Chemical industrial park-oriented safety hazard intelligent identification method and system

Abstract

The application relates to the field of safety hazard identification, and specifically discloses a safety hazard intelligent identification method and system for a chemical industrial park. First, the multi-channel voltage response of a distributed sensor array is time-domain aligned, tensor spliced and baseline normalized with temperature and humidity parameters to form a comparable mixed response matrix. Then, differences are extracted in the frequency domain, and low-frequency weak fluctuations are moderately strengthened to highlight early leakage signs. Subsequently, blind source decoupling separation is performed on the mixed features to obtain mutually independent background features and target features. Then, only based on the target features, pure signal strength quantitative inversion is completed, and the background features are used to depict interference strength. According to the inhibition trend of the competition and occupation of the position with the change of the concentration, adaptive reverse compensation correction is performed on the inversion result, so that the suppressed low-concentration signal can be restored and overcompensation in the high-concentration section can be avoided. Finally, the concentration output is subjected to spike suppression and smoothing, and is compared with a safety threshold rule to trigger the generation of a hazard warning report.

Description

Technical Field

[0001] This application relates to the field of safety hazard identification, and more specifically, to a method and system for intelligent identification of safety hazards in chemical industrial parks. Background Technology

[0002] With the rapid development of the chemical industry and the continuous expansion of industrial parks, the types of toxic, harmful, flammable and explosive gases involved in production facilities, storage and transportation corridors and tank areas are numerous and widely distributed. Once a small leak occurs, it can easily trigger a chain of accidents. Therefore, the construction of a set of intelligent safety hazard identification solutions for chemical industrial parks is of profound practical significance for achieving early risk perception, accurate early warning and ensuring the safety of production in the park.

[0003] However, existing technologies often employ single-sensor threshold determination or simple linear feature fusion for concentration estimation. In complex industrial park environments where multiple aromatic volatile gases coexist, the weak response signal of the target gas is easily masked by background cross-interference factors, leading to significant misjudgments of low concentration thresholds. Specifically, because the active surface of the sensor is constantly exposed to the coexistence of background gases such as benzene, toluene, and xylene with target hazardous gases (such as hydrogen sulfide and chlorine), existing solutions typically treat interference as an independent constant for fixed compensation, neglecting the concentration-dependent competitive adsorption coupling relationship of gas molecules vying for active sites on the sensor surface. When the target gas is in the early micro-leakage stage at the ppb level, its adsorption site occupancy is severely suppressed by background molecules, and the response intensity exhibits a nonlinear and drastic decay. This causes the linear compensation method to undercompensate in the critical warning range, resulting in missed detections, or overcompensate in the medium-to-high concentration range, leading to false alarms. This fundamental bottleneck severely restricts the reliable identification of safety hazards in chemical industrial parks.

[0004] Therefore, an optimized intelligent identification solution for safety hazards in chemical industrial parks is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method and system for intelligent identification of safety hazards in chemical industrial parks.

[0006] According to one aspect of this application, a method for intelligent identification of safety hazards in chemical industrial parks is provided, comprising: S1. The raw sensor data stream, which includes multi-channel voltage response sequences and environmental temperature and humidity parameter sequences, collected by the distributed gas sensor array in the park, is processed by time-domain alignment, tensor splicing, and normalization to obtain a hybrid response matrix. S2, perform frequency domain difference feature extraction and low-frequency weak signal enhancement coding on the hybrid response matrix to obtain the hybrid feature tensor; S3, perform blind source decoupling and separation of background interference components and target anomaly components on the mixed feature tensor to obtain background feature vector and target feature vector; S4, quantitatively invert the target gas pure signal intensity of the target feature vector, and perform inverse concentration compensation correction of cross-interference attenuation on the inversion result to obtain the target gas concentration value. S5 performs transient spike elimination processing on the target gas concentration value and compares the smoothed concentration value with the preset safety threshold rule. When the target gas concentration value exceeds the safety lower limit, a hidden danger alarm report is generated.

[0007] According to another aspect of this application, a smart safety hazard identification system for chemical industrial parks is provided, comprising: The raw data preprocessing module is used to perform time-domain alignment, tensor splicing, and normalization on the raw sensor data stream, which includes multi-channel voltage response sequences and environmental temperature and humidity parameter sequences, collected by the distributed gas sensor array in the park, in order to obtain a hybrid response matrix. The frequency domain processing module is used to extract frequency domain difference features and encode low-frequency weak signals to obtain a hybrid feature tensor from the hybrid response matrix. The blind source decoupling and separation module is used to perform blind source decoupling and separation of background interference components and target anomaly components on the mixed feature tensor to obtain background feature vectors and target feature vectors; The quantitative inversion and compensation correction module is used to quantitatively invert the target gas purity signal intensity from the target feature vector and perform inverse concentration compensation correction of the inversion result by cross-interference attenuation to obtain the target gas concentration value. The safety identification module is used to perform transient spike elimination processing on the target gas concentration value and compare the smoothed concentration value with the preset safety threshold rules. When the target gas concentration value exceeds the safety lower limit, a hidden danger alarm report is generated.

[0008] Compared with existing technologies, this application provides a method and system for intelligent identification of safety hazards in chemical industrial parks. First, it performs time-domain alignment, tensor splicing, and baseline normalization on the multi-channel voltage response and temperature and humidity parameters of a distributed sensor array to form a comparable hybrid response matrix. Then, it extracts differences in the frequency domain and appropriately enhances weak low-frequency fluctuations to highlight early leakage signs. Next, it performs blind-source decoupling and separation of the hybrid features to obtain independent background and target features, fundamentally reducing the aliasing of background fluctuations on target anomalies. Then, it performs quantitative inversion of the pure signal intensity based solely on the target features, and uses background features to characterize the interference intensity. It then performs adaptive inverse compensation correction on the inversion results according to the suppression trend of competition and occupancy with concentration changes, allowing suppressed low-concentration signals to be recovered and avoiding overcompensation in high-concentration segments. Finally, it performs peak suppression smoothing on the concentration output and compares it with safety threshold rules to trigger the generation of hazard alarm reports, thereby reducing missed and false alarms. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a method for intelligent identification of safety hazards in chemical industrial parks according to an embodiment of this application; Figure 2 This is a data flow diagram illustrating the intelligent identification method for safety hazards in chemical industrial parks according to an embodiment of this application; Figure 3 This is a flowchart of step S4 of the intelligent safety hazard identification method for chemical industrial parks according to an embodiment of this application; Figure 4 This is a flowchart illustrating the intelligent identification method for safety hazards in chemical industrial parks based on a compensation factor, which performs reverse concentration amplification compensation and absolute purity concentration calculation on the original concentration intensity value to obtain the target gas concentration value. Figure 5 This is a block diagram of a safety hazard intelligent identification system for chemical industrial parks according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0016] Chemical industrial parks have long experienced overlapping fluctuations in background volatiles such as benzene, toluene, and xylene. This causes the sensor response to the target hazardous gas in the early stages of a trace leak to be masked by cross-interference and exhibit nonlinear attenuation. As a result, there is a technical problem that low concentration thresholds are prone to missed detections, while medium and high concentrations are prone to false alarms due to compensation imbalances. To address this, this solution starts with multi-channel voltage and temperature / humidity data from a distributed sensor array. First, it completes time-domain alignment of asynchronous data, feature tensor splicing, and normalization based on historical alarm-free baselines to form a comparable hybrid response matrix. Then, it extracts the differential features of the adsorption / desorption process in the frequency domain and appropriately enhances low-frequency weak fluctuations to highlight subtle changes in early leakage. Subsequently, the hybrid features are introduced into a dual-branch blind source decoupling network and geometric orthogonal constraints are applied to forcibly separate background interference and target anomalies in the latent space, obtaining background and target features. Then, it performs quantitative inversion of the pure signal intensity based solely on the target features and uses the intensity quantification results of the background features to construct adaptive interference compensation, inversely correcting the inverted values ​​to recover suppressed low-concentration information. Simultaneously, it avoids excessive amplification in the high-concentration range through limited dynamic gain. Finally, after peak suppression and smoothing of the concentration output, it performs a hard comparison with safety threshold rules. When the safety lower limit is exceeded, a structured alarm report is generated, thereby correspondingly improving the early detection capability in complex backgrounds while maintaining stability across the entire measurement range.

[0017] The technical solution of this application proposes an intelligent identification method for safety hazards in chemical industrial parks. Figure 1 This is a flowchart of a method for intelligent identification of safety hazards in chemical industrial parks according to an embodiment of this application. Figure 2 This is a data flow diagram illustrating the intelligent identification method for safety hazards in chemical industrial parks according to an embodiment of this application. Figure 1 and Figure 2 As shown, the intelligent identification method for safety hazards in chemical industrial parks according to an embodiment of this application includes the following steps: S1, performing time-domain alignment, tensor splicing, and normalization on the original sensor data stream containing multi-channel voltage response sequences and environmental temperature and humidity parameter sequences collected by a distributed gas sensor array in the park to obtain a hybrid response matrix; S2, performing frequency domain difference feature extraction and low-frequency weak signal enhancement encoding on the hybrid response matrix to obtain a hybrid feature tensor; S3, performing blind source decoupling and separation of background interference components and target abnormal components on the hybrid feature tensor to obtain background feature vectors and target feature vectors; S4, performing quantitative inversion of the target feature vector to obtain the pure signal intensity of the target gas, and performing inverse concentration compensation correction of the inversion result by cross-interference attenuation to obtain the target gas concentration value; S5: performing transient spike elimination processing on the target gas concentration value, and comparing the smoothed concentration value with a preset safety threshold rule for judgment, and generating a hazard alarm report when the target gas concentration value exceeds the safety lower limit.

