Multi-dimensional dynamic monitoring and early warning system and method for dangerous environment of explosion-proof unit

Abstract

The invention discloses a multi-dimensional dynamic monitoring and early warning system and method for a dangerous environment of an explosion-proof unit, and the method comprises the steps: collecting node detection data and historical data in real time, generating a node feature vector, and completing the initialization; in each early warning period, fusing the node feature vector and the periodic feedback feature vector to form a periodic feature vector, inputting the periodic feature vector into an improved MobileNetV3 regression model, and outputting a danger deviation evolution trend fitting parameter; generating a multi-scale risk deviation curve family based on the fitting parameters, fusing the multi-scale risk deviation curve family into a comprehensive risk deviation curve, and dynamically adjusting each scale sliding window; scheduling a plurality of evaluation and prediction paths according to the node state difference quantity to output an early warning state; and after the period is ended, counting a path response contribution degree, resource consumption and a fitting error, constructing a path response evaluation tensor, calculating a path update weight and a hierarchical optimization model weight, and generating a period feedback feature vector to be recharged to the next period, so that global risk real-time prediction and adaptive closed-loop optimization of the explosion-proof unit in a dangerous environment are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial safety, in particular to a multi-dimensional dynamic monitoring and early warning system and method for an explosion-proof unit in a dangerous environment. BACKGROUND

[0002] In the dynamic evolution process of dangerous behaviors of the explosion-proof unit, the monitoring nodes are subject to performance degradation, and the resource redundancy is limited (more nodes or devices cannot be added), making it difficult for the system to ensure long-term stable operation through traditional physical redundancy. At the same time, the randomness and complexity of environmental disturbances also make the traditional prediction and feedback mechanism based on a single model gradually ineffective. The existing dangerous early warning system usually relies on a state evaluation model based on node performance weight adjustment and a single prediction model feedback correction mechanism. When the system resources are sufficient and the environmental disturbances are relatively stable, a certain degree of early warning deviation correction and system stability can be achieved by continuously optimizing the weight distribution and adjusting the model parameters. However, there are still the following technical defects: 1. The monitoring data is limited in dimension and lacks fusion capability, making it difficult to identify the chain risk transmission process across nodes. Generally, risk judgment is based on fixed periods and static thresholds, and the analysis window or weight distribution cannot be dynamically adjusted according to the changes in the running state, resulting in insufficient accuracy and timeliness of risk assessment, and prone to early warning lag or false alarms; 2. The model structure and parameter configuration are usually static and fixed, and mostly stay in the data collection and simple alarm stage, lacking dynamic closed-loop management capability based on model optimization and strategy adjustment driven by prediction results, making it difficult to cope with multi-dimensional risk changes in complex environments, and unable to dynamically adjust the model structure and parameters in combination with the path contribution, resource consumption level and prediction error, resulting in a decrease in prediction accuracy in the long run.

[0003] In view of this, the present application provides a multi-dimensional dynamic monitoring and early warning system and method for an explosion-proof unit in a dangerous environment to solve the above problems. SUMMARY

[0004] The purpose of the present application is to provide a multi-dimensional dynamic monitoring and early warning system and method for an explosion-proof unit in a dangerous environment to solve the problem that multi-dimensional monitoring data of the explosion-proof unit in a dangerous environment cannot be analyzed in real time, dynamically predicted and adaptively optimized in terms of early warning strategy.

[0005] In order to achieve the above purpose, the present application provides the following technical solutions: In the first aspect, a multi-dimensional dynamic monitoring and early warning method for an explosion-proof unit in a dangerous environment includes the following steps: Real-time collection of detection data and historical data of each monitoring node, analysis of running state and resource consumption level, generation of node feature vector and completion of initialization configuration; At the beginning of each early warning period, the node feature vector is fused with the period feedback feature vector generated by the model optimization module in the last early warning period to form a period feature vector, which is input into the improved MobileNetV3 regression model to output the fitting parameters of the dangerous deviation evolution trend; Based on the fitting parameters, a multi-scale risk deviation curve family is generated and fused into a comprehensive risk deviation curve, and the sliding window of each scale is dynamically adjusted according to the change rate of the comprehensive risk deviation curve; Based on the performance state feature and the comprehensive risk deviation curve, the node state difference is extracted, the path scheduling priority of multiple evaluation and prediction paths is generated based on the node state difference, and the early warning state of the current early warning period is output by scheduling each path. After the period ends, the response contribution degree, resource consumption level and fitting error of each path are counted, the path response evaluation tensor is constructed, the path update weight is calculated, the weights of the full connection layer and the bottleneck layer of the improved MobileNetV3 regression model are optimized in layers, and the period feedback feature vector for the next early warning period is extracted.

[0006] As a preferred technical solution of the first aspect of the application, the improved MobileNetV3 regression model comprises: The MobileNetV3 network architecture is used as the basic framework, the convolution layer and the bottleneck layer are retained, the convolution layer is used for time series feature extraction, the bottleneck layer is used for resource state modeling, and the residual connection and the nested fusion mechanism of graph convolution are introduced; The output layer is a regression structure, L1 loss optimization is used to fit the deviation evolution trend, and fitting parameters representing the system dangerous deviation evolution trend are output, the fitting parameters include deviation trend slope, deviation fluctuation amplitude and deviation convergence parameter.

[0007] As a preferred technical solution of the first aspect of the application, the generation logic of the risk deviation curve family is: Based on the preset time scale, a multi-scale nested risk deviation curve family is constructed, wherein: the time scale is constructed by using a sliding window, and is divided into a short-term scale for capturing instantaneous fluctuations, a medium-term scale for reflecting trend changes, and a long-term scale for accumulating risks; Deviation trend time series are constructed on the short-term scale, the medium-term scale and the long-term scale, and corresponding risk deviation curves are generated under each scale; the risk deviation curves accumulate to generate the deviation trend time series in the time sequence of the early warning period, and the fitting parameter vectors of each period are mapped to the same time axis based on the deviation trend time series, so as to realize the risk trend connection and change trajectory recording between continuous periods. On the basis of the deviation trend time series, a risk deviation curve representing a change trend of a dangerous state in a current early warning period is generated by using a trend fitting function, a curve horizontal axis of the risk deviation curve is time, a curve vertical axis is a risk deviation value, and a curve shape can reflect both short-term fluctuation characteristics and long-term evolution trends.

