A gas safety prediction method and system based on time series modeling

Abstract

The application provides a gas safety prediction method and system based on time series modeling, and belongs to the field of industrial control. The method comprises the following steps: using a plurality of sliding analysis windows with different fixed widths to slide on a time series of a normalized data matrix to traverse the normalized data matrix and generate a fusion feature vector; using a radial basis kernel function as a hidden layer neuron to construct a feedforward neural network prediction model; using the fusion feature vector to perform precision accounting on the trained feedforward neural network prediction model, and adjusting a prediction result and / or a prediction step of the feedforward neural network prediction model in response to a result of the precision accounting being unacceptable; outputting a cumulative deviation warning level in a targeted manner according to a correlation between an error accumulation and a dynamic judgment boundary; and adding information of the known event to the fusion feature vector and / or adjusting a size of the dynamic judgment boundary.

Description

Technical Field

[0001] This invention relates to the fields of industrial control and agricultural technology, and specifically to a gas safety prediction method based on time-series modeling. Background Technology

[0002] In modern intensive agricultural production and bulk agricultural product logistics and warehousing, the safety indicators of the production environment are directly related to crop growth and yield, livestock health, and the safety and stability of national grain reserves. Due to the complex characteristics of agricultural sites, such as high humidity, high dust levels, rich microorganisms, and interference from biological activities, environmental indicators (such as temperature, humidity, and concentrations of specific harmful gases) often exhibit strong nonlinear and time-varying characteristics. Passive monitoring schemes, which typically trigger alarms only when fixed thresholds are exceeded, suffer from severe response delays in the face of sudden safety incidents such as localized smoldering fires in grain silos or acute poisoning of livestock from harmful gases due to ventilation interruptions in modern pigsties. Alarms are often only triggered after irreversible and significant economic losses have already occurred.

[0003] Existing technical solutions typically focus on a specific type of model or scenario: they may rely on complex deep models, require comprehensive spatial, meteorological, and leakage source parameters, primarily focus on the spatial distribution of concentration fields and leakage source location, or only retrain the model after the prediction deviation exceeds a threshold. For multi-source sensor time-series data deployed on resource-constrained industrial edge nodes, it is difficult to construct a unified technical framework that meets the requirements for automatic feature extraction, lightweight prediction, robust data quality processing, self-accuracy verification, conditional online correction, and tiered early warning capabilities. Related technical solutions still require improvement. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a gas safety prediction method and system based on time-series modeling. This method and system can perform general feature extraction, lightweight and high-precision reasoning, online closed-loop prediction with embedded continuous self-evaluation, and gas safety prediction by combining prior knowledge of structured events in the external environment.

[0005] In a first aspect, this invention provides a gas safety prediction method based on time-series modeling, comprising the following steps: Convert raw data acquired from multi-source and heterogeneous industrial sensors into a normalized data matrix; Multiple sliding analysis windows with different fixed widths are used to slide across the time series of the normalized data matrix to traverse the normalized data matrix. Various statistical measures are performed on the data subsets that fall within the sliding analysis windows, and a fusion representation vector is generated based on the results of the statistical measures. A feedforward neural network prediction model is constructed using radial basis function kernel functions as hidden layer neurons; The first part of the fused representation vector is used to train the feedforward neural network prediction model. The second part of the fused representation vector, which is later than the first part in time series, is used to perform accuracy calculation on the trained feedforward neural network prediction model. The prediction results and / or prediction step size of the feedforward neural network prediction model are adjusted based on the accuracy calculation results. The results of the accuracy calculation are accumulated into a residual sequence. The cumulative error is determined based on the residual sequence. The cumulative deviation warning level is output in a targeted manner based on the correlation between the cumulative error and the dynamic judgment boundary. Information about known events is stored in a structured event knowledge base. In response to the current moment falling within the influence time domain of any known event, the information of that known event is added to the fusion representation vector and / or the size of the dynamic decision boundary is adjusted. A feedforward neural network prediction model is used to predict gas safety and output the prediction results. Gas safety information is determined based on the prediction results and the cumulative deviation warning level.

[0006] In some embodiments, converting raw collected data into a normalized data matrix includes: The independent timestamps used by the original collected data were adjusted to a unified timestamp based on Coordinated Universal Time (UTC). The node identifiers and unified timestamps of industrial sensors are used as composite primary keys to hierarchically sort the raw collected data, thereby constructing a normalized data matrix with monotonically increasing time series. In response to the original acquired data being a slowly varying signal with large inertia, piecewise linear interpolation is used to supplement the missing data in the normalized data matrix; in response to the original acquired data being a fast-responding transient signal, forward filling is used to supplement the missing data in the normalized data matrix. The faulty industrial sensor is identified based on the physically feasible range of the raw acquired data, and any data provided by the faulty industrial sensor is removed from the normalized data matrix.

[0007] In some embodiments, multiple statistical measures are applied to the subset of data falling within the sliding analysis window, and a fusion representation vector is generated based on the results of the statistical measures, including: A dispersion measure is applied to a subset of data to obtain an unbiased sample variance estimate, which is then used as a volatility index. Covariance correlation measures are performed on a subset of data to obtain the sample covariance matrix as a coupling index; By measuring the disorder of a subset of data, we obtain information entropy as a disorder index. A fusion representation vector is generated based on the volatility index, coupling index, and disorder index.

[0008] In some embodiments, performing multiple statistical measures on the data subsets falling within the sliding analysis window, and generating a fusion representation vector based on the results of the statistical measures, further includes: Perform statistical analysis on a subset of data to obtain one or more of the following: arithmetic mean, maximum and minimum values, fourth-order normalized moments of the distribution, third-order normalized moments of the distribution, sine transform of the hour index of the daily period, and cosine transform of the hour index of the daily period. A fusion representation vector is generated based on one or more of the following: volatility index, coupling index, disorder index, arithmetic mean, maximum and minimum values, fourth-order normalized moments of the distribution, third-order normalized moments of the distribution, sine transform of the hour index of the daily cycle, and cosine transform of the hour index of the daily cycle.

[0009] In some embodiments, multiple statistical measures are applied to the subset of data falling within the sliding analysis window, and a fusion representation vector is generated based on the results of the statistical measures, including: Locality-sensitive hashing is applied to subsets of data to filter out redundant data with the same locality-sensitive hashing in adjacent subsets, and then the redundant data is removed from the normalized data matrix. The mean and standard deviation of the entire sequence are calculated for all sampled data in the normalized data matrix. In response to the detection that the absolute value of the normalized deviation of a sampled point is greater than 3 times the standard deviation, the sampled point is marked as a candidate outlier observation point. In response to the existence of multiple candidate anomaly observation points, the system determines whether the multiple candidate anomaly observation points fall within the time domain of the influence of any known event recorded in the structured event knowledge base based on the structured event knowledge base. The anomaly detection sensitivity coefficient is reduced when the anomaly falls within the time domain of influence, and noise reduction filtering is applied to multiple candidate anomaly observation points when the anomaly does not fall within the time domain of influence. The fused representation vector is generated based on the results of statistical measures performed on subsets of data in the normalized data matrix after removing redundant data and performing noise reduction filtering.

[0010] In some embodiments, constructing a feedforward neural network prediction model using radial basis function kernel functions as hidden layer neurons includes: The K-Means clustering algorithm is used to perform unsupervised clustering in the feature space of a feedforward neural network prediction model to determine the optimal spatial location of the center point of the kernel function. The width parameter of the kernel function is determined based on the feature distance between the center points of the kernel function and the preset scaling factor; The radial basis kernel function of the feedforward neural network prediction model is determined based on the center point and width parameters of the kernel function. The optimal solution for the output layer weights in the least-squares sense is obtained using the closed-form formula of the Moore-Penrose generalized inverse matrix. The feedforward neural network prediction model responds to the received fused representation vector as model input, and generates a scalar prediction as model output based on the activation function of the hidden layer neurons composed of radial basis kernel functions and the linear relationship with the output layer weights.

[0011] In some embodiments, the second part is used to perform accuracy calculation, and the prediction results and / or prediction step size of the feedforward neural network prediction model are adjusted based on the results of the accuracy calculation. The reference window interval and the prediction window interval are determined. The real data with a length equal to the reference window interval is truncated from the second part and input into the feedforward neural network prediction model. The prediction data with a length equal to the prediction window interval is obtained by forward extrapolation. The symmetric mean absolute percentage error is calculated between the predicted data and the actual data corresponding to the predicted data in the second part. In response to the symmetric mean absolute percentage error exceeding the acceptable accuracy threshold, the average residual within the prediction window interval is added to the predicted data of the next prediction window interval, and / or the length of the prediction window interval is shortened.

[0012] In some embodiments, the cumulative error sum is determined based on the residual sequence, and the cumulative deviation warning level is output selectively based on the correlation between the cumulative error sum and the dynamic decision boundary, including: The residual mean and residual standard deviation are determined based on the residual sequence. Using the residual mean, residual standard deviation and preset acceptable offset reference value, the positive cumulative sum and negative cumulative sum of the residuals are calculated for each symmetrical mean absolute percentage error. The maximum value of the positive cumulative sum and negative cumulative sum is used as the anomaly discrimination statistic. A low-priority review notification is output in response to the anomaly detection statistic falling within the first interval of the dynamic decision boundary; In response to an anomaly detection statistic falling within a second interval that is higher than the first interval at the dynamic judgment boundary, a medium-priority inspection alarm is output. A high-priority emergency response plan is output in response to an anomaly detection statistic falling within a third interval that is higher than the second interval on the dynamic judgment boundary.

