Industrial safety cross-domain collaborative analysis system based on time series data and visual data

Abstract

The application discloses an industrial safety cross-domain collaborative analysis system based on time series data and visual data, and relates to the technical field of cross-domain data collaborative processing. The system comprises a collaborative data acquisition and processing module, an infrared thermal image time series response extraction module, a cross-domain time difference compensation alignment module and a cross-domain collaborative mutation analysis module. The collaborative data acquisition and processing module is used for collecting multi-source monitoring data in real time, extracting a hot area characteristic parameter set and performing data preprocessing. The infrared thermal image time series response extraction module is used for quantifying a thermal response change rate and performing signal smoothing. The cross-domain time difference compensation alignment module is used for eliminating a trend item and scanning an offset of the multi-source monitoring data, quantifying the correlation of the thermal response change rate, obtaining the most accurate offset and realizing time alignment of multi-modal signals. The cross-domain collaborative mutation analysis module is used for quantifying the mutation synchronization strength of the multi-modal signals and identifying potential abnormalities. The leakage judgment grading feedback module is used for performing grading judgment and implementing response measures. The problems that early leakage features are difficult to be synchronously identified due to time difference of multi-modal responses in a dangerous chemical pipeline and a storage tank area are solved.

Description

Technical Field

[0001] This invention relates to the field of cross-domain data collaborative processing technology, specifically to an industrial safety cross-domain collaborative analysis system based on time-series data and visual data. Background Technology

[0002] With the continued development of high-risk industries such as energy, chemicals, and storage and transportation, industrial production systems are gradually moving towards a stage of high automation and intelligence. Along with the expansion of production facilities and the increasing complexity of processes, the connotation of industrial safety management is also constantly expanding, shifting from the safety protection of single equipment to systemic risk perception and dynamic pre-control driven by multi-source data. At the same time, the types of data generated in industrial settings are becoming increasingly diverse. How to establish connections between multi-domain and multi-modal data to form a unified basis for safety perception and decision-making has become an important direction for the development of intelligent industry.

[0003] For example, the invention patent with publication number CN119989154A discloses a two-step discrimination strategy for cross-domain generalized zero-shot industrial fault diagnosis, belonging to the field of industrial fault diagnosis technology. The method includes the following steps: S1: Constructing a CNN-based deep feature extractor for feature extraction, and using a feature alignment method based on the feature-level maximum mean difference strategy for cross-domain alignment; S2: Constructing a fault detector based on Kullback-Leibler divergence (KLD) to measure the difference between two probability distributions; S3: Using a semantic similarity-based detector to further identify the results of step S2; S4: Evaluating the classification of marked visible faults using cross-entropy loss (Lc).

[0004] For example, invention patent CN119989060A discloses a cross-source domain industrial fault diagnosis method based on generalized zero-shot learning, belonging to the field of industrial fault diagnosis. This method achieves fault diagnosis in both the source and target domains through joint learning of multiple classifiers and domain alignment consistency. It also constructs a latent hypersphere space with orthogonal constraints to connect the feature space and semantic attribute space, thereby extracting discriminative information and diagnosing both visible and unseen faults. This invention achieves efficient and accurate diagnosis of both visible and unseen faults even with missing unseen samples, improving the model's generalization performance and practicality.

[0005] However, in hazardous chemical pipelines and storage tank areas, the early stages of a leak often only manifest as a weak thermal shift in the infrared thermogram or a slight oscillation in the pressure timing signal, which cannot be identified by a single mode. Different modes have time differences in response; for example, the thermal imaging response is fast while the pressure change is lagging, leading to asynchronous timing and visual perception, and ambiguity in event determination.

[0006] Therefore, in order to address the above issues, there is an urgent need for an industrial safety cross-domain collaborative analysis system based on time-series data and visual data. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides an industrial safety cross-domain collaborative analysis system based on time-series data and visual data, which solves the problem that the time difference in multimodal responses in hazardous chemical pipelines and storage tank areas makes it difficult to synchronously identify early leakage characteristics.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: an industrial safety cross-domain collaborative analysis system based on time-series data and visual data, comprising: a collaborative data acquisition and processing module, used to acquire multi-source monitoring data of hazardous chemical storage tanks and pipeline areas in real time, and extract a set of thermal zone feature parameters based on the multi-source monitoring data, and perform data preprocessing on the multi-source monitoring data and the set of thermal zone feature parameters; an infrared thermal image time-series response extraction module, used to receive the set of thermal zone feature parameters, quantify the comprehensive change rate and diffusion behavior to obtain the thermal response change rate, and perform time synchronization and signal smoothing on the thermal response change rate; and a cross-domain time difference compensation and alignment module, used to process the multi-source monitoring data... The system performs trend term elimination and offset scanning, quantifies the correlation between the multi-source monitoring data after trend term elimination and offset and the thermal response change rate, and obtains the offset corresponding to the maximum correlation by traversing the allowable offset range, thus achieving time alignment of multimodal signals. The cross-domain collaborative mutation analysis module is used to comprehensively analyze the aligned multimodal signals, quantify the mutation synchronization strength in the time domain, and identify potential anomalies based on the spatiotemporal evolution trend of mutation synchronization strength. The leakage judgment and hierarchical feedback module is used to perform hierarchical judgment based on the mutation synchronization strength of potential anomalies, and implement response measures in conjunction with it. At the same time, it uses historical data to optimize the threshold, forming a closed-loop response mechanism.

[0011] Furthermore, the process of real-time acquisition of multi-source monitoring data for hazardous chemical storage tanks and pipeline areas, and extraction of thermal characteristic parameter sets based on the multi-source monitoring data, involves the following preprocessing steps: Real-time acquisition of multi-source monitoring data for hazardous chemical storage tanks and pipeline areas, including surface temperature field, infrared thermal image frames, and pipe pressure; millisecond-level alignment of the multi-source monitoring data based on PTP; correction of temperature measurement deviation using emissivity correction and blackbody calibration for infrared temperature field data; removal of random noise by sliding median filtering for pipe pressure data; setting reasonable physical range thresholds for the multi-source monitoring data; and data repair using the K-nearest neighbor interpolation algorithm when exceeding limits or abrupt changes are detected; and adaptive background segmentation within the infrared thermal image frames. The algorithm identifies hot zones by averaging the pixel temperature values ​​within the hot zone based on the surface temperature field. It then counts the number of pixels in the hot zone and calculates the area by combining this with the camera's imaging resolution. The algorithm calculates the centroid displacement of the hot zone by differentiating the centroid positions of two consecutive frames. Global affine registration is performed using feature points outside the hot zone to counteract camera jitter. A dense optical flow field is calculated within the hot zone, and the gradient is obtained using Sobel calculus to obtain the local divergence field. The spatial average of the divergence field within the hot zone is then used to obtain the hot zone divergence. Dimensionless normalization is applied to the multi-source monitoring data, average hot zone temperature, hot zone area, hot zone centroid displacement, and hot zone divergence. Finally, a collaborative analysis database is established, storing the original and pre-processed multi-source monitoring data, average hot zone temperature, hot zone area, hot zone centroid displacement, and hot zone divergence.

[0012] Further, the specific process of receiving the set of thermal zone characteristic parameters and quantifying the comprehensive change rate and diffusion behavior to obtain the thermal response change rate is as follows: Obtain the average temperature, area, centroid displacement, and divergence of the thermal zone; subtract the average temperature of the thermal zone from the average temperature of the previous frame and divide by the frame time interval to obtain the average temperature change rate of the thermal zone; subtract the area of ​​the thermal zone from the area of ​​the previous frame and divide by the frame time interval to obtain the area change rate of the thermal zone; multiply the area change rate by the area change weighting factor and divide by the area of ​​the thermal zone in the current frame to obtain the area change value; add the average temperature change rate and the area change value to obtain the thermal change rate value; use the negative of the thermal divergence as an exponent for natural exponentiation, add the result of the natural exponentiation to a constant and take the reciprocal to obtain the diffusion suppression value; multiply the thermal change rate value and the diffusion suppression value to obtain the thermal response change rate value.

