Intelligent detection and analysis method for leakproofness of sewage treatment tank in coal chemical industry

Abstract

The invention relates to the technical field of chemical equipment detection, in particular to an intelligent detection and analysis method for the sealing performance of a coal chemical sewage treatment tank, which comprises the following steps: S1, performing static multi-angle scanning through a lightweight distributed CT module, and obtaining tank body image data in combination with an iterative reconstruction algorithm; s2, historical data and real-time meteorological data are obtained through the sensors, when the method is used, a dynamic threshold vector containing a plurality of sensor judgment standards is generated, the accuracy of detecting the sealing performance of the tank body is improved, false alarm or missing alarm caused by environmental changes is reduced, the detection condition can be optimized in real time through generation of the dynamic threshold, and the detection efficiency is improved. The strain capacity to different monitoring environments is improved, various emergencies can be more flexibly dealt with, the radiation risk can be conveniently reduced, the personnel and environment safety is guaranteed, the structural complexity and maintenance cost can be reduced, the sampling speed and coverage are improved through distributed parallel scanning, and the device is particularly suitable for detection of large or special-shaped tanks.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical equipment detection, and in particular to an intelligent detection and analysis method for the sealing of a coal chemical wastewater treatment tank. Background Art

[0002] Testing the leak tightness of coal chemical wastewater treatment tanks is a crucial step in ensuring the proper functioning of wastewater treatment systems and preventing environmental pollution. Since wastewater generated during the coal chemical process contains a variety of hazardous substances, such as benzene, phenols, and ammonia nitrogen, leaks in wastewater treatment tanks can not only pollute water resources but also pose a threat to the surrounding environment and human health. Therefore, regular leak tightness testing is essential.

[0003] The patent publication number is CN118885784A, which states in its specification that "the present invention relates to the field of chemical industry, specifically to a method for intelligent detection and analysis of the sealing of a sewage treatment tank, comprising the following steps: S1. soil moisture detection; S2. soil moisture analysis; S3. pressure detection; S4. pressure analysis; S5. water level analysis; S6. appearance data acquisition; S7. appearance data analysis; S8. tank body evaluation and analysis; S9. leakage risk analysis. The present invention determines whether there is a possibility of leakage in the sewage treatment tank based on soil moisture, and then performs pressure analysis on the sewage treatment tank to obtain the pressure uniformity. The water level at each time point in the sewage treatment tank is monitored by a water level sensor to obtain the degree of water level fluctuation. By obtaining a three-dimensional image of the sewage treatment tank body and combining it with the degree of appearance qualification, the method can be used to determine whether the leakage is likely to occur in the sewage treatment tank. The tank evaluation coefficient of the sewage treatment tank is obtained, and finally the leakage risk index of the sewage treatment tank is comprehensively analyzed. The leakage situation of the sewage treatment tank is secondary verified, avoiding the adverse impact of the leakage on the environment. Although the above technology determines whether the sewage treatment tank is leaking by detecting the soil moisture in the inner and outer circles, it can comprehensively consider the soil moisture conditions in different areas around the sewage treatment tank to achieve the purpose of timely detection of potential leaks. However, the preset soil moisture anomaly threshold in the above technology leads to fixed thresholds for soil moisture and pressure parameters, which are easily affected by environmental interference and lead to misjudgment. In addition, the CT scanner in the above technology rotates around the tank to emit X-rays, resulting in high cost and radiation risks in large tank detection. In addition, the data processing delay is high, making it difficult to meet real-time requirements.

[0004] In summary, the development of an intelligent detection and analysis method for the sealing of coal chemical wastewater treatment tanks is still a key issue that needs to be urgently addressed in the field of chemical equipment testing technology. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem in the prior art that although the above-mentioned technology judges whether the sewage treatment tank is leaking by detecting the soil moisture of the inner and outer circles, it can comprehensively consider the soil moisture conditions in different areas around the sewage treatment tank to achieve the purpose of timely detection of potential leakage. However, the preset soil moisture abnormality threshold in the above-mentioned technology leads to the fixed threshold of soil moisture and pressure parameters, which is easily affected by environmental interference and leads to misjudgment, and is falsely reported as a leak when rainfall occurs. At the same time, the CT scanner in the above-mentioned technology rotates around the tank body to emit X-rays, resulting in high cost and radiation risk in large tank detection, and high data processing delay, making it difficult to meet real-time requirements.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention provides an intelligent detection and analysis method for the sealing of a coal chemical wastewater treatment tank, comprising the following steps: S1, performing static multi-angle scanning using a lightweight distributed CT module and combining iterative reconstruction algorithm to obtain tank image data;

[0008] S2. Obtain historical data and real-time meteorological data through sensors, pre-process them together with tank image data, and output high-quality data;

[0009] S3: Deploy a lightweight AI model on the sensor side to identify damage features in the high-quality data and offload the 3D reconstruction task to the cloud.

[0010] S4. Using a time series prediction model to generate a dynamic threshold based on the high-quality data;

[0011] S5. Dynamically adjust sensor weights based on feedback from sensor false alarm rates in high-quality data through reinforcement learning.