[0018] Specifically, in step S1, the original sensor data stream, which includes multi-channel voltage response sequences and environmental temperature and humidity parameter sequences, collected by the distributed gas sensor array within the park, undergoes time-domain alignment, tensor splicing, and normalization to obtain a hybrid response matrix. It should be understood that, due to the multi-point deployment of the distributed gas sensor array within the park in areas such as the equipment area, tank area, and pipe gallery, the multi-channel voltage response sequences and environmental temperature and humidity parameter sequences of each sensor node exhibit differences in sampling frequency, communication link jitter, and asynchronous arrival of reporting codes at the acquisition end. Furthermore, temperature and humidity changes introduce sensor baseline drift and sensitivity drift, making the original sensor data streams incomparable in terms of time axis and dimensional scale. This leads to input distribution mismatch in subsequent frequency domain difference feature extraction, blind source decoupling and separation, and concentration inversion compensation, amplifying the risk of misjudgment of low concentration thresholds caused by background cross-interference. Therefore, in the technical solution of this application, the original sensor data stream, which includes multi-channel voltage response sequences and environmental temperature and humidity parameter sequences, collected by the distributed gas sensor array within the park, is subjected to time-domain alignment, tensor splicing, and standardization. This constrains the asynchronously arriving multi-source data to a unified time reference and forms a learnable and comparable hybrid representation within a unified feature space. This simultaneously eliminates transient feature distortions caused by temporal misalignment and weakens scale drift and inter-node differences caused by environmental disturbances on the input side. This provides a stable and consistent data foundation for subsequent extraction of frequency domain difference features from the hybrid response matrix, blind source decoupling and separation of the hybrid feature tensor, and quantitative inversion and cross-interference compensation of the target feature vector. Ultimately, this improves the detection rate of extremely low concentration warning ranges and reduces false alarms.

[0019] More specifically, in the embodiments of this application, step S1 includes: synchronously extracting and aligning the asynchronously arriving multi-channel voltage response sequence and environmental temperature and humidity parameter sequence in the original sensing data stream to obtain aligned data frames; within each aligned data frame, extracting the peak maximum slope characteristic and steady-state amplitude characteristic from the voltage response sequence, and concatenating the slope characteristic, steady-state amplitude characteristic, and environmental temperature and humidity parameters at the same time stamp using multi-dimensional tensors to obtain the original feature tensor; and standardizing and normalizing the original feature tensor to obtain the hybrid response matrix.

[0020] Accordingly, the asynchronously arriving multi-channel voltage response sequences and environmental temperature and humidity parameter sequences in the original sensor data stream are synchronously truncated and time-domain aligned to obtain aligned data frames. It should be understood that, due to the distributed gas sensor arrays deployed in different locations within the chemical industrial park, such as the plant area, tank area, and pipe gallery, the multi-channel voltage response sequences and environmental temperature and humidity parameter sequences of each sensor node are affected by sampling frequency differences, communication transmission delays, and network jitter, exhibiting asynchronous arrival characteristics in the original sensor data stream. This results in significant misalignment on the time axis, making it difficult to accurately correlate adsorption and desorption dynamics in subsequent frequency domain feature extraction. The blind source decoupling and separation process is easily interfered with by time mismatch, ultimately amplifying the masking effect of background cross-gas on the low-concentration signal of the target gas, increasing the risk of threshold misjudgment. Therefore, in the technical solution of this application, the asynchronously arriving multi-channel voltage response sequences and environmental temperature and humidity parameter sequences in the original sensor data stream are synchronously truncated and time-domain aligned to constrain the asynchronous multi-source sequences to the same time base and form a time-synchronized frame structure. In this way, transient feature distortion and cross-sequence correspondence bias induced by time-series asynchrony can be completely eliminated, providing a time-domain consistent and directly comparable data foundation for subsequent splicing of original feature tensors, standardization of hybrid response matrices, and the entire blind source decoupling and compensation link, significantly improving the reliable detection capability of extremely low concentration hazards.

[0021] In a specific example of this application, a unified system timestamp reference is first established for the original sensing data stream, and a sliding time window with a fixed step size and window width is configured. Then, based on the timestamp, the multi-channel voltage response sequence and the environmental temperature and humidity parameter sequence are synchronously truncated. The set of continuous sampling points in the same start and end time intervals is synchronously extracted from the two types of sequences. For sampling missing points appearing in the truncated window, a linear interpolation algorithm is used to fill in the missing values. For redundant oversampling points, equal-interval downsampling compression processing is performed to ensure that each generated aligned data frame has a completely consistent frame length, start timestamp, and end timestamp. The final output is an aligned data frame that can be directly used for subsequent intra-frame feature extraction.

[0022] Accordingly, within each aligned data frame, the maximum slope characteristic and steady-state amplitude feature of the peak are extracted from the voltage response sequence. The slope characteristic, steady-state amplitude feature, and environmental temperature and humidity parameters at the same time stamp are then concatenated using multi-dimensional tensors to obtain the original feature tensor. It is understandable that, since the voltage response sequence within the aligned data frame simultaneously contains the transient rate change of gas molecules during the initial adsorption stage on the sensor's active surface and the steady-state equilibrium level after adsorption saturation, and the environmental temperature and humidity parameters directly modulate the sensor's baseline drift and sensitivity coefficient, relying solely on the original voltage sequence makes it difficult to extract discriminative physical characteristics that differentiate between normal fluctuations in background aromatic hydrocarbons and weak leaks of the target hazardous gas. This leads to the subsequent blind source decoupling and separation failing to effectively remove cross-interference components and amplifies the problem of low-concentration threshold misjudgment where the early ppb-level signal of the target gas is masked under conditions of coexistence of multiple aromatic volatile gases. Therefore, in the technical solution of this application, the maximum slope characteristic and steady-state amplitude characteristics of the peak are further extracted from the voltage response sequence within each aligned data frame. The slope characteristic, steady-state amplitude characteristics, and environmental temperature and humidity parameters at the same timestamp are then multi-dimensionally spliced ​​to obtain the original feature tensor. This allows for the extraction and fusion of transient dynamics, steady-state levels, and environmental modulation factors from the electrochemical response mechanism level. This provides a multi-modal physical fusion feature foundation rich in early leakage signs for the standardized normalization processing of the original feature tensor and for subsequent frequency domain enhancement and blind source decoupling of the hybrid response matrix. Ultimately, this improves the accuracy of target gas purity signal inversion and the reliability of hidden danger alarms under complex background interference.

[0023] In a specific example of this application, a first-order time difference operation is first performed on the voltage response sequence within each aligned data frame to generate a slope sequence. The element with the largest absolute value is extracted as the peak maximum slope characteristic to accurately characterize the most intense response rate during the gas adsorption or desorption stage. Simultaneously, an arithmetic mean aggregation operation is performed on the voltage response sequence within the steady-state time interval of the aligned data frame to obtain the steady-state amplitude characteristics, thereby capturing the stable signal level after adsorption equilibrium. Then, the extracted peak maximum slope characteristics, steady-state amplitude characteristics, and environmental temperature and humidity parameters with the same timestamp as the aligned data frame are subjected to a multi-dimensional tensor splicing operation along the feature dimension. The dimensional structure of the spliced ​​tensor includes the time frame index, sensor channels, and fused physical feature components, thereby ensuring that the electrical transient and steady-state responses and environmental modulation information are consistently expressed in a unified tensor space. Finally, an original feature tensor that can be directly input into the standardized normalization process is generated.

[0024] Accordingly, the original feature tensor is standardized and normalized to obtain a hybrid response matrix. It is understandable that, due to differences in factory calibration, long-term aging drift, and cumulative environmental degradation, the numerical benchmarks for the same feature dimension vary significantly across different nodes in the chemical industrial park. Furthermore, the maximum slope characteristics of the peak, the steady-state amplitude characteristics, and the dimensions and dynamic range of the environmental temperature and humidity parameters differ considerably. Directly inputting the original feature tensor into subsequent processing would lead to features with large numerical magnitudes dominating the model training and inference process, suppressing the weak features corresponding to ppb-level micro-leakage of the target gas, further exacerbating the risk of misjudging low-concentration thresholds under background cross-gas interference. Therefore, in the technical solution of this application, the original feature tensor is further standardized and normalized to obtain a hybrid response matrix, thereby mapping the features of different nodes and different feature dimensions to a unified zero-mean unit variance distribution space, eliminating the distribution mismatch caused by individual hardware differences and dimensional differences. This avoids the problem of weak leakage features being submerged due to numerical scale imbalance, and provides a consistent input basis for the subsequent extraction of frequency domain difference features of the hybrid response matrix and the blind source decoupling and separation of the hybrid feature tensor, effectively improving the identifiability of low-concentration target signals and the accuracy of subsequent inversion compensation.

[0025] Specifically, in the embodiments of this application, the original feature tensor is standardized and normalized to obtain a hybrid response matrix, including: based on the background baseline mean and background baseline variance that are adaptively updated within a long-term historical alarm-free period, the original feature tensor is standardized and normalized to obtain a hybrid response matrix.

[0026] More specifically, in a concrete example of this application, firstly, for each feature dimension of each sensing node, the corresponding background baseline mean and variance are adaptively updated statistically from historical samples collected during long-term alarm-free periods. The statistical process employs a sliding window dynamic update mechanism to adapt to baseline drift caused by the slow aging of the sensors. Subsequently, pointwise standardization and normalization are performed on each feature element in the original feature tensor, using the following formula: in, This represents the original feature tensor element corresponding to the i-th sensing node, the j-th feature dimension, and the k-th time frame, i.e., the input feature component to be normalized; This represents the background baseline mean that is adaptively updated for the i-th sensor node and the j-th feature dimension during a long-term historical period without alarms. This represents the background baseline variance of the i-th sensor node and the j-th feature dimension, which is adaptively updated during a long-term historical period without alarms. This represents a very small constant introduced to avoid the denominator being zero, with a fixed value of 1 × 10. -7 ; The normalized eigenele represents the basic unit constituting the mixed response matrix. It's worth noting that the numerator part of the formula... This is the baseline drift calibration term, which eliminates the historical baseline offset of the corresponding feature dimension for each sensing node. For example, a hydrogen sulfide sensor deployed in a tank area may experience a slow increase in baseline voltage due to long-term exposure to a low-concentration aromatic hydrocarbon environment. By subtracting the historical baseline mean of the corresponding feature for that node, the features of different nodes can be unified to a zero baseline, eliminating the influence of individual hardware differences. The denominator in the formula... As a scale unification term, its function is to map features of different dimensions to a consistent fluctuation range. For example, the dynamic range of the maximum slope characteristic of the wave crest is 0~12V / s, the dynamic range of the ambient temperature is -20~60℃, and the dynamic range of the relative humidity is 0~100%RH. By dividing by the standard deviation of the background baseline of the corresponding feature, the fluctuation range of all features is unified to the unit variance interval, preventing a certain type of feature from dominating subsequent calculations and avoiding the suppression of small feature fluctuations corresponding to low-concentration leaks by large-scale features; among which, the minimal constant... Its purpose is to mitigate the risk of numerical overflow in extreme scenarios. For example, if a sensor node deployed in an area with extremely stable background gas shows no fluctuation in a certain feature dimension during a historical no-alarm period, and the background baseline variance is zero, then adding... This avoids system anomalies caused by a zero denominator, ensuring operational stability across all scenarios. After flattening all normalized feature elements along the feature dimension, a hybrid response matrix is ​​obtained that can be directly input into the frequency domain difference feature extraction stage, fully meeting the requirements of data flow closure.

