A gas extraction multi-parameter monitoring method and system based on edge computing

Abstract

The application provides a gas extraction multi-parameter monitoring method and system based on edge computing, and relates to the technical field of coal mine gas extraction. The method comprises the following steps: arranging multi-parameter sensor nodes on a drill hole, a pipeline and a gas collection station and connecting the multi-parameter sensor nodes to an edge computing unit; performing time synchronization, zero drift correction, outlier elimination and normalization on gas concentration, negative pressure, flow rate, temperature, humidity and hydrogen sulfide concentration data to generate a feature vector; calling a lightweight neural network model to infer and output an extraction efficiency score and a leakage risk degree; triggering an audible and light alarm and issuing a speed reduction instruction when the risk degree exceeds a limit, and adjusting the negative pressure when the efficiency is insufficient; compressing and encrypting the result and reporting the result to a cloud platform through a wireless network at a fixed time; and performing incremental training on the model based on historical data and feedback on the cloud, and issuing an updated model by hot replacement. The application realizes local intelligent analysis, rapid early warning and dynamic self-adaptive optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal mine gas extraction, in particular to a gas extraction multi-parameter monitoring method and system based on edge computing. BACKGROUND

[0002] Coal mine gas extraction is the core link of reducing the risk of underground gas disasters and realizing green and safe mining. The existing monitoring system usually arranges single or a few parameter sensors such as gas concentration, negative pressure (orifice vacuum degree) and flow rate near the extraction pipeline or borehole, and uploads the raw data directly to the ground information center for centralized processing through wired or industrial Wi-Fi. Due to long transmission link and strong underground interference, the monitoring system generally has problems such as large data time delay, high packet loss rate and lack of local early warning, which is difficult to support the real-time needs of integrated dynamic regulation of "extraction-ventilation-hole sealing".

[0003] With the gradual deployment of mine 5G, industrial Ethernet and low-power wide-area network (LPWAN), multi-parameter sensing, distributed computing and artificial intelligence analysis are becoming the mainstream trend of coal mine Internet of Things. Edge computing sinks lightweight inference models to the underground, which can complete data denoising, fusion and discrimination nearby, significantly reducing bandwidth occupation and cloud load; multi-modal sensor fusion and federated learning promote the evolution of gas extraction monitoring towards "advanced perception-local autonomy-global coordination", laying a foundation for advanced warning and adaptive extraction strategy.

[0004] The existing gas extraction monitoring schemes at this stage mostly adopt the "upload-centralize-issuing" mode, which has the following problems: (1) relying on ground servers, the edge lacks real-time decision-making ability and is invalid once the communication is interrupted; (2) limited parameter dimensions, which cannot comprehensively evaluate extraction efficiency and leakage risk; (3) non-uniform sensor manufacturers, incompatible data formats and time references, making it difficult to flexibly expand; (4) lack of model online updating and security encryption mechanism, making the system difficult to cope with changes in working conditions and network attacks. Therefore, an innovative method is needed that can complete multi-parameter fusion analysis on the edge side, with model self-updating and secure transmission capabilities. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a gas extraction multi-parameter monitoring method and system based on edge computing, which overcomes the defects of relying on the cloud, delayed response and untimely model updating in the prior art, and realizes local intelligent analysis, rapid warning and dynamic adaptive optimization of multi-parameter monitoring in the gas extraction process.

[0006] To achieve the above purpose, the present application provides the following solutions:

[0007] A gas extraction multi-parameter monitoring method based on edge computing, comprising:

[0008] A plurality of sensor nodes are arranged in a gas extraction borehole, pipeline and gas collection station of a target area, and the sensor nodes are used to collect original monitoring parameters composed of multiple parameters respectively; each sensor node is connected to an edge computing unit arranged around the sensor node through an industrial field bus;

[0009] In the edge computing unit, the received original monitoring parameters are subjected to time synchronization, zero drift correction, outlier elimination and dimensionless normalization to form a multi-parameter feature vector;

[0010] A pre-deployed lightweight deep neural network model is called to infer the multi-parameter feature vector, and an extraction efficiency score and a leakage risk degree are output;

[0011] When the leakage risk degree is greater than a first threshold value, the edge computing unit triggers a local audible and visual alarm and issues a speed reduction instruction to a nearby extraction pump inverter; when the extraction efficiency score is lower than a second threshold value, the edge computing unit adjusts the extraction negative pressure set value;

[0012] In the edge computing unit, an elliptic curve encryption algorithm is used to compress and encrypt the multi-parameter feature vector, the extraction efficiency score and the leakage risk degree, and the encrypted data is sent to a central cloud platform through a wireless network according to a set upload period;

[0013] In the central cloud platform, the lightweight deep neural network model is subjected to incremental training based on historical monitoring data and downhole feedback results, an updated model is generated, and the updated model is distributed to each edge computing unit through a secure channel for hot replacement.

[0014] Preferably, the original monitoring parameters include gas concentration, extraction negative pressure, gas flow, temperature, humidity and hydrogen sulfide concentration.

[0015] Preferably, the historical monitoring data include a time series constructed by the multi-parameter feature vector, the extraction efficiency score and the leakage risk degree; and the downhole feedback results include alarm response records and extraction adjustment efficiency change trends.

[0016] Preferably, in the edge computing unit, the received original monitoring parameters are subjected to time synchronization, zero drift correction, outlier elimination and dimensionless normalization to form a multi-parameter feature vector, including:

[0017] Based on a unified time service mechanism, a parameter sequence with consistent time stamps is generated for each type of original monitoring parameter;

[0018] The parameter sequence of gas concentration C, extraction negative pressure P and gas flow Q is looked up to obtain corresponding zero point offset value respectively, and the parameter value is corrected item by item combined with environmental temperature compensation factor, to obtain corrected parameter sequence C', P' and Q';

[0019] The parameter sequence of C', P' and Q' after correction and the original parameter sequence of temperature T, humidity H and hydrogen sulfide concentration S are detected and outliers are removed by using sliding window median standard deviation method, to obtain processed six types of parameter sequence; each sliding window contains N time points, if the parameter value deviates from the median of the sliding window by more than the preset multiple standard deviation threshold, the adjacent valid data is linearly interpolated to repair;

[0020] The six types of parameter sequence after processing are normalized and mapped to the interval [-1, 1], and the six-dimensional data at time point t is used to form the multi-parameter feature vector at time t.

[0021] Preferably, the calculation formula of the sliding window median standard deviation method is:

[0022]

[0023] Wherein, x t is the monitoring value at the current time; μ t = median(W t ) is the median of the data in the sliding window W t with t as the center and length N; σ t = 1.4826 * median{|w-μ t |:w∈W t} is the robust standard deviation estimate converted by the normalizing coefficient 1.4826 from the absolute deviation of the relative median of the window data; N is the sample point number of the sliding window, n is an odd number, and the typical value is 11 or 21); is an adaptive multiplication factor derived from the universal threshold of Gaussian noise, which monotonically increases with the window size.

[0024] Preferably, a pre-deployed lightweight deep neural network model is called to infer the multi-parameter feature vector, and the extraction efficiency score and leakage risk degree are output, including:

[0025] Before loading the lightweight deep neural network model M, the edge computing unit calculates the SHA-256 hash value of the model file and compares it with the pre-stored reference value, and performs the inference process after verification;

[0026] The multi-parameter feature vector obtained by normalization is packaged as a 1x6 floating point tensor and written into