, generates path scheduling priorities for multiple evaluation and prediction paths based on the node status difference, and schedules each path to output the early warning status for the current early warning period. The model optimization module calculates the contribution of each path response, resource consumption level, and fitting error after the cycle ends, constructs a path response evaluation tensor, calculates path update weights, performs hierarchical optimization on the weights of the fully connected layer and bottleneck layer of the improved MobileNetV3 regression model, and extracts residual feedback factors and trend disturbance factors to generate the periodic feedback feature vector for the next early warning cycle, which is then fed back into the fitting parameter generation module.

Images