[0013] In some embodiments, storing information about known events in a structured event knowledge base, and adding information about known events to the fusion representation vector and / or adjusting the size of the dynamic decision boundary in response to the current time falling within the influence time domain of any known event; The impact factors, time-domain range of impact, and anomaly detection sensitivity coefficients of industrial sensors affected by known events are stored in a structured event knowledge base; In response to the current moment falling within the influence time domain of any known event, the influence factor, influence time domain range, and anomaly detection sensitivity coefficient of the industrial sensor are concatenated to the end of the fused characterization vector, and / or the anomaly detection sensitivity coefficient is used to amplify the size of the dynamic decision boundary.

[0014] A second aspect of the present invention provides a gas safety prediction system based on time-series modeling, comprising: The preprocessing module converts the raw data acquired from multi-source and heterogeneous industrial sensors into a normalized data matrix; The feature extraction module uses multiple sliding analysis windows with different fixed widths to slide across the time series of the normalized data matrix to traverse the normalized data matrix. It performs various statistical measures on the data subsets that fall within the sliding analysis windows and generates a fusion representation vector based on the results of the statistical measures. The model building module uses radial basis function kernel functions as hidden layer neurons to build a feedforward neural network prediction model; An adaptive module trains a feedforward neural network prediction model using the first part of the fused representation vector, performs accuracy calculation on the trained feedforward neural network prediction model using the second part of the fused representation vector that is later in time than the first part, and adjusts the prediction results and / or prediction step size of the feedforward neural network prediction model based on the accuracy calculation results. The early warning module accumulates the results of accuracy calculation into a residual sequence, determines the cumulative error sum based on the residual sequence, and outputs the cumulative deviation early warning level based on the correlation between the cumulative error sum and the dynamic judgment boundary. The prior knowledge module stores information about known events in a structured event knowledge base. In response to the current moment falling within the influence time domain of any known event, it adds the information of that known event to the fusion representation vector and / or adjusts the size of the dynamic decision boundary. The implementation module uses a feedforward neural network prediction model to predict gas safety and output the prediction results. Based on the prediction results and the cumulative deviation warning level, it determines the gas safety information.

[0015] This invention converts raw data acquired from multi-source and heterogeneous industrial sensors into a normalized data matrix based on a unified standard. Multiple sliding analysis windows with different fixed widths are used to slide across the time series of the normalized data matrix to traverse it. Various statistical measures are applied to the data subsets falling within each sliding analysis window. A fusion representation vector is generated based on the results of these statistical measures. A feedforward neural network prediction model is constructed using radial basis function kernels as hidden layer neurons. The first part of the fusion representation vector is used to train the feedforward neural network prediction model. The second part of the fusion representation vector, which is later in time than the first part, is used to perform accuracy verification on the trained feedforward neural network prediction model. If the accuracy verification result is unacceptable, the prediction results of the feedforward neural network prediction model are adjusted accordingly. Alternatively, the prediction step size can be adjusted, and the results of the accuracy calculation can be accumulated into a residual sequence. The cumulative error sum can be determined based on the residual sequence. Based on the correlation between the cumulative error sum and the dynamic judgment boundary, a targeted cumulative deviation warning level can be output. The information of known events can be stored in a structured event knowledge base. In response to the current moment falling within the influence time domain of any known event, the information of the known event can be added to the fusion representation vector, and / or the size of the dynamic judgment boundary can be adjusted. A feedforward neural network prediction model can be used to predict gas safety and output the prediction results. The technical solution for determining gas safety information based on the prediction results and the cumulative deviation warning level can be implemented. This allows for the execution of general feature extraction, a trade-off between lightweight and high-precision reasoning, online closed-loop prediction with embedded continuous self-evaluation, and gas safety prediction combined with prior knowledge of structured events in the external environment. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the gas safety prediction method based on time-series modeling provided in this embodiment of the invention; Figure 2 A block diagram of a gas safety prediction system based on time-series modeling provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to specific examples and the accompanying drawings.

[0019] These embodiments are provided to make the disclosure thorough and complete, and to fully express the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specifically stated, the relative arrangement of components and steps, material composition, numerical expressions, and values ​​set forth in these embodiments should be interpreted as exemplary only and not as limiting.

[0020] Furthermore, words such as "include" or "contain" mean that the element preceding the word covers the element listed after the word, and do not exclude the possibility that it may also cover other elements.

[0021] It should also be noted that, in the description of this disclosure, unless otherwise expressly specified and limited, the specific meaning of each term in this disclosure can be understood by those skilled in the art as appropriate. All terms used in this disclosure have the same meaning as understood by those skilled in the art to which this disclosure pertains, unless otherwise specifically defined. It should also be understood that, unless expressly defined herein, terms defined in a general dictionary should be interpreted as having a meaning consistent with their meaning in the context of the relevant art, and not as having an idealized or highly formalized meaning.

[0022] In industrial safety production, occupational health protection, and process industry risk management scenarios, continuous online monitoring and trend prediction of combustible gases, toxic and harmful gases, and early fire characteristic parameters in the work environment are important proactive technical means to prevent fires, explosions, poisonings, and equipment malfunctions. Typical monitoring targets include, but are not limited to, changes in the concentrations of methane, propane, hydrogen, hydrogen sulfide, carbon monoxide, ammonia, chlorine, and oxygen, smoke particle concentration, and the rate of change of ambient temperature. Combustible gas detectors and process industry safety instrumented systems already have corresponding standards or specifications, indicating that these monitoring targets are an important part of industrial safety and functional safety.

[0023] Regarding the concentration prediction, leakage early warning, and anomaly detection of the aforementioned objects, existing technologies have generally formed the following representative technical routes.

[0024] The first category comprises gas diffusion prediction routes based on physical mechanisms and numerical simulations. These methods are typically grounded in fluid mechanics, mass transfer and diffusion theory, turbulent diffusion models, or heavy gas diffusion models. Representative methods include computational fluid dynamics simulations, Gaussian plume or Gaussian cloud models, SLAB heavy gas diffusion models, and Dense Gas Atmospheric Dispersion (DEGADIS) models. Existing patented solutions also include techniques that utilize meteorological parameters, topographic parameters, leak source parameters, and machine learning models to determine the concentration distribution of leaked gas. The advantage of this type of method lies in its strong physical interpretability, enabling spatial extrapolation of gas concentration fields or diffusion paths when the leak source location, leakage rate, meteorological boundary conditions, topographic and architectural information, and material physicochemical parameters are relatively clear. However, in actual industrial settings, information such as leak source location, instantaneous release volume, gas composition, ventilation conditions, and local obstacle layout is often incomplete or highly uncertain. High-fidelity computational fluid dynamics (CFD) models or complex diffusion simulations may also face challenges such as complex modeling, high parameter calibration costs, and significant real-time computational pressure. Therefore, these methods are more suitable for accident consequence analysis, risk assessment, drills, and detailed modeling of high-value scenarios, rather than being suitable as the core for second- or minute-level online predictions at all industrial edge gas monitoring points.

[0025] The second category comprises prediction approaches based on classical statistics and traditional time series analysis. These methods typically use historical observation sequences as a foundation, employing autoregression, moving averages, exponential smoothing, seasonal decomposition, or other statistical prediction models to characterize the changes in the target variable over time. In the field of gas concentration prediction, existing research has combined traditional time series models such as the Autoregressive Integrated Moving Average (ARIMA) with neural network models such as Long Short-Term Memory (LSTM) networks for gas or gas concentration prediction tasks, demonstrating that traditional statistical models remain an important baseline and component of this field. The advantages of this type of method lie in its relatively clear model structure, well-defined parameter meanings, and lower computational resource consumption. It is suitable for scenarios with limited data scale, relatively stable system dynamics, and where linear or weakly nonlinear assumptions are largely valid. For industrial safety monitoring systems, this type of method has a certain degree of engineering feasibility in edge computing, embedded controllers, and real-time alarm scenarios. However, industrial field sensor time-series data are typically affected by factors such as equipment start-up and shutdown, valve operation, changes in ventilation status, seasonal temperature and humidity cycles, sensor drift, operating condition switching, and multi-source coupling diffusion, exhibiting obvious non-stationarity, nonlinearity, and multivariate coupling characteristics. Simply relying on linear parameter models or fixed statistical assumptions often fails to fully express the temporal structure changes under complex operating conditions, and also makes it difficult to distinguish between real hazardous events, sensor artifacts, and predictable operating condition disturbances.

[0026] The third category comprises complex nonlinear modeling approaches based on machine learning and deep neural networks. In recent years, deep learning models have been widely used for tasks such as complex temporal anomaly detection, multivariate sensor anomaly detection, gas concentration prediction, multi-step gas prediction, and pipeline leak early warning. Among these, recurrent neural networks, long short-term memory networks, and graph neural networks can learn complex temporal dependencies and multivariate correlations from historical data; graph neural network methods can also characterize the correlation structure between multiple sensor nodes, providing new modeling ideas for anomaly detection in multi-point industrial sensor systems. Existing patented solutions also include multi-sensor fusion gas concentration multi-step prediction methods based on deep learning, and natural gas pipeline leak early warning methods based on machine learning. These methods typically achieve high prediction accuracy when there are sufficient samples, good training conditions, and a stable deployment environment. However, industrial safety accidents are low-probability events, and real-world accident scenario samples are often scarce and extremely imbalanced. Deep models typically require a large number of training samples, significant computing resources, and complex model maintenance processes. Furthermore, their internal representation is relatively implicit, necessitating additional safety justification and engineering constraints to address the interpretability, verifiability, traceability, and failure mode analysis requirements of safety-critical systems. Therefore, while deep models represent an important development direction, they still face limitations in deployment cost, maintenance complexity, and auditability under conditions of low-power edge computing, unattended sites, and small-to-medium-scale industrial samples.