[0027] Specifically, in step S2, the mixed response matrix is ​​subjected to frequency domain difference feature extraction and low-frequency weak signal enhancement encoding to obtain a mixed feature tensor. It should be understood that the random fluctuations generated by background gases such as benzene, toluene, and xylene, which are normally present in chemical industrial parks, are mostly concentrated in the high-frequency range. In contrast, the early micro-leakage of the target hazardous gas is characterized by a long duration and small fluctuation amplitude of low-frequency weak signals. These two types of signals exhibit frequency domain aliasing characteristics in the mixed response matrix. If blind source decoupling and separation are directly performed, the weak low-frequency leakage signal will be submerged by high-frequency background interference, leading to deviations in subsequent concentration inversion and further exacerbating the risk of misjudging the low-concentration threshold. Therefore, in the technical solution of this application, the mixed response matrix is ​​further subjected to frequency domain difference feature extraction and low-frequency weak signal enhancement encoding to obtain a mixed feature tensor. This is used to capture the hidden frequency domain difference features of the mixed gas adsorption and desorption stages, while simultaneously directionally enhancing the low-frequency weak signal component corresponding to low-concentration leakage and suppressing background high-frequency redundant fluctuations. This can improve the signal-to-noise ratio of early micro-leakage signals during the feature encoding stage, providing a more discriminative feature basis for blind source decoupling and separation of subsequent mixed feature tensors, and effectively reducing the false negative probability of low-concentration leakage.

[0028] More specifically, in the embodiments of this application, step S2 includes: extracting hidden frequency domain difference features generated by the mixed gas during the cross-desorption and adsorption stages from the mixed response matrix based on a one-dimensional convolutional kernel with learnable parameters to obtain a convolutional feature tensor; performing multi-head self-attention dynamic weighting on any two frequency domain feature frames in the convolutional feature tensor for weak leakage fluctuations to obtain an attention feature tensor; and performing deep nonlinear mapping reconstruction and compression shaping on the attention feature tensor based on a multilayer perceptron feedforward network to obtain a mixed feature tensor.

[0029] Accordingly, a convolutional feature tensor is obtained by extracting the hidden frequency domain difference features generated by the mixed gas during the cross-desorption and adsorption stages from the mixed response matrix based on a one-dimensional convolutional kernel with learnable parameters. It should be understood that, due to the aliasing characteristics of the adsorption-desorption response of the early micro-leakage of the target hazardous gas in the chemical industrial park and the normal fluctuations of the background aromatic gases in the time domain of the mixed response matrix, the frequency domain differences between the two types of responses are highly hidden. Relying solely on the original mixed response matrix cannot effectively extract discriminative features that distinguish the target leak signal from the background interference signal, making it difficult to accurately remove background interference components in the subsequent blind source decoupling and separation process, further exacerbating the risk of misjudgment of low concentration thresholds. Therefore, in the technical solution of this application, a convolutional feature tensor is obtained by extracting the hidden frequency domain difference features generated by the mixed gas during the cross-desorption and adsorption stages from the mixed response matrix based on a one-dimensional convolutional kernel with learnable parameters. This allows for the capture of cross-time domain frequency domain correlation characteristics through a local sliding receptive field, filtering out redundant time domain noise, and condensing a specific frequency domain characterization of the target leak. In this way, the frequency domain components of the target signal and background interference can be initially separated during the feature extraction stage, improving the discriminative power of the features, providing a basic input for the attention-enhanced coding of weak low-frequency signals, and reducing the difficulty of subsequent blind source decoupling.

[0030] In a specific example of this application, the hyperparameters of the one-dimensional convolutional kernel are first configured. The kernel size is set to cover three consecutive time frames to capture the dynamic response correlation between adjacent time steps, and the stride is set to 1. An equal-length padding strategy is used to ensure that the time dimension of the convolutional output is consistent with the time dimension of the input mixed response matrix. The kernel parameters are adaptively updated through backpropagation during model training. Subsequently, the mixed response matrix is ​​input into the one-dimensional convolutional layer to perform a sliding convolution operation. After the convolution operation, nonlinear negative noise is filtered out using the ReLU activation function, and the final output is a convolutional feature tensor. The formula for this convolution operation is as follows: in, For the t-th time frame and the k-th convolutional channel, there are output feature elements, which are the basic units that constitute the convolutional feature tensor; This represents the ReLU nonlinear activation function, used to achieve non-negative truncation to filter out negative noise; This represents the learnable weight parameter corresponding to the position with offset i for the k-th convolutional kernel. The offset range is from -1 to 1, corresponding to a convolutional kernel of size 3. This represents a vector representing all feature dimensions of the input mixture response matrix at the (t+i)th time frame; This represents the learnable bias parameters corresponding to the k-th convolutional kernel. The summation term in the formula... This is a locally weighted aggregation term, which performs weighted fusion of features from three consecutive time frames, where the learnable weights... During training, the system adaptively learns the response patterns of different frequency domain components. In the context of a chemical industrial park, some convolutional kernels learn high weights for the low-frequency, slowly changing frequency domain features corresponding to the micro-leakage of the target gas, while others learn low weights for the high-frequency fluctuation features of the background. This allows for the initial screening of frequency domain features related to the target leak during the aggregation process. (Bias parameters) Its function is to adjust the output baseline of each convolutional channel to adapt to the response offset characteristics of different sensor nodes. The ReLU activation function... The function is to filter out the negative noise components generated by the convolution operation. These negative components mostly correspond to random perturbations in the background gas. After filtering, the signal-to-noise ratio of the target leakage feature can be further improved. The final convolution feature tensor has initially separated the frequency domain difference components between the target and the background, and can be directly input into the subsequent multi-head self-attention dynamic weighting stage, which meets the data flow closed-loop requirements.

[0031] Accordingly, an attention feature tensor is obtained by dynamically assigning multi-head self-attention weights to any two frequency domain feature frames in the convolutional feature tensor to address weak leakage fluctuations. It should be understood that the low-frequency weak signals generated by early micro-leaks of target hazardous gases in chemical industrial parks exhibit stable and similar response patterns across multiple time frames of the convolutional feature tensor, while the high-frequency random fluctuations of background aromatic gases show low inter-frame similarity. Although the convolutional feature tensor has initially separated the frequency domain difference components, the importance weights of each feature frame remain equal, failing to specifically strengthen the feature frames corresponding to the low-frequency weak leakage signals. This results in the target abnormal components still being easily masked by background interference components during subsequent blind source decoupling and separation, further exacerbating the risk of misjudging low concentration thresholds. Therefore, in the technical solution of this application, an attention feature tensor is obtained by dynamically assigning multi-head self-attention weights to any two frequency domain feature frames in the convolutional feature tensor to address weak leakage fluctuations. This dynamically increases the weight of feature frames corresponding to low-frequency stable signals by calculating inter-frame similarity, while suppressing the weight of feature frames corresponding to high-frequency random fluctuations. In this way, the characterization intensity of early micro-leakage signals can be significantly enhanced in the feature encoding stage, providing a feature basis for the differential weights of target signals and background interference for the subsequent blind source decoupling and separation of hybrid feature tensors, and effectively reducing the false negative probability of low-concentration leakage.

[0032] In a specific example of this application, the convolutional feature tensor is first input into three independent linear mapping layers to generate a query matrix, a key matrix, and a value matrix. The query and key matrices are used to calculate inter-frame similarity scores, while the value matrix carries the actual feature information. Subsequently, matrix multiplication is performed on the transposes of the query and key matrices to obtain the inter-frame similarity score matrix. This matrix is ​​then scaled by the square root of the key matrix's feature dimension to prevent gradient vanishing. The similarity score is then converted into an attention weight matrix ranging from 0 to 1 using the Softmax normalization function. Finally, matrix multiplication is performed on the attention weight matrix and the value matrix to obtain the single-head attention output. The outputs of multiple attention heads are concatenated along the feature dimension and linearly mapped to obtain the attention feature tensor. The core calculation formula for this multi-head self-attention dynamic weighting is as follows: in, This represents the attention-weighted feature component of the single attention head output, which is the basic unit that constitutes the attention feature tensor; It is a query matrix obtained by linear mapping of the convolutional feature tensor, representing the query information of the current feature frame; It is the transpose of the key matrix obtained by linear mapping of the convolutional feature tensor, used to represent the matching reference information of all feature frames; It is the feature dimension of the key matrix, which is a fixed-configuration hyperparameter; A scaling factor used to prevent gradient vanishing due to excessively large inner product values; This is a normalized activation function used to convert the inner product score into a weight value in the range of 0 to 1; It is a value matrix obtained by linear mapping from the convolutional feature tensor, carrying the actual feature information. The numerator part of the formula... This is the inter-frame similarity calculation term, which measures the correlation between any two frequency domain feature frames in the convolutional feature tensor. In the chemical industrial park scenario, the low-frequency signals generated by the early micro-leakage of the target hazardous gas exhibit stable similar response patterns across multiple time frames. For example, the peak maximum slope and steady-state amplitude characteristics of hydrogen sulfide micro-leakage show a slow upward trend over 10 consecutive time frames, corresponding to a higher inter-frame inner product score. Conversely, the high-frequency random fluctuations of background aromatic gases have low inter-frame similarity, resulting in a smaller inner product score. (The denominator is missing from the original text.) The gradient stabilizing term scales the inner product score to a suitable range, preventing extreme large values ​​from saturating the Softmax function output and ensuring the stability of model training and inference. After Softmax normalization, the weights corresponding to low-frequency leakage features are significantly increased; for example, the weights corresponding to 10 consecutively and steadily increasing time frames are increased from an equal 0.1 to the range of 0.3 to 0.5, while the weights corresponding to high-frequency background fluctuations are suppressed to the range of 0.01 to 0.05. This is then compared with the value matrix. Multiplication yields a single-head attention output that enhances weak low-frequency signals. Multiple attention heads capture similar patterns from different frequency domain components in different subspaces. The attention feature tensor obtained by splicing and linear mapping achieves targeted enhancement of early micro-leakage signals and can be directly input into the subsequent multilayer perceptron feedforward network for deep nonlinear mapping reconstruction, meeting the data flow closed-loop requirements.