[0012] Furthermore, in step S1, a method for acquiring tank image data by performing static multi-angle scanning through a lightweight distributed CT module and combining iterative reconstruction algorithm is as follows:

[0013] The CT module is based on a carbon nanotube cold cathode X-ray source and achieves static multi-angle scanning through synchronous triggering, expressed as:

[0014] where Q i (t,θ i ) is the i-th CT module at time t and angle θ i The received X-ray intensity is is the local linear attenuation coefficient distribution function, α(x(W), y(W)ds) is the material density function, exp(·) is the natural exponential function, L i (θ i ) is the angle θ of the i-th modulei The ray path under the path L is dl. i The infinitesimal length is L, and L is the total length of the penetration path. All CT modules sample data in parallel to form a projection set. The radiation dose is controlled by the dose control function. The parallel sampled data are reconstructed using the compressed sensing algorithm. The alternating direction multiplier method is used to iteratively solve the reconstructed image and obtain the tank image data.

[0015] Furthermore, in step S2, the method for obtaining historical data and real-time meteorological data through sensors, pre-processing them together with the tank image data, and outputting high-quality data is as follows:

[0016] The historical data includes soil moisture historical data: E(t)∈R T×m , where E(t) represents the historical data matrix of soil moisture, T represents the time length of 30 days, and R T×m This means that this is a real number matrix with T rows and m columns. The pressure history data is: Y(t)∈R T×m , where Y(t) represents the historical data matrix of pressure and the humidity data structure is consistent, and the real-time meteorological data is:

[0017] Where U(t) represents the real-time meteorological data vector at the current time t, I(t) represents the rainfall at the current time t, and T(t) represents the temperature at the current time t. Indicates that rainfall and temperature are combined into a 1×2 vector, R 1×2 Indicates that this is a vector with one row and two columns containing two real values. The feature tensor of the tank image data is: O(t)∈R c×h×w , where T is the length of the historical window (30 days), m is the number of deployment points for each type of sensor, c, h, w are the number of image channels and resolution, and the sliding window average convolution reconstruction model is used for time window resampling, expressed as: Among them, P i (t)∈{E,Y,U} is the weighted average of the i-th sensor at the current time t, A(χ-t) is the time Gaussian kernel, Δt is the uniform sampling interval of 1 hour, χ is the integral variable representing any moment in the time window, P i (χ) is the original observation value of the i-th sensor at time χ, It is the sum of the times before and after the current time t. Integrate over the time period Normalize the integral result.

[0018] Furthermore, in step S2, the method for obtaining historical data and real-time meteorological data through sensors, pre-processing them together with the tank image data, and outputting high-quality data is as follows:

[0019] The feature cross-attention mechanism is used to align the heterogeneous information of the image tensor and the meteorological sequence, and projected into a unified dimension through a fully connected layer to generate a fused tensor. Then, the joint outlier detection and discriminative attention mechanism is used to perform high-quality screening on the fused tensor. The expression formula is: in is the Laplace operator of the soil moisture data, is the historical average rainfall, ε edge (O(t)) is the edge loss of the image, δ1, δ2, δ3 are task weight parameters, S(t) is the meteorological correction factor at the current time t for dynamic threshold calculation, is the square of the norm of the second-order derivative, I(t) is the real-time rainfall at the current time t, is the average rainfall in the past 30 days, when S(t)>φ ou , then the data at that moment is removed and the rest of the data is normalized and output as high-quality data D fu ∈R T′×d″ .

[0020] Furthermore, in step S3, a lightweight AI model is deployed on the sensor side to identify damage features in the high-quality data, and the 3D reconstruction task is offloaded to the cloud. The method is as follows:

[0021] The lightweight AI model uses a compressed lightweight convolutional network (MobileNetV3-SSDLite) to perform real-time detection based on the damage features in the high-quality data, infer and output a crack area mask, and estimates the crack area in physical space based on the crack area mask, expressed as follows: Among them F cr is the total area of ​​the crack region, is the sum of all pixels x′, y′ in the entire image, G(x′, y′) is the crack detection result at the pixel point x′, y′, Δx′ is the actual physical length represented by each pixel in the horizontal direction, Δy′ is the actual physical length represented by each pixel in the vertical direction, and in F cr >F th =1mm 2 Trigger an alarm.

[0022] Furthermore, in step S3, a lightweight AI model is deployed on the sensor side to identify damage features in the high-quality data, and the 3D reconstruction task is offloaded to the cloud. The method is as follows:

[0023] The 3D reconstruction task is offloaded to the cloud. Only the edge crack feature tensor is extracted and uploaded to the cloud using a 5G link. After receiving the edge crack feature tensor, the cloud performs the 3D reconstruction task. The 3D reconstruction task includes CT slice stacking and 3D volume reconstruction. The expression formula is: Among them G i is the i-th CT slice, K i is the pose registration transformation, J i is the weight coefficient of the i-th CT module, H(x″,y″,z) is the final reconstructed value at the 3D reconstructed voxel coordinates x″,y″,z, is the accumulation of the reconstruction results of the i-th CT module, G i (x′, y′) is the crack detection result of the i-th CT module at pixel x′, y′, K i (·) is the two-dimensional image data G collected by the i-th CT module i A function that back-projects into three-dimensional space, where x″, y″, and z are the coordinate positions of the voxel in three-dimensional space, generating a voxel model.