[0027] The fourth category is rule-based or statistical monitoring approaches oriented towards sensor alarm discrimination. In existing industrial gas alarm systems, fixed threshold over-limit alarms, tiered threshold alarms, 3σ anomaly detection, and alarm rules based on historical residuals or statistical biases are still quite common. The advantages of this type of method are its simple engineering implementation, fast response speed, and clear alarm logic, making it easy for field personnel to understand and for regulatory agencies to review. However, this type of method usually relies on preset thresholds or relatively stable normal operating condition baselines, and its adaptability to contextual factors such as sensor drift, environmental interference, communication anomalies, planned maintenance, holiday shutdowns and restarts, extreme weather, and seasonal process changes is limited. When the alarm threshold is set too low, frequent false alarms are likely to occur; when the threshold is set too high, early warning signs may be missed. Therefore, how to introduce data-driven prediction, context recognition, and adaptive discrimination mechanisms while maintaining rule transparency and computational lightweightness remains a problem that needs to be solved in the field of industrial safety monitoring.

[0028] Based on publicly available patent technologies, existing solutions cover areas such as multi-sensor fusion for multi-step prediction of gas concentration, distributed gas concentration prediction, real-time prediction using dynamic neural networks, machine learning-based early warning for natural gas pipeline leaks, and prediction of leaked gas concentration distribution. These solutions improve existing gas monitoring and early warning systems from the perspectives of deep learning, multi-source fusion, distributed computing, dynamic neural networks, leak source localization, concentration field prediction, or machine learning-based early warning. For example, existing distributed gas concentration prediction solutions involve prediction deviation judgment and secondary training mechanisms, while existing dynamic neural network solutions involve real-time gas concentration prediction and dynamic network adjustment.

[0029] Based on the above-mentioned existing technical approaches and related patent disclosures, existing industrial gas safety monitoring and fire precursor prediction technologies have at least the following four common shortcomings.

[0030] First, there remains a contradiction between automation, interpretability, and cross-scenario generalization in feature construction and input representation. Traditional statistical models and alarm systems often rely on manual selection of raw concentration values, differential values, moving averages, or several empirical thresholds as input. Feature design is closely related to specific gas types, sensor placement, workshop structure, and production conditions, requiring parameter retuning or redesign when migrating across scenarios. While deep neural networks can automatically learn high-dimensional implicit representations from raw data, their features often lack direct physical meaning, hindering understanding by on-site safety engineers and regulatory review. Existing research on automatic temporal feature extraction shows that temporal data can be structurally represented through a large number of statistical features or a small number of selected general features. Furthermore, a compact, low-redundancy, and interpretable set of temporal features is of significant value for cross-task modeling and engineering deployment. Therefore, for industrial edge monitoring scenarios, a compact feature representation method that does not rely on specific scenario expert experience, has low computational cost, and possesses clear statistical meaning is still needed.

[0031] Secondly, existing methods are insufficient for systematically addressing long-term sensor drift, data quality degradation, and predictable operational disturbances. Industrial gas sensors may be exposed to high humidity, high temperature, dust, oil mist, corrosive atmospheres, electromagnetic interference, and cross-sensitive gases for extended periods. Electrochemical, semiconductor, metal oxide, infrared absorption, and catalytic combustion sensors may all experience baseline drift, sensitivity decay, cross-interference, transient saturation, communication packet loss, or duplicate data uploads. Gas sensor drift has become a publicly available benchmark dataset and a long-standing research issue, indicating that this problem is not an isolated device anomaly but a common engineering challenge in long-term online gas monitoring systems. Furthermore, planned maintenance, purging and replacement, concentrated shutdowns and restarts during holidays, extreme weather, seasonal process adjustments, and regular equipment maintenance can also cause short-term fluctuations or localized anomalies in sensor readings. Existing systems, relying solely on fixed threshold rejection, simple smoothing filters, or manual verification, struggle to distinguish between "abnormal concentrations of genuine hazardous gases," "sensor performance degradation," "data artifacts caused by communication link anomalies," and "predictable fluctuations caused by known operational events." Therefore, it is necessary to combine data deduplication, outlier screening, sensor drift identification, context event discrimination, and robust filtering into a more complete data quality control mechanism.

[0032] Third, model deployment methods are generally static, lacking continuous performance self-evaluation and conditional adaptive correction mechanisms. Many predictive models adopt a "historical data training—parameter freezing—long-term deployment" approach, relying mainly on manual inspection, periodic offline retraining, or post-event maintenance to restore performance after the model goes online. However, the data distribution and input-output relationships in industrial settings continuously change with equipment aging, raw material batch variations, process parameter adjustments, seasonal environmental changes, ventilation condition alterations, and sensor drift. These changes can be categorized as data drift or concept drift problems. When drift occurs, the predictive performance of fixed-parameter models may gradually degrade, and this degradation process may not be detected by on-site personnel in a timely manner. Existing research has begun to focus on automatic detection of model prediction quality degradation and model update triggering mechanisms, indicating that self-evaluation, self-monitoring, and adaptive maintenance are important directions for deployed machine learning systems. Therefore, for unattended or low-maintenance-cost-requirement industrial safety monitoring systems, prediction error calculation, performance degradation identification, and conditional parameter correction mechanisms need to be embedded in the online operation process, rather than relying entirely on manual offline retraining.

[0033] Fourth, existing input features do not adequately utilize the internal statistical structure of multi-source sensor data. The safety status of industrial gases is not merely manifested as an increase or decrease in concentration values ​​from a single sensor; it may also exhibit various statistical structural changes, such as increased local fluctuations, altered synergistic relationships between multiple sensors, a shift in concentration distribution from a stable to a chaotic state, and deviations in periodic operating rhythms. Traditional methods, which only use raw sample values, first-order differences, second-order differences, or univariate moving averages as input, easily overlook the dispersion scale, synergistic relationships between variables, and the degree of system uncertainty within data segments. Statistical quantities such as variance, covariance, correlation, skewness, kurtosis, and Shannon entropy can describe the volatility, coupling, distribution pattern, and disorder of time series segments from different perspectives. Among these, Shannon entropy provides a classic theoretical basis for characterizing uncertainty and disorder. Therefore, it is necessary to construct a multi-dimensional statistical feature extraction mechanism oriented towards sliding time windows to release the discriminative information contained in multi-source sensor time series data without significantly increasing model complexity.

[0034] In summary, while existing technologies have accumulated certain expertise in areas such as physical simulation, traditional time series prediction, deep learning, multi-sensor fusion, statistical alarms, sensor drift compensation, and automatic time series feature extraction, a unified and lightweight solution for industrial safety monitoring scenarios is still lacking. This solution should be able to automatically extract compact features with clear statistical meaning from multi-source sensor time series data, achieve effective prediction on small-to-medium scale samples and resource-constrained edge devices, robustly handle sensor data quality degradation and known event disturbances, continuously monitor its own prediction performance and conditionally correct it when performance degrades, and achieve graded safety early warning based on prediction residuals.

[0035] In response to the aforementioned four common shortcomings, Figure 1 A flowchart illustrating a gas safety prediction method based on time-series modeling, according to an embodiment of the present invention, is shown. Figure 1 As shown, the gas safety prediction method based on time-series modeling includes the following steps: Step S101: Convert the raw data collected from multi-source and heterogeneous industrial sensors into a normalized data matrix based on a unified standard.

[0036] Step S103: Use multiple sliding analysis windows with different fixed widths to slide across the time series of the normalized data matrix to traverse the normalized data matrix. Perform various statistical measures on the data subsets that fall within the sliding analysis windows, and generate a fusion representation vector based on the results of the statistical measures.

[0037] Step S105: Construct a feedforward neural network prediction model using radial basis function kernel functions as hidden layer neurons.

[0038] Step S107: Train the feedforward neural network prediction model using the first part of the fused representation vector, perform accuracy calculation on the trained feedforward neural network prediction model using the second part of the fused representation vector that is later in time than the first part, and adjust the prediction results and / or prediction step size of the feedforward neural network prediction model in response to the result of the accuracy calculation being unacceptable.

[0039] Step S109: Accumulate the results of the accuracy calculation into a residual sequence, determine the cumulative error sum based on the residual sequence, and output the cumulative deviation warning level in a targeted manner based on the correlation between the cumulative error sum and the dynamic judgment boundary.

[0040] Step S111: Store the information of known events in the structured event knowledge base. In response to the current moment falling within the influence time domain of any known event, add the information of the known event to the fusion representation vector and / or adjust the size of the dynamic decision boundary.

[0041] Step S113: Use a feedforward neural network prediction model to predict gas safety and output the prediction results. Determine gas safety information based on the prediction results and the cumulative deviation warning level.

[0042] This invention aims to construct a general feature extraction framework that is decoupled from specific application scenarios (specific crops, gas combinations, or sensor topologies) and can automatically generate compact and highly physically / biologically interpretable features from any multivariate time-series data based solely on general statistical principles. Under small- to medium-scale training sample sets (the level of thousands to tens of thousands of records collected by typical agricultural / industrial gateways), this invention achieves a lightweight prediction model that significantly surpasses classical statistical models in prediction accuracy while being far less computationally efficient and less complex than deep neural networks, enabling efficient operation at edge devices without GPUs. This invention also enables the prediction model to sensitively perceive its own predictive performance degradation trends during long-term unattended autonomous operation and autonomously initiate conditional adaptive recalibration and repair procedures. Furthermore, this invention incorporates parameterized representations of pre-known structured environmental disturbances (such as scheduled ventilation and manure removal periods in pigsties, automatic irrigation / fertilization windows in greenhouses, planned maintenance in factories, and extreme weather warning windows for cold waves / rainstorms) into the prediction loop as beneficial prior knowledge, significantly reducing the probability of false alarms caused by known and expected fluctuations.

[0043] In some embodiments, converting raw collected data into a normalized data matrix includes: The independent timestamps used by the original collected data were adjusted to a unified timestamp based on Coordinated Universal Time (UTC). The node identifiers and unified timestamps of industrial sensors are used as composite primary keys to hierarchically sort the raw collected data, thereby constructing a normalized data matrix with monotonically increasing time series. In response to the original acquired data being a slowly varying signal with large inertia, piecewise linear interpolation is used to supplement the missing data in the normalized data matrix; in response to the original acquired data being a fast-responding transient signal, forward filling is used to supplement the missing data in the normalized data matrix. The faulty industrial sensor is identified based on the physically feasible range of the raw acquired data, and any data provided by the faulty industrial sensor is removed from the normalized data matrix.