[0033] Accordingly, a hybrid feature tensor is obtained by performing deep nonlinear mapping reconstruction and compression shaping on the attention feature tensor based on a multilayer perceptron feedforward network. It should be understood that although the attention feature tensor achieves directional enhancement of low-frequency weak leakage signals after multi-head self-attention dynamic weighting, the feature components output by multiple attention heads are still in relatively independent subspace states. The deep nonlinear interaction correlation between the features of each subspace has not been fully explored. Furthermore, the attention feature tensor has a high feature dimension; directly inputting it into the subsequent blind source decoupling and separation stage would lead to excessive computational complexity and easily introduce redundant noise, thereby reducing the separation accuracy between background interference components and target abnormal components and exacerbating the risk of misjudgment at low concentration thresholds. Therefore, in the technical solution of this application, a hybrid feature tensor is further obtained by performing deep nonlinear mapping reconstruction and compression shaping on the attention feature tensor based on a multilayer perceptron feedforward network. This allows for the full fusion of heterogeneous feature components from each attention head subspace through deep fully connected mapping, and the elimination of redundant information through dimensionality compression, generating a compact and discriminative hybrid feature representation. In this way, the deep fusion and information condensation of multi-subspace features can be completed in the final stage of feature encoding, providing a dimensionally adapted and discriminative input basis for the blind source decoupling and separation of subsequent mixed feature tensors, effectively improving the separation accuracy of background and target components and the accuracy of subsequent concentration inversion.

[0034] In a specific example of this application, the attention feature tensor is first input into the first fully connected mapping layer of the multilayer perceptron feedforward network. This layer maps the input feature dimension to a higher-dimensional hidden space to enhance feature representation. After mapping, a nonlinear transformation capability is introduced through the ReLU activation function. Subsequently, the activated hidden features are input into the second fully connected mapping layer. This layer compresses the hidden feature dimension to the target output dimension to remove redundant information and generate a compact representation. The compressed feature is the hybrid feature tensor. To prevent gradient vanishing and feature degradation during deep network training, a residual connection path is established between the input and output of the multilayer perceptron feedforward network. The attention feature tensor is dimensionally aligned and then fused with the multilayer perceptron output through element-wise addition. After fusion, the feature distribution is stabilized through layer normalization. The final output can be directly used as a hybrid feature tensor for blind source decoupling and separation, completing the complete feature encoding link from the hybrid response matrix to the hybrid feature tensor, which meets the data flow closure requirement.

[0035] Specifically, in step S3, the background interference component and the target anomalous component are decoupled and separated by blind source in the mixed feature tensor to obtain the background feature vector and the target feature vector. It should be understood that, due to the competitive adsorption of background aromatic gases such as benzene, toluene, and xylene with the target hazardous gas on the active surface of the metal oxide semiconductor sensor in the chemical industrial park, the response characteristics of the two types of gases exhibit a deeply nonlinear coupling state in the mixed feature tensor. If concentration inversion is directly performed on the mixed feature tensor, the background interference component will continuously mask the weak characteristics corresponding to the target anomalous component, resulting in the inability to effectively separate the early leakage signal of the target gas at the ppb level from the background noise, thus causing misjudgment of the low concentration threshold. Therefore, in the technical solution of this application, the background interference component and the target anomalous component are further decoupled and separated by blind source in the mixed feature tensor to forcibly sever the nonlinear coupling relationship between the background interference and the target signal in the high-dimensional hidden layer space, projecting the two types of components into mutually orthogonal independent subspaces. In this way, the masking effect of background aromatic gases on the characteristics of target hazardous gases can be eliminated from the source, providing a pure target feature base that is not contaminated by background interference for the subsequent quantitative inversion of target feature vectors. At the same time, it provides a background feature vector that accurately quantifies the intensity of background interference for the inverse concentration compensation correction of cross-interference attenuation, ultimately significantly improving the detection rate of low-concentration hazards and reducing the false alarm rate.

[0036] More specifically, in the embodiments of this application, step S3 includes: synchronously channeling the hybrid feature tensor to two independent branch networks that do not share weights, and performing parallel splitting of the feature basis through fully connected mapping and nonlinear activation projection to obtain a background projection matrix and a target projection matrix respectively; performing geometric spatial forced decoupling on the background projection matrix and the target projection matrix to obtain an orthogonal background matrix and an orthogonal target matrix respectively; and performing global average pooling and one-dimensional flattening compression on the orthogonal background matrix and the orthogonal target matrix along the time dimension to obtain a background feature vector and a target feature vector respectively.

[0037] In a specific example of this application, the hybrid feature tensor is first synchronously fed into two independent branch networks that do not share weights. The background branch network focuses on extracting the interference feature basis of normal background aromatic gases in the park, while the target branch network focuses on extracting the abnormal feature basis of micro-leaks of target hazardous gases. The two branch networks perform parallel splitting of the feature basis through fully connected mapping and ReLU nonlinear activation projection, respectively, outputting the background projection matrix and the target projection matrix. Subsequently, a geometrically forced decoupling constraint is applied to the background projection matrix and the target projection matrix. By introducing an orthogonal constraint loss function during the model training phase to penalize the non-zero projection overlap of the two projection matrices in the vector space, the background projection matrix and the target projection matrix are forced to tend to be mutually perpendicular in Euclidean space, thus obtaining the orthogonal background matrix and the orthogonal target matrix, respectively. The formula for this orthogonal constraint loss function is as follows: in, This represents the orthogonal constraint loss function value, which is used to quantify the degree of coupling between the background projection matrix and the target projection matrix in the vector space. The smaller the value, the more orthogonal the two matrices tend to be. This represents the background projection matrix output by the background branch network; This represents the target projection matrix output by the target branch network; This represents the transpose of the target projection matrix; The Frobenius norm is used to calculate the overall scale of a matrix to perform normalization. This indicates that the square of the Frobenius norm is calculated over the normalized inner product matrix, serving as the final orthogonality constraint penalty. The numerator part of the formula... This is the coupling degree calculation term, which measures the degree of directional overlap between the background projection matrix and the target projection matrix in the vector space. In a chemical industrial park scenario, if the interference characteristics of the background aromatic gases and the abnormal characteristics of the target hazardous gases overlap directionally in the hidden layer space, then the absolute value of the elements of this inner product matrix is ​​relatively large, indicating that the two types of components still have coupling and aliasing; the denominator part The scaling term eliminates the inner product scaling bias caused by differences in feature amplitudes between the two projection matrices, ensuring that the orthogonal constraint loss function reflects only the degree of directional coupling rather than the magnitude of amplitudes. The outer Frobenius norm square sums the squares of all elements in the normalized inner product matrix, serving as a penalty signal for backpropagation to drive the weights of the two branch networks to update in the orthogonal direction. This ultimately forces the decoupling of the background projection matrix and the target projection matrix in geometric space during the inference phase, yielding orthogonal background and target matrices, respectively. Finally, global average pooling is performed on the orthogonal background and target matrices along the time dimension, compressing the time-dimensional feature sequence into a single time-step feature representation. A one-dimensional flattening operation further compresses the matrix structure into a one-dimensional vector structure, yielding the background and target feature vectors, respectively. These can be directly input into the subsequent quantitative inversion of the target gas purity signal intensity and the inverse concentration compensation correction stage for cross-interference attenuation, meeting the data flow closed-loop requirements.

[0038] Specifically, in step S4, the target feature vector is quantitatively inverted to obtain the target gas purity signal intensity, and the inversion result is corrected by inverse concentration compensation through cross-interference attenuation to obtain the target gas concentration value. It should be understood that although the target feature vector output from the upstream blind source decoupling and separation stage has been stripped of direct aliasing from background interference components, this vector is still in the high-dimensional abstract feature space of the deep neural network and has not yet been mapped to the physical concentration dimensions that can be directly used for safety assessment at the chemical industrial park site. Therefore, in the technical solution of this application, the target feature vector is further quantitatively inverted to obtain the target gas purity signal intensity, and the inversion result is corrected by inverse concentration compensation through cross-interference attenuation to obtain the target gas concentration value.

[0039] Figure 3 This is a flowchart of step S4 of the intelligent safety hazard identification method for chemical industrial parks according to an embodiment of this application. Figure 3 As shown, step S4 includes: S41, performing nonlinear mapping of the target original intensity under isolated interference state on the target feature vector to obtain the original concentration intensity value; S42, determining the compensation factor based on the Euclidean norm of the background feature vector and the preset physical compensation coefficient; S43, based on the compensation factor, performing reverse concentration amplification compensation and absolute purity concentration calculation on the original concentration intensity value to obtain the target gas concentration value.

[0040] Accordingly, in step S41, the target feature vector is subjected to a nonlinear mapping of the original target intensity under isolated interference state to obtain the original concentration intensity value. It should be understood that although the target feature vector output by the upstream blind source decoupling and separation link has achieved geometric orthogonal separation from the background interference components in the high-dimensional hidden layer space, the vector is still in the tensor representation form of the abstract feature space of the deep neural network. Its numerical dimension has an essential semantic gap with the physical concentration dimension required for on-site safety monitoring in chemical industrial parks. It cannot be directly used as a concentration criterion at the ppm or ppb level to input into the subsequent threshold comparison link. At the same time, the target feature vector carries the comprehensive response characteristics of the sensor to the target hazardous gas in the multi-component gas coexistence environment. It has not yet separated the sensor's own nonlinear response characteristics and the complex mapping relationship in the signal transmission path. If a quantitative inversion mapping from the abstract feature space to the physical concentration space is not performed on it, the entire intelligent identification link will experience data type mismatch in the last link and will be unable to complete the closed-loop decision. Therefore, in the technical solution of this application, the original target intensity is further nonlinearly mapped under isolated interference state to obtain the original concentration intensity value. This is then used to project the high-dimensional target feature vector to a single scalar apparent concentration intensity domain via a deep nonlinear regression network, establishing a deterministic mapping relationship from neural network abstract features to sensor physical response intensity. This transforms the pure target features after blind source decoupling into preliminary intensity estimates that can be directly used for concentration quantification, providing a foundational input for subsequent compensation and correction based on the background competitive adsorption physical model, and ensuring a complete data flow closed loop from feature extraction to concentration output.