[0024] Furthermore, in step S4, the method of using the time series prediction model to generate a dynamic threshold based on the high-quality data is:

[0025] The time series prediction model performs trend modeling on the environmental series, and the expression formula is:

[0026] in is the predicted value at the next moment, is the input historical high-quality data sequence, is the length of the time window, It is a time series forecasting model Parameters, the predicted value and the historical mean are integrated to form a dynamic threshold, expressed as: in is the historical mean of soil moisture, γ∈[0,1] is the time smoothing factor, ι is the meteorological sensitivity factor, C wea is the meteorological correction factor, Indicates that the threshold adjusted according to the current conditions is used for judgment at future moments. is the predicted value at the next moment.

[0027] Furthermore, in step S4, the method of using the time series prediction model to generate a dynamic threshold based on the high-quality data is:

[0028] Assuming that the meteorological influencing factor vector is affected by rainfall increment and temperature change, a linear regression model is established to calculate the meteorological correction factor, which is expressed as: C wea =κ ⊥ ·Vt +λ, where κ=[κ1,κ2,κ3,κ4] ⊥ is the regression weight coefficient, is a Gaussian disturbance term used to model unobservable external disturbances, and κ is a ridge regression for history fitting optimization. When there are multiple monitoring variables including humidity, pressure, and water level, an overall dynamic threshold is generated, expressed as: Where η is the corresponding historical mean vector, is the unit column vector used to broadcast the meteorological factor, X a is the final output dynamic threshold vector, The dynamic threshold of the humidity dimension corresponds to the data judgment standard of the soil moisture sensor. The dynamic threshold of the pressure dimension corresponds to the data judgment standard of the pressure sensor. It is the judgment standard of the water level or liquid level sensor corresponding to the dynamic threshold of the real-time meteorological dimension.

[0029] Furthermore, in step S5, the method for dynamically adjusting the sensor weights based on the feedback of the sensor false alarm rate in the high-quality data through reinforcement learning is as follows:

[0030] The sensor false alarm rate in the high-quality data is assumed to include n sensors, and the observation data at time t is: Where V t is the multi-sensor observation data vector at time t, is the data value collected by the first sensor at time t, n represents the total number of sensors, and the false alarm feedback label is formed based on the alarm result and manual confirmation result. The expression formula is: Where ρ is the false positive penalty factor, FP t (i) is the false alarm mark of the i-th sensor at time t, It is a missed mark, TP t (i) , is the correct mark, is a positive number, r t is the current state e t The action selected is, n is the number of sensors, is the weight of the i-th sensor at time t, It is the comprehensive discrimination accuracy index of the i-th sensor at time t.

[0031] Furthermore, in step S5, the method for dynamically adjusting the sensor weights based on the feedback of the sensor false alarm rate in the high-quality data through reinforcement learning is as follows:

[0032] The dynamic adjustment of sensor weights uses Q-Learning to iteratively update the policy value function to output the optimal action, expressed as:

[0033] in is the learning rate, τ is the discount factor, is the current state e t Next action r t Value, ← means Update of value, M t is the immediate reward obtained at the current time t, In the next state e t+1 The maximum of all possible actions r′ value, r t is the current state e t Next, select the action you want to take. In the current state t Next Select The action with the largest value, υ is the exploration rate, 1-υ represents the probability of selecting the currently known optimal action, SJ is to randomly select an action from the action space, the dynamic adjustment of sensor weights, and the correction of the current sensor weight vector according to the optimal action. At the same time, when the h sensor shows a high false alarm rate, that is, long-term If the value is low, the h sensor will gradually decrease, automatically increasing the influence of other sensors in the overall judgment.

[0034] Beneficial effects

[0035] Compared with the known public technology, the technical solution provided by the present invention has the following advantages:

[0036] Beneficial effects:

[0037] When in use, the present invention generates a dynamic threshold vector containing multiple sensor judgment criteria, which is beneficial to improving the accuracy of detecting the sealing of the tank body and reducing false alarms or missed alarms caused by environmental changes. The generation of dynamic thresholds can optimize the detection conditions in real time, improve the adaptability to different monitoring environments, make it more flexible to deal with various emergencies, facilitate reducing radiation risks, ensure the safety of personnel and the environment, and help reduce structural complexity and maintenance costs. Distributed parallel scanning improves sampling speed and coverage, and is particularly suitable for large or special-shaped tank body detection.