[0044] The step of converting raw acquired data into a normalized data matrix aims to shield the heterogeneity between the raw outputs of industrial sensors from different manufacturers, using different communication protocols, with different sampling frequencies, and different encoding formats, transforming them into a unified normalized data matrix required for subsequent analysis. Specifically, this invention implements tentative automatic identification and decoding of multiple candidate encoding schemes, aligning the independent timestamps of each sensor node to the Coordinated Universal Time (UTC) standard. This invention uses a hierarchical sorting method with sensor node identifiers and sampling times as composite primary keys to construct a strictly time-monotonically increasing sequence. This invention employs differentiated missing value imputation strategies based on the physical characteristics of different types of sensor signals. Specifically, piecewise linear interpolation is used for slowly varying signals with high inertia (such as ambient temperature and humidity), while forward imputation is used for rapidly changing transient signals (such as the abrupt response of catalytic combustion combustible gas sensors) to avoid introducing artificial smoothing. This invention also performs validity checks based on the physically feasible value ranges of each measured parameter. For example, the lower flammability limit volume fraction of methane in air is about 4.9%, and the upper flammability limit is about 15.4%. When the monitoring reading exceeds these physical limits, it is determined that the sensor itself is faulty and is marked or eliminated.

[0045] In some embodiments, multiple statistical measures are applied to the subset of data falling within the sliding analysis window, and a fusion representation vector is generated based on the results of the statistical measures, including: A dispersion measure is applied to a subset of data to obtain an unbiased sample variance estimate, which is then used as a volatility index. Covariance correlation measures are performed on a subset of data to obtain the sample covariance matrix as a coupling index; By measuring the disorder of a subset of data, we obtain information entropy as a disorder index. A fusion representation vector is generated based on the volatility index, coupling index, and disorder index.

[0046] Instead of directly feeding the original high-dimensional sensor sampling vector sequence into the prediction model, this embodiment of the invention uses a sliding analysis window with a fixed width W to traverse the multivariate time-series signal matrix segment by segment along the time axis. For each subset of data falling into the window, three types of statistical measures with complementary physical interpretations are calculated. These measures are then vertically concatenated into a fusion representation vector that is much more compact than the original data but rich in discriminative structural information. This vector forms the input feature space for the subsequent prediction inference engine.

[0047] Let the width of the sliding analysis window be W sampling intervals (in this embodiment of the invention, it is set to have three scales, W = 3 days / 7 days / 14 days / 30 days), and the multi-sensor time series matrix be denoted as... Where T is the total number of sampling times and D is the number of sensor channels. For the t-th window slice starting from time index t... Extract the corresponding statistical measures according to the following three dimensions: Dimension 1, Scattering Measure (Volatility Index): Employs an unbiased sample variance estimator. The variance represents the average squared magnitude of the deviation of data points within a window from their local sample mean. This measure is highly sensitive to instantaneous fluctuations in sensor signals—when a sudden gas leak or abnormal equipment venting occurs in an industrial setting, the variance will exhibit a detectable step increase.

[0048] Dimension 2, Covariance Measure (Coupling Index): When the number of sensor channels D≥2, calculate the sample covariance matrix between each pair of the D channels within the window. The arithmetic mean of the absolute values ​​of the non-main diagonal elements of the matrix is ​​taken as a scalar summary of the intensity of covariance among multivariate sensor signals. This measure effectively captures changes in the covariance patterns among different gas types: under normal production conditions, the readings of each sensor follow a relatively stable correlation structure (e.g., methane concentration and oxygen concentration are usually negatively correlated, and combustion products CO and CO2 are positively correlated); when abnormal events cause changes in the location of the leak source, the composition of the leaked substance, or the diffusion path, the covariance pattern will produce a statistically detectable structural shift.

[0049] Dimension Three, Disorder Measurement (Chaos Index): Drawing on Shannon's entropy framework in information theory for quantitatively characterizing system uncertainty, the dynamic range of continuous data within the window is equally divided into K non-overlapping sub-intervals (K=12 in this embodiment). The relative frequency of the number of sample points contained in each sub-interval relative to the total number of samples in the window is used as the probability estimate. Substituting into Shannon's definition of entropy Calculate information entropy. This measure reflects the degree of uniformity in the distribution of sensor signal data: when the production equipment is operating in a steady state, the sensor readings are concentrated in a relatively narrow range of values ​​(exhibiting a low-entropy ordered state); while when multiple abnormal physical and chemical processes occur in combination, the sensor signals exhibit large-scale disordered fluctuations (exhibiting a high-entropy chaotic state).

[0050] In some embodiments, performing multiple statistical measures on the data subsets falling within the sliding analysis window, and generating a fusion representation vector based on the results of the statistical measures, further includes: Perform statistical analysis on a subset of data to obtain one or more of the following: arithmetic mean, maximum and minimum values, fourth-order normalized moments of the distribution, third-order normalized moments of the distribution, sine transform of the hour index of the daily period, and cosine transform of the hour index of the daily period. A fusion representation vector is generated based on one or more of the following: volatility index, coupling index, disorder index, arithmetic mean, maximum and minimum values, fourth-order normalized moments of the distribution, third-order normalized moments of the distribution, sine transform of the hour index of the daily cycle, and cosine transform of the hour index of the daily cycle.

[0051] Based on the aforementioned three core statistical measures, this embodiment of the invention further extracts four auxiliary descriptive statistics (arithmetic mean of data within the window, maximum value of data within the window, fourth-order standardized moment of the distribution—kurtosis coefficient, third-order standardized moment of the distribution—skewness coefficient) and two cyclic transformation features for encoding periodic time-varying patterns (sine and cosine transformations according to the hourly index h of the daily cycle).

[0052] In this embodiment of the invention, 15 scalar values ​​are multiplied by 4 time scales, and 7 AR lags, 3 rates of change, 1 EMA, and 5 time codes are added to form a d=76-dimensional fused feature vector. Compared to the original window data matrix, this fused representation achieves more effective information condensation than the original data. At the same time, it condenses and retains the essential structural information in the original data that is most valuable for the prediction task in a structured statistical measure with clear physical interpretation.

[0053] In some embodiments, multiple statistical measures are applied to the subset of data falling within the sliding analysis window, and a fusion representation vector is generated based on the results of the statistical measures, including: Locality-sensitive hashing is performed on a subset of data. If adjacent subsets of data have the same locality-sensitive hashing, it is determined whether there is redundant data. If redundant data is determined to exist, the redundant data is removed from the normalized data matrix. The mean and standard deviation of the entire sequence are calculated for all sampled data in the normalized data matrix. In response to the detection that the absolute value of the normalized deviation of a sampled point is greater than 3 times the standard deviation, the sampled point is marked as a candidate outlier observation point. In response to the existence of multiple candidate anomaly observation points, the system determines whether the multiple candidate anomaly observation points fall within the influence time domain of any known event recorded in the structured event knowledge base. In response to falling within the influence time domain, the anomaly detection sensitivity coefficient acting on the influence time domain is reduced. In response to not falling within the influence time domain, the system performs noise reduction filtering on the multiple candidate anomaly observation points. The fused representation vector is generated based on the results of statistical measures performed on subsets of data in the normalized data matrix after removing redundant data and performing noise reduction filtering.

[0054] This invention also includes data quality control and anti-interference processing. First, redundant segments are identified and removed based on Locality-Sensitive Hashing (LSH) fingerprints. The LSH encoding of continuous window segments sliding along the time axis is calculated. By comparing the hash fingerprints of adjacent analysis windows, redundant data transmission caused by abnormal cache refresh logic in the data acquisition system or duplicate forwarding mechanisms in the communication gateway is detected. Once a duplicate segment is identified, it is excluded from subsequent analysis links, and an exclusion log record containing a timestamp and sensor identifier is generated for traceability and auditing by equipment maintenance personnel.

[0055] Then, the PauTa Criterion is used for initial screening of suspected outliers. After identifying and removing redundant segments, the mean μ and standard deviation σ of the entire sequence are calculated. If the absolute value of the standardized deviation of a sampling point exceeds... (Right now The data point is then marked as a "candidate abnormal observation" and retained instead of being immediately removed.

[0056] Finally, a context-aware noise-event binary adaptive discriminator is used. When N consecutive sampling points (N≥4 in this embodiment, this threshold can be adjusted according to the actual sensor sampling frequency and the dynamic response characteristics of the measured process) are all marked as candidate anomalies, the context discrimination logic is triggered. The system retrieves a pre-built structured event knowledge base (covering known event types and their impact parameter templates, such as planned maintenance shutdown windows, concentrated shutdown-resumption transition periods during statutory holidays, high temperature / cold wave / rainstorm warning windows issued by meteorological departments, and seasonal process switching nodes), and determines whether the current candidate anomaly data segment falls within the time domain of the impact of a known event pre-recorded in the knowledge base. If the match is successful, the data segment is classified as "expected normal fluctuations associated with the event" rather than a true anomaly and is retained. At the same time, the anomaly discrimination sensitivity coefficient within this time period is proportionally reduced according to the parameter template of this event type (i.e., the warning trigger threshold is increased). If the match fails, it is determined to be an artifact caused by sensor hardware failure or electromagnetic interference from an unknown environment, and adaptive median filtering or soft threshold noise reduction based on wavelet coefficients is performed.