[0041] More specifically, in a concrete example of this application, the target feature vector is first input into a pre-trained multi-layer feedforward neural network regression module. The first layer of this regression module is a fully connected mapping layer, which projects the input high-dimensional target feature vector to the latent feature space through a linear transformation of the weight matrix. The dimension of the latent feature space is set to half the dimension of the target feature vector to achieve feature compression. The projected latent features are then processed by the ReLU nonlinear activation function to introduce nonlinear transformation capability and filter out negative response components. Subsequently, the activated latent features are input into the second fully connected mapping layer of the regression module. This layer further compresses the latent feature space to a single scalar output node. Through a linear combination of the weight matrix and the bias vector, the final dimensionality reduction projection from multi-dimensional latent features to a one-dimensional concentration intensity scalar is completed. The output scalar value is the original concentration intensity value, which reflects the apparent response intensity of the target gas on the sensor surface without considering the background competitive adsorption inhibition effect. This value can be directly used as the basic input data for subsequent compensation and correction stages.

[0042] Accordingly, in step S42, the compensation factor is determined based on the Euclidean norm of the background feature vector and the preset physical compensation coefficient. It should be understood that because background aromatic volatile organic compounds such as benzene, toluene, and xylene, which are long-term diffused in the tank area, pipe gallery, and equipment area of ​​the chemical industrial park, compete with the target hazardous gas for adsorption on the active surface of the metal oxide semiconductor sensor, the higher the concentration of the background interfering gas, the more active sites it occupies, and the more severe the suppression of the target gas response signal. Although the background feature vector output from the upstream blind source decoupling and separation stage has achieved orthogonal separation from the target feature vector in the hidden layer space, this vector is still in an abstract feature representation form and has not yet been transformed into a physical parameter that can directly quantify the intensity of the background interference. If the background feature vector is not numerically evaluated, it will be impossible to construct a compensation basis reflecting the degree of suppression of the target signal by the background interference, resulting in a lack of targeted correction benchmarks and systematic bias in the subsequent concentration inversion process. Therefore, in the technical solution of this application, a compensation factor is further determined based on the Euclidean norm of the background feature vector and a preset physical compensation coefficient. This is used to quantify the comprehensive competitive adsorption intensity of the background interfering gas by calculating the geometric modulus of the background feature vector in high-dimensional space. Furthermore, by combining the sensor material characteristics and the physical compensation coefficient of the background gas composition in the park, the abstract feature intensity is mapped to a physically meaningful compensation factor. In this way, the background feature vector can be transformed into a quantitative parameter that can be directly used for concentration compensation correction, providing a precise numerical basis for quantifying the background interference intensity for subsequent reverse concentration amplification compensation, ensuring that the compensation strength matches the actual degree of background interference.

[0043] More specifically, in a concrete example of this application, the Euclidean norm is first calculated on the background feature vector to extract its geometric modulus in the high-dimensional feature space. This norm calculation involves summing the squares of each dimensional component of the background feature vector and then taking the square root. The resulting scalar value intuitively reflects the comprehensive intensity of the background interference features in the abstract space. The larger the norm value, the stronger the competitive adsorption intensity of the background aromatic gases on the active surface of the sensor. Subsequently, the calculated background feature vector norm is substituted into an exponential amplification function. This exponential function uses the natural constant as its base, and the exponent is the product of the background feature vector norm and a preset physical compensation coefficient. The physical compensation coefficient is pre-calibrated based on the adsorption affinity characteristics of the metal oxide semiconductor material used in the sensor for the mixed gas and the actual background gas composition ratio in the park. The calculation formula for this exponential function is as follows: in, The compensation factor is used to calculate the output and quantify the degree to which background interference gases suppress the target gas signal. It is an exponential function with the natural constant as its base, used to convert the linear norm measure into a nonlinear amplification factor to reflect the exponential inhibition characteristics of competitive adsorption; The preset physical compensation coefficient is a hyperparameter pre-calibrated based on the sensor material and the background gas composition of the park. The Euclidean norm, or L2 norm, of the background feature vector represents the geometric modulus of the background interference features in high-dimensional space. This is the background feature vector output by the upstream blind source decoupling and separation stage. The norm term in the formula... As a fundamental quantitative indicator of background interference intensity, in chemical industrial park scenarios, when the concentration of background aromatic gases such as benzene, toluene, and xylene increases in tank areas or pipe corridors, the modulus of the background feature vector in high-dimensional space increases accordingly. This modulus value directly reflects the overall competitive ability of background gases to occupy sensor active sites. Physical compensation coefficient. For calibration parameters related to sensor materials, in chemical industrial park scenarios, different models of metal oxide semiconductor sensors exhibit varying adsorption affinities for background aromatic hydrocarbons and target hazardous gases such as hydrogen sulfide and chlorine. This coefficient, obtained through laboratory calibration or on-site calibration, is used to map abstract normative quantities to physically meaningful compensation scales. (Exponential function) The function is to transform the linear norm value into a nonlinear compensation factor. In the chemical industrial park scenario, when the background interfering gas concentration is low, the norm value is small, and the exponential function output is close to 1, keeping the compensation factor at a low level. However, when the background interfering gas concentration increases significantly, the norm value increases, and the exponential function output grows exponentially, causing the compensation factor to increase significantly. This nonlinear mapping relationship conforms to the exponential suppression characteristic of the background gas on the target gas signal during the multi-component competitive adsorption process. The final output compensation factor can be directly used as the basic parameter for subsequent reverse concentration amplification compensation.

[0044] Furthermore, after obtaining the compensation factor, the scalar multiplication formula is used to complete the final pure concentration calculation, that is, the original concentration intensity value output by the upstream regression module is directly multiplied by the compensation factor. The calculation formula is as follows: in, The target gas concentration value output in the first embodiment; The original target concentration intensity value was not compensated. This is a compensation factor.

[0045] Specifically, in the actual operating environment of chemical industrial parks, tank areas, pipe corridors, and equipment areas are permeated with various aromatic volatile organic compounds (such as benzene, toluene, xylene, etc.). These background interfering gases, along with the target hazardous gases (such as hydrogen sulfide, chlorine, etc.), are exposed to the active surface of the metal oxide semiconductor (MOS) gas sensor. The first embodiment described above uses a simple scalar multiplication formula to calculate the final pure concentration, directly multiplying the uncompensated original target concentration intensity value output by the upstream regression module by the background interference penalty compensation factor, aiming to inversely recover the true concentration masked by cross-interference. However, this multiplication operation implicitly assumes a serious disconnect from the physicochemical processes on the sensor surface: the degree of suppression of the target gas signal by the background interfering gases is an independent constant completely unrelated to the concentration of the target gas itself.

[0046] On the surface of a real electrochemical sensor, the competition between target gas molecules and background interfering gas molecules for the sensor's active sites follows the physical law of the multi-component Langmuir competitive adsorption isotherm. This law clearly reveals a special relationship completely ignored in the first embodiment—a nonlinear interactive coupling relationship between the target gas's own concentration level and the degree of inhibition by the background interfering gas, i.e., a concentration-dependent competitive adsorption coupling relationship. Specifically, when the target hazardous gas is at an extremely low concentration (ppb level) in the early micro-leakage stage, the sensor's active sites are almost completely occupied by a large number of background aromatic hydrocarbon molecules, and the effective adsorption occupancy fraction of the target gas approaches zero. At this time, the inhibition effect is exponentially amplified, far exceeding the linear range that can be compensated by simple multiplication. However, when the target gas is at a moderate concentration level, it begins to effectively compete for adsorption sites, and the inhibition effect is relatively mild. The simple multiplication in the first embodiment crudely approximates the nonlinear suppression curve, which dynamically changes with the target concentration, as a straight line with a constant slope. This results in severe undercompensation in the critical early warning range of extremely low concentrations, causing the trace leakage signals that should be identified to remain submerged in background noise and be missed. In the medium and high concentration range, overcompensation may occur, transmitting false ultra-high concentration signals to the downstream threshold determination module and causing false alarms.

[0047] To address the imbalance in full-range compensation accuracy caused by neglecting concentration-dependent competitive adsorption coupling in the first embodiment, an improved mechanism is proposed. Accordingly, in step S43, based on the compensation factor, the original concentration intensity value is subjected to reverse concentration amplification compensation and absolute purity concentration calculation to obtain the target gas concentration value. Figure 4 This is a flowchart illustrating the intelligent safety hazard identification method for chemical industrial parks based on a compensation factor, which performs reverse concentration and amplification compensation on the original concentration intensity value and calculates the absolute purity concentration to obtain the target gas concentration value. (See attached flowchart.) Figure 4As shown, step S43 includes: S431, estimating the competitive adsorption equilibrium occupancy fraction of the original concentration intensity value and the compensation factor to obtain the occupancy fraction value; S432, constructing a gating ratio based on the occupancy fraction value, and performing nonlinear amplification synthesis of the compensation factor under occupancy fraction gating through an exponential dynamic amplification function to obtain the dynamic compensation coefficient; S433, performing saturation clamping normalization on the dynamic compensation coefficient and multiplying it with the original concentration intensity value to obtain the target gas concentration value.