[0038] When the present invention is in use, it dynamically adjusts the degree of influence of each sensor on the overall decision, realizing an intelligent evaluation mechanism of "survival of the fittest", which is conducive to enhancing the adaptive ability of the present invention, does not require frequent manual intervention, and has the ability of continuous learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The present invention provides a flow chart of an intelligent detection and analysis method for the sealing performance of a coal chemical wastewater treatment tank. DETAILED DESCRIPTION

[0040] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0041] It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the numbers used in this manner are interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatuses.

[0042] The present invention is described in further detail below with reference to the accompanying drawings:

[0043] Example:

[0044] like Figure 1 As shown, the present invention provides an intelligent detection and analysis method for the sealing of coal chemical wastewater treatment tanks, comprising the following steps: S1, performing static multi-angle scanning by a lightweight distributed CT module, and combining iterative reconstruction algorithm to obtain tank image data;

[0045] Furthermore, in step S1, a method for acquiring tank image data by performing static multi-angle scanning through a lightweight distributed CT module and combining iterative reconstruction algorithm is as follows:

[0046] The CT module is based on a carbon nanotube cold cathode X-ray source and achieves static multi-angle scanning through synchronous triggering, expressed as:

[0047] where Q i (t,θ i ) is the i-th CT module at time t and angle θ i The received X-ray intensity is is the local linear attenuation coefficient distribution function, α(x(W), y(W)ds) is the material density function, exp(·) is the natural exponential function, L i (θ i ) is the angle θ of the i-th module i The ray path under the path L is dl. i The infinitesimal length is L, and L is the total length of the penetration path. All CT modules sample data in parallel to form a projection set. The radiation dose is controlled by the dose control function. The parallel sampled data are reconstructed using the compressed sensing algorithm. The alternating direction multiplier method is used to iteratively solve the reconstructed image and obtain the tank image data.

[0048] In this embodiment, the present invention utilizes a lightweight distributed CT module to perform static multi-angle image acquisition on the sewage treatment tank, and combines it with advanced image reconstruction algorithms to generate high-quality tank image data. The CT module has a built-in carbon nanotube-based cold cathode X-ray source. Through a unified time synchronization trigger mechanism, it simultaneously emits X-rays from multiple fixed angles and receives the signal intensity after penetration. The dose control function dynamically adjusts the intensity of each exposure, which facilitates reducing radiation risks, ensures the safety of personnel and the environment, and helps reduce structural complexity and maintenance costs. Distributed parallel scanning improves sampling speed and coverage, and is particularly suitable for large or special-shaped tank inspections.

[0049] S2. Obtain historical data and real-time meteorological data through sensors, pre-process them together with tank image data, and output high-quality data;

[0050] Furthermore, in step S2, the method for obtaining historical data and real-time meteorological data through sensors, pre-processing them together with the tank image data, and outputting high-quality data is as follows:

[0051] The historical data includes soil moisture historical data: E(t)∈R T×m , where E(t) represents the historical data matrix of soil moisture, T represents the time length of 30 days, and R T×m This means that this is a real number matrix with T rows and m columns. The pressure history data is: Y(t)∈R T×m , where Y(t) represents the historical data matrix of pressure and the humidity data structure is consistent, and the real-time meteorological data is:

[0052] Where U(t) represents the real-time meteorological data vector at the current time t, I(t) represents the rainfall at the current time t, and T(t) represents the temperature at the current time t. Indicates that rainfall and temperature are combined into a 1×2 vector, R 1×2 Indicates that this is a vector with one row and two columns containing two real values. The feature tensor of the tank image data is: O(t)∈Rc×h×w , where T is the length of the historical window (30 days), m is the number of deployment points for each type of sensor, c, h, w are the number of image channels and resolution, and the sliding window average convolution reconstruction model is used for time window resampling, expressed as: Among them, P i (t)∈{E,Y,U} is the weighted average of the i-th sensor at the current time t, A(χ-t) is the time Gaussian kernel, Δt is the uniform sampling interval of 1 hour, χ is the integral variable representing any moment in the time window, P i (χ) is the original observation value of the i-th sensor at time χ, It is the sum of the times before and after the current time t. Integrate over the time period Normalize the integral result.

[0053] Furthermore, in step S2, the method for obtaining historical data and real-time meteorological data through sensors, pre-processing them together with the tank image data, and outputting high-quality data is as follows:

[0054] The feature cross-attention mechanism is used to align the heterogeneous information of the image tensor and the meteorological sequence, and projected into a unified dimension through a fully connected layer to generate a fused tensor. Then, the joint outlier detection and discriminative attention mechanism is used to perform high-quality screening on the fused tensor. The expression formula is: in is the Laplace operator of the soil moisture data, is the historical average rainfall, ε edge (O(t)) is the edge loss of the image, δ1, δ2, δ3 are task weight parameters, S(t) is the meteorological correction factor at the current time t for dynamic threshold calculation, is the square of the norm of the second-order derivative, I(t) is the real-time rainfall at the current time t, is the average rainfall in the past 30 days, when S(t)>φ ou , then the data at that moment is removed and the rest of the data is normalized and output as high-quality data D fu ∈R T′×d″ .