[0057] In some embodiments, constructing a feedforward neural network prediction model using radial basis function kernel functions as hidden layer neurons includes: The K-Means clustering algorithm is used to perform unsupervised clustering in the feature space of a feedforward neural network prediction model to determine the optimal spatial location of the center point of the kernel function. The width parameter of the kernel function is determined based on the feature distance between the center points of the kernel function and the preset scaling factor; The radial basis kernel function of the feedforward neural network prediction model is determined based on the center point and width parameters of the kernel function. The optimal solution for the output layer weights in the least-squares sense is obtained using the closed-form formula of the Moore-Penrose generalized inverse matrix. The feedforward neural network prediction model responds to the received fused representation vector as model input, and generates a scalar prediction as model output based on the activation function of the hidden layer neurons composed of radial basis kernel functions and the linear relationship with the output layer weights.

[0058] This invention adopts a feedforward architecture using the Radial Basis Kernel Function (RBF, also known as the Gaussian kernel function) as the activation mechanism for hidden neurons, instead of the more popular deep convolutional or gated recursive architectures in the machine learning community. This prudent choice of technical approach aims to achieve perfect implementation in agricultural IoT and low-cost industrial edge hardware.

[0059] First, the model capacity of the feedforward architecture is highly adaptable to small sample sizes. The total number of freely adjustable parameters in the RBF network is mainly determined by the number of hidden kernel center points K (typically ranging from K=40 to 120, with the optimal configuration in this embodiment being K=80). Under typical accumulation scales of industrial sensor monitoring data (thousands to tens of thousands of valid records), this parameter scale can achieve a good bias-variance trade-off. In contrast, deep recurrent networks or Transformer-like architectures typically carry tens of thousands to millions of trainable parameters. In industrial scenarios with limited training data and extremely imbalanced positive and negative samples (extremely scarce accident data), they are prone to severe overfitting and degraded generalization performance due to over-parameterization.

[0060] Secondly, the training computation process of the feedforward architecture has low resource consumption. The parameter identification process of the RBF network is divided into two sub-stages: first, the optimal spatial positions of the center vectors of the K kernel functions are automatically located in the feature space using the K-Means clustering analysis algorithm; then, the optimal solution of the output layer weight vector in the least squares sense is obtained in a one-time closed-form calculation using Moore-Penrose generalized inverse matrix operations. The entire training process does not rely on the iterative gradient descent optimization of the error back-propagation algorithm, and the computational complexity is O(log n). The training can be completed in seconds to tens of seconds on a general-purpose central processing unit (CPU) platform, without the need for expensive acceleration hardware such as graphics processing units (GPUs) or dedicated tensor processors (TPUs).

[0061] Furthermore, the feedforward architecture possesses the ability to match local receptive field response characteristics with operating conditions. The activation function of the RBF hidden layer neurons adopts the mathematical form of a Gaussian kernel: This activation function endows each hidden layer neuron with a "spatially localized" response—at its nucleus center. Strong activation outputs are generated within a finite feature space region surrounding the central region, while the response intensity to input patterns far from this central region decays rapidly to near zero at an exponential rate. This local receptive field characteristic makes RBF networks naturally suitable for performing condition-prototype matching industrial time series predictions: the network can learn the typical prototype locations in the feature space for different types of operating conditions (such as normal steady-state operation of equipment, gradual heating phase, rapid pressure increase phase, emergency discharge phase, etc.) and perform adaptive weighted fusion prediction based on the kernel space similarity between the current input feature vector and each operating condition prototype.

[0062] Kernel width parameter The determination strategy is as follows: for each determined center point Calculate its distance to other K The minimum value in the set of Euclidean distances to a given center point is taken. The preset scaling factor α (in this embodiment of the invention, it is set to 0.5 to 1.5, preferably) is used. The product of ), i.e. This heuristic strategy ensures that each kernel function has an appropriate activation response radius within its responsible feature space sub-region, avoiding both response ambiguity (pattern confusion) caused by excessive overlap between kernels and feature space coverage blind spots (missed detections) caused by excessive discretization between kernels.

[0063] Final scalar prediction output The linear weighted combination of the activation outputs of neurons in each hidden layer: ,in Let b be the output weight of the k-th kernel function, and b be the global bias scalar.

[0064] In some embodiments, the second part is used to perform accuracy calculation, and adjusting the prediction results and / or prediction step size of the feedforward neural network prediction model in response to the result of the accuracy calculation being unacceptable includes: The reference window interval and the prediction window interval are determined. The real data with a length equal to the reference window interval is truncated from the second part and input into the feedforward neural network prediction model. The prediction data with a length equal to the prediction window interval is obtained by forward extrapolation. The symmetric mean absolute percentage error is calculated between the predicted data and the actual data corresponding to the predicted data in the second part. In response to the symmetric mean absolute percentage error exceeding the acceptable accuracy threshold, the average residual within the prediction window interval is added to the predicted data of the next prediction window interval, and / or the length of the prediction window interval is shortened.

[0065] Existing prediction systems generally employ a static operational paradigm: single offline model training → model parameter freezing → long-term fixed deployment. Once deployed, the model enters a "parameter freeze" state, losing any ability for autonomous evolution or adaptive adjustment. This invention fundamentally changes this inherent paradigm, endowing the system with self-evolution and continuous self-awareness and repair capabilities under unattended conditions. This invention's operational sequence is divided into two stages: Phase 1: Base Model Pre-training with Historical Accumulated Data. The first 75% of all available historical data records, ordered chronologically (strictly maintaining temporal causality constraints to prevent future information leakage into the training process), are used as the base model pre-training dataset. An enhanced RBF network model with K=80 kernel center points and embedded linear skip connections is trained on this dataset to enhance linear pattern capture capabilities. Ridge regularization α=1.0 is used to prevent overfitting. The optimization objective of this phase is to enable the model to fully learn the long-term statistical regularities of the multi-sensor coupling temporal evolution process in the target industrial scenario, covering intraday periodic work shift-rest shift fluctuation patterns, weekly workday-restday differentiated emission characteristics, the systematic modulation effect of annual seasonal temperature and humidity cycles on sensor baseline readings, and the regular and predictable signal disturbance patterns caused by typical process operations (such as material input, product output, equipment switching, and maintenance purging).

[0066] Phase Two: Online Sequential Sliding Window Prediction and Adaptive Verification Loop. On the remaining 25% of recent data (representing the actual online prediction task environment faced after model deployment), the following recursive iterative procedure is executed: (a) Window positioning. Define the reference analysis window interval at the current time index τ. and the forecast forward window interval In this embodiment of the invention, it is configured as follows: Each sampling step size, The sampling step size. When the sampling period is 1 minute, this configuration corresponds to extrapolating the gas concentration evolution trend for the next approximately 15 minutes using historical observation data from the past approximately 45 minutes. The 45:15 ratio is a compromise optimal value determined after systematic cross-validation screening of different ratio schemes—the shorter sampling step. Insufficient to capture sufficient temporal dependency structures, longer This can also cause the model to respond slowly to recent local changes.

[0067] (b) Forward extrapolation. Using the base RBF network model trained in the first stage, extrapolate the predictions within the look-ahead window. Perform forward inference calculations one by one in consecutive future moments to obtain a sequence of predicted values. .

[0068] (c) Accuracy verification. Compare predicted values ​​point by point. Compared with the actual recorded value The deviation between them. The Symmetric Mean Absolute Percentage Error (SMAPE) is used as the core accuracy evaluation index, defined as: Furthermore, by employing the normalized root mean square error (nRMSE) as an auxiliary evaluation metric, SMAPE demonstrates good numerical stability when the true value is close to zero.

[0069] (d) Adaptive Decision Branching. The calculated SMAPE value is compared with the preset acceptable accuracy upper limit δ (set to δ=45% in this embodiment), and branching is performed according to the following differentiated processing logic. If SMAPE<δ, it indicates that the current model's tracking and prediction performance of recent data evolution patterns remains within an acceptable tolerance range, and the system accepts the prediction results of this window with confidence. The window reference time index τ is advanced forward by a full step. The system updates the bias correction term based on the exponentially weighted moving average of recent residuals (set to EMA coefficient = 0.3 in this embodiment) and enters the next prediction cycle. If SMAPE ≥ δ, it indicates that the model's adaptability to recent data patterns has degraded, triggering the conditional bias correction mechanism. The system calculates the average bias of the prediction residuals within the current reference analysis window and adds the bias correction term to the prediction output of the next window. If the accuracy after correction is still insufficient, the window reference time index τ is advanced step = max( After 5 steps, only the first step values ​​of the predicted results are added to the summary. This bias correction mechanism has the advantage of completely retraining the RBF network: extremely low computational cost (only matrix multiplication is required, without needing to solve K-Means clustering and pseudo-inverse matrices), and correction can be completed in a very short time, making it particularly suitable for industrial edge deployment scenarios with stringent real-time requirements. Simultaneously, this bias correction event is recorded in the system operation log.

[0070] Thus, a complete real-time feedback control loop has been constructed, from "prediction generation → accuracy auditing → decision branching → parameter adaptive adjustment → new round of prediction generation". The model is no longer a static function mapper that passively accepts fixed parameters and unidirectionally outputs prediction results, but has been upgraded to an autonomous intelligent agent with continuous self-examination and conditional self-correction capabilities. From a cybernetics perspective, this mechanism has a profound structural homology with the classic "perception → decision → execution" adaptive control closed-loop paradigm. Mathematically, it can be understood as a piecewise local steady-state approximation and dynamic online tracking of the slowly time-varying statistical characteristics in non-stationary time-series signals.

[0071] The "prediction-verification-recalibration" closed-loop adaptive mechanism proposed in this step differs from all currently known industrial / agricultural gas monitoring and prediction methods. Existing methods either lack adaptive update capabilities entirely (static models) or can only achieve passive, non-intelligent model updates through manually scheduled periodic offline retraining. The method of this invention endows the model with metacognitive capabilities to continuously monitor its own predictive performance and autonomously take corrective actions when unacceptable performance degradation is detected, without any external human intervention. This autonomous characteristic has extremely important engineering practical value for industrial safety monitoring systems deployed at remote, unattended sites (such as desert oil and gas wellheads, high-altitude weather stations, and deep-sea oil platforms).