[0048] Specifically, in step S431, the competitive adsorption equilibrium occupancy fraction is estimated based on the original concentration intensity value and the compensation factor to obtain the occupancy fraction value. It should be understood that, since target hazardous gas molecules and background interfering gas molecules simultaneously compete for limited active adsorption sites on the sensor surface in chemical industrial parks, the competitive relationship between them directly determines how much target gas the sensor can detect. The first embodiment did not model this competitive process at all. Therefore, the primary step in the improved mechanism is to introduce a multi-component Langmuir competitive adsorption isotherm physical model, simultaneously substituting the upstream-transferred original concentration intensity value and the compensation factor into the competitive equilibrium equation to solve for the effective adsorption occupancy fraction of target gas molecules on the sensor's active surface. This fraction directly quantifies the core physical quantity—how many effective sensing sites the target gas can actually compete for under the current background interference intensity—and its value ranges from 0 to 1. The smaller the value, the more severely the target gas is suppressed by the background interfering gas on the sensor surface; conversely, the larger the value, the more sufficient the target gas has adsorption competitive ability. Therefore, in the technical solution of this application, the competitive adsorption equilibrium occupancy fraction is estimated by comparing the original concentration intensity value with the compensation factor to obtain the occupancy fraction value. This is used to establish an explicit physical correlation between the target gas concentration and the degree of background interference suppression based on a multi-component Langmuir isotherm physical model. The proportion of active sites actually successfully occupied by the target gas in competition with background interference molecules is solved through a competitive equilibrium equation. In this way, an explicit physical correlation between the target gas concentration and the degree of background interference suppression can be established in the inference chain, providing a precise gating basis for subsequent adaptive dynamic compensation. This ensures that the compensation strategy is no longer blindly amplified, but precisely controlled according to the physical competition situation.

[0049] More specifically, in a particular example of this application, the original concentration intensity value and compensation factor are first substituted as input parameters into the multi-component Langmuir competitive adsorption isotherm physical model. This model is based on the competitive occupancy mechanism of limited adsorption sites on the active surface of the sensor. By using the adsorption equilibrium constants of the target gas and the background interference gas and the corresponding concentration proxy, the actual number of active sites occupied by the target gas in the competitive adsorption equilibrium state is calculated. The calculation formula is as follows: in, To calculate the competitive adsorption equilibrium occupancy fraction of the target gas on the active surface of the sensor, i.e., the competitive adsorption equilibrium occupancy fraction value, its physical meaning is the proportion of active sites actually successfully occupied by the target gas molecules in competition with background interference molecules, and the value range is between 0 and 1. The input is the apparent target signal intensity without coupling compensation, i.e., the original target concentration intensity value without compensation, which is equivalent to the normalized partial pressure proxy of the target gas in the Langmuir model. The Langmuir adsorption equilibrium constant of the target hazardous gas on the surface of a specific sensor material reflects the affinity strength between the target gas and the active site, and is a pre-calibrated material property parameter. The input penalty factor characterizing the competitive adsorption intensity of the background interfering gas, i.e., the cross-interference penalty compensation factor, is equivalent to the normalized partial pressure proxy of the background interfering gas in the Langmuir model. The equivalent Langmuir adsorption equilibrium constant of the background interfering mixed gas on the same sensor material surface reflects the competitive affinity of the interfering gas group for the active site and is a pre-calibrated environmental property parameter.

[0050] It is worth mentioning that the constant term 1 in the denominator of the formula represents the normalized baseline of the total available adsorption sites on the sensor's active surface. In the context of chemical industrial parks, this baseline value characterizes the normalized representation of the total number of active sites on the surface of metal oxide semiconductor sensors available for gas molecule adsorption. This item represents the contribution of the target gas's own competitive ability to compete for active sites. In a chemical industrial park scenario, when the target hazardous gas, such as hydrogen sulfide or chlorine, is in the early leakage stage at an extremely low concentration (ppb level), it represents the original concentration intensity value. The extremely small value indicates that the target gas can hardly compete with the background gas for effective adsorption sites, and its adsorption equilibrium constant is very small. This reflects the chemical affinity between the target gas molecules and the active sites on the sensor surface; a larger constant indicates that the target gas is more easily adsorbed onto the sensor surface. The denominator of the formula contains... The term represents the contribution of background aromatic interfering gases to the competitive ability of active sites. In environments where background gases such as benzene, toluene, and xylene are present for extended periods in tank areas and pipe corridors of chemical industrial parks, the compensation factor is [not specified]. The large value indicates that the background gas severely suppresses the adsorption of the target gas, thus affecting its adsorption equilibrium constant. This reflects the overall competitive affinity of the background interfering gas group for the active site. (Molecular part of the formula) The ratio to the denominator represents the actual occupancy fraction obtained by the target gas in the fierce competition. The smaller the fraction, the more severely the target gas is suppressed by the background and the stronger the compensation required. The closer the fraction is to 1, the more sufficient the target gas has adsorption competition ability without the need for overcompensation. This step establishes an explicit physical correlation between the target gas concentration and the degree of background interference suppression in the inference link, providing a precise gating basis for subsequent adaptive dynamic compensation. This makes the compensation strategy no longer blindly amplified, but precisely controlled according to the physical competition situation.

[0051] Specifically, in step S432, a gating ratio is constructed based on the occupancy fraction value. The compensation factor is then nonlinearly amplified and synthesized under occupancy fraction gating using an exponential dynamic amplification function to obtain the dynamic compensation coefficient. It should be understood that after obtaining the competitive adsorption occupancy fraction, this physical quantity needs to be converted into a dynamic compensation coefficient that can directly affect concentration calculation. However, the static compensation factor in the first embodiment applies the same compensation force regardless of the degree to which the target gas is suppressed. This is particularly fatal in early warning scenarios of ppb-level micro-leaks in chemical industrial parks. It is precisely at the moment when the target gas is most severely suppressed and requires the strongest compensation that the compensation force provided by the first embodiment is no different from that in the medium-to-high concentration range. This results in severe undercompensation in the extremely low concentration early warning range, causing the micro-leakage signal to remain submerged in background noise and be missed. In the medium-to-high concentration range, overcompensation may occur, transmitting false ultra-high concentration signals to the downstream threshold determination module and causing false alarms. Therefore, in the technical solution of this application, a gating ratio is further constructed based on the placeholder fraction value. The compensation factor is nonlinearly amplified and synthesized under placeholder fraction gating using an exponential dynamic amplification function to obtain dynamic compensation coefficients. This is used to construct a placeholder fraction-gated exponential dynamic compensation amplifier. The compensation factor serves as the basic amplification base, and the placeholder fraction value serves as the gating signal, synthesizing dynamic compensation coefficients with adaptive nonlinear amplification characteristics. This achieves an adaptive nonlinear amplification characteristic where the more suppressed the signal, the stronger the compensation, accurately compensating for the fatal flaw of insufficient compensation in the extremely low concentration range of the first embodiment. This allows the ppb-level micro-leakage signal to be effectively enhanced from background noise to a level recognizable by the downstream threshold determination module after dynamic amplification.

[0052] More specifically, in a specific example of this application, a occupancy fraction gating ratio is first constructed based on the occupancy fraction value. This gating ratio is obtained by calculating the ratio of the complementary fraction of the occupancy fraction value to the occupancy fraction value itself. The numerator is 1 minus the occupancy fraction value, representing the proportion of active sites that the target gas failed to occupy. The denominator is the occupancy fraction value itself plus a very small normal value to prevent the denominator from being zero and causing numerical overflow. The numerical characteristic of this gating ratio is that when the occupancy fraction value approaches zero, the ratio tends to a maximum value, thereby driving the subsequent compensation coefficient to soar exponentially. When the occupancy fraction value approaches 1, the ratio tends to zero, thereby causing the compensation coefficient to naturally fall back. Subsequently, the compensation factor is used as the basic amplification basis. The gating ratio is multiplied by an adjustable hyperparameter to obtain the exponential part of the exponential function. The compensation factor is nonlinearly amplified and synthesized under occupancy fraction gating through an exponential dynamic amplification function. The calculation formula is as follows: in, The nonlinear compensation coefficient after dynamic modulation of the competitive adsorption coupling relationship, i.e., the dynamic coupling compensation coefficient, increases exponentially with the increasing degree of suppression of the target gas. This is the static penalty factor that serves as the basic amplification basis, i.e., the cross-interference penalty compensation factor; To control the adjustable hyperparameter of dynamic compensation amplification sensitivity, the larger the value, the more intense the compensation amplification effect in the low position fraction range. It is necessary to perform engineering calibration according to the specific sensor model and the background gas composition of the park. This is the competitive adsorption equilibrium occupancy fraction obtained in the previous step, i.e., the competitive adsorption equilibrium occupancy fraction value. To prevent numerical overflow caused by a denominator of zero, extremely small normal quantities (such as 1×10⁻⁶) are used. -8 ); For the occupancy score gating ratio, when →0 (the target gas is completely suppressed) This ratio tends to a maximum value, thus driving the compensation coefficient to soar exponentially. →1 (the target gas is almost unsuppressed) When the ratio approaches zero, the compensation coefficient naturally falls back to a level close to the original static penalty factor.

[0053] It is worth mentioning that the base of the exponential function in the formula This provides a baseline compensation factor. In the context of chemical industrial parks, this compensation factor has quantified the competitive adsorption intensity of background aromatic gases, serving as the base amplification factor for dynamic compensation. The exponential part of the formula... This constitutes an adaptive gated amplifier, used in chemical industrial park scenarios when target hazardous gases such as hydrogen sulfide or chlorine are in the ppb-level micro-leakage stage, with a placeholder fraction... A very small threshold indicates that the target gas can hardly compete with the background gas for effective adsorption sites; at this point, the gating ratio is... Approaching the maximum value, adjustable hyperparameter After sensitivity modulation of this ratio, the exponential function output a drastic amplification factor, which affects the dynamic compensation coefficient. Far exceeding the static compensation factor This powerfully enhances severely suppressed weak signals, precisely compensating for the fatal flaw of insufficient compensation in the extremely low concentration range in the first embodiment. This allows ppb-level micro-leakage signals, after dynamic amplification, to be effectively enhanced from background noise to a level recognizable by the downstream threshold determination module. Furthermore, when the target gas is in a medium-to-high concentration range, the occupancy fraction... A threshold close to 1 indicates that the target gas has sufficient adsorption competition ability, at which point the gating ratio is... As the exponential function approaches zero, its output approaches 1, causing the dynamic compensation coefficient to... Naturally falling back to the static compensation factor By maintaining similar levels, over-amplification is avoided, preventing false alarms caused by transmitting false ultra-high concentration signals to the downstream threshold determination module. Through this step, the improved mechanism achieves an adaptive nonlinear amplification characteristic where the more suppressed the signal, the stronger the compensation. An explicit physical correlation between the target gas concentration and the degree of background interference suppression is established in the inference link, so that the compensation strategy is no longer blindly amplified, but precisely controlled according to the physical competition situation.