[0055] In this embodiment, the present invention collects soil moisture and pressure data from the past 30 days to construct a two-dimensional time series matrix to record daily environmental fluctuations. At the same time, the current rainfall and temperature are obtained in real time and integrated into a meteorological feature vector, which together with the historical environmental data constitutes a comprehensive input. On the other hand, the tank image data is expressed in tensor form, containing multi-dimensional information such as image channels, pixel resolution, and sensor deployment locations within the historical time window. The unified fusion processing of multi-source heterogeneous data is conducive to improving the accuracy of spatiotemporal alignment. The dynamic rainfall correction mechanism effectively suppresses the impact of environmental interference on the misjudgment rate.

[0056] S3: Deploy a lightweight AI model on the sensor side to identify damage features in the high-quality data and offload the 3D reconstruction task to the cloud.

[0057] Furthermore, in step S3, a lightweight AI model is deployed on the sensor side to identify damage features in the high-quality data, and the 3D reconstruction task is offloaded to the cloud. The method is as follows:

[0058] The lightweight AI model uses a compressed lightweight convolutional network (MobileNetV3-SSDLite) to perform real-time detection based on the damage features in the high-quality data, infer and output a crack area mask, and estimates the crack area in physical space based on the crack area mask, expressed as follows: Among them F cr is the total area of ​​the crack region, is the sum of all pixels x′, y′ in the entire image, G(x′, y′) is the crack detection result at the pixel point x′, y′, Δx′ is the actual physical length represented by each pixel in the horizontal direction, Δy′ is the actual physical length represented by each pixel in the vertical direction, and in F cr >F th =1mm 2 Trigger an alarm.

[0059] Furthermore, in step S3, a lightweight AI model is deployed on the sensor side to identify damage features in the high-quality data, and the 3D reconstruction task is offloaded to the cloud. The method is as follows:

[0060] The 3D reconstruction task is offloaded to the cloud. Only the edge crack feature tensor is extracted and uploaded to the cloud using a 5G link. After receiving the edge crack feature tensor, the cloud performs the 3D reconstruction task. The 3D reconstruction task includes CT slice stacking and 3D volume reconstruction. The expression formula is: Among them G i is the i-th CT slice, K i is the pose registration transformation, J iis the weight coefficient of the i-th CT module, H(x″,y″,z) is the final reconstructed value at the 3D reconstructed voxel coordinates x″,y″,z, is the accumulation of the reconstruction results of the i-th CT module, G i (x′, y′) is the crack detection result of the i-th CT module at pixel x′, y′, K i (·) is the two-dimensional image data G collected by the i-th CT module i A function that back-projects into three-dimensional space, where x″, y″, and z are the coordinate positions of the voxel in three-dimensional space, generating a voxel model.

[0061] In this embodiment, the present invention realizes real-time recognition of damage features (such as cracks) in high-quality data by deploying a lightweight artificial intelligence model on the sensor side, and offloads the computationally intensive three-dimensional reconstruction tasks to the cloud to achieve efficient and low-latency monitoring and response, thereby greatly reducing the detection delay and meeting the second-level warning requirements. 5G communication is used to offload the three-dimensional reconstruction tasks, significantly reducing the edge computing load and transmission costs.

[0062] S4. Using a time series prediction model to generate a dynamic threshold based on the high-quality data;

[0063] Furthermore, in step S4, the method of using the time series prediction model to generate a dynamic threshold based on the high-quality data is:

[0064] The time series prediction model performs trend modeling on the environmental series, and the expression formula is:

[0065] in is the predicted value at the next moment, is the input historical high-quality data sequence, is the length of the time window, It is a time series forecasting model Parameters, the predicted value and the historical mean are integrated to form a dynamic threshold, expressed as: in is the historical mean of soil moisture, γ∈[0,1] is the time smoothing factor, ι is the meteorological sensitivity factor, C wea is the meteorological correction factor, Indicates that the threshold adjusted according to the current conditions is used for judgment at future moments. is the predicted value at the next moment.

[0066] Furthermore, in step S4, the method of using the time series prediction model to generate a dynamic threshold based on the high-quality data is:

[0067] Assuming that the meteorological influencing factor vector is affected by rainfall increment and temperature change, a linear regression model is established to calculate the meteorological correction factor, which is expressed as: C wea =κ ⊥ ·V t +λ, where κ=[κ1,κ2,κ3,κ4] ⊥ is the regression weight coefficient, is a Gaussian disturbance term used to model unobservable external disturbances, and κ is a ridge regression for history fitting optimization. When there are multiple monitoring variables including humidity, pressure, and water level, an overall dynamic threshold is generated, expressed as: Where η is the corresponding historical mean vector, is the unit column vector used to broadcast the meteorological factor, X a is the final output dynamic threshold vector, The dynamic threshold of the humidity dimension corresponds to the data judgment standard of the soil moisture sensor. The dynamic threshold of the pressure dimension corresponds to the data judgment standard of the pressure sensor. It is the judgment standard of the water level or liquid level sensor corresponding to the dynamic threshold of the real-time meteorological dimension.