[0072] In some embodiments, the cumulative error sum is determined based on the residual sequence, and the cumulative deviation warning level is output selectively based on the correlation between the cumulative error sum and the dynamic decision boundary, including: The residual mean and residual standard deviation are determined based on the residual sequence. Using the residual mean, residual standard deviation and preset acceptable offset reference value, the positive cumulative sum and negative cumulative sum of the residuals are calculated for each symmetrical mean absolute percentage error. The maximum value of the positive cumulative sum and negative cumulative sum is used as the anomaly discrimination statistic. A low-priority review notification is output in response to the anomaly detection statistic falling within the first interval of the dynamic decision boundary; A medium-priority inspection alarm is output in response to the anomaly detection statistic falling within the second interval of the dynamic judgment boundary; A high-priority emergency response plan is output in response to the anomaly detection statistic falling within the third interval of the dynamic judgment boundary.

[0073] Obtain the predicted residual sequence Subsequently, this embodiment of the invention employs the CUSUM (Cumulative Sum) control chart methodology framework, which has been widely validated in engineering practice within the field of Statistical Process Control (SPC), for online detection of abnormal states. The core advantage of the CUSUM method lies in its ability to amplify small but persistent (rather than occasional instantaneous spikes) offsets through targeted summation of standardized residuals. This allows gradual, progressive anomalies to be effectively identified in their early stages—that is, before their absolute values ​​are sufficient to trigger a fixed threshold alarm.

[0074] First, estimate the mean of the residuals based on historical residual samples (or initial adaptation period samples from online operation). and standard deviation The prediction residual for each new arrival time t Then, the CUSUM statistics in both directions are updated recursively: Positive cumulative sum (detecting anomalies and upward deviation):

[0075] Negative cumulative sum (detecting anomalies and downward deviation):

[0076] in, .

[0077] in The Allowable Slack is a reference value for acceptable offsets. Its physical meaning is to define the boundary between acceptable normal random fluctuations and systematic offsets that need to be addressed. The value ranges from 0.02 to 0.50. The specific value needs to be calibrated based on the noise level of the actual data and the desired detection sensitivity. Lower values ​​are preferred. Higher values ​​make detection more sensitive but may increase the false alarm rate. The values ​​are the opposite. The configuration of this embodiment of the invention takes... At each time t, take the maximum value of the two-sided cumulative sum. As a comprehensive anomaly detection statistic. Based on Relative to dynamic decision boundary To determine the degree of exceedance, a three-tiered progressive early warning strategy will be implemented: Attention Level: Early signs of a systematic accumulation of prediction bias have emerged, prompting the system to issue a low-priority notification, recommending that on-site maintenance personnel arrange for subsequent manual review.

[0078] Warning level: The cumulative deviation is showing a continuous expansion trend, indicating that a medium-to-high probability industrial anomaly is occurring. The system has issued a medium-priority alarm, prompting an on-site inspection to be arranged as soon as possible.

[0079] Operational level: The cumulative deviation has significantly exceeded the safety tolerance boundary, and there is a very high probability that there are signs of an impending accident that are continuously deteriorating (such as continuous gas leakage and escalating smoldering). The system has issued the highest priority emergency alarm, and it is recommended to activate the emergency plan immediately.

[0080] Compared to the fixed threshold over-limit alarm method and 3σ statistical rule still widely used in industry, the CUSUM framework has the following significant advantages: significantly higher sensitivity to detecting gradual shifts and persistent small deviations (stronger statistical power); and through bidirectional statistics (C... + and C -The design can simultaneously capture two distinctly different abnormal states: positive abnormal events (abnormally high gas concentration, which may correspond to leakage or combustion) and negative abnormal events (sudden decrease or failure of sensor response sensitivity, which may correspond to sensor hardware failure and require maintenance and replacement). The three-level progressive classification enables safety response measures to be reasonably matched with the severity of the abnormality, avoiding overreaction or underreaction caused by a one-size-fits-all response.

[0081] In some embodiments, storing information about known events in a structured event knowledge base, and adding information about known events to the fusion representation vector and / or adjusting the size of the dynamic decision boundary in response to the current time falling within the influence time domain of any known event; The impact factors, time-domain range of impact, and anomaly detection sensitivity coefficients of industrial sensors affected by known events are stored in a structured event knowledge base; In response to the current moment falling within the influence time domain of any known event, the influence factor, influence time domain range, and anomaly detection sensitivity coefficient of the industrial sensor are concatenated to the end of the fused characterization vector, and / or the anomaly detection sensitivity coefficient is used to amplify the size of the dynamic decision boundary.

[0082] This invention integrates structured event information that can be predicted in advance in the external environment (such as planned shutdown and maintenance periods registered in the equipment management system, concentrated shutdowns and post-holiday resumption periods recorded in the personnel attendance system, extreme weather warnings issued by the meteorological department, and seasonal process switching schedules arranged by the production planning department) into a data-driven prediction-detection process in the form of parameterized structured prior knowledge, thereby realizing a hybrid enhancement paradigm of domain prior knowledge guidance and data-driven learning.

[0083] Specifically, during the long-term operation of the system, a structured event type knowledge base that can be dynamically added, deleted, and modified is maintained. Each event type record in this knowledge base contains the following parameter tuples: (a) a list of target sensor channels affected by the event and the expected impact multiplier factor for each channel (e.g., a purging operation during planned equipment maintenance is expected to cause a temporary increase in combustible gas concentration to several times the normal level); (b) an estimated duration of the typical impact of the event (based on statistical analysis of historical records or judgment by domain experts); and (c) an adjustment coefficient γ for anomaly detection sensitivity, with a value range of (0,1]. γ=1.0 indicates no sensitivity adjustment (maintaining baseline sensitivity completely), and a smaller γ value indicates a greater decrease in sensitivity (i.e., a greater tendency to not trigger high-confidence alarms within the event window).

[0084] When the operating system detects that the current time falls within the time window of a registered external event (for example, when the pigsty is undergoing planned manure cleaning and powerful ventilation, during which the concentrations of ammonia and hydrogen sulfide will inevitably experience a predictable and normal brief increase), two adjustment operations are executed in parallel: (1) The structured parameter feature vector of the event (containing numerical features such as the expected impact multiplier of each affected channel, the duration of the event, and the remaining estimated duration of the event) is appended to the statistical fusion feature vector. At the end, the input vector dimension of the RBF network is dynamically expanded from d=76, enabling the network to incorporate auxiliary information from the event context during prediction; (2) Multiply the dynamic decision boundary H of the CUSUM control chart by 1 / γ to amplify the value (because γ<1), thereby appropriately reducing the sensitivity of the system to issue high-confidence abnormal alarms during the event period, so as to avoid repeated false alarms due to predictable known environmental disturbances, and prevent maintenance personnel from experiencing alarm fatigue due to frequent false alarms.

[0085] At the methodological level, its essence lies in injecting the structured prior knowledge of domain experts into the data-driven statistical reasoning process in the form of regularized constraints: during known events, the system is given an appropriate judgment tolerance, retaining a margin of tolerance for expected fluctuations; while outside the known event window, the system maintains a high-sensitivity alert state, ensuring that it does not lose its ability to detect truly sudden and unknown dangerous events. This differentiated sensitivity control strategy embodies the core concept of risk-based hierarchical control in safety management practice.

[0086] The specific embodiments of the present invention are further described below with reference to specific examples.

[0087] In step S101, the system's underlying input interface module accesses the multi-source heterogeneous environmental monitoring sequence via network protocol or local file system. For massive historical observation data records (sampling frequency once a day) containing national control monitoring stations at the prefecture level and above, the system automatically uses a specific character set encoding (such as GBK) to perform penetrating reading of the comma-separated values ​​(CSV) data stream.

[0088] Specify the concentration of a certain type of harmful particulate matter in a specific target city (e.g., Beijing). Daily concentration, in units ) as the target dependent variable for prediction And construct a continuous, equally spaced one-dimensional time series using timestamps as the unique index.

[0089] In step S103, in order to completely solve the defects of traditional RBF neural networks in that they are prone to "severe underfitting" or "temporal lag" for nonlinear environmental signals under a single sliding window, the system constructs a multi-scale 3-layer feature space mapping through the low-level feature engine, transforming the original one-dimensional time series... Powerful upgrade to A high-value feature matrix X. The strategy for constructing feature operators is as follows: 1. Autoregressive Lag Operator Layer (AR-Lags): Extracts the physical inertia of the preceding term at the current time step, and explicitly extracts historical observations from the past 14 independent sampling steps as explicit basis features:

[0090] 2. Multi-scale Windows: Configures a diverse set of time sliding window widths. Within each scale window, the system computes two types of dynamic statistical features in parallel: the rolling mean operator, which captures the macroscopic evolutionary trends of both long and short periods; and the rolling standard deviation operator, which captures the intensity and uncertainty of local environmental fluctuations.

[0091] 3. Dynamic Difference and Smoothing Trend Operator Layer (Delta & EMA): Calculates first-order one-dimensional temporal differences ( To capture the instantaneous acceleration of timing signals; a smoothing coefficient is introduced. The exponentially weighted moving average (EMA) is used as a low-pass filter to isolate stable environmental basis components; discrete periodic features such as year, month, day, day of the week, and day of the year are extracted from the timestamp to establish a comprehensive time context association mapping, and finally spliced ​​to generate a 76-dimensional wide table feature vector.