[0008] The risk deviation curves corresponding to respective time scales are aligned according to a time axis and combined in a nested manner to form a multi-scale nested risk deviation curve family.

[0009] As a preferred technical solution of the first aspect of the present application, the fusion logic of the comprehensive risk deviation curve is: Short-term fluctuation intensity changes, medium-term trend slope change rates and long-term baseline drift amounts of the risk deviation curves are respectively calculated under short-term, medium-term and long-term scales; The short-term fluctuation intensity changes, the medium-term trend slope change rates and the long-term baseline drift amounts are differentially calculated with corresponding short-term fluctuation intensity changes, medium-term trend slope change rates and long-term baseline drift amounts of a previous early warning period to obtain change amplitudes; The change amplitudes of respective scales are fused based on preset risk assessment weights to generate a comprehensive risk deviation curve; when the short-term fluctuation intensity change of the comprehensive risk deviation curve exceeds a first threshold value, a short-term sliding window is shortened and a sliding step is reduced; when the long-term baseline drift amount is lower than a second threshold value and the medium-term trend slope change rate tends to be stable, a long-term sliding window is lengthened and a sliding step is increased.

[0010] As a preferred technical solution of the first aspect of the present application, the extraction logic of the node state difference amount is: Performance state characteristics of the current early warning period are arranged according to monitoring node numbers to form a performance characteristic matrix; Deviation trend slopes, fluctuation amplitudes and convergence parameters of the comprehensive risk deviation curve in the current period are mapped to a deviation characteristic matrix according to nodes; Values of the same monitoring node in corresponding dimensions of the performance characteristic matrix and the deviation characteristic matrix are subjected to difference operations to obtain a node difference vector; The node difference vectors of all monitoring nodes are compared with a safety benchmark state to obtain a baseline deviation measure, the baseline deviation measure is stacked by rows to form a baseline deviation matrix, and a norm value of each node baseline deviation measure in the baseline deviation matrix is marked as a node state difference amount.

[0011] As a preferred technical solution of the first aspect of the present application, the output logic of the early warning state of the current early warning period is: According to the topological mapping relationship between the path and the monitoring node, the average difference of each path covering node is calculated; the average difference is taken as the path risk influence factor, combined with the historical performance fluctuation index of the path, and the path correlation score is calculated; According to the path correlation score from high to low, the scheduling priority value is allocated to the sorted path in sequence, and a priority list is generated; According to the priority order, the risk threshold condition of the corresponding path is started; the risk warning state output of each path is collected, and the warning state output of the current warning period is synthesized by a comprehensive voting method.

[0012] As a preferred technical solution of the first aspect of the application, the acquisition logic of the path response evaluation tensor is: During the operation of each path in the current period, the weighted voting proportion of the warning state output of the corresponding path in the current warning period is calculated, and the weighted voting proportion is taken as the response contribution degree; The CPU occupancy rate, memory occupancy and execution time of the path in the operation period are recorded, and the resource consumption rate is synthesized after normalization according to a unified dimension; The mean square error between the prediction output and the real risk state of the path in the current period is calculated as the fitting error index; The response contribution degree, resource consumption rate and fitting error of each path are arranged in sequence according to the path number; the three types of indexes are stacked in three-dimensional tensor in the path dimension, and the dimensions correspond to the path number, the index type and the index value respectively; the minimum-maximum normalization is performed on the three-dimensional tensor along the index type dimension, and the path response evaluation tensor is obtained.

[0013] As a preferred technical solution of the first aspect of the application, the driving logic for hierarchical optimization of the full connection layer and the bottleneck layer weight of the improved MobileNetV3 regression model is: According to the path combination parameters, the resource consumption upper limit and parallel scheduling limit of each path are read, and the resource gating factor of the path is calculated; The response contribution degree, fitting error and node state difference of the path are weighted according to a preset proportion to obtain the importance score of the path; The importance scores of all paths are sorted by value, and the path set allowed to participate in model updating is selected in combination with the parallel scheduling limit; the scores of the paths in the path set are multiplied by the corresponding resource gating factors and normalized to obtain the path update weight; The full connection layer update coefficient is obtained by summing the path update weight in the path dimension, and the path update weight is distributed to the corresponding channel and summed according to the mapping relationship between the path and the bottleneck layer channel to obtain the update coefficient of each channel of the bottleneck layer; the update coefficients are used to scale the weight update amplitude of the corresponding layer respectively, and hierarchical optimization is performed.

[0014] As a preferred technical solution of the first aspect of the application, the acquisition logic of the periodic feedback feature vector is: For each detection node in the path set allowed to participate in model updating, the residual error between the model prediction value and the actual observation value of the detection node in the current early warning period is calculated; a weighted average is performed according to the distribution of the detection node in the path set, wherein the weight is the product of the path update weight and the node state difference amount, to obtain the path residual error value of each path; All path residual error values are combined in the order of path numbers into a residual feedback factor.

[0015] The short-term scale part is extracted from the comprehensive risk deviation curve, and the change rate between the current period and the last early warning period is calculated; the change rate sequence is smoothed and filtered to remove high-frequency noise, and the short-term trend disturbance value of each node is obtained; The path average of the node trend disturbance value is calculated according to the path mapping relationship, and is combined into a trend disturbance factor; The residual feedback factor and the trend disturbance factor are spliced according to the corresponding positions of the path numbers to obtain the periodic feedback feature vector as the next early warning period.

[0016] In the second aspect, the application provides a multi-dimensional dynamic monitoring and early warning system for an explosion-proof unit in a dangerous environment, which is based on the implementation of the first aspect and includes a data preprocessing module, a fitting parameter generation module, a risk situation analysis module, an early warning state output module, and a model optimization module, which are connected through wired and / or wireless connections; The data preprocessing module is used to collect node detection data and node historical detection data of each monitoring node in real time after the explosion-proof unit is equipped with monitoring nodes, to analyze the running state and resource consumption based on the node detection data and the node historical detection data to generate a node feature vector, and to complete the initialization configuration of each monitoring node; The fitting parameter generation module fuses the node feature vector and the periodic feedback feature vector generated by the model optimization module in the last early warning period into a periodic feature vector at the beginning of each early warning period, inputs the periodic feature vector into an improved MobileNetV3 regression model, and outputs fitting parameters of the evolution trend of the risk deviation. The risk situation analysis module generates a multi-scale risk deviation curve family based on the fitting parameters and fuses them into a comprehensive risk deviation curve, and dynamically adjusts the sliding windows of each scale according to the change rate of the comprehensive risk deviation curve. The early warning state output module extracts the node state difference amount based on the performance state feature and the comprehensive risk deviation curve, generates path scheduling priorities of multiple evaluation and prediction paths based on the node state difference amount, and schedules each path to output the early warning state of the current early warning period. The model optimization module is used for statistically counting the path response contribution degree, resource consumption level and fitting error after the period ends, constructing a path response evaluation tensor, calculating path update weights, performing hierarchical optimization on the full connection layer and bottleneck layer weights of the improved MobileNetV3 regression model, and extracting residual feedback factors and trend disturbance factors to generate a period feedback feature vector of the next early warning period and then feeding back to the fitting parameter generation module.