[0054] Specifically, in step S433, the dynamic compensation coefficient is saturated and clamped to normalize before being multiplied by the original concentration intensity value to obtain the target gas concentration value. It should be understood that in the final concentration calculation stage, the exponential amplifier in the previous step may overcompensate numerically when the target gas concentration is extremely low, meaning the dynamic compensation coefficient tends to a maximum value. If the original concentration intensity value is directly multiplied by the dynamic compensation coefficient, it will cause a false spike in the output concentration value without physical meaning, transmitting an incorrect ultra-high concentration signal to the downstream threshold judgment module and triggering a false alarm. In the actual operation of a chemical industrial park, any false high concentration alarm may trigger an emergency shutdown procedure for the entire plant, causing huge economic losses. Therefore, in the technical solution of this application, the dynamic compensation coefficient is further saturated and clamped to normalize before being multiplied by the original concentration intensity value to obtain the target gas concentration value. This introduces a hyperbolic tangent saturated convergence constraint mechanism, utilizing the natural saturation characteristics of the hyperbolic tangent function to apply an upper bound convergence clamp to the dynamic compensation coefficient, ensuring that the final output target gas concentration value is always within a physically reasonable range. In this way, an upper bound convergence clamp can be applied to the dynamic compensation coefficient, which can effectively curb the risk of numerical explosion under extreme conditions while ensuring strong compensation capability in the extremely low concentration range. This ensures that the final output target gas concentration value has physical rationality and numerical stability across the entire range from ppb-level micro-leakage to ppm-level significant leakage.

[0055] More specifically, in a specific example of this application, the dynamic compensation coefficient is first subjected to saturation clamping normalization. The dynamic compensation coefficient is then divided by the maximum reliable concentration saturation constant pre-calibrated based on the physicochemical properties of the target hazardous gas and the upper limit of the sensor range to obtain the normalized amplification ratio. This ratio maps the dynamic compensation coefficient to the effective input domain of the hyperbolic tangent function. Subsequently, this normalized amplification ratio is substituted into the hyperbolic tangent saturated convergence function. The output range of the hyperbolic tangent function is naturally constrained to the interval between -1 and +1. In this scenario, its combined effect with the maximum reliable concentration saturation constant smoothly clamps the effective amplification factor of the dynamic compensation coefficient within the interval from 0 to the maximum reliable concentration saturation constant. Finally, the compensation coefficient after saturation clamping normalization is multiplied by the original concentration intensity value to obtain the target gas concentration value. The calculation formula is as follows: in, The final output result after joint calculation by the three mechanisms of competitive adsorption coupling modeling, dynamic gating compensation and saturation convergence constraint is the concentration value of the pure target gas after removing interference. Its value is strictly constrained within a physically reasonable concentration range. The apparent target signal baseline from the upstream regression module is the uncompensated original target concentration intensity value. The compensation coefficient with adaptive nonlinear amplification characteristics synthesized in the previous step, namely the dynamic coupling compensation coefficient, is the maximum reliable concentration saturation constant pre-calibrated based on the physicochemical properties of the target hazardous gas and the upper limit of the sensor range. It is used to define the convergence asymptotic upper bound of the hyperbolic tangent function to prevent the compensated concentration value from exceeding the physical detection limit of the sensor. As a hyperbolic tangent saturated convergent function, its output range is naturally constrained within the interval (-1, 1). In this scenario, it is similar to... The combined effect smoothly clamps the effective magnification of the dynamic compensation coefficient at [the desired value]. Within the range; To normalize the amplification ratio, the dynamic compensation coefficients are mapped to the effective input domain of the hyperbolic tangent function, ensuring that when... much smaller Time-compensated approximate linear ( (preserving sensitivity), while when The output naturally saturates and converges when approaching the maximum value. (to prevent overflow).

[0056] in the formula This provides an uncompensated baseline concentration. In a chemical industrial park setting, this raw concentration intensity reflects the apparent response intensity of the target gas on the sensor surface without considering background competitive adsorption inhibition effects. In the formula... The normalized compensation coefficients constituting the saturation convergence constraint, in the chemical industrial park scenario, are the dynamic compensation coefficients. Much smaller than the saturation constant At that time, normalized magnification ratio It lies within the approximately linear interval of the hyperbolic tangent function, at this time ≈ This makes the compensation coefficient close to This provides strong compensation in the extremely low concentration range, precisely compensating for the fatal flaw of insufficient compensation in the ppb-level micro-leakage stage of the first embodiment, effectively enhancing the weak signal that is severely submerged by the background to a level that can be recognized by the downstream threshold determination module. And when the dynamic compensation coefficient... When approaching the maximum value, the normalized amplification ratio Beyond the linear interval of the hyperbolic tangent function, the output of the hyperbolic tangent function naturally saturates and converges to 1, clamping the compensation coefficients within this range. The upper bound prevents false spikes of non-physical significance, avoiding the transmission of erroneous ultra-high concentration signals to downstream threshold judgment modules and thus preventing false alarms. In the actual operation of chemical industrial parks, this effectively curbs the risk of false high concentration alarms triggering emergency shutdown procedures for the entire plant, resulting in huge economic losses. The target gas concentration value output by the final multiplication operation has physical rationality and numerical stability across the entire range from ppb-level micro-leakage to ppm-level significant leakage. Through the synergistic effect of the above three-level linkage improvement mechanism, in the complex background environment of multiple aromatic volatile gases coexisting in chemical industrial parks, the improvement mechanism for the first time embeds the multi-component Langmuir competitive adsorption isotherm physical model into the deep learning inference link, establishing an explicit nonlinear interactive coupling relationship between the target gas concentration level and the degree of background interference suppression. This fundamentally overcomes the essential defect of the original simple multiplication mechanism in crudely linearizing the concentration-dependent competitive adsorption effect, and can be directly input into the subsequent transient spike elimination and safety threshold comparison judgment stages.

[0057] In the extremely low concentration early warning range, the fractional gating exponential dynamic compensation amplifier adaptively provides compensation strength far exceeding that of the first embodiment, effectively boosting the ppb-level micro-leakage signal, which is severely overwhelmed by background interference, to a level recognizable by the threshold judgment module, significantly reducing the false alarm rate. In the medium-to-high concentration range, the hyperbolic tangent saturation convergence constraint mechanism automatically suppresses overcompensation, preventing the transmission of false ultra-high concentration signals downstream, effectively curbing the false alarm rate. This improved mechanism enables the pure concentration inversion results to possess high-precision compensation consistency and numerical stability across the entire range from ppb-level micro-leakage to ppm-level significant leakage, thereby enhancing the early micro-leakage detection sensitivity and full-range early warning reliability of the intelligent identification system for safety hazards in chemical industrial parks under complex multi-component gas coexistence environments.

[0058] Specifically, in step S5, transient spike elimination processing is performed on the target gas concentration value, and the smoothed concentration value is compared with the preset safety threshold rule. When the target gas concentration value exceeds the safety lower limit, a hidden danger alarm report is generated. It should be understood that although the target gas concentration value output by the upstream concentration inversion and compensation correction link has completed the mapping from the abstract feature space to the physical concentration space and undergone dynamic compensation through competitive adsorption coupling modeling, the concentration value may still produce occasional high-frequency calculation spikes and small oscillations due to factors such as electromagnetic interference, network inference errors, or sensor transient noise. If the concentration value containing spikes is directly compared with the safety threshold, it will lead to false transient over-limit triggering of false alarms. In the actual operation of chemical industrial parks, each false alarm may trigger unnecessary emergency responses or even plant shutdowns, causing huge economic losses. At the same time, if the concentration value is not stabilized and the threshold is directly determined, the judgment result may frequently jump between safe and dangerous states due to signal jitter, reducing the reliability of the early warning system. Therefore, in the technical solution of this application, transient spike elimination processing is further performed on the target gas concentration value, and the smoothed concentration value is compared with the preset safety threshold rule. When the target gas concentration value exceeds the safety lower limit, a hazard alarm report is generated. This firstly eliminates the glitch oscillation generated by the deep learning network when inferring at extremely low concentrations through state transition smoothing filtering. Then, the stabilized concentration value is compared with the extremely low warning threshold extracted from the safety threshold rule through hard boundary comparison and binarization truncation. Finally, when the judgment flag is in an active state, a structured hazard alarm report is generated according to the industrial Internet of Things standard protocol. In this way, while maintaining high sensitivity early warning capability, false alarms caused by false transient spikes can be effectively suppressed, ensuring that the output hazard alarm report has high confidence and traceability, and providing a reliable basis for safety production decisions in chemical industrial parks.

[0059] More specifically, in this embodiment, step S5 includes: performing state transition smoothing filtering on the target gas concentration value to eliminate transient calculation spikes and obtain a smoothed concentration value; comparing the smoothed concentration value with the extremely low warning threshold extracted from the safety threshold rules using hard boundary comparison and binarization truncation to obtain a hazard determination flag and event context data; when the hazard determination flag is active, performing structured encoding, encapsulation, and solidification of the event context data according to the Industrial Internet of Things standard protocol to obtain a hazard alarm report.