[0092] Before the feature matrix is ​​fed into the network, deduplication (Dedup) is also performed. Statistical scalar filtering (3-SigmaFilter) and event-triggered cleaning. Deduplication refers to the automatic removal of overlapping records in a sequence based on the principle of monotonically increasing timestamps. Statistical limiting filtering calculates the residual distribution within the sliding segment and automatically truncates and replaces faulty pixels in hardware sensors that deviate from the mean by more than three standard deviations. Event-triggered cleaning automatically triggers linear or cubic spline interpolation to complete the data when a large area of ​​time sequence is detected. After this processing, data contamination is removed, and the sample set is accurately divided into the first 75% (2251 days) of the independent historical evolution set and the latter 25% (751 days) of the online independent test set.

[0093] In step S105, the nonlinear kernel function space is mapped to a feedforward neural network, introducing a hidden layer self-pruning mechanism and a Ridge Regression (RBF) regularized pseudo-inverse solution mechanism onto the traditional RBF structure. The network topology is as follows: the input layer receives a D=76-dimensional feature vector, and the hidden layer uses a Gaussian radial basis function kernel.

[0094] in For the hidden core center, The radius is the core radius.

[0095] In the historical pre-training, the number of hidden layer centroids was preset to K=80. The K-Means clustering algorithm was used to iteratively cluster in the X space, and the resulting 80 cluster centroids were directly used as the primary kernel centers. The system calculates each center The average Euclidean distance to the five nearest neighboring nucleus centers is used as an adaptive parameter for selecting the nucleus radius of a specific neuron. During the forward propagation of the training matrix, the system monitors the average activation level of each neuron in the hidden layer in real time. If the output activation level of a neuron remains below a set confidence threshold for an extended period across the entire sample, ... (That is, neuronal death occurs. In this embodiment, due to the high-dimensional sparsity of features in the initial base model, 6 dead neurons were automatically and accurately detected and pruned, leaving 74 active neurons.) The system will perform online structural pruning to clear its output weights to zero. To eliminate overfitting caused by collinearity of hidden layer matrices, the output layer weight matrix... It abandons the traditional unstable gradient descent update and instead uses a method with... Regularization penalty term (ridge regression coefficient) The pseudo-inverse closed-form solution can be obtained directly in one step:

[0096] In step S107, the system executes a unique confidence verification control loop based on a sliding window controller. This control loop operates fully automatically during the 751-day online blind test. The initial window occupies 56 days, with the remaining 695 days dedicated to blind testing. In this unattended state, the system autonomously adopts 13 full-window standard advances (each advance lasting 15 days, totaling 13 × 15 = 195 days) and sensitively captures 100 instances of operating condition variations triggering local recalibration (each cautious advance lasting 5 days, totaling 100 × 5 = 500 days). The total number of advance days perfectly matches: 195 days + 500 days = 695 days. Through this non-overlapping slicing and continuous extraction, a final summary of 695 high-fidelity, completely uncontaminated, independent sequential prediction points is generated.

[0097] In step S109, based on 695 high-fidelity, misalignment-free prediction residual sequences ( The system cascades the CUSUM module for statistical process control. Parameter configuration sets the standard value for the average residual drift. residual standard deviation The cumulative sum of the positive and negative biases calculated in parallel is:

[0098] The system uses a pyramid-style three-level hierarchical interception output. The system determines the control threshold based on residual deviation. Spatial interception depth, automatically triggering graded responses: Attention Level: Triggered 43 times in total. This indicates a slight perturbation in the residuals, suggesting a potential slight drift in the environmental state.

[0099] Warning level: Triggered 22 times in total. This indicates that the residual deviation exhibits significant linear accumulation, prompting the system to issue an early warning and request intervention.

[0100] Action level: Triggered a total of 66 times. This indicates that the cumulative effect and indicators have strongly broken through the critical boundary. The system accurately intercepted high-risk periods of change caused by severe weather or sudden leaks of hazardous gases, and immediately issued alarm action commands to the security terminal.

[0101] During online operation, the system detected data drift and initiated recalibration ([RECAL]) with local incremental adjustments and step size reduction a total of 100 times. Among these, 66 instances showed severe residual fluctuations that strongly breached the CUSUM red critical boundary, directly triggering the highest-level red action warning and initiating retraining. The remaining 34 recalibration instances occurred when the residuals did not breach the red boundary, but the sliding window accuracy index (SMAPE) temporarily exceeded the acceptable upper limit. The system operates during a slow, drifting phase. This dual interception mechanism ensures that both "sudden, strong mutations" and "gradual, slow drifts" can be automatically refreshed and captured by the system.

[0102] In step S111, the event trigger is activated upon receiving an alert or action signal. The system sends a control interrupt command to the sliding window controller, forcibly changing the number of hidden kernel functions in the next stage of online adaptive operation from the base. Fine-tuning to achieve high specificity Meanwhile, the sliding window fine-tuning bias correction coefficient was temporarily increased from 0.3 to 0.6 to enhance the model's tracking accuracy for extreme value fluctuations and form a complete control closed loop.

[0103] The embodiments of the present invention have achieved significant technical effects. First, in an independent blind test lasting 751 days, the determination coefficient of the present invention was [missing information]. It surged strongly to +0.4432. In time series forecasting with high dynamic noise, this was achieved by completely eliminating spurious "temporal overlap self-contamination". It can reliably approach 0.45, and The powerful compression to 23.3 objectively proves that the combination of the multi-scale feature space and the two-stage RBF network in this invention can capture the nonlinear multi-temporal variation features hidden in the time series with extremely high accuracy.

[0104] Second, in the embodiment of the present invention, under the same 45 / 15 sliding window architecture, replacing the RBF network with a linear ridge regression sliding window baseline (LR-Sliding) completely collapses on the test set. The value plummeted to negative (-1.7270), indicating that the linear model completely lacks long-term forward forecasting capabilities; simultaneously, the coefficient of determination compared to the strongest physical inertia baseline (Persistence)... It achieved a significant relative increase of 41.5%.

[0105] Third, the full-process data alignment closed loop of the embodiments of the present invention ( With 100 high-fidelity prediction points, it perfectly produces 695 residual inputs to CUSUM, captures 131 reasonable early warning points in a graded manner (with a trigger rate of 18.8%), and has excellent graphic effects. This fully demonstrates that as a time series modeling solution for general industrial safety monitoring scenarios, it has extremely high generalization survivability and engineering deployment value.

[0106] Figure 2 A block diagram of a gas safety prediction system based on time-series modeling, according to an embodiment of the present invention, is shown. Figure 2 As shown, the gas safety prediction system based on time-series modeling includes: The preprocessing module 21 converts the raw data acquired from multi-source and heterogeneous industrial sensors into a normalized data matrix based on a unified standard. The feature extraction module 22 uses multiple sliding analysis windows with different fixed widths to slide across the time series of the normalized data matrix to traverse the normalized data matrix. It performs various statistical measures on the data subsets falling within the sliding analysis windows and generates a fusion representation vector based on the results of the statistical measures. Model building module 23 constructs a feedforward neural network prediction model using radial basis kernel functions as hidden layer neurons. The adaptive module 24 trains the feedforward neural network prediction model using the first part of the fused representation vector, performs accuracy calculation on the trained feedforward neural network prediction model using the second part of the fused representation vector that is later in time than the first part, and adjusts the prediction results and / or prediction step size of the feedforward neural network prediction model in response to the result of the accuracy calculation being unacceptable. The early warning module 25 accumulates the results of accuracy calculation into a residual sequence, determines the cumulative error sum based on the residual sequence, and outputs the cumulative deviation early warning level based on the correlation between the cumulative error sum and the dynamic judgment boundary. Prior knowledge module 26 stores information about known events in a structured event knowledge base. In response to the current moment falling within the influence time domain of any known event, it adds information about that known event to the fusion representation vector and / or adjusts the size of the dynamic decision boundary; and Module 27 implements a gas safety prediction model using a feedforward neural network and outputs the prediction results. Based on the prediction results and the cumulative deviation warning level, it determines the gas safety information.

[0107] Finally, it should be noted that those skilled in the art will understand that all or part of the processes in the time-series modeling-based gas safety prediction method can be implemented by a computer program instructing related hardware. The program for the time-series modeling-based gas safety prediction method can be stored in a computer-readable storage medium. When executed, the program can include the processes described in the embodiments of the methods above. The storage medium for the program can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. The embodiments of the computer program described above can achieve the same or similar effects as any of the corresponding foregoing method embodiments.

[0108] Furthermore, the gas safety prediction method based on time-series modeling disclosed in the embodiments of the present invention can also be implemented as a computer program executed by a processor, which can be stored in a computer-readable storage medium. When the computer program is executed by the processor, it performs the functions defined in the method disclosed in the embodiments of the present invention.

[0109] In one or more exemplary designs, functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, functionality may be stored as one or more instructions or code on or transmitted via a computer-readable medium. Computer-readable media include computer storage media and communication media, including any medium that facilitates the transfer of a computer program from one location to another. Storage media may be any available medium accessible to a general-purpose or special-purpose computer. By way of example, and not limitation, computer-readable media may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage devices, disk storage devices or other magnetic storage devices, or any other medium that may be used to carry or store the required program code in the form of instructions or data structures and is accessible to a general-purpose or special-purpose computer or a general-purpose or special-purpose processor. Furthermore, any connection may be appropriately referred to as computer-readable media. For example, if software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DOL), or wireless technologies such as infrared, radio, and microwave, then the aforementioned coaxial cable, fiber optic cable, twisted pair, DOL, or wireless technologies such as infrared, radio, and microwave are all included in the definition of media. As used herein, disks and optical discs include compact discs (CDs), laser discs, optical discs, digital multifunction discs (DVDs), floppy disks, and Blu-ray discs, where disks typically reproduce data magnetically, while optical discs reproduce data optically using lasers. Combinations of the above should also be included within the scope of computer-readable media.