[0017] In the above technical solution, the present application provides technical effects and advantages: The present application collects the running state and resource consumption data of the monitoring node in real time, and performs multi-dimensional feature fusion to ensure the completeness and diversity of the input information; then, the improved MobileNetV3 regression model is used to realize high-precision fitting of the dangerous deviation evolution trend by taking advantage of the multi-scale fusion and long-range dependence modeling capability, so as to obtain a multi-scale risk deviation curve family of short-term, medium-term and long-term, and then fuse them into a comprehensive risk deviation curve to reflect the overall situation change.

[0018] The sliding window of each time scale is dynamically adjusted based on the change rate of the comprehensive curve, so that the system can sensitively respond to emergencies while stably tracking long-term trends; the path scheduling priority is generated by using the node state difference, so that the key path is preferentially executed under resource constraints, improving the timeliness of early warning and resource utilization efficiency; The model full connection layer and bottleneck layer are hierarchically optimized by constructing a path response evaluation tensor and calculating path update weights, and the residual feedback factors and trend disturbance factors are fed back to the next period to realize cross-period adaptive optimization, thereby maintaining the accuracy of monitoring, the timeliness of early warning and the stability of system operation in a complex dangerous environment for a long time. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0020] Figure 1 The structural block diagram of the explosion-proof unit dangerous environment multi-dimensional dynamic monitoring and early warning system of the present application; Figure 2 The application logic diagram of the improved MobileNetV3 regression model of the present application; Figure 3 The multi-scale risk deviation curve family and comprehensive risk deviation curve diagram of the present application; Figure 4 The multi-scale dynamic weight change over time diagram of the present application; Figure 5 It is a flow chart of the multi-dimensional dynamic monitoring and early warning method for hazardous environments of explosion-proof units of the present invention. DETAILED DESCRIPTION

[0021] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be more comprehensive and complete, and will fully convey the concepts of the example embodiments to those skilled in the art. The accompanying drawings are merely schematic illustrations of the present disclosure and are not necessarily drawn to scale. Identical reference numerals in the figures indicate identical or similar parts, and thus repeated descriptions thereof will be omitted.

[0022] In addition, the described features, structures or characteristics can be combined in any suitable manner in one or more example embodiments. In the following description, many specific details are provided to provide a full understanding of the example embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure can be practiced while omitting one or more of the specific details, or other methods, components, steps, etc. can be adopted. In other cases, well-known structures, methods, implementations or operations are not shown or described in detail to avoid obscuring various aspects of the present disclosure.

[0023] Example 1 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, wherein each module is connected via wired and / or wireless connections; The data preprocessing module is used to collect node detection data and node historical 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 node historical detection data, the module generates node feature vectors by analyzing the operating status and resource consumption. At the same time, it completes the initialization configuration of each monitoring node. It should be noted that the explosion-proof unit is equipped with multiple monitoring nodes, including temperature nodes, pressure nodes and vibration nodes. Each node collects node detection data and node historical detection data in real time; among them: Node detection data includes real-time collected operating status data and node resource redundancy data; among them: 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 cycles are extracted to characterize the node's operating trends and performance fluctuation characteristics. The node resource redundancy data includes resource consumption rate and residual capacity of the current monitored node in the early warning period, and the resource consumption rate and the residual capacity are taken as resource state feature inputs of the system; Specifically, the node feature vector acquisition logic is: The node feature vector includes performance state features and resource state features. Synchronize the sampling data of multiple types of monitoring nodes according to a unified timestamp, eliminate feature misalignment caused by different node sampling delays, and use interpolation or padding strategy to process missing data to ensure time series integrity. Based on historical reference interval and sliding window statistics, identify outliers caused by single-point mutation, excessive range sampling and sensor failure; perform outlier rejection or replacement (such as adjacent mean replacement, trend extrapolation replacement) to prevent their influence on trend modeling and threshold determination.

[0024] Map different physical quantities to comparable scales by standard normalization processing for each performance indicator; normalize resource consumption rate and residual capacity features to form resource state features, and maintain numerical distribution consistency of cross-node features. Calculate statistical indicators corresponding to the running state data in the sliding window; the statistical indicators include performance state mean, performance state standard deviation, fluctuation amplitude, trend slope and abnormal frequency; combine the statistical results with real-time collection values to form performance state features; each monitoring node and its features are added with a unique node identifier and a feature category label to facilitate global information propagation and differential modeling of the model in a multi-node environment.

[0025] Further, the initialization setting includes setting the safety reference state, path combination parameter and risk threshold condition of each monitoring node; wherein: The safety reference state is dynamically constructed by sliding window statistics, which extracts the performance state features of the closest historical early warning periods of each monitoring node, calculates the performance state mean and performance state standard deviation, and forms the reference interval: Wherein: represents the performance state mean of the th monitoring node in the sliding window, represents the performance state standard deviation of the th monitoring node in the sliding window, is a sensitivity adjustment coefficient. represents the reference interval of the th monitoring node in the sliding window, represents the maximum value of the reference interval, represents the minimum value of the reference interval.​ The path combination parameters include path initial weights, resource consumption upper limits and parallel scheduling limits, used to guide path selection and resource scheduling in the model cold start phase. ; Wherein: is the path number, represents the total number of paths, represents the initial selection weight of the th path (set in the model cold start phase), and represents the probability or priority of the path initially selected among all available paths; represents the sum of the initial weights of all paths, which is 1, constituting a path selection probability distribution; represents the current actual resource consumption of the th path, which measures the system resources occupied in the path evaluation or prediction process, and according to the system resource definition, it can be CPU%, memory MB, bandwidth Mbps, execution time, etc. represents the resource consumption upper limit allowed by the th path, which is used to prevent a path from occupying too much resource due to complex logic, i.e. the current resource consumption of each path cannot exceed its allowed maximum consumption upper limit; represents the maximum number of paths allowed to be executed in parallel at the same time, which controls the number of paths started at the same time in each early warning period, to prevent system overload.