[0060] Accordingly, in a specific example of this application, the target gas concentration value is first subjected to state transition smoothing filtering to eliminate transient calculation spikes and obtain a smooth concentration value. A smoothing window based on a local time series hidden Markov model is established. The forward maximum a posteriori probability estimation algorithm is used to combine the observed emission probability and the prior state transition probability matrix to obtain the most likely true hidden concentration state at the current moment. The algorithm evolves from the smoothest trust state of the previous observation sample to the current candidate state through historical inertia constraints. When traversing all possible continuous hidden state sets, the value of the independent variable that makes the posterior probability objective function reach the global maximum value is obtained, thereby eliminating false high-frequency calculation spikes and small oscillations and outputting an absolutely stable smooth concentration value. Subsequently, the smoothed concentration value is compared with the extremely low warning threshold extracted from the safety threshold rules using a hard boundary comparison and binary truncation judgment to obtain the hazard judgment flag and event context data. The preset safety threshold rules are deeply analyzed to extract the trace early absolute lower limit alarm line applicable to the current specific time period and specific target hazardous gas type. The Herveside step function is used to perform a hard comparison and binary classification truncation mapping between the smoothed and continuous stable concentration value and the extracted extremely low warning threshold. Once the safety lower limit is exceeded, it is immediately set to zero and activated instantaneously, and a high-confidence Boolean trigger flag, i.e., the hazard judgment flag, is output. At the same time, the specific overflow amount, the current timestamp, and the node coordinates are sealed as an accompanying tensor, i.e., event context data. Finally, when the hazard judgment flag is active, the event context data is structured, encoded, encapsulated, and solidified according to the Industrial Internet of Things (IIoT) standard protocol to obtain a hazard alarm report. The system listens to the Boolean state of the hazard judgment flag. Only when the flag is high, indicating that the event is true and there is a weak leakage hazard, does the serialization encapsulation action occur. The hazard value, three-dimensional spatial node number, and alarm confidence level in the event context data are structured, encoded, and packaged according to the IIoT standard protocol such as JSON or MQTT packet format. This generates a tamper-proof, standardized electronic file and outputs the final hazard alarm report, completing the highest-level macroscopic closed loop of this edge computing inference cycle.

[0061] In summary, the intelligent safety hazard identification method for chemical industrial parks according to the embodiments of this application is explained. First, the multi-channel voltage response and temperature / humidity parameters of the distributed sensor array are time-domain aligned, tensor-stitched, and baseline-normalized to form a comparable hybrid response matrix. Then, differences are extracted in the frequency domain, and low-frequency weak fluctuations are appropriately amplified to highlight early leakage signs. Subsequently, the hybrid features are decoupled and separated by blind sources to obtain independent background and target features, fundamentally reducing the aliasing of background fluctuations on target anomalies. Next, a quantitative inversion of the pure signal intensity is performed based solely on the target features, and the interference intensity is characterized by background features. Adaptive inverse compensation correction is applied to the inversion results according to the suppression trend of competition and occupancy with concentration changes, allowing suppressed low-concentration signals to be recovered and avoiding overcompensation in high-concentration segments. Finally, peak suppression smoothing is applied to the concentration output and compared with safety threshold rules to trigger the generation of a hazard alarm report, thereby reducing missed and false alarms.

[0062] Furthermore, an intelligent safety hazard identification system for chemical industrial parks is also provided.

[0063] Figure 5 This is a block diagram of an intelligent safety hazard identification system for chemical industrial parks, according to an embodiment of this application. Figure 5 As shown, the intelligent safety hazard identification system 100 for chemical industrial parks according to an embodiment of this application includes: a raw data preprocessing module 110, used to perform time-domain alignment, tensor splicing, and normalization processing on the raw sensor data stream containing multi-channel voltage response sequences and environmental temperature and humidity parameter sequences collected by a distributed gas sensor array in the park to obtain a hybrid response matrix; a frequency domain processing module 120, used to perform frequency domain difference feature extraction and low-frequency weak signal enhancement encoding on the hybrid response matrix to obtain a hybrid feature tensor; and a blind source decoupling and separation module 130, used to process the hybrid feature tensor... The system performs blind source decoupling and separation of background interference components and target abnormal components to obtain background feature vectors and target feature vectors; the quantitative inversion and compensation correction module 140 is used to quantitatively invert the target feature vector to obtain the target gas pure signal intensity, and performs inverse concentration compensation correction of cross-interference attenuation on the inversion result to obtain the target gas concentration value; the safety identification module 150 is used to perform transient spike elimination processing on the target gas concentration value, and compare the smoothed concentration value with the preset safety threshold rule to determine the result. When the target gas concentration value exceeds the safety lower limit, a hidden danger alarm report is generated.

[0064] As described above, the intelligent safety hazard identification system 100 for chemical industrial parks according to the embodiments of this application can be implemented in various wireless terminals, such as servers with intelligent safety hazard identification algorithms for chemical industrial parks. In one possible implementation, the intelligent safety hazard identification system 100 for chemical industrial parks according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the intelligent safety hazard identification system 100 for chemical industrial parks can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the intelligent safety hazard identification system 100 for chemical industrial parks can also be one of many hardware modules of the wireless terminal.

[0065] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for intelligent identification of safety hazards in chemical industrial parks, characterized in that, include: S1. The raw sensor data stream, which includes multi-channel voltage response sequences and environmental temperature and humidity parameter sequences, collected by the distributed gas sensor array in the park, is processed by time-domain alignment, tensor splicing, and normalization to obtain a hybrid response matrix. S2, perform frequency domain difference feature extraction and low-frequency weak signal enhancement coding on the hybrid response matrix to obtain the hybrid feature tensor; S3, perform blind source decoupling and separation of background interference components and target anomaly components on the mixed feature tensor to obtain background feature vector and target feature vector; S4, quantitatively invert the target gas pure signal intensity of the target feature vector, and perform inverse concentration compensation correction of cross-interference attenuation on the inversion result to obtain the target gas concentration value. S5 performs transient spike elimination processing on the target gas concentration value and compares the smoothed concentration value with the preset safety threshold rule. When the target gas concentration value exceeds the safety lower limit, a hidden danger alarm report is generated.

2. The intelligent identification method for safety hazards in chemical industrial parks according to claim 1, characterized in that, Step S1 includes: Synchronously extract and align the asynchronously arriving multi-channel voltage response sequence and ambient temperature and humidity parameter sequence from the original sensor data stream to obtain aligned data frames. Within each aligned data frame, the peak maximum slope characteristic and steady-state amplitude characteristic are extracted from the voltage response sequence. The slope characteristic, steady-state amplitude characteristic and the ambient temperature and humidity parameters at the same time stamp are then spliced ​​together as multi-dimensional tensors to obtain the original feature tensor. The original feature tensor is standardized and normalized to obtain the mixed response matrix.

3. The intelligent identification method for safety hazards in chemical industrial parks according to claim 2, characterized in that, The original feature tensor is standardized and normalized to obtain the hybrid response matrix, including: the original feature tensor is standardized and normalized based on the background baseline mean and background baseline variance that are adaptively updated within a long-term historical alarm-free period to obtain the hybrid response matrix.

4. The intelligent identification method for safety hazards in chemical industrial parks according to claim 1, characterized in that, Step S2 includes: Based on a one-dimensional convolutional kernel with learnable parameters, the hidden frequency domain difference features generated by the cross-desorption and adsorption stages of the mixed gas are extracted from the mixed response matrix to obtain the convolutional feature tensor; Multi-head self-attention dynamic weighting for weak leakage fluctuations is performed on any two frequency domain feature frames in the convolutional feature tensor to obtain the attention feature tensor. A hybrid feature tensor is obtained by performing deep nonlinear mapping reconstruction and compression shaping on the attention feature tensor based on a multilayer perceptron feedforward network.

5. The intelligent identification method for safety hazards in chemical industrial parks according to claim 1, characterized in that, Step S3 includes: The hybrid feature tensor is synchronously directed to two independent branch networks that do not share weights. The feature basis is split in parallel through fully connected mapping and nonlinear activation projection to obtain the background projection matrix and the target projection matrix, respectively. Geometric spatial forced decoupling is performed on the background projection matrix and the target projection matrix to obtain the orthogonal background matrix and the orthogonal target matrix, respectively. Global average pooling and one-dimensional flattening compression are performed on the orthogonal background matrix and orthogonal target matrix along the time dimension to obtain the background feature vector and target feature vector, respectively.

6. The intelligent identification method for safety hazards in chemical industrial parks according to claim 1, characterized in that, Step S4 includes: The original concentration intensity value is obtained by performing a nonlinear mapping of the target feature vector under an isolated interference state. The compensation factor is determined based on the Euclidean norm of the background feature vector and the preset physical compensation coefficient. Based on the compensation factor, the original concentration intensity value is subjected to reverse concentration amplification compensation and absolute purity concentration calculation to obtain the target gas concentration value.

7. The intelligent identification method for safety hazards in chemical industrial parks according to claim 1, characterized in that, Step S5 includes: The target gas concentration value is subjected to state transition smoothing filter to eliminate transient calculation spikes and obtain a smoothed concentration value. The smoothed concentration value is compared with the extremely low warning threshold extracted from the safety threshold rule by hard boundary comparison and binarization truncation to obtain the hazard judgment flag and event context data. When the hazard assessment flag is active, the event context data is structured, encoded, encapsulated, and solidified according to the Industrial Internet of Things standard protocol to obtain a hazard alarm report.

8. The intelligent identification method for safety hazards in chemical industrial parks according to claim 6, characterized in that, Based on the compensation factor, the original concentration intensity value is subjected to reverse concentration amplification compensation and absolute purity concentration calculation to obtain the target gas concentration value, including: The competitive adsorption equilibrium occupancy fraction was estimated by comparing the original concentration intensity value with the compensation factor to obtain the occupancy fraction value; Based on the occupancy fraction value, a gating ratio is constructed, and the compensation factor is nonlinearly amplified and synthesized under the occupancy fraction gating by an exponential dynamic amplification function to obtain the dynamic compensation coefficient. The dynamic compensation coefficient is saturated clamped and normalized, and then multiplied with the original concentration intensity value to obtain the target gas concentration value.

9. A smart safety hazard identification system for chemical industrial parks, characterized in that, include: The raw data preprocessing module is used to perform time-domain alignment, tensor splicing, and normalization on the raw sensor data stream, which includes multi-channel voltage response sequences and environmental temperature and humidity parameter sequences, collected by the distributed gas sensor array in the park, in order to obtain a hybrid response matrix. The frequency domain processing module is used to extract frequency domain difference features and encode low-frequency weak signals to obtain a hybrid feature tensor from the hybrid response matrix. The blind source decoupling and separation module is used to perform blind source decoupling and separation of background interference components and target anomaly components on the mixed feature tensor to obtain background feature vectors and target feature vectors; The quantitative inversion and compensation correction module is used to quantitatively invert the target gas purity signal intensity from the target feature vector and perform inverse concentration compensation correction of the inversion result by cross-interference attenuation to obtain the target gas concentration value. The safety identification module is used to perform transient spike elimination processing on the target gas concentration value and compare the smoothed concentration value with the preset safety threshold rules. When the target gas concentration value exceeds the safety lower limit, a hidden danger alarm report is generated.

Images