[0068] In this embodiment, the present invention uses a time series prediction model to perform trend modeling on historical high-quality data, analyzing the periodicity or variation patterns in the data. When monitoring the health of a sewage treatment plant tank, the present invention predicts the state at the next moment based on the historical sequence of environmental data such as soil moisture and pressure. The predicted value is then integrated with the historical mean to form a dynamic threshold. When the historical soil moisture data fluctuates significantly, the present invention adjusts the normal range of soil moisture by combining meteorological data (such as rainfall and temperature). A meteorological correction factor is calculated based on real-time meteorological conditions (such as temperature and rainfall) to adjust the dynamic threshold. When rainfall has increased recently, the normal threshold of soil moisture may rise to avoid false alarms or missed alarms due to sudden weather events. The present invention uses a linear regression model to calculate the meteorological correction factor and optimizes the predicted value according to meteorological changes (such as rainfall increments and temperature changes). The meteorological correction factor is calculated using a ridge regression algorithm to perform historical data fitting and external interference (such as meteorological changes) modeling to ultimately generate a dynamic threshold vector containing multiple sensor judgment criteria, providing real-time adaptive thresholds for different sensors, facilitating more accurate judgment of whether the alarm criteria have been met, and helping to improve the accuracy of the present invention in detecting the sealing of the tank body, reducing false alarms or missed alarms due to environmental changes. The generation of dynamic thresholds can optimize the detection conditions in real time, improve the adaptability to different monitoring environments, and enable it to respond more flexibly to various emergencies.

[0069] S5. Dynamically adjust sensor weights based on feedback from sensor false alarm rates in high-quality data through reinforcement learning;

[0070] Furthermore, in step S5, the method for dynamically adjusting the sensor weights based on the feedback of the sensor false alarm rate in the high-quality data through reinforcement learning is as follows:

[0071] The sensor false alarm rate in the high-quality data is assumed to include n sensors, and the observation data at time t is: Where V t is the multi-sensor observation data vector at time t, is the data value collected by the first sensor at time t, n represents the total number of sensors, and the false alarm feedback label is formed based on the alarm result and manual confirmation result. The expression formula is: Where ρ is the false positive penalty factor, FP t (i) is the false alarm mark of the i-th sensor at time t, It is a missed mark, TP t (i) , is the correct mark, is a positive number, r t is the current state e t The action selected is, n is the number of sensors, is the weight of the i-th sensor at time t, It is the comprehensive discrimination accuracy index of the i-th sensor at time t.

[0072] Furthermore, in step S5, the method for dynamically adjusting the sensor weights based on the feedback of the sensor false alarm rate in the high-quality data through reinforcement learning is as follows:

[0073] The dynamic adjustment of sensor weights uses Q-Learning to iteratively update the policy value function to output the optimal action, expressed as:

[0074] in is the learning rate, τ is the discount factor, is the current state e t Next action r t Value, ← means Update of value, M t is the immediate reward obtained at the current time t, In the next state e t+1 The maximum of all possible actions r′ value, r t is the current state e t Next, select the action you want to take. In the current state t Next Select The action with the largest value, υ is the exploration rate, 1-υ represents the probability of selecting the currently known optimal action, SJ is to randomly select an action from the action space, the dynamic adjustment of sensor weights, and the correction of the current sensor weight vector according to the optimal action. At the same time, when the h sensor shows a high false alarm rate, that is, long-term If the value is low, the h sensor will gradually decrease, automatically increasing the influence of other sensors in the overall judgment.

[0075] In this embodiment, the present invention collects observation data from multiple sensors at each moment. Combining alarm results with manual verification, the system determines whether each sensor's warning is a false alarm, a missed alarm, or a correct alarm. If a sensor frequently issues an alarm but no abnormality is detected upon manual verification, it is marked as a false alarm. Conversely, if no alarm is issued but an abnormality is present, it is marked as a missed alarm, while normal alarms are marked as correct. False alarm feedback is incorporated into the present invention as a reward mechanism for reinforcement learning. The Q-Learning algorithm within reinforcement learning learns the optimal decision-making strategy for different sensor combinations and environmental conditions. For example, the Q-value function determines which data from multiple sensors is most reliable. The present invention continuously iteratively updates the Q-value, gradually learning the optimal action to take in response to different sensor performances. Specifically, it adjusts the weight of each sensor. If a sensor exhibits a high false alarm rate over a long period of time, its corresponding weight is gradually reduced, while the weights of other, more reliable sensors are automatically increased. In this way, the present invention dynamically adjusts the degree of influence of each sensor on the overall decision, realizing an intelligent evaluation mechanism of "survival of the fittest", which is conducive to enhancing the adaptive ability of the present invention, without the need for frequent human intervention, and having the ability of continuous learning.

[0076] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A coal chemical wastewater treatment tank sealing intelligent detection and analysis method, characterized in that: The following steps are involved: S1. Static multi-angle scanning is performed through a lightweight distributed CT module, combined with an iterative reconstruction algorithm to obtain tank image data; S2. Obtain historical data and real-time meteorological data through sensors, pre-process them together with tank image data, and output high-quality data; S3: Deploy a lightweight AI model on the sensor side to identify damage features in the high-quality data and offload the 3D reconstruction task to the cloud. S4. Using a time series prediction model to generate a dynamic threshold based on the high-quality data; S5. Dynamically adjust sensor weights based on feedback from sensor false alarm rates in high-quality data through reinforcement learning.