[0026] The risk threshold condition includes a risk early warning state determined based on the path correlation score, and the risk early warning state includes a normal operation state, a fluctuation operation state and an abnormal operation state, which will be dynamically optimized in combination with a feedback mechanism. The determination logic is as follows: When the path correlation score is less than the fluctuation abnormality determination threshold minimum value, it is determined that the current risk early warning state is the normal operation state; When the path correlation score is between the fluctuation abnormality determination threshold minimum value and the maximum value, it is determined that the current risk early warning state is the fluctuation operation state; When the path correlation score is greater than the fluctuation abnormality determination threshold maximum value, it is determined that the current risk early warning state is the abnormal operation state; Wherein: the fluctuation abnormality determination threshold minimum value and the maximum value can be adaptively adjusted through the feedback mechanism.

[0027] The fitting parameter generation module is used to fuse the node feature vectors of each monitoring node and the cycle feedback feature vectors of the last early warning period to form a cycle feature vector at the beginning of each early warning period, and input it into the improved MobileNetV3 regression model to output the fitting parameter representing the evolution trend of the system risk deviation. Specifically, the periodic feedback feature vector comprises 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 observation value of each monitoring node in the last early warning period, and in the case of a multi-dimensional feature vector, the residual feedback factor adopts the mean square error of the predicted value and the actual observation value in the last early warning period; the trend disturbance factor is used to represent the trend change amplitude of each monitoring node in the adjacent early warning period, and is defined as the difference of the smoothed trend function in the adjacent period.

[0028] The periodic feature vector is spliced by the performance state feature and the resource state feature of each monitoring node in the current early warning period, and the residual feedback factor and the trend disturbance factor in the last early warning period in a preset order, that is: The performance state feature and the resource state feature of the monitoring node with node number are spliced by channel to obtain a node feature vector, and the node feature vector is ; wherein: , the monitoring node has , and is a positive integer, indicates the node number corresponding to the monitoring node, and the performance state feature of the thmonitoring node is and the resource state feature is ; the periodic feedback feature vector of the last early warning period is ; ; ; ; The feedback feature vector of the last early warning period is spliced at the end of the node feature vector of the current early warning period to form a node expansion vector, and the node expansion vectors of all monitoring nodes are spliced in turn according to the node number, and are flattened into a one-dimensional vector of the periodic feature vector ; .

[0029] It should be further pointed out that the improved MobileNetV3 classification network model is used to realize global joint modeling of node performance state and resource redundancy features and fitting of dangerous deviation evolution trend. As shown in Figure 2 , the improved MobileNetV3 regression model comprises: Using the MobileNetV3 network architecture as the foundation, the convolutional and bottleneck layers are retained. By sliding the convolutional layers in the time dimension, feature extraction is performed on the node feature vectors of each monitoring node, capturing both local transient fluctuations and long-term degradation trends. Within the bottleneck layer, performance status features and resource status features are concatenated according to a preset channel sequence, and multi-scale fusion expression is achieved through linear bottleneck compression. The nested fusion mechanism introduces channel-level attention weights on the fused feature tensor to enhance the coupling relationship between performance state features and resource state features, highlighting the dynamic impact of resource changes on performance degradation; Introducing node feature vectors and periodic feedback feature vectors into residual paths at different levels ensures long-range transmission of node feature vectors and periodic feedback feature vectors in the network, enabling cross-node information dissemination and global feature modeling, and improving the model's ability to perceive the chain-like evolution trend of systemic risks. The original classification fully connected layer of MobileNetV3 is replaced with a regression output layer, which outputs fitting parameters that characterize the evolution trend of the system's dangerous deviations. The fitting parameters include the deviation trend slope, deviation fluctuation amplitude, and deviation convergence parameter; where: Deviation trend slope: measures the direction and speed of changes in system risk indicators; Deviation fluctuation amplitude: quantifies the intensity of short-term fluctuations and is used to reflect operational stability; Deviation convergence parameter: assesses whether the risk change tends to be stable or continues to deteriorate.

[0030] The L1 loss function is introduced to enhance the model's ability to fit slight performance changes and abnormal samples. The L1 loss function is used to optimize the regression output results, suppress the excessive interference of outliers on the model prediction results, and improve the accuracy and stability of deviation trend fitting.

[0031] It should be noted that the improved MobileNetV3 classification network model internally integrates a feature extraction mechanism and a fusion coding mechanism to enhance the model's ability to express complex performance-resource coupling relationships. Global joint modeling can not only improve the recognition accuracy of local node deviations, but also realize dynamic prediction of the chain evolution trend of systematic deviations, providing high-reliability model support for subsequent risk deviation curve generation and warning status judgment. The improved MobileNetV3 regression model can realize global joint modeling of the multi-node performance status and resource redundancy characteristics of explosion-proof units while ensuring low computing resource consumption, and output high-precision and high-robustness dangerous deviation evolution trend prediction results, effectively supporting the dynamic warning and resource adaptive optimization decision-making of the explosion-proof system.

[0032] The risk situation analysis module generates a multi-scale nested risk deviation curve family based on the fitting parameters, fuses the change amplitudes to generate a comprehensive risk deviation curve, and dynamically adjusts the sliding windows of the short-term scale, the medium-term scale, and the long-term scale according to the comprehensive risk deviation curve.

[0033] Specifically, the generation logic of the risk deviation curve family is as follows: A multi-scale nested risk deviation curve family is constructed based on preset time scales, wherein the time scales are constructed by using sliding windows, and are divided into a short-term scale for capturing instantaneous fluctuations, a medium-term scale for reflecting trend changes, and a long-term scale for accumulating risks based on the time scales. Deviation trend time series are constructed on the short-term scale, the medium-term scale, and the long-term scale, and corresponding risk deviation curves are generated under each scale. More specifically, the acquisition logic of the risk deviation curve is as follows: The deviation trend time series reflecting the system risk changes are constructed by accumulating in the time order of the warning periods, and the risk trend connection and change trajectory recording between consecutive periods are realized by mapping the fitting parameter vectors of each period to the same time axis based on the deviation trend time series. Based on the deviation trend time series, a risk deviation curve representing the change trend of the dangerous state in the current warning period is generated by using a trend fitting function (such as polynomial fitting, exponential smoothing, or spline interpolation), wherein the horizontal axis of the risk deviation curve is time, the vertical axis is the risk deviation value, and the curve shape can reflect both short-term fluctuation characteristics and long-term evolution trend.