[0110] The above are exemplary embodiments disclosed in this invention. However, it should be noted that various changes and modifications can be made without departing from the scope of the embodiments of this invention as defined by the claims. The functions, steps, and / or actions of the methods according to the disclosed embodiments described herein do not need to be performed in any particular order. Furthermore, although the elements disclosed in the embodiments of this invention may be described or claimed individually, they may be understood as multiple unless explicitly limited to a singular number.

[0111] It should be understood that, as used herein, the singular form “a” is intended to include the plural form as well, unless the context clearly supports an exception. It should also be understood that, as used herein, “and / or” refers to any and all possible combinations of one or more of the associated listed items.

[0112] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples. Within the framework of the invention, technical features of the above embodiments or different embodiments can be combined, and many other variations of different aspects of the invention exist, which are not provided in the details for the sake of brevity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. A gas safety prediction method based on time-series modeling, characterized in that, Perform the following steps: Convert raw data acquired from multi-source and heterogeneous industrial sensors into a normalized data matrix; Multiple sliding analysis windows with different fixed widths are used to slide over the time series of the normalized data matrix to traverse the normalized data matrix. Various statistical measures are performed on the data subsets falling within the sliding analysis windows, and a fusion representation vector is generated based on the results of the statistical measures. A feedforward neural network prediction model is constructed using radial basis function kernel functions as hidden layer neurons; The feedforward neural network prediction model is trained using the first part of the fusion representation vector, and the accuracy of the trained feedforward neural network prediction model is calculated using the second part of the fusion representation vector that is later than the first part in the time series. The prediction results and / or prediction step size of the feedforward neural network prediction model are adjusted according to the accuracy calculation results. The results of the accuracy calculation are accumulated into a residual sequence. The cumulative error is determined based on the residual sequence. The cumulative deviation warning level is output in a targeted manner based on the correlation between the cumulative error and the dynamic judgment boundary. Information about known events is stored in a structured event knowledge base. In response to the current moment falling within the influence time domain of any known event, information about that known event is added to the fusion representation vector and / or the size of the dynamic decision boundary is adjusted. The feedforward neural network prediction model is used to predict gas safety and output the prediction results. Gas safety information is determined based on the prediction results and the cumulative deviation warning level.

2. The method according to claim 1, characterized in that, Converting the raw collected data into the normalized data matrix includes: The independent timestamps used by the original collected data are adjusted to a unified timestamp based on Coordinated Universal Time (UTC). The node identifier of the industrial sensor and the unified timestamp are used as a composite primary key to perform hierarchical sorting of the original collected data, thereby constructing the normalized data matrix with monotonically increasing time series. In response to the original acquired data being a slowly varying signal with large inertia, piecewise linear interpolation is used to supplement the missing data in the normalized data matrix; in response to the original acquired data being a fast-responding transient signal, forward filling is used to supplement the missing data in the normalized data matrix. The faulty industrial sensor is identified based on the physically feasible range of the original acquired data, and any data provided by the faulty industrial sensor is removed from the normalized data matrix.

3. The method according to claim 1, characterized in that, Multiple statistical measures are applied to the subset of data falling within the sliding analysis window, and the fusion representation vector is generated based on the results of the statistical measures, including: A dispersion measure is applied to the data subset to obtain an unbiased sample variance estimate, which is used as a volatility index. Perform a covariance correlation measure on the data subset to obtain the sample covariance matrix as a coupling index; The disorder of the data subset is measured to obtain the information entropy as the disorder index; The fusion representation vector is generated based on the volatility index, the coupling index, and the disorder index.

4. The method according to claim 3, characterized in that, Performing various statistical measures on the data subsets falling within the sliding analysis window, and generating the fusion representation vector based on the results of the statistical measures, further includes: Data statistics are performed on the data subset to obtain one or more of the following: arithmetic mean, maximum and minimum value, fourth-order normalized moment of distribution, third-order normalized moment of distribution, sine transform of hourly index of daily period, and cosine transform of hourly index of daily period. The fusion representation vector is generated based on one or more of the following: the volatility index, the coupling index, the disorder index, the arithmetic mean, the maximum and minimum values, the fourth normalized moment of the distribution, the third normalized moment of the distribution, the sine transform of the hour index of the daily cycle, and the cosine transform of the hour index of the daily cycle.

5. The method according to claim 1, characterized in that, Multiple statistical measures are applied to the subset of data falling within the sliding analysis window, and the fusion representation vector is generated based on the results of the statistical measures, including: Locality-sensitive hashing is performed on the data subset, and redundant data with the same locality-sensitive hashing in adjacent data subsets are filtered out and removed from the normalized data matrix; The mean and standard deviation of the entire sequence are calculated for all sampled data in the normalized data matrix. In response to the detection that the absolute value of the normalized deviation of a sampled point is greater than 3 times the standard deviation, the sampled point is marked as a candidate outlier observation point. In response to the existence of multiple candidate anomaly observation points, it is determined, based on the structured event knowledge base, whether the multiple candidate anomaly observation points fall within the influence time domain range of any known event recorded in the structured event knowledge base; The anomaly detection sensitivity coefficient is reduced in response to the anomaly falling within the time domain of influence, and noise reduction filtering is performed on the multiple candidate anomaly observation points in response to the anomaly not falling within the time domain of influence. The fused representation vector is generated based on the results of statistical measures performed on the subset of data in the normalized data matrix after removing the redundant data and performing noise reduction filtering.

6. The method according to claim 1, characterized in that, Feedforward neural network prediction models constructed using radial basis function kernel functions as hidden layer neurons include: The K-Means clustering algorithm is used to perform unsupervised clustering in the feature space of the feedforward neural network prediction model to determine the optimal spatial location of the center point of the kernel function. The width parameter of the kernel function is determined based on the feature distance between the center points of the kernel function and a preset scaling factor; The radial basis kernel function of the feedforward neural network prediction model is determined based on the center point and the width parameter of the kernel function. The optimal solution for the output layer weights in the least-squares sense is obtained using the closed-form formula of the Moore-Penrose generalized inverse matrix. The feedforward neural network prediction model responds to receiving the fused representation vector as model input, and generates a scalar prediction as model output based on the activation function of the hidden layer neurons composed of the radial basis kernel function and the linear relationship with the output layer weights.

7. The method according to claim 1, characterized in that, Using the second part to perform accuracy calculation, and adjusting the prediction results and / or prediction step size of the feedforward neural network prediction model based on the results of the accuracy calculation includes: The reference window interval and the prediction window interval are determined. The real data with a length equal to the reference window interval is extracted from the second part and input into the feedforward neural network prediction model. The prediction data with a length equal to the prediction window interval is obtained by forward extrapolation. Calculate the symmetrical mean absolute percentage error between the predicted data and the actual data corresponding to the predicted data in the second part; In response to the symmetric mean absolute percentage error exceeding an acceptable accuracy threshold, the average residual within the prediction window interval is added to the prediction data of the next prediction window interval, and / or the length of the prediction window interval is shortened.

8. The method according to claim 1, characterized in that, The cumulative error sum is determined based on the residual sequence, and the cumulative deviation warning level is output selectively based on the correlation between the cumulative error sum and the dynamic judgment boundary, including: Based on the residual sequence, the residual mean and residual standard deviation are determined. Using the residual mean, the residual standard deviation, and a preset acceptable offset reference value, the positive cumulative sum and negative cumulative sum of the residuals are calculated for each symmetrical mean absolute percentage error. The maximum value of the positive cumulative sum and the negative cumulative sum is used as the anomaly discrimination statistic. A low-priority review notification is output in response to the anomaly detection statistic falling within the first interval of the dynamic determination boundary; In response to the anomaly detection statistic being in a second interval above the first interval of the dynamic judgment boundary, a medium-priority inspection alarm is output; A high-priority emergency plan is output in response to the anomaly detection statistic being in a third interval above the second interval of the dynamic judgment boundary.

9. The method according to claim 1, characterized in that, Storing information about known events in the structured event knowledge base, and adding information about the known event to the fusion representation vector and / or adjusting the size of the dynamic decision boundary in response to the current time falling within the influence time domain of any known event; The impact factors, time-domain range of impact, and anomaly detection sensitivity coefficients of the industrial sensors affected by known events are stored in the structured event knowledge base; In response to the current moment falling within the influence time domain of any known event, the influence factor, influence time domain range, and anomaly detection sensitivity coefficient of the industrial sensor are concatenated to the end of the fused characterization vector, and / or the anomaly detection sensitivity coefficient is used to amplify the size of the dynamic decision boundary.

10. A gas safety prediction system based on time-series modeling, characterized in that, For implementing the method of any one of claims 1 to 9, comprising: The preprocessing module converts the raw data acquired from multi-source and heterogeneous industrial sensors into a normalized data matrix; The feature extraction module uses multiple sliding analysis windows with different fixed widths to slide across the time series of the normalized data matrix to traverse the normalized data matrix. It performs various statistical measures on the data subsets falling within the sliding analysis windows and generates a fusion representation vector based on the results of the statistical measures. The model building module uses radial basis function kernel functions as hidden layer neurons to build a feedforward neural network prediction model; An adaptive module trains the feedforward neural network prediction model using the first part of the fused representation vector, performs accuracy calculation on the trained feedforward neural network prediction model using the second part of the fused representation vector that is later than the first part in the time series, and adjusts the prediction results and / or prediction step size of the feedforward neural network prediction model based on the accuracy calculation results. The early warning module accumulates the results of the accuracy calculation into a residual sequence, determines the cumulative error sum based on the residual sequence, and outputs the cumulative deviation early warning level based on the correlation between the cumulative error sum and the dynamic judgment boundary. The prior knowledge module stores information about known events in a structured event knowledge base. In response to the current moment falling within the influence time domain of any known event, it adds information about the known event to the fusion representation vector and / or adjusts the size of the dynamic decision boundary. The implementation module uses the feedforward neural network prediction model to perform gas safety prediction and outputs the prediction results. Based on the prediction results and the cumulative deviation warning level, it determines the gas safety information.

Images