[0013] Furthermore, the specific process of time synchronization and signal smoothing of the thermal response change rate is as follows: the thermal response change rate value sequence of consecutive frames is subjected to time interpolation and unified sampling rate processing to keep the sampling interval synchronized with the pressure data in the pipe, and the thermal response change rate value sequence is denoised and smoothed by exponential moving average, and the original and processed thermal response change rate values ​​are written into the collaborative analysis database.

[0014] Furthermore, the specific process of trend term elimination and offset scanning of multi-source monitoring data is as follows: After completing sampling synchronization, based on the sliding time window, the thermal response rate of change value sequence and the pipe pressure sequence are obtained. The trend term of the pipe pressure sequence is eliminated by the moving average method to obtain the pressure dynamic fluctuation value sequence, and the mean of thermal response rate of change and the mean of pressure dynamic fluctuation are calculated. The allowable offset range is obtained. For each offset, the pressure dynamic fluctuation value sequence is shifted backward by the current offset to align with the thermal response rate of change value sequence.

[0015] Furthermore, the specific process for quantifying the correlation between the multi-source monitoring data after trend term elimination and offset and the thermal response change rate is as follows: Calculate the difference between each thermal response change rate value and the mean thermal response change rate within the window; simultaneously calculate the difference between each pressure dynamic fluctuation value and the mean pressure dynamic fluctuation within the window; multiply the two differences to obtain the cross-domain coordinated fluctuation value; accumulate the cross-domain coordinated fluctuation values ​​of each sampling point within the window to obtain the cross-domain coordinated correlation integral value; simultaneously, accumulate the squares of the differences between each thermal response change rate value and the mean thermal response change rate within the window, and then take the square root to obtain the thermal response energy value; accumulate the squares of the differences between each pressure dynamic fluctuation value and the mean pressure dynamic fluctuation within the window, and then take the square root to obtain the pressure fluctuation energy value; multiply the thermal response energy value and the pressure fluctuation energy value to obtain the cross-domain energy benchmark value; divide the cross-domain coordinated correlation integral value by the cross-domain energy benchmark value to obtain the cross-domain dynamic correlation value.

[0016] Furthermore, the specific process of traversing the allowable offset range to obtain the offset corresponding to the maximum correlation and achieving time alignment of multimodal signals is as follows: Based on the allowable offset range, the cross-domain dynamic correlation value is calculated for each offset, and the offset corresponding to the maximum value is selected as the dynamic time difference correction value. Based on the dynamic time difference correction value, the pressure dynamic fluctuation value sequence is shifted along the time axis and synchronized with the thermal response rate of change value sequence in the time domain. The instantaneous phase difference and synchronization residual are calculated for the synchronized pressure dynamic fluctuation value and thermal response rate of change value within a sliding window. When the absolute value of the instantaneous phase difference is greater than the phase synchronization threshold, or the synchronization residual exceeds the residual threshold, the dynamic time difference correction value is iteratively adjusted. The dynamic time difference correction value, the synchronized pressure dynamic fluctuation value sequence, and the thermal response rate of change value sequence are written into the collaborative analysis database.

[0017] Further, the specific process of comprehensively analyzing the aligned multi-modal signals and quantifying the mutation synchronization intensity in the time domain is as follows: Based on a sliding time window, obtain the sequences of synchronized pressure dynamic fluctuation values and heat response change rate values, and calculate the derivatives with respect to time to obtain the pressure dynamic change rate and heat response change rate respectively. Calculate the standard deviation of the pressure dynamic change rate within the window to obtain the local fluctuation standard deviation. Obtain the synchronization residual sequence of the pressure dynamic fluctuation value and heat response change rate value within the window, and calculate the synchronization residual standard deviation; Divide the pressure dynamic change rate by the sum of the local fluctuation standard deviation and a minimum constant value to obtain the relative pressure rate value, and take the natural logarithm after adding the pressure rate value to a constant one to obtain the pressure rate gain value; Calculate the absolute difference between the current heat response change rate value and the synchronized pressure dynamic fluctuation value corresponding to the time, and divide it by the synchronization residual standard deviation to obtain the synchronization difference ratio. Take the negative of the synchronization difference ratio as the exponential power for natural exponential operation to obtain the synchronization suppression value; Multiply the absolute value of the heat response change rate, the pressure rate gain value, and the synchronization suppression value to obtain the cross-domain collaborative mutation intensity value.

[0018] Further, the specific process of identifying potential anomalies based on the spatio-temporal evolution trend of the mutation synchronization intensity is as follows: Based on a sliding time window, calculate the time change rate of the cross-domain collaborative mutation intensity value, and extract the intensity rising segment, peak point, and duration; At the same time, in terms of space, combine the centroid displacement trajectory of the hot area in the infrared thermal image frame, and conduct a correlation analysis on the time change rate and the centroid displacement rate to determine whether the mutation has spatial aggregation and persistence. When the duration of the rising cross-domain collaborative mutation intensity value exceeds the allowable threshold and forms a relevant aggregation area in space, it is marked as a potential abnormal area, and the analysis result is synchronously written into the collaborative analysis database together with the cross-domain collaborative mutation intensity value.

[0019] Further, based on the mutation synchronization intensity of potential anomalies, conduct hierarchical determination, and联动实施响应措施, and at the same time use historical data to optimize the threshold to form a closed-loop response mechanism. The specific process is as follows: Receive the cross-domain collaborative mutation intensity value corresponding to the potential abnormal area, and compare it with the multi-level risk thresholds E1 and E2; When < E1, it is determined to be in a normal state, and the conventional monitoring frequency is maintained; Periodically check the calibration status of the sensor; When E1 ≤ < E2, it is determined to be in a suspicious state, start local encrypted sampling, and mark the current hot area to trigger a low-level audible and visual prompt signal; When When the value is ≥E2, a leak warning state is determined, an audible and visual alarm is immediately issued, the valves in the corresponding area are closed and the pump is stopped, the inert gas local isolation and negative pressure exhaust device is activated, the sampling frame rate of the infrared thermal image frame is switched to a higher resolution mode, an emergency linkage command is sent to the central control system and the abnormal area number is locked; the trend of cross-domain collaborative mutation intensity value change and control action execution results are continuously tracked and written into the collaborative analysis database, and multi-level risk thresholds are periodically optimized using exponential smoothing sliding quantile and kernel density estimation algorithms.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) This invention achieves a unified spatiotemporal benchmark for visual and temporal heterogeneous data through PTP millisecond-level time alignment, blackbody calibration and Z-score normalization, fundamentally solving the problem that different modal data cannot be analyzed synchronously in traditional industrial safety monitoring, and providing a quantifiable and traceable data foundation for subsequent cross-domain collaborative computing.

[0023] (2) This invention constructs a set of characteristic parameters of the hot zone by introducing four indicators: average temperature, area, centroid displacement and divergence of the hot zone, and proposes a calculation model for the thermal response change rate. It comprehensively considers energy change and diffusion behavior, realizes the fine quantification of dynamic changes in the thermal field, and can reflect the evolution law of weak thermal characteristics in the early stage of leakage, thereby improving the sensitivity and interpretability of thermal anomaly detection.

[0024] (3) This invention achieves automatic time alignment of multimodal signals through trend term elimination, offset scanning and dynamic correlation integration methods, and proposes a cross-domain dynamic correlation value index, which effectively compensates for the time difference between fast thermal imaging response and slow pressure response, and significantly improves the synergy between multimodal data and the accuracy of early anomaly identification.

[0025] (4) This invention proposes a cross-domain collaborative mutation intensity value as the core criterion by utilizing the derivative characteristics and residual information of thermal response change rate and pressure dynamic fluctuation. It also establishes a closed-loop mechanism of hierarchical judgment, linkage control and threshold optimization by combining multi-level thresholds. This can realize early warning of leakage, automatic isolation and threshold self-learning update, significantly improving real-time performance, adaptability and industrial safety intelligence.

[0026] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0027] Figure 1 This is a structural diagram of an industrial safety cross-domain collaborative analysis system based on time-series data and visual data.