[0034] The risk deviation curves corresponding to each time scale are aligned according to the time axis and combined in a nested manner to form a multi-scale nested risk deviation curve family.

[0035] As shown in FIG. 1, Figure 3 In the risk situation analysis process, short-term scale risk deviation curves, medium-term scale risk deviation curves, and long-term scale risk deviation curves are generated based on the fitting parameters, which correspond to different time window lengths, respectively: the short-term scale risk deviation curve is used to reflect transient fluctuations and sudden abnormalities in a short time; the medium-term scale risk deviation curve is used to describe trend changes in a medium time span, taking into account fluctuations and trends; and the long-term scale risk deviation curve is used to depict the risk baseline changes in a long time running, reflecting the health evolution trend of the equipment. The comprehensive risk deviation curve is obtained by weighting and fusing the three curves based on the risk evaluation weight, which is used as the total output of the risk situation analysis and drives the decision logic of the warning state output module and the model optimization module.

[0036] The fusion logic of the comprehensive risk deviation curve is as follows: The short-term fluctuation intensity change, the medium-term trend slope change rate, and the long-term baseline drift amount are calculated under a short-term scale, a medium-term scale, and a long-term scale respectively. The short-term fluctuation intensity change, the medium-term trend slope change rate, and the long-term baseline drift amount are calculated under a short-term scale, a medium-term scale, and a long-term scale respectively. The short-term fluctuation intensity change, the medium-term trend slope change rate, and the long-term baseline drift amount are calculated under a short-term scale, a medium-term scale, and a long-term scale respectively.

[0037] Exemplarily, the comprehensive risk deviation curve is: The short-term fluctuation intensity change, the medium-term trend slope change rate, and the long-term baseline drift amount of the current early warning period are respectively: The short-term fluctuation intensity change, the medium-term trend slope change rate, and the long-term baseline drift amount of the previous early warning period are respectively: The risk assessment weights of the short-term scale, the medium-term scale, and the long-term scale are respectively: .

[0038] As shown in FIG. 1, the change of the risk assessment weights of the three types of risk deviation curves with time is shown. The change rate and the trend characteristics of the comprehensive risk deviation curve are used to trigger the adaptive adjustment of the sliding window length: when the short-term fluctuation amplitude exceeds a threshold, the short-term time window is shortened to improve the sensitivity to mutations; when the medium-term curve change tends to be stable, the medium-term time window is lengthened to enhance the trend smoothness; when the long-term baseline occurs a rapid shift, the long-term time window is shortened to speed up the response to the long-term trend change; when the short-term fluctuation is lower than a threshold for a long time, the short-term time window is lengthened to avoid over-response to small fluctuations. Figure 4 The risk assessment weights are determined by the system running state, the risk trend characteristics, and the resource allocation strategy, and are updated in each early warning period, and are used to allocate the importance of each time scale in the comprehensive curve calculation, so as to realize the closed-loop control of multi-scale risk perception and adaptive weight adjustment.

[0039]

[0040] ​​​​​​​​​​The early warning state output module extracts node state difference based on the performance state feature and the comprehensive risk deviation curve, generates path scheduling priorities of multiple evaluation and prediction paths based on the node state difference, and schedules each path to output the early warning state of the current early warning period; Specifically, the extraction logic of the node state difference is as follows: The performance state features of the current early warning period are arranged according to the monitoring node number to form a performance feature matrix; The deviation trend slope, fluctuation amplitude and convergence parameters of the comprehensive risk deviation curve in the current period are mapped to a deviation feature matrix according to the nodes; The numerical values of the same monitoring node in the corresponding dimensions of the performance feature matrix and the deviation feature matrix are subjected to difference operation to obtain a node difference vector; The node difference vectors of all monitoring nodes are compared with the safety benchmark state to obtain a baseline deviation measure, and the baseline deviation measure is stacked by row to form a baseline deviation matrix. The norm value of each node baseline deviation measure in the baseline deviation matrix is marked as the node state difference.

[0041] Further, the output logic of the early warning state of the current early warning period is as follows: According to the topological mapping relationship between the path and the monitoring node, the average difference of the nodes covered by each path is calculated. The average difference is used as a path risk influence factor, combined with the historical performance fluctuation index of the path, to calculate the path correlation score; The paths are sorted in descending order of the path correlation score, and the scheduling priority value (such as 1 for the highest priority) is assigned to the sorted paths in turn, and a priority list is generated; The risk threshold condition of the corresponding path is started according to the priority order. The risk early warning state output by each path is collected, and the early warning state of the current early warning period is synthesized by comprehensive voting.

[0042] The model optimization module, after the end of the early warning period, calculates the response contribution degree, resource consumption rate and fitting error of each path, constructs a path response evaluation tensor, and performs hierarchical optimization on the full connection layer and the bottleneck layer weight of the improved MobileNetV3 regression model based on the path response evaluation tensor. At the same time, the residual feedback factor and the trend disturbance factor are extracted as the feature fusion input of the next early warning period.

[0043] Specifically, the acquisition logic of the path response evaluation tensor is as follows: For each path in the current period, the weighted voting proportion of the early warning state output of the corresponding path in the current early warning period is calculated, and the weighted voting proportion is used as the response contribution degree; The CPU occupancy rate, memory occupancy and execution time of the path in the running period are recorded, and the resource consumption rate is synthesized after normalization according to a unified dimension. Calculate the mean square error of the path between the current period prediction output and the true risk state as the fitting error indicator; Arrange the response contribution, resource consumption rate and fitting error of each path in order according to the path number; stack the three indicators in the path dimension into a three-dimensional tensor, and the dimensions correspond to the path number, the indicator type and the indicator value respectively; perform min-max normalization on the three-dimensional tensor along the indicator type dimension to obtain the path response evaluation tensor.