2. The method for intelligent detection and analysis of the sealing performance of a coal chemical wastewater treatment tank according to claim 1, characterized in that: In step S1, a method for acquiring tank image data is as follows: static multi-angle scanning is performed by a lightweight distributed CT module and combined with an iterative reconstruction algorithm. The CT module is based on a carbon nanotube cold cathode X-ray source and achieves static multi-angle scanning through synchronous triggering, expressed as: where Q i (t,θ i ) is the i-th CT module at time t and angle θ i The received X-ray intensity is is the local linear attenuation coefficient distribution function, α(x(W), y(W)ds) is the material density function, exp(·) is the natural exponential function, L i (θ i ) is the angle θ of the i-th module i The ray path under the path L is dl. i The infinitesimal length is L, and L is the total length of the penetration path. All CT modules sample data in parallel to form a projection set. The radiation dose is controlled by the dose control function. The parallel sampled data are reconstructed using the compressed sensing algorithm. The alternating direction multiplier method is used to iteratively solve the reconstructed image and obtain the tank image data.

3. The method for intelligent detection and analysis of the sealing performance of a coal chemical wastewater treatment tank according to claim 2, characterized in that: In step S2, the method for obtaining historical data and real-time meteorological data through sensors, pre-processing them together with the tank image data, and outputting high-quality data is as follows: The historical data includes soil moisture historical data: E(t)∈R T×m , where E(t) represents the historical data matrix of soil moisture, T represents the time length of 30 days, and R T×m This means that this is a real number matrix with T rows and m columns. The pressure history data is: Y(t)∈R T×m , where Y(t) represents the historical data matrix of pressure and the humidity data structure is consistent, and the real-time meteorological data is: Where U(t) represents the real-time meteorological data vector at the current time t, I(t) represents the rainfall at the current time t, and T(t) represents the temperature at the current time t. Indicates that the rainfall and temperature are combined into a 1×2 vector, R 1×2 This indicates that this is a vector with one row and two columns containing two real values. The feature tensor of the tank image data is: O(t)∈R c×h×w , where T is the length of the historical window (30 days), m is the number of deployment points for each type of sensor, c, h, w are the number of image channels and resolution, and the sliding window average convolution reconstruction model is used for time window resampling, expressed as: Among them, P i (t)∈{E,Y,U} is the weighted average of the i-th sensor at the current time t, A(χ-t) is the time Gaussian kernel, Δt is the uniform sampling interval of 1 hour, χ is the integral variable representing any moment in the time window, P i (χ) is the original observation value of the i-th sensor at time χ, It is the sum of the times before and after the current time t. Integrate over the time period Normalize the integral result.

4. The method for intelligent detection and analysis of the sealing performance of a coal chemical wastewater treatment tank according to claim 3, characterized in that: In step S2, the method for obtaining historical data and real-time meteorological data through sensors, pre-processing them together with the tank image data, and outputting high-quality data is as follows: The feature cross-attention mechanism is used to align the heterogeneous information of the image tensor and the meteorological sequence, and projected into a unified dimension through a fully connected layer to generate a fused tensor. Then, the joint outlier detection and discriminative attention mechanism is used to perform high-quality screening on the fused tensor. The expression formula is: in is the Laplace operator of the soil moisture data, is the historical average rainfall, ε edge (O(t)) is the edge loss of the image, δ1, δ2, δ3 are task weight parameters, S(t) is the meteorological correction factor at the current time t for dynamic threshold calculation, is the square of the norm of the second-order derivative, I(t) is the real-time rainfall at the current time t, is the average rainfall in the past 30 days, when S(t)>φ ou , then the data at that moment is removed and the rest of the data is normalized and output as high-quality data D fu ∈R T′×d″ .

5. The method for intelligent detection and analysis of the sealing performance of a coal chemical wastewater treatment tank according to claim 4, characterized in that: In step S3, a lightweight AI model is deployed on the sensor side to identify damage features in the high-quality data and offload the 3D reconstruction task to the cloud. The method is as follows: The lightweight AI model uses a compressed lightweight convolutional network (MobileNetV3-SSDLite) to perform real-time detection based on the damage features in the high-quality data, infer and output a crack area mask, and estimates the crack area in physical space based on the crack area mask, expressed as follows: Among them F cr is the total area of ​​the crack region, is the sum of all pixels x′, y′ in the entire image, G(x′, y′) is the crack detection result at the pixel point x′, y′, Δx′ is the actual physical length represented by each pixel in the horizontal direction, Δy′ is the actual physical length represented by each pixel in the vertical direction, and in F cr >F th =1mm 2 Trigger an alarm.