[0028] Figure 2A comparison chart of trends in cross-domain collaborative mutation intensity values;

[0029] Figure 3 This is a flowchart of cross-domain collaborative mutation detection and closed-loop response. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. As those skilled in the art will understand, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figures 1-3 This invention provides a technical solution: an industrial safety cross-domain collaborative analysis system based on time-series data and visual data, such as... Figure 1 As shown, it includes: a collaborative data acquisition and processing module, used to acquire multi-source monitoring data of hazardous chemical storage tanks and pipeline areas in real time, and extract thermal zone feature parameter sets based on multi-source monitoring data, and perform data preprocessing on multi-source monitoring data and thermal zone feature parameter sets; an infrared thermal image time-series response extraction module, used to receive thermal zone feature parameter sets, quantify the comprehensive change rate and diffusion behavior to obtain the thermal response change rate, and perform time synchronization and signal smoothing on the thermal response change rate; a cross-domain time difference compensation and alignment module, used to eliminate trend terms from multi-source monitoring data and perform offset scanning, quantify the correlation between the multi-source monitoring data after trend term elimination and offset and the thermal response change rate, traverse the allowable offset range to obtain the offset corresponding to the maximum correlation, and realize the time alignment of multi-modal signals; a cross-domain collaborative mutation analysis module, used to comprehensively analyze the aligned multi-modal signals, quantify the mutation synchronization intensity in the time domain, and identify potential anomalies based on the spatiotemporal evolution trend of mutation synchronization intensity; and a leakage judgment and graded feedback module, used to perform graded judgment based on the mutation synchronization intensity of potential anomalies, and implement response measures in conjunction with them, while using historical data for threshold optimization to form a closed-loop response mechanism.

[0032] Specifically, the process of real-time acquisition of multi-source monitoring data of hazardous chemical storage tanks and pipeline areas, and extraction of thermal characteristic parameter sets based on multi-source monitoring data, and data preprocessing of multi-source monitoring data and thermal characteristic parameter sets is as follows: real-time acquisition of multi-source monitoring data of hazardous chemical storage tanks and pipeline areas, including surface temperature field, infrared thermal image frames, and pipe pressure; data acquisition is achieved through temperature sensors, pressure sensors, and infrared imaging equipment deployed on tank walls, pipeline nodes, and thermal image monitoring points, and timestamp annotation and caching are performed by edge acquisition units. Millisecond-level alignment of multi-source monitoring data is performed based on PTP to ensure a one-to-one correspondence between infrared thermal image frames and pressure signals on the time axis. Emissivity correction and blackbody calibration are used to correct temperature measurement deviations in the infrared temperature field data. Emissivity correction is based on standard parameters set according to the material and reflectivity of the measured surface, while blackbody calibration improves the accuracy of thermal imaging by periodically comparing the data with standard blackbody temperature source signals. Sliding median filtering is used to remove random noise from the pipe pressure data, with the sliding median filter window length automatically adjusted according to the sampling frequency and signal noise level. Reasonable physical range thresholds are set for the multi-source monitoring data. When out-of-range or abrupt changes are detected, data repair is performed using the K-nearest neighbor interpolation algorithm. These physical range thresholds are set based on equipment nameplate parameters and actual operating data, such as the pressure sensor's range and the infrared thermal imager's temperature detection range. The K-nearest neighbor interpolation algorithm simultaneously searches for the nearest valid sample points in both the time and spatial domains, smoothly interpolating missing and outlier points to ensure signal continuity. Simultaneously, a background adaptive segmentation algorithm is used to identify hot zones within the infrared thermal image frame. The average temperature of the hot zone is obtained by averaging the pixel temperature values ​​within the hot zone based on the surface temperature field. The number of pixels in the hot zone is counted, and the area of ​​the hot zone is calculated by combining the camera imaging resolution. The resolution parameter is obtained through the camera calibration process, ensuring that the area of ​​the hot zone accurately reflects the size of the actual physical region. The centroid displacement of the hot zone is calculated by differential calculation of the centroid position of the hot zone in two consecutive frames. Global affine registration is performed using feature points outside the hot zone to counteract camera jitter. The affine registration uses feature point matching and least squares optimization algorithms to calculate the translation and rotation transformation matrices, ensuring that the spatial correspondence between adjacent frames is stable and reliable. Dense optical flow field is calculated within the hot zone, and Sobel gradient is used to obtain the local divergence field. The spatial average of the divergence field within the hot zone is taken to obtain the thermal zone divergence. Dense optical flow is solved using light intensity consistency constraints and local smoothness constraints. The Sobel operator is used to calculate the image gradient field, reflecting the direction of heat conduction or diffusion. The multi-source monitoring data, average temperature of the hot zone, area of ​​the hot zone, centroid displacement of the hot zone, and divergence are normalized using the minimum-maximum linear mapping method. A collaborative analysis database is established to store the original and preprocessed multi-source monitoring data, average temperature of the hot zone, area of ​​the hot zone, centroid displacement of the hot zone, and divergence. The database supports time indexing and multimodal feature indexing structures, enabling time-based data retrieval, batch querying, and parallel computation.

[0033] In this implementation scheme, multi-source sensors and infrared imaging equipment are used to achieve multi-modal data acquisition and time synchronization for hazardous chemical storage tanks and pipeline areas. Emissivity correction, blackbody calibration, and filtering algorithms are combined to eliminate temperature measurement errors and noise. Abnormal data is repaired using physical range thresholds and K-nearest neighbor interpolation. Thermal image segmentation, optical flow analysis, and affine registration are used to extract thermal zone temperature, area, displacement, and divergence features, which are then normalized and written into a collaborative analysis database. This significantly improves the temporal consistency and feature reliability of the data, providing high-precision data support for subsequent cross-domain analysis and leak risk identification.

[0034] Specifically, the process of receiving the set of characteristic parameters of the hot zone and quantifying the comprehensive change rate and diffusion behavior to obtain the thermal response change rate is as follows: The average temperature, area, centroid displacement, and divergence of the hot zone are acquired; the average temperature of the hot zone in the current frame is subtracted from the average temperature of the hot zone in the previous frame, and then divided by the frame time interval to obtain the average temperature change rate of the hot zone; the area of ​​the hot zone in the current frame is subtracted from the area of ​​the hot zone in the previous frame, and then divided by the frame time interval to obtain the area change rate of the hot zone, which reflects the expansion or contraction trend of the hot zone in the time series and can reflect the energy release and diffusion behavior of the leakage area; the frame time interval is calculated from the frame rate of the thermal imaging acquisition device, preferably within the range of 0.02 seconds to 0.1 seconds to ensure the time comparability of the change rate calculation; the area change rate of the hot zone is multiplied by the area change weighting factor and then divided by the current frame time interval. The area change value is obtained by measuring the area of ​​the hot zone. The thermal change rate value is obtained by adding the average temperature change rate of the hot zone to the area change value. The thermal change rate value reflects the intensity of the comprehensive energy change of the hot zone per unit time and is an intermediate indicator for thermal response calculation. The diffusion suppression value is obtained by using the negative of the thermal zone divergence as the exponent for natural exponentiation, adding the result of the natural exponentiation to a constant and taking the reciprocal. This value describes the spatial diffusion trend of heat. When the divergence is positive, i.e., diffusion is enhanced, the suppression value decreases. When the divergence is negative, i.e., the diffusion is concentrated, the suppression value increases, thereby suppressing the false thermal response fluctuations caused by diffusion. The thermal change rate value is multiplied by the diffusion suppression value to obtain the thermal response change rate value, which comprehensively reflects the energy evolution characteristics of the hot zone in time and space. This value is the core physical quantity used for subsequent cross-domain time difference compensation and leakage anomaly identification.

[0035] The specific formula for the thermal response rate of change is as follows:

[0036] ;

[0037] In the formula, It represents the thermal response rate of change value, which is used to quantify the overall dynamic change intensity of the hot zone in infrared thermography and is a core indicator for identifying early thermal anomalies in hazardous chemical leaks. Indicates the current frame. Indicates the previous frame; This indicates the average temperature of the hot zone in the current frame; This indicates the average temperature of the hot zone in the previous frame; The area of ​​the hot zone in the current frame represents the projected area of ​​the thermal anomaly region in the infrared frame, which is directly related to the leaked gas or the extent of heat diffusion. Indicates the area of ​​the hot zone; Indicates the area of ​​the hot zone in the previous frame; Indicates the frame time interval; It represents the thermal divergence, characterizing the diffusion trend within the thermal zone, and is used to distinguish between real leak diffusion and random fluctuations; The area change weighting factor is used to balance the relative contributions of temperature change and area change to the rate of energy change. Based on the historical change data of the average temperature and area of ​​the hot zone in continuous frames of infrared thermal images, the standard deviations are calculated respectively. The ratio of the standard deviation of the average temperature of the hot zone to the standard deviation of the area of ​​the hot zone is taken as the area change weighting factor, with a value range between 0.5 and 3.0. The result of the natural exponent calculation with the negative of the thermal divergence as the exponent is used to dynamically modulate the rate of change of thermal response based on the spatial diffusion trend of heat in the thermal zone: when the divergence is positive, it indicates that the heat in the thermal zone diffuses outward, and its exponent result is less than 1, which can suppress false responses caused by jitter, noise or background fluctuations; when the divergence is negative, it indicates that the heat accumulates, and its exponent result is greater than 1, which can amplify the energy concentration effect caused by real leakage.

[0038] In this implementation scheme, the rate of energy change and diffusion behavior of the hot zone are quantitatively characterized by dynamic calculation of characteristic parameters such as the average temperature, area, and divergence of the hot zone. By introducing an area change weighting factor and a diffusion suppression mechanism, the influence of hot zone expansion and thermal energy accumulation is balanced, eliminating random thermal noise and spurious diffusion responses. This accurately reflects the true evolution of thermal anomalies in the leak area, providing a highly sensitive and robust core thermal response indicator for cross-domain time-series alignment and early leak identification.

[0039] Specifically, the process of time synchronization and signal smoothing of the thermal response rate of change is as follows: The thermal response rate of change value sequence of consecutive frames undergoes time interpolation and sampling rate unification processing. Time interpolation uses a linear interpolation method, compensating for missing and non-uniform sampling points on the time axis to ensure the thermal response rate of change sequence remains smooth and continuous in the continuous time domain. Sampling rate unification processing involves setting a target sampling interval and performing time resampling based on the pressure signal sampling frequency, ensuring consistency between the thermal response rate of change and the pressure signal on the time scale. This avoids sampling offset between multimodal signals and is a sampling density standardization process based on completed timestamp alignment, ensuring that multimodal data... Subsequent window calculations employ comparable and consistent time steps to synchronize the sampling interval with the pipe pressure data, achieving millisecond-level synchronization accuracy. This is specifically achieved through global timestamp alignment, ensuring that the thermal response rate of change and pressure fluctuation values ​​correspond to the same physical state change within the same time window. An exponential moving average is used to denoise and smooth the thermal response rate of change value sequence using a parameterized approach. The exponential moving average algorithm assigns higher weights to new data and exponentially decaying weights to older data, effectively suppressing random impulse noise while maintaining the true trend. The smoothing factor is preferably set between 0.1 and 0.3, and can be adaptively adjusted based on the sampling frequency and noise level. The original and processed thermal response rate of change values ​​are written into a collaborative analysis database. This database uses a time-series structure, recording the original and smoothed values ​​at each moment, along with synchronization identifiers and sampling timestamps. This data is used for subsequent cross-domain alignment and dynamic correlation analysis, enabling data traceability and multimodal joint computation.

[0040] In this implementation scheme, precise alignment and noise suppression of multimodal signals in the time domain are achieved by performing time interpolation, sampling rate unification, and exponential moving average smoothing on the thermal response rate of change sequence. Further sampling density standardization is performed on top of the already synchronized timestamps to ensure complete consistency between the infrared thermal imaging signal and the pressure signal on the time scale, thereby avoiding calculation errors caused by sampling offsets. Exponential moving average smoothing filters out sudden noise, preserving true dynamic changes while improving signal stability and comparability, providing high-precision, low-noise time-series input for subsequent cross-domain collaborative analysis.

[0041] Specifically, the process of trend term elimination and offset scanning for multi-source monitoring data is as follows: After sampling synchronization is completed, a sliding time window structure is established through a unified time reference window. The window length is adaptively set according to the sampling frequency and response characteristics, with a preferred value range of 2 to 10 seconds. A fixed step size sliding method is used to ensure continuous data coverage. Based on the sliding time window, the thermal response change rate value sequence and the pipe pressure sequence are obtained. The sliding mean method is used to eliminate the trend term of the pipe pressure sequence to obtain the pressure dynamic fluctuation value sequence. That is, by calculating the average value within the sliding window and subtracting the window average value from the original pressure value, long-term trend terms are eliminated, and only short-term fluctuations are retained. A portion is used as the dynamic fluctuation value of pressure, so that the residual signal only reflects the short-term dynamic disturbance characteristics. The length of the moving average window can be automatically adjusted according to the periodic characteristics of the pressure signal and the sampling rate to maintain the integrity of the dynamic response information while filtering out trends. To avoid the phase delay effect introduced by the moving average method during signal smoothing, this embodiment introduces a phase delay compensation mechanism after the trend term is eliminated: when the moving average window adopts a front-to-back symmetrical structure, i.e., a centrally symmetrical moving window, zero-phase filtering is achieved through symmetrical sampling, thereby eliminating the trend term and avoiding signal lag. When a causal, i.e., an asymmetrical moving window is adopted, the time delay compensation amount is automatically calculated based on the window length and the frame time interval. ;in, This indicates the amount of time delay compensation. Indicates the frame time interval. The window length is specified, with an odd number of sampling points preferred. The time index is shifted forward in subsequent offset scans to ensure that the pressure dynamic fluctuation value sequence, after trend term elimination, remains strictly synchronized with the thermal response rate of change sequence in the time domain. Within the same time window, the mean thermal response rate of change and the mean pressure dynamic fluctuation are statistically calculated for normalization and correlation calculations in subsequent cross-domain collaborative analysis, ensuring energy scale consistency. An allowable offset range is obtained, adaptively set based on the physical response time difference of different modal signals, preferably between 0 and 3 seconds. Time offset scans are performed within this range to identify the optimal time alignment point between different modes. For each offset, the entire pressure dynamic fluctuation value sequence is shifted backward by the current offset to align with the thermal response rate of change value sequence, specifically achieved through time index translation. The sequence correspondence is recalculated at each offset step for subsequent cross-domain dynamic correlation value calculation, thereby determining the optimal time difference correction value.

[0042] In this implementation scheme, by constructing an adaptive sliding time window, the sliding mean trend term of the pressure sequence in the pipe is eliminated, removing low-frequency changes and retaining only short-term dynamic fluctuation signals; and the mean of thermal response change rate and pressure fluctuation is calculated simultaneously to achieve energy scale unification of multimodal data; by performing time offset scanning through the allowable offset range, the optimal alignment point between different modes is identified, thereby improving the timing matching accuracy and dynamic coordination of cross-domain signals.

[0043] Specifically, the process of quantifying the correlation between multi-source monitoring data and thermal response change rate after trend term elimination and offset is as follows: The difference between each thermal response change rate value and the mean thermal response change rate within the window is calculated, and the difference between each pressure dynamic fluctuation value and the mean pressure dynamic fluctuation value within the window is also calculated. These two differences represent the instantaneous deviation of the two signals from their average levels, reflecting the local fluctuation characteristics of the signals. Multiplying the two differences yields the cross-domain coordinated fluctuation value, which characterizes the instantaneous coupling state of whether the thermal response change rate and pressure fluctuation show the same or opposite changes at the current moment. The cross-domain coordinated fluctuation values ​​of each sampling point within the window are accumulated to obtain the cross-domain coordinated correlation integral value. The cross-domain coordinated correlation integral value quantifies the overall coordinated change intensity of the multimodal signals within the window; a larger value indicates a more significant coupling between the two signals. Simultaneously... Within the window, the squares of the differences between each thermal response rate of change value and the mean thermal response rate of change are summed, and the square root is taken to obtain the thermal response energy value. Similarly, within the window, the squares of the differences between each pressure dynamic fluctuation value and the mean pressure dynamic fluctuation are summed, and the square root is taken to obtain the pressure fluctuation energy value. These two energy values ​​represent the activity level of the corresponding modal signals within the time window and are standardized measures of their respective change intensity. The thermal response energy value and the pressure fluctuation energy value are multiplied to obtain the cross-domain energy benchmark value, which serves as the energy normalization coefficient for different modal signals. The cross-domain dynamic correlation value is obtained by dividing the cross-domain co-correlation integral value by the cross-domain energy benchmark value. This value is used to quantitatively characterize the dynamic coupling strength and synchronicity between the thermal response signal and the pressure fluctuation signal within the time window. The result ranges from -1 to 1, with positive values ​​indicating positive correlation coupling and negative values ​​indicating negative correlation changes.

[0044] The specific formula for the cross-domain dynamic correlation value is as follows:

[0045] ;

[0046] In the formula, This represents the cross-domain dynamic correlation value, which measures the degree of synchronization and coupling strength of time-series signals from two different physical modes by performing normalized cross-correlation calculations, thereby achieving coordinated alignment of cross-modal signals in the time domain; among which, This represents the current calculation time of the cross-domain dynamic correlation value, i.e., the right endpoint of the sliding time window. This indicates the starting time of the sliding time window, that is, going back from the current time. The point in time; Represents a sequence of thermal response rate of change values; This represents the average rate of change of thermal response, serving as the central reference value for the thermal response signal and used to remove the DC component. This represents a sequence of dynamic pressure fluctuation values, specifically the rapid fluctuation portion of the pressure signal, which, after time offset, is used to match the thermal response signal. It represents the average value of dynamic pressure fluctuations and serves as the central reference for pressure fluctuation signals, used to remove low-frequency drift.

[0047] In this implementation scheme, dynamic collaborative quantification of multimodal data is achieved by performing normalized cross-correlation calculation on the thermal response change rate and pressure dynamic fluctuation signal within a sliding time window; dimensional differences are eliminated by energy normalization, making signals of different physical modes comparable; the final cross-domain dynamic correlation value can accurately reflect the synchronous coupling strength and change direction of the thermo-pressure signal, providing a highly reliable quantitative basis for subsequent time difference correction and anomaly identification.

[0048] Specifically, the process of traversing the allowable offset range to obtain the offset corresponding to the maximum correlation and achieving time alignment of multimodal signals is as follows: Based on the allowable offset range, cross-domain dynamic correlation values ​​are calculated for each offset. That is, under a given time offset step, cross-domain dynamic correlation value calculation is performed on the thermal response change rate and pressure dynamic fluctuation signals to obtain the cross-domain dynamic correlation value sequence under the current offset condition. The allowable offset range is set according to the physical response delay characteristics between multimodal signals, ranging from 0 to 3 seconds. The offset step is adaptively determined according to the sampling rate to ensure the accuracy and computational efficiency of time scanning. The offset corresponding to the maximum value is selected as the dynamic time difference correction value. The point where the cross-domain dynamic correlation value reaches its global maximum indicates the strongest synchronization between the two signals at this time delay, making it the optimal correction parameter for subsequent time alignment. Based on the dynamic time difference correction value, the pressure dynamic fluctuation value sequence is shifted along the time axis to synchronize with the thermal response rate of change value sequence in the time domain. This shift operation ensures that the two sequences reflect the same physical state at the same sampling time, achieving global consistency of cross-modal signals in the time domain. The instantaneous phase difference and synchronization residual are calculated within a sliding window for the synchronized pressure dynamic fluctuation value and thermal response rate of change value. The instantaneous phase difference is obtained by extracting the envelope phase through Hilbert transform, used to measure the degree of local phase deviation between the pressure dynamic fluctuation value and the thermal response rate of change value. The synchronization residual is obtained by calculating the amplitude difference between the pressure dynamic fluctuation value and the thermal response rate of change value. The system is used to reflect the instantaneous intensity difference after alignment, and the synchronization residual energy is obtained by calculating the sum of squares of the amplitude difference. The instantaneous phase difference and the synchronization residual are used together for dynamic consistency evaluation. When the absolute value of the instantaneous phase difference is greater than the phase synchronization threshold, or the synchronization residual exceeds the residual threshold, the dynamic time difference correction value is iteratively adjusted. That is, the offset is recalculated by minimizing the synchronization error to achieve adaptive time difference update. To ensure the controllability and convergence of the iteration process, convergence and stopping conditions are set for the iterative adjustment: when the change in synchronization error between two consecutive iterations is less than the convergence threshold, or when the number of iterations reaches the maximum upper limit, preferably 5 to 10 times, it is considered convergent. If the instantaneous phase difference and the synchronization residual energy are both lower than the corresponding allowable threshold range, the iteration is terminated early and the current dynamic time difference correction value is output. Through the convergence control mechanism, iterative oscillation or over-correction can be effectively prevented, ensuring the stability and reliability of the dynamic time difference correction value. Finally, the dynamic time difference correction value, the synchronized pressure dynamic fluctuation value sequence, and the thermal response change rate value sequence are written into the collaborative analysis database.

[0049] In this implementation scheme, by traversing and calculating cross-domain dynamic correlation values ​​within the allowable offset range, the offset corresponding to the highest correlation is selected as the dynamic time difference correction value, achieving high-precision time alignment of multimodal signals. Instantaneous phase difference is calculated by extracting phase information and combined with synchronization residuals for dynamic consistency evaluation. An iterative correction mechanism is employed to minimize synchronization error, achieving adaptive optimization of alignment accuracy. This effectively eliminates the physical response time difference between different modal signals, enabling global synchronization of the thermal response change rate and pressure dynamic fluctuation signals in the time domain, providing a precise temporal consistency foundation for subsequent cross-domain collaborative mutation analysis.

[0050] Specifically, the process of comprehensively analyzing the aligned multimodal signals and quantifying the abrupt synchronization strength in the time domain is as follows: Based on a sliding time window, the synchronized pressure dynamic fluctuation value sequence and thermal response change rate value sequence are obtained. The derivative calculation is implemented using the central difference algorithm to reduce the influence of differential amplification noise while ensuring time resolution. The derivatives with respect to time are calculated to obtain the pressure dynamic change rate and thermal response change rate, respectively. The standard deviation of the pressure dynamic change rate within the window is calculated to obtain the local fluctuation standard deviation, which is used to characterize the short-time fluctuation intensity of the pressure signal within the window and provides a benchmark scale for subsequent normalization and amplification suppression. The synchronization residual sequence of pressure dynamic fluctuation values ​​and thermal response change rate values ​​within the window is obtained, which can reflect the instantaneous synchronization deviation of cross-modal signals. The synchronization residual standard deviation is calculated to measure the overall synchronization stability of the two signals within the sliding window. The relative pressure rate value is obtained by dividing the pressure dynamic change rate by the sum of the local fluctuation standard deviation and the minimum constant value, where the minimum constant value is used to prevent... When the standard deviation approaches zero, the denominator becomes unstable. The pressure rate gain is obtained by adding the pressure rate value to a constant and taking the natural logarithm. This natural logarithm maps the linear change in the pressure response to a logarithmic gain response, thus increasing the sensitivity of weak fluctuations to sudden events. The absolute difference between the current thermal response rate of change and the corresponding time-synchronized pressure dynamic fluctuation value is calculated and divided by the standard deviation of the synchronization residual to obtain the synchronization difference ratio. This ratio measures the normalized deviation amplitude of the thermal and pressure signals and is a dimensionless indicator of the degree of instantaneous asynchrony. The inverse of the synchronization difference ratio is used as the exponent for natural exponential operation to obtain the synchronization suppression value. The synchronization suppression value decays exponentially as the difference between the two signals increases, and is used to suppress the weight of asynchronous regions. The cross-domain cooperative mutation intensity value is obtained by multiplying the absolute value of the thermal response rate of change, the pressure rate gain value, and the synchronization suppression value. These three values ​​reflect time orientation, local gain, and synchronization consistency, respectively, and their product comprehensively characterizes the degree of mutational cooperation of multimodal signals within a given time window.

[0051] The specific formula for the cross-domain collaborative mutation intensity value is as follows:

[0052] ;

[0053] In the formula, It represents the cross-domain coordinated mutation intensity value, which comprehensively reflects the mutation synchronization and energy response intensity of the thermo-pressure coupled signal in the time domain. When there is an anomaly, the thermal response change rate and pressure dynamic fluctuation will simultaneously experience high-frequency disturbances and phase mismatches. The instantaneous change rate, fluctuation intensity and synchronization deviation of the two are normalized and coupled to output a quantitative index, which is used to determine whether a cross-domain mutation event has occurred. It represents the rate of change of thermal response, quantifies the instantaneous intensity of temperature and area changes in the hot zone in infrared thermography, and reflects the rate of thermal anomaly in the leak area. This represents the rate of change of thermal response; This indicates the dynamic rate of change of pressure, representing the speed at which the pressure signal changes within a local time interval. This indicates the dynamic pressure fluctuation value after time offset synchronization; This represents the standard deviation of local fluctuations, which indicates the baseline noise or normal fluctuation amplitude of the pressure signal. This represents a very small constant value, used to avoid instability caused by the denominator approaching zero, and takes the value of... ; This represents the standard deviation of the synchronous residual, and the range of acceptable normal amplitude differences.

[0054] In this embodiment, Table 1 is a data table of cross-domain cooperative mutation intensity values. The table records in detail the thermal response change rate, thermal response change rate, pressure dynamic fluctuation value, pressure dynamic change rate, local fluctuation standard deviation, synchronization residual standard deviation, and cross-domain cooperative mutation intensity value under five time windows. Specifically, time window 1 corresponds to a thermal response change rate of 0.10, a thermal response change rate of 0.02, a pressure dynamic fluctuation of 0.08, a pressure dynamic change rate of 0.015, a local fluctuation standard deviation of 0.050, a synchronous residual standard deviation of 0.080, and a cross-domain cooperative mutation intensity of 0.0041; time window 2 corresponds to a thermal response change rate of 0.18, a thermal response change rate of 0.03, a pressure dynamic fluctuation of 0.16, a pressure dynamic change rate of 0.022, a local fluctuation standard deviation of 0.053, a synchronous residual standard deviation of 0.081, and a cross-domain cooperative mutation intensity of 0.0081; time window 3 corresponds to a thermal response change rate of 0.35, a thermal response change rate of 0.06, a pressure dynamic fluctuation of 0.33, and a pressure dynamic change rate of 0.0041. The rate of change of thermal response corresponding to time window 4 is 0.045, the standard deviation of local fluctuation is 0.060, the standard deviation of synchronous residual is 0.075, and the cross-domain cooperative mutation intensity is 0.0257; the thermal response change rate corresponding to time window 4 is 0.55, the thermal response change rate is 0.09, the pressure dynamic fluctuation is 0.52, the pressure dynamic change rate is 0.070, the standard deviation of local fluctuation is 0.062, the standard deviation of synchronous residual is 0.072, and the cross-domain cooperative mutation intensity is 0.0448; the thermal response change rate corresponding to time window 5 is 0.62, the thermal response change rate is 0.085, the pressure dynamic fluctuation is 0.48, the pressure dynamic change rate is 0.055, the standard deviation of local fluctuation is 0.070, the standard deviation of synchronous residual is 0.070, and the cross-domain cooperative mutation intensity is 0.0067.

[0055] Table 1. Data on cross-domain collaborative mutation intensity values

[0056]

[0057] like Figure 2 The figure shows a trend comparison of cross-domain collaborative mutation intensity values. The figure illustrates the trend of cross-domain collaborative mutation intensity values ​​changing with time windows. The horizontal axis represents the time window number, and the vertical axis represents the cross-domain collaborative mutation intensity value. The broken lines in the figure represent the trend of cross-domain collaborative mutation intensity values ​​calculated within each time window. The two dashed lines correspond to multi-level risk thresholds, used to distinguish the boundary between normal, suspicious, and leak warning states. (Based on Table 1 and...) Figure 2It can be seen that the cross-domain coordinated mutation intensity value exhibits obvious stage-wise changes within different time windows: In the first two time windows, the cross-domain coordinated mutation intensity value is low, below E1, indicating a normal monitoring state; in the third time window, the cross-domain coordinated mutation intensity value exceeds E1 but does not reach E2, indicating the entry into a suspicious state range, with the coupling mutation between thermo-pressure signals beginning to strengthen; in the fourth time window, the cross-domain coordinated mutation intensity value further increases and exceeds E2, reaching the leakage warning range, indicating significant synchronous mutations in the thermal response change rate and pressure dynamic fluctuations, potentially indicating an actual leakage risk; in the fifth time window, the cross-domain coordinated mutation intensity value drops back below E1, indicating a return to a stable state after linkage control or self-recovery. Therefore, the cross-domain coordinated mutation intensity value can effectively reflect the dynamic coordinated relationship and abnormal evolution process between multimodal signals, providing quantitative basis and visual support for leakage risk identification and classification.

[0058] In this implementation scheme, the aligned multimodal signals are jointly analyzed within a sliding time window. The central difference algorithm is used to calculate the rate of change of pressure and thermal response, and the short-term fluctuations and synchronization deviations are quantified and normalized using the local fluctuation standard deviation and the synchronization residual standard deviation. Based on this, three quantitative indicators—relative pressure rate value, pressure rate gain value, and synchronization suppression value—are introduced to couple the time orientation, energy amplification effect, and synchronization consistency of the multimodal signals, obtaining the cross-domain cooperative mutation intensity value. This approach maintains high temporal resolution and stability under noise interference, enabling sensitive identification of early thermo-pressure cooperative mutations in hazardous chemical leaks, and significantly improving the mutation detection accuracy and robustness of multimodal signals in complex industrial scenarios.

[0059] Specifically, the process of identifying potential anomalies based on the spatiotemporal evolution trend of mutation synchronization intensity is as follows: Based on a sliding time window, the time rate of change of the cross-domain cooperative mutation intensity value is calculated. This time rate of change is calculated using a central difference algorithm to determine the rate of change of the cross-domain cooperative mutation intensity value within adjacent windows, reflecting the dynamic trend of cross-modal signal energy and synchronization. The intensity rise segment, peak point, and duration are extracted. The intensity rise segment is used to identify the initial stage of anomaly evolution, the peak point corresponds to the moment of concentrated mutation energy, and the duration reflects the stability and evolution cycle of the anomaly event. Simultaneously, spatially, the centroid displacement trajectory of the thermal zone in the infrared thermal image frame is combined with the centroid displacement rate calculated through continuous frame centroid coordinate changes to quantify the spatial distribution of the thermal anomaly region. The movement trend and diffusion direction were analyzed by correlation analysis of the time change rate and the centroid displacement rate. The Pearson correlation coefficient was used to measure the coupling degree between the time change rate and the centroid displacement rate, thereby determining whether there is synchronicity between the mutation energy change and spatial diffusion, and whether the mutation has spatial aggregation and persistence. When the duration of the cross-domain cooperative mutation intensity value increase exceeds the allowable threshold and a related aggregation area is formed in space, it indicates that the thermal anomaly and pressure disturbance show a strong coupling relationship in both time and space, characterizing the occurrence of potential leakage and abnormal diffusion events, and marking it as a potential anomaly area. The analysis results and the cross-domain cooperative mutation intensity value are synchronously written into the collaborative analysis database to provide data support for subsequent risk classification and threshold optimization.

[0060] This implementation scheme achieves spatiotemporal consistency identification of anomalous events by jointly analyzing the temporal change rate and spatial distribution characteristics of cross-domain collaborative mutation intensity values. By calculating the intensity rise segment, peak point, and duration, the initiation and peak stages of anomalous evolution can be accurately captured; combined with infrared thermographic centroid displacement rate analysis, the spatial aggregation and diffusion characteristics of the anomaly can be further determined. It can effectively distinguish between transient noise and real leakage events, enabling automatic marking and dynamic identification of potential anomalous areas, providing highly reliable input data support for subsequent risk classification and early warning decisions, and significantly improving early perception and judgment capabilities.

[0061] Specifically, based on the mutation synchronization intensity of potential anomalies, hierarchical determination is carried out, and response measures are implemented in a linked manner. At the same time, historical data is used to optimize the threshold, and the specific process of forming a closed-loop response mechanism is as follows: Receive the cross-domain collaborative mutation intensity value corresponding to the potential anomaly area, and compare the cross-domain collaborative mutation intensity value E with the multi-level risk thresholds E1 and E2; when E < E1, it is determined to be in a normal state, in a steady-state monitoring mode, and the conventional monitoring frequency is maintained; periodically check the calibration status of the sensor to prevent misjudgment caused by measurement drift; when E1 ≤ E < E2, it is determined to be in a suspicious state, indicating that there are weak mutations or high synchronization fluctuations in a local area. Enter the enhanced monitoring mode, start local encrypted sampling, and increase the sampling frequency by 1.5 to 2 times to capture finer-grained dynamic changes, and mark the current hot zone to trigger a low-level audible and visual prompt signal; when E ≥ E2, it is determined to be in a leakage warning state, indicating that there is a significant synchronous mutation in the cross-domain signal and the energy response exceeds the safety limit, inferring that there may be an actual leakage event. Immediately issue an audible and visual alarm, close the valves in the corresponding area and stop the pump, and at the same time start the local inert gas isolation and negative pressure exhaust device to achieve emergency isolation and dilution control of the leakage area, prevent the diffusion of dangerous gases, switch the sampling frame rate of the infrared thermal image to a higher-resolution mode, and increase the sampling frame rate of the infrared thermal image to 1.5 to 3 times the current frame rate to improve the thermal field capture accuracy to track the diffusion dynamics of the abnormal heat source, send an emergency linkage instruction to the central control system and lock the abnormal area number, trigger the central control linkage module to execute automatic responses, including pipeline switching, alarm log recording, and safety interlock actions; continuously track the change trend of the cross-domain collaborative mutation intensity value and the execution result of the control action, monitor the change in the mutation intensity after the implementation of the linkage measures to verify the disposal effect, and write it into the collaborative analysis database, record the cross-domain collaborative mutation intensity value, disposal action, and feedback result of each event to form a closed-loop learning sample set, and use the exponential smoothing sliding quantile and kernel density estimation algorithms to regularly optimize the multi-level risk thresholds. Among them, the exponential smoothing algorithm is used to perform trend weighting on the historical cross-domain collaborative mutation intensity value sequence to improve the responsiveness of the threshold to long-term changes; the sliding quantile algorithm is based on a fixed time window to statistically analyze the distribution interval of the cross-domain collaborative mutation intensity value and dynamically adjust the hierarchical boundary of the threshold; the kernel density estimation algorithm is used to fit the probability density of the mutation intensity, and the threshold boundary is adaptively corrected by identifying the distribution peak and inflection point, so as to maintain the stability and sensitivity of the risk classification under different working conditions.

[0062] Such as Figure 3As shown in the figure, it is a flowchart of cross - domain collaborative mutation detection and closed - loop response. It shows the whole process of hazardous chemical leakage detection based on multi - modal monitoring and collaborative analysis, forming a complete closed - loop from data acquisition to intelligent response. First, the on - site monitoring data is synchronously collected by temperature, pressure and infrared thermal image multi - dimensional sensors, and abnormal detection, interpolation repair and normalization processing are carried out to ensure data quality and time synchronization. On this basis, the hot - zone features in the infrared thermal image, including temperature, area, centroid displacement, and divergence, are identified, the thermal response change rate is calculated, and the corresponding relationship with the pressure fluctuation in the pipe is extracted. Subsequently, long - term drift errors are eliminated through trend - term elimination and offset scanning to complete the accurate time alignment of multi - modal signals. Then, the cross - domain dynamic correlation value is calculated to obtain the strength of the collaborative relationship between different signals, and the optimal timing deviation correction is calculated; further, combined with derivative calculation, relative rate and synchronous suppression terms, the cross - domain collaborative mutation strength is quantified; and then, based on the spatio - temporal evolution trend of the mutation strength, potential anomalies are identified and whether there are spatial aggregation characteristics is determined. If an anomaly is detected, risk grading is determined according to multi - level thresholds and corresponding linkage responses are triggered, such as encrypted sampling, audible and visual alarms, valve closing, and negative - pressure exhaust. All process data and response results are continuously written into the collaborative analysis database, and the thresholds are continuously optimized through exponential smoothing, moving quantiles and kernel density estimation methods to achieve a closed - loop safety control mechanism from monitoring, identification, response to self - learning.

[0063] In this implementation plan, through the hierarchical determination of the cross - domain collaborative mutation strength value and the dynamic threshold optimization, an adaptive closed - loop control from anomaly identification to linkage disposal is achieved. When a suspicious state is detected, the sampling frequency can be automatically increased to refine the timing resolution to capture early leakage characteristics; in the leakage warning stage, through the adaptive increase of the infrared frame rate and multi - module linkage control, the heat source can be locked and the area can be isolated under millisecond - level response; at the same time, the self - learning and dynamic optimization of the risk threshold are realized by combining the three algorithms of exponential smoothing, moving quantiles and kernel density estimation, so that the threshold can be automatically adjusted according to the working conditions, significantly improving the robustness, sensitivity and long - term stability in complex scenarios.

[0064] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non - exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0065] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. As those skilled in the art will understand, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An industrial safety cross-domain collaborative analysis system based on time-series data and visual data, characterized in that, include: The collaborative data acquisition and processing module is used to collect multi-source monitoring data of hazardous chemical storage tanks and pipeline areas in real time, extract thermal zone feature parameter sets based on multi-source monitoring data, and perform data preprocessing on multi-source monitoring data and thermal zone feature parameter sets. The infrared thermal image time-series response extraction module is used to receive the set of thermal region feature parameters, quantify and synthesize the rate of change and diffusion behavior to obtain the thermal response change rate, and perform time synchronization and signal smoothing on the thermal response change rate. The cross-domain time difference compensation and alignment module is used to eliminate trend terms from multi-source monitoring data and perform offset scanning. It quantifies the correlation between the multi-source monitoring data after trend term elimination and offset and the thermal response change rate, and traverses the allowable offset range to obtain the offset corresponding to the maximum correlation, thereby realizing the time alignment of multi-modal signals. The cross-domain collaborative mutation analysis module is used to comprehensively analyze aligned multimodal signals, quantify mutation synchronization strength in the time domain, and identify potential anomalies based on the spatiotemporal evolution trend of mutation synchronization strength. The specific process of quantifying the abrupt synchronization strength in the time domain of the integrated and aligned multimodal signal is as follows: Based on a sliding time window, the synchronized pressure dynamic fluctuation value sequence and thermal response change rate value sequence are obtained, and the derivative with respect to time is calculated to obtain the pressure dynamic change rate and thermal response change rate, respectively. The standard deviation of the pressure dynamic change rate within the window is calculated to obtain the local fluctuation standard deviation. The synchronized residual sequence of pressure dynamic fluctuation value and thermal response change rate value within the window is obtained, and the synchronized residual standard deviation is calculated. The relative pressure rate value is obtained by dividing the pressure dynamic change rate by the sum of the local fluctuation standard deviation and the minimum constant value. The pressure rate gain value is obtained by adding the pressure rate value to the constant and taking the natural logarithm. The absolute difference between the current thermal response change rate value and the corresponding time-synchronized pressure dynamic fluctuation value is calculated and divided by the synchronization residual standard deviation to obtain the synchronization difference ratio. The synchronization suppression value is obtained by using the negative of the synchronization difference ratio as the exponent for natural exponential operation. The cross-domain cooperative mutation intensity value is obtained by multiplying the absolute value of the thermal response change rate, the pressure rate gain value, and the synchronization inhibition value. The leakage detection and feedback module is used to classify and determine the potential anomalies based on the mutation synchronization strength, and to implement response measures in a coordinated manner. At the same time, it uses historical data to optimize the threshold, forming a closed-loop response mechanism.

2. The industrial safety cross-domain collaborative analysis system based on time-series data and visual data according to claim 1, characterized in that, The specific process of real-time acquisition of multi-source monitoring data of hazardous chemical storage tanks and pipeline areas, extraction of thermal zone feature parameter sets based on multi-source monitoring data, and data preprocessing of multi-source monitoring data and thermal zone feature parameter sets is as follows: Real-time acquisition of multi-source monitoring data for hazardous chemical storage tanks and pipeline areas, including surface temperature field, infrared thermal image frames, and internal pressure; Based on PTP, multi-source monitoring data is aligned at the millisecond level. For infrared temperature field data, emissivity correction and blackbody calibration are used to correct temperature measurement deviations. Sliding median filtering is used to remove random noise from pipe pressure data. Reasonable physical range thresholds are set for multi-source monitoring data. When out-of-limit and abrupt change points are detected, data repair is completed through K-nearest neighbor interpolation algorithm. Simultaneously, a background adaptive segmentation algorithm is used to identify hot zones within the infrared thermal image frame. The average temperature of the hot zone is obtained by averaging the pixel temperature values ​​within the hot zone based on the surface temperature field. The number of pixels in the hot zone is counted, and the area of ​​the hot zone is calculated by combining the camera imaging resolution. The centroid displacement of the hot zone is calculated by differential calculation of the centroid position of the hot zone in two consecutive frames. Global affine registration is performed using feature points outside the hot zone to counteract camera jitter. A dense optical flow field is calculated within the hot zone, and the local divergence field is obtained by calculating the gradient using Sobel. The spatial average of the divergence field within the hot zone is taken to obtain the hot zone divergence. Dimensionless normalization processing is performed on the multi-source monitoring data, average temperature of the hot zone, area of ​​the hot zone, centroid displacement of the hot zone, and hot zone divergence. A collaborative analysis database is established, storing the original and pre-processed multi-source monitoring data, average temperature of the hot zone, area of ​​the hot zone, centroid displacement of the hot zone, and hot zone divergence in the collaborative analysis database.

3. The industrial safety cross-domain collaborative analysis system based on time-series data and visual data according to claim 1, characterized in that, The specific process of quantifying the combined rate of change and diffusion behavior of the received thermal region characteristic parameter set to obtain the thermal response change rate is as follows: Obtain the average temperature, area, centroid displacement, and divergence of the hot zone; subtract the average temperature of the hot zone from the previous frame's average temperature in the current frame, and then divide by the frame time interval to obtain the average temperature change rate of the hot zone; subtract the area of ​​the hot zone from the previous frame's area in the current frame, and then divide by the frame time interval to obtain the area change rate of the hot zone; multiply the area change rate by the area change weighting factor and divide by the area of ​​the hot zone in the current frame to obtain the area change value; add the average temperature change rate and the area change value to obtain the thermal change rate value; use the negative of the thermal divergence as an exponent for natural exponentiation, add the result of the natural exponentiation to a constant, and take the reciprocal to obtain the diffusion suppression value; multiply the thermal change rate value and the diffusion suppression value to obtain the thermal response change rate value.

4. The industrial safety cross-domain collaborative analysis system based on time-series data and visual data according to claim 1, characterized in that, The specific process of time synchronization and signal smoothing of the thermal response change rate is as follows: The sequence of thermal response rate of change values ​​of consecutive frames is processed by time interpolation and sampling rate unification to keep the sampling interval synchronized with the pressure data in the pipe. The sequence of thermal response rate of change values ​​is then denoised and smoothed by exponential moving average. The original and processed thermal response rate of change values ​​are written into the collaborative analysis database.

5. The industrial safety cross-domain collaborative analysis system based on time-series data and visual data according to claim 1, characterized in that, The specific process of eliminating trend terms and performing offset scanning on multi-source monitoring data is as follows: After completing the sampling synchronization, based on the sliding time window, the thermal response rate of change value sequence and the pipe pressure sequence are obtained. The trend term is eliminated by the sliding mean method on the pipe pressure sequence to obtain the pressure dynamic fluctuation value sequence, and the mean of thermal response rate of change and the mean of pressure dynamic fluctuation are calculated. The allowable offset range is obtained. For each offset, the pressure dynamic fluctuation value sequence is shifted backward by the current offset to align with the thermal response rate of change value sequence.

6. The industrial safety cross-domain collaborative analysis system based on time-series data and visual data according to claim 1, characterized in that, The specific process for quantifying the correlation between the multi-source monitoring data after trend term elimination and offset and the rate of change of thermal response is as follows: Calculate the difference between each thermal response rate of change value and the mean thermal response rate of change within the window, and simultaneously calculate the difference between each pressure dynamic fluctuation value and the mean pressure dynamic fluctuation within the window. Multiply the two differences to obtain the cross-domain coordinated fluctuation value, and sum the cross-domain coordinated fluctuation values ​​of each sampling point within the window to obtain the cross-domain coordinated correlation integral value. Meanwhile, within the window, the squares of the differences between each thermal response rate of change value and the mean thermal response rate of change are accumulated, and the square root is taken to obtain the thermal response energy value. Within the window, the squares of the differences between each dynamic pressure fluctuation value and the mean dynamic pressure fluctuation value are accumulated, and the square root is taken to obtain the pressure fluctuation energy value; the thermal response energy value is multiplied by the pressure fluctuation energy value to obtain the cross-domain energy benchmark value. The cross-domain dynamic correlation value is obtained by dividing the cross-domain collaborative correlation integral value by the cross-domain energy benchmark value.

7. The industrial safety cross-domain collaborative analysis system based on time-series data and visual data according to claim 1, characterized in that, The specific process of obtaining the offset corresponding to the maximum correlation by traversing the allowable offset range to achieve time alignment of multimodal signals is as follows: Based on the allowable offset range, cross-domain dynamic correlation values ​​are calculated for each offset, and the offset corresponding to the maximum value is selected as the dynamic time difference correction value. Based on the dynamic time difference correction value, the pressure dynamic fluctuation value sequence is shifted along the time axis and synchronized with the thermal response change rate value sequence in the time domain. The instantaneous phase difference and synchronization residual are calculated within a sliding window for the dynamic fluctuation value of pressure and the rate of change of thermal response after synchronization. When the absolute value of the instantaneous phase difference is greater than the phase synchronization threshold, or the synchronization residual exceeds the residual threshold, the dynamic time difference correction value is iteratively adjusted. The dynamic time difference correction value, the synchronized pressure dynamic fluctuation value sequence, and the thermal response change rate value sequence are written into the collaborative analysis database.

8. The industrial safety cross-domain collaborative analysis system based on time-series data and visual data according to claim 1, characterized in that, The specific process for identifying potential anomalies based on the spatiotemporal evolution trend of mutation synchronization strength is as follows: Based on a sliding time window, the temporal rate of change of the cross-domain collaborative mutation intensity value is calculated, and the intensity rise segment, peak point, and duration are extracted. Simultaneously, spatially, the temporal rate of change and the centroid displacement rate are correlated with the displacement trajectory of the hot zone in the infrared thermal image frame to determine whether the mutation has spatial clustering and persistence. When the duration of the rise in the cross-domain collaborative mutation intensity value exceeds the allowable threshold and a related clustering area is formed in space, it is marked as a potential abnormal area. The analysis results and the cross-domain collaborative mutation intensity value are synchronously written into the collaborative analysis database.

9. The industrial safety cross-domain collaborative analysis system based on time-series data and visual data according to claim 1, characterized in that, The specific process of classifying and implementing response measures based on the mutation synchronization strength of potential anomalies, and simultaneously optimizing thresholds using historical data to form a closed-loop response mechanism is as follows: Receive the cross-domain collaborative mutation intensity value corresponding to the potential abnormal area, and compare the cross-domain collaborative mutation intensity value with the multi-level risk thresholds E1 and E2; when it is less than E1, it is determined to be in a normal state, and the conventional monitoring frequency is maintained; periodically check the calibration status of the sensor; When E1 ≤ < E2, it is determined as a suspicious state, local encrypted sampling is started, and the current hot zone is marked to trigger a low-level audible and visual prompt signal; when When the value is ≥E2, it is determined to be a leak warning state. An audible and visual alarm is immediately issued, the valves in the corresponding area are closed and the pump is stopped. At the same time, the inert gas local isolation and negative pressure exhaust device is activated, the sampling frame rate of the infrared thermal image frame is switched to a higher resolution mode, an emergency linkage command is sent to the central control system and the abnormal area number is locked. We continuously track the changing trends of cross-domain collaborative mutation intensity values ​​and the results of control actions, write them into the collaborative analysis database, and periodically optimize multi-level risk thresholds using exponential smoothing moving quantile and kernel density estimation algorithms.

Images