[0044] Further, the driving logic for hierarchical optimization of the full connection layer and the bottleneck layer weight of the improved MobileNetV3 regression model is as follows: According to the path combination parameters, read the resource consumption upper limit and parallel scheduling limit of each path, and calculate the resource gating factor of the path; Weight the response contribution, fitting error and node state difference of the path according to the preset proportion to obtain the importance score of the path; Sort the importance scores of all paths according to the numerical value, and select a path set allowed to participate in model updating in combination with the parallel scheduling limit; multiply the scores of the paths in the path set by the corresponding resource gating factors and normalize to obtain the path update weight; Sum the path update weights in the path dimension to obtain the full connection layer update coefficient, and distribute the path update weights to the corresponding channels according to the mapping relationship between the path and the bottleneck layer channel and sum them to obtain the update coefficient of each channel of the bottleneck layer; the update coefficients are used to scale the weight update amplitude of the corresponding layer respectively to perform hierarchical optimization.

[0045] Further, the acquisition logic of the period feedback feature vector is as follows: For each detection node in the path set allowed to participate in model updating, calculate the residual between the model prediction value and the actual observation value of the detection node in the current warning period; perform weighted average according to the distribution of the detection node in the path set, wherein: the weight is the product of the path update weight and the node state difference, to obtain the path residual value of each path; Combine all path residual values in order of path number into a residual feedback factor.

[0046] Extract the short-term scale part from the comprehensive risk deviation curve, and calculate the change rate between the current period and the last warning period; smooth filter the change rate sequence to remove high-frequency noise to obtain the short-term trend disturbance value of each node; Map the node trend disturbance value to the path to obtain the path mean value and combine it into a trend disturbance factor; Splice the residual feedback factor and the trend disturbance factor according to the corresponding position of the path number to obtain the period feedback feature vector as the next warning period.

[0047] Embodiment 2 As Figure 5 shown, the embodiments not detailed as described in embodiment 1, the embodiments provide a hazardous environment multi-dimensional dynamic monitoring and early warning method for explosion-proof units, comprising the following steps: Real-time acquisition of detection data and historical data of each monitoring node, analysis of running state and resource consumption level, generation of node feature vector and completion of initialization configuration; At the beginning of each early warning period, the node feature vector and the period feedback feature vector generated by the model optimization module in the last early warning period are fused into a period feature vector, which is input into the improved MobileNetV3 regression model, and the fitting parameters of the dangerous deviation evolution trend are output. Based on the fitting parameters, a multi-scale risk deviation curve family is generated and fused into a comprehensive risk deviation curve, and the sliding window of each scale is dynamically adjusted according to the change rate of the comprehensive risk deviation curve; Based on the performance state features and the comprehensive risk deviation curve, the node state difference is extracted, the path scheduling priority of multiple evaluation and prediction paths is generated based on the node state difference, and the early warning state of the current early warning period is output by scheduling each path. After the period ends, the response contribution degree, resource consumption level and fitting error of each path are counted, the path response evaluation tensor is constructed, the path update weight is calculated, the weights of the fully connected layer and the bottleneck layer of the improved MobileNetV3 regression model are optimized, and the period feedback feature vector for the next early warning period is extracted.

[0048] The improved MobileNetV3 regression model comprises: MobileNetV3 network architecture as the basic framework, retaining the convolutional layer and the bottleneck layer, the convolutional layer is used for time series feature extraction, the bottleneck layer is used for resource state modeling, and the residual connection and the nested fusion mechanism of graph convolution are introduced; The output layer is a regression structure, which uses L1 loss optimization to fit the deviation evolution trend, and outputs the fitting parameters representing the system dangerous deviation evolution trend, including the deviation trend slope, the deviation fluctuation amplitude and the deviation convergence parameter.

[0049] The generation logic of the risk deviation curve family is: Based on the preset time scale, a multi-scale nested risk deviation curve family is constructed, wherein: the time scale is constructed by sliding window, and is divided into short-term scale for capturing instantaneous fluctuations, medium-term scale for reflecting trend changes, and long-term scale for cumulative risk based on the time scale. constructing a deviation trend time sequence on a short-term scale, a medium-term scale and a long-term scale, and generating a corresponding risk deviation curve under each scale; the risk deviation curve accumulates to generate a deviation trend time sequence in the order of early warning periods, and the risk trend connection and change trajectory record between continuous periods are realized by mapping the fitting parameter vector of each period to the same time axis based on the deviation trend time sequence; On the basis of the deviation trend time sequence, a risk deviation curve representing the change trend of the dangerous state in the current early warning period is generated using a trend fitting function. The curve of the risk deviation curve has a time axis as the horizontal axis and a risk deviation value as the vertical axis, and the curve form can reflect both short-term fluctuation characteristics and long-term evolution trend.

[0050] The risk deviation curves corresponding to each time scale are aligned according to the time axis and combined in a nested manner to form a multi-scale nested risk deviation curve family.

[0051] The fusion logic of the comprehensive risk deviation curve is: The short-term fluctuation intensity change, medium-term trend slope change rate and long-term baseline drift of the risk deviation curve are calculated under the short-term scale, medium-term scale and long-term scale respectively; The short-term fluctuation intensity change, medium-term trend slope change rate and long-term baseline drift are difference calculated with the corresponding short-term fluctuation intensity change, medium-term trend slope change rate and long-term baseline drift of the last early warning period to obtain the change amplitude; The change amplitudes of each scale are fused based on a preset risk assessment weight to generate a comprehensive risk deviation curve. When the short-term fluctuation intensity change in the comprehensive risk deviation curve exceeds a first threshold value, the short-term sliding window is shortened and the sliding step is reduced. When the long-term baseline drift is lower than a second threshold value and the medium-term trend slope change rate tends to be stable, the long-term sliding window is lengthened and the sliding step is increased.

[0052] The extraction logic of the node state difference quantity is: The performance state characteristics of the current early warning period are arranged according to the monitoring node number to form a performance characteristic matrix; The deviation trend slope, fluctuation amplitude and convergence parameter of the comprehensive risk deviation curve in the current period are mapped to the deviation characteristic matrix according to the node; The numerical values of the same monitoring node in the corresponding dimensions of the performance characteristic matrix and the deviation characteristic matrix are difference calculated to obtain a node difference vector; The node difference vectors of all monitoring nodes are compared with the safety benchmark state to obtain a baseline deviation measure, and the baseline deviation measure is stacked by row to form a baseline deviation matrix. The norm value of each node baseline deviation measure in the baseline deviation matrix is marked as a node state difference quantity.