6. The method for intelligent detection and analysis of the sealing performance of a coal chemical wastewater treatment tank according to claim 5, characterized in that: In step S3, a lightweight AI model is deployed on the sensor side to identify damage features in the high-quality data and offload the 3D reconstruction task to the cloud. The method is as follows: The 3D reconstruction task is offloaded to the cloud. Only the edge crack feature tensor is extracted and uploaded to the cloud using a 5G link. After receiving the edge crack feature tensor, the cloud performs the 3D reconstruction task. The 3D reconstruction task includes CT slice stacking and 3D volume reconstruction. The expression formula is: Among them G i is the i-th CT slice, K i is the pose registration transformation, J i is the weight coefficient of the i-th CT module, H(x″,y″,z) is the final reconstructed value at the 3D reconstructed voxel coordinates x″,y″,z, is the accumulation of the reconstruction results of the i-th CT module, G i (x′, y′) is the crack detection result of the i-th CT module at pixel x′, y′, K i (·) is the two-dimensional image data G collected by the i-th CT module i A function that back-projects into three-dimensional space, where x″, y″, and z are the coordinate positions of the voxel in three-dimensional space, generating a voxel model.

7. The method for intelligent detection and analysis of the sealing performance of a coal chemical wastewater treatment tank according to claim 6, characterized in that: In step S4, the method of generating a dynamic threshold based on the high-quality data using a time series prediction model is as follows: The time series prediction model performs trend modeling on the environmental series, and the expression formula is: in is the predicted value at the next moment, is the input historical high-quality data sequence, is the length of the time window, It is a time series forecasting model Parameters, the predicted value and the historical mean are integrated to form a dynamic threshold, expressed as: in is the historical mean of soil moisture, γ∈[0,1] is the time smoothing factor, ι is the meteorological sensitivity factor, C wea is the meteorological correction factor, Indicates that the threshold adjusted according to the current conditions is used for judgment at future moments. is the predicted value at the next moment.

8. The method for intelligent detection and analysis of the sealing performance of a coal chemical wastewater treatment tank according to claim 7, characterized in that: In step S4, the method of generating a dynamic threshold based on the high-quality data using a time series prediction model is as follows: Assuming that the meteorological influencing factor vector is affected by rainfall increment and temperature change, a linear regression model is established to calculate the meteorological correction factor, which is expressed as: C wea =κ ⊥ ·V t +λ, where κ=[κ1,κ2,κ3,κ4] ⊥ is the regression weight coefficient, is a Gaussian disturbance term used to model unobservable external disturbances, and κ is a ridge regression for history fitting optimization. When there are multiple monitoring variables including humidity, pressure, and water level, an overall dynamic threshold is generated, expressed as: Where η is the corresponding historical mean vector, is the unit column vector used to broadcast the meteorological factor, X a is the final output dynamic threshold vector, The dynamic threshold of the humidity dimension corresponds to the data judgment standard of the soil moisture sensor. The dynamic threshold of the pressure dimension corresponds to the data judgment standard of the pressure sensor. It is the judgment standard of the water level or liquid level sensor corresponding to the dynamic threshold of the real-time meteorological dimension.

9. The method for intelligent detection and analysis of the sealing performance of a coal chemical wastewater treatment tank according to claim 8, characterized in that: In step S5, the method for dynamically adjusting the sensor weights based on the feedback of the sensor false alarm rate in the high-quality data through reinforcement learning is as follows: The sensor false alarm rate in the high-quality data is assumed to include n sensors, and the observation data at time t is: Where V t is the multi-sensor observation data vector at time t, is the data value collected by the first sensor at time t, n represents the total number of sensors, and the false alarm feedback label is formed based on the alarm result and manual confirmation result. The expression formula is: where ρ is the false positive penalty factor, is the false alarm mark of the i-th sensor at time t, It is a missed mark. is the correct mark, is a positive number, r t is the current state e t The action selected is, n is the number of sensors, is the weight of the i-th sensor at time t, It is the comprehensive discrimination accuracy index of the i-th sensor at time t.

10. The method for intelligent detection and analysis of the sealing performance of a coal chemical wastewater treatment tank according to claim 8, characterized in that: In step S5, the method for dynamically adjusting the sensor weights based on the feedback of the sensor false alarm rate in the high-quality data through reinforcement learning is as follows: The dynamic adjustment of sensor weights uses Q-Learning to iteratively update the policy value function to output the optimal action, expressed as: in is the learning rate, τ is the discount factor, is the current state e t r t Value, ← means Update of value, M t is the immediate reward obtained at the current time t, In the next state e t+1 The maximum of all possible actions r′ Value, r t is the current state e t Next, select the action you want to take. In the current state t Next Select The action with the largest value, υ is the exploration rate, 1-υ represents the probability of selecting the currently known optimal action, SJ is to randomly select an action from the action space, the dynamic adjustment of sensor weights, the current sensor weight vector is corrected according to the optimal action, and at the same time, the y sensor shows a high false alarm rate, that is, long-term If the value is low, the corresponding y sensor will gradually decrease, automatically increasing the influence of other sensors in the overall judgment.

Images