[0053] The output logic of the early warning state of the current early warning period is: According to the topological mapping relationship between the path and the monitoring node, the average difference of the node covered by each path is calculated; the average difference is taken as the path risk influence factor, combined with the historical performance fluctuation index of the path, to calculate the path correlation score; According to the path correlation score from high to low, the scheduling priority value is assigned to the sorted path in turn, and the priority list is generated; According to the priority order, the risk threshold condition of the corresponding path is started; the risk early warning state output of each path is collected, and the early warning state output of the current early warning period is synthesized by comprehensive voting.

[0054] The acquisition logic of the path response evaluation tensor is: During the running of each path in the current period, the weighted voting proportion of the early warning state output of the corresponding path in the current early warning period is calculated, and the weighted voting proportion is taken as the response contribution degree; The CPU occupancy rate, memory occupancy and execution time of the path in the running period are recorded, and the resource consumption rate is synthesized after normalization according to the unified dimension; The mean square error between the prediction output and the real risk state of the path in the current period is calculated as the fitting error index; The response contribution degree, resource consumption rate and fitting error of each path are arranged in order according to the path number; the three types of indexes are stacked in the path dimension to form a three-dimensional tensor, and the dimensions correspond to the path number, the index type and the index value; the minimum-maximum normalization is performed on the three-dimensional tensor along the index type dimension to obtain the path response evaluation tensor.

[0055] The driving logic for hierarchical optimization of the full connection layer and the bottleneck layer weight of the improved MobileNetV3 regression model is: According to the path combination parameters, the resource consumption upper limit and the parallel scheduling limit of each path are read, and the resource gating factor of the path is calculated; The response contribution degree, fitting error and node state difference of the path are weighted according to the preset proportion to obtain the importance score of the path; The importance scores of all paths are sorted by numerical value, and the path set allowed to participate in model updating is selected in combination with the parallel scheduling limit; the scores of the paths in the path set are multiplied by the corresponding resource gating factors and normalized to obtain the path update weight; The full connection layer update coefficient is obtained by summing the path update weight in the path dimension, and the path update weight is distributed to the corresponding channel and summed according to the mapping relationship between the path and the bottleneck layer channel to obtain the update coefficient of each channel of the bottleneck layer; the update coefficients are used to scale the weight update amplitude of the corresponding layer respectively to perform hierarchical optimization.

[0056] The acquisition logic of the periodic feedback feature vector is: For each detection node in the path set allowed to participate in model updating, the residual error between the model prediction value and the actual observation value of the detection node in the current early warning period is calculated, and a weighted average is performed according to the distribution of the detection node in the path set, wherein the weight is the product of the path update weight and the node state difference, to obtain the path residual value of each path; All path residual values are combined into residual feedback factors in the order of path numbers.

[0057] The short-term scale part is extracted from the comprehensive risk deviation curve, the change rate between the current period and the last early warning period is calculated, the change rate sequence is smoothed and filtered to remove high-frequency noise, and the short-term trend disturbance value of each node is obtained; The path average of the node trend disturbance value is calculated according to the path mapping relationship, and the trend disturbance factor is combined; The residual feedback factor and the trend disturbance factor are spliced according to the corresponding positions of the path numbers to obtain the periodic feedback feature vector as the next early warning period.

[0058] The explosion-proof unit dangerous environment multi-dimensional dynamic monitoring and early warning method provided by the embodiment is used to execute the explosion-proof unit dangerous environment multi-dimensional dynamic monitoring and early warning system provided by each embodiment of the present application. The specific method and process of each structure realizing the corresponding function included in the explosion-proof unit dangerous environment multi-dimensional dynamic monitoring and early warning method are described in detail in the above embodiments of the explosion-proof unit dangerous environment multi-dimensional dynamic monitoring and early warning system, and will not be repeated here.

[0059] The above only describes some exemplary embodiments of the present application by way of illustration. It is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above figures and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the present application.

Claims

1. A multi-dimensional dynamic monitoring and early warning method for hazardous environments of explosion-proof units, characterized by: The following steps are included: Collect detection data and historical data of each monitoring node in real time, analyze the operating status and resource consumption level, generate node feature vectors and complete initialization configuration; At the beginning of each warning cycle, the node feature vector is fused with the cycle feedback feature vector generated by the model optimization module in the previous warning cycle to form a cycle feature vector. This vector is then fed into the improved MobileNetV3 regression model, which outputs the fitting parameters of the dangerous deviation evolution trend. Generate a multi-scale risk deviation curve family based on the fitting parameters and fuse them into a comprehensive risk deviation curve, and dynamically adjust the sliding windows of each scale according to the change rate of the comprehensive risk deviation curve; Extract node status differences based on performance status characteristics and comprehensive risk deviation curves, generate path scheduling priorities for multiple evaluation and prediction paths based on node status differences, and schedule each path to output the warning status of the current warning cycle; After the cycle ends, the response contribution, resource consumption level and fitting error of each path are counted, the path response evaluation tensor is constructed, the path update weight is calculated, the weights of the fully connected layer and bottleneck layer of the improved MobileNetV3 regression model are layered optimized, and the cycle feedback feature vector for the next warning cycle is extracted.

2. The multi-dimensional dynamic monitoring and early warning method for explosion-proof unit hazardous environment according to claim 1 is characterized in that: The improved MobileNetV3 regression model includes: The MobileNetV3 network architecture is used as the basic framework, retaining the convolutional layer and bottleneck layer. The convolutional layer is used for temporal feature extraction, and the bottleneck layer is used for resource status modeling. A nested fusion mechanism of residual connections and graph convolution is introduced. The output layer is a regression structure, which uses L1 loss optimization to fit the deviation evolution trend and outputs fitting parameters that characterize the system's dangerous deviation evolution trend. The fitting parameters include the deviation trend slope, deviation fluctuation amplitude, and deviation convergence parameter.

3. The multi-dimensional dynamic monitoring and early warning method for explosion-proof unit hazardous environment according to claim 2 is characterized in that: The generation logic of the risk deviation curve family is: A multi-scale nested family of risk deviation curves is constructed based on a preset time scale, where 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 risks. Deviation trend time series are constructed at short-term, medium-term, and long-term scales, and corresponding risk deviation curves are generated at each scale. The risk deviation curves are accumulated in the chronological order of the warning cycles to generate deviation trend time series. Based on the deviation trend time series, the fitting parameter vector of each cycle is mapped onto the same time axis to achieve the connection of risk trends and the recording of change trajectories between consecutive cycles. Based on the deviation trend time series, a trend fitting function is used to generate a risk deviation curve that characterizes the trend of changes in the dangerous state within the current warning period. The horizontal axis of the risk deviation curve is time, and the vertical axis is 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 multi-dimensional dynamic monitoring and early warning method for explosion-proof unit hazardous environment according to claim 3 is characterized in that: The fusion logic of the comprehensive risk deviation curve is: The short-term volatility intensity change, medium-term trend slope change rate and long-term baseline drift of the risk deviation curve are calculated at the short-term scale, medium-term scale and long-term scale respectively; The change amplitude is obtained by performing differential calculation on the change of short-term fluctuation intensity, medium-term trend slope change rate and long-term baseline drift with the corresponding short-term fluctuation intensity, medium-term trend slope change rate and long-term baseline drift of the previous warning period; The change amplitudes of each scale are integrated based on the preset risk assessment weights to generate a comprehensive risk deviation curve; when the short-term fluctuation intensity change in the comprehensive risk deviation curve exceeds a 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 medium-term trend slope change rate tends to be stable, the long-term sliding window is extended and the sliding step size is increased.

5. The multi-dimensional dynamic monitoring and early warning method for explosion-proof unit hazardous environment according to claim 4 is characterized in that: The logic for extracting the node status difference is: Arrange the performance status characteristics of the current warning cycle into a performance characteristic matrix according to the monitoring node number; Map the deviation trend slope, fluctuation range and convergence parameters of the comprehensive risk deviation curve in the current period to the deviation feature matrix by node; Perform difference calculation on the values ​​of corresponding dimensions of the same monitoring node in the performance characteristic matrix and the deviation characteristic matrix to obtain the node difference vector; The node difference vectors of all monitored nodes are compared with the safety baseline state to obtain the baseline deviation measurement, and the baseline deviation measurement is stacked row by row to form a baseline deviation matrix; the norm value of the baseline deviation measurement of each node in the baseline deviation matrix is ​​marked as the node state difference.

6. The multi-dimensional dynamic monitoring and early warning method for explosion-proof unit hazardous environment according to claim 5 is characterized in that: The output logic of the warning status of the current warning cycle is: Based on the topological mapping relationship between paths and monitoring nodes, the average difference in the nodes covered by each path is calculated. The average difference is used as a path risk influencing factor and combined with the path's historical performance fluctuation index to calculate the path relevance score. Sort the paths by their relevance scores from high to low, assign scheduling priority values ​​to the sorted paths, and generate a priority list. Start the risk threshold conditions of the corresponding paths in order of priority; Collect the risk warning status output by each path and synthesize the warning status output of the current warning cycle by comprehensive voting.

7. The multi-dimensional dynamic monitoring and early warning method for explosion-proof unit hazardous environment according to claim 6 is characterized in that: The logic for obtaining the path response evaluation tensor is as follows: For each path in the current cycle, the weighted voting ratio of the warning status output of 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 run cycle, and normalize them to a unified dimension to synthesize the resource consumption rate; Calculate the mean square error between the predicted output of the path in the current period and the actual risk state as the fitting error indicator; The response contribution, resource consumption rate, and fitting error of each path are arranged in sequence according to the path number. The three types of indicators are stacked into a three-dimensional tensor in the path dimension, where the dimensions correspond to the path number, indicator type, and indicator value, respectively. The three-dimensional tensor is normalized min-max along the indicator type dimension to obtain the path response evaluation tensor.

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

9. The multi-dimensional dynamic monitoring and early warning method for explosion-proof unit hazardous environment according to claim 8 is characterized in that: The acquisition logic of the periodic feedback feature vector is: For each detection node in the path set that is allowed to participate in the model update, the residual between the model prediction value and the actual observation value of the detection node in the current warning cycle is calculated; Perform weighted averaging based on the distribution of detection nodes in the path set, 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; Combine all path residual values ​​into residual feedback factors in the order of path numbers; Extract the short-term scale portion from the comprehensive risk deviation curve and calculate the rate of change between the current cycle and the previous warning cycle; smooth and filter the rate of change sequence to remove high-frequency noise and obtain the short-term trend disturbance value of each node; The node trend disturbance value is calculated according to the path mapping relationship to obtain the path mean and combined into the trend disturbance factor; The residual feedback factor and the trend disturbance factor are spliced ​​according to the corresponding position of the path number to obtain the periodic feedback feature vector of the next warning period.

10. A multi-dimensional dynamic monitoring and early warning system for explosion-proof units in hazardous environments, based on the implementation of a multi-dimensional dynamic monitoring and early warning method for explosion-proof units in hazardous environments according to any one of claims 1 to 9, characterized in that: It includes a data preprocessing module, a fitting parameter generation module, a risk situation analysis module, a warning status output module and a model optimization module, and each module is connected by wired and / or wireless connections; The data preprocessing module is used to collect node detection data and node historical 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 node historical detection data, the module generates node feature vectors by analyzing the operating status and resource consumption. At the same time, it completes the initialization configuration of each monitoring node. The fitting parameter generation module, at the beginning of each warning cycle, fuses the node feature vector with the cycle feedback feature vector generated by the model optimization module in the previous warning cycle to form a cycle feature vector, inputs it into the improved MobileNetV3 regression model, and outputs the fitting parameters of the dangerous deviation evolution trend; The risk situation analysis module generates a family of multi-scale risk deviation curves based on fitting parameters and fuses them into a comprehensive risk deviation curve. The sliding windows of each scale are dynamically adjusted according to the change rate of the comprehensive risk deviation curve. The warning status output module extracts node status differences based on performance status characteristics and comprehensive risk deviation curves, generates path scheduling priorities for multiple evaluation and prediction paths based on the node status differences, and schedules each path to output the warning status of the current warning cycle; The model optimization module calculates the response contribution, resource consumption level, and fitting error of each path after the cycle ends, constructs a path response evaluation tensor, calculates the path update weight, performs hierarchical optimization on the fully connected layer and bottleneck layer weights of the improved MobileNetV3 regression model, and extracts the residual feedback factor and trend disturbance factor to generate the periodic feedback feature vector for the next warning cycle and feed it back to the fitting parameter generation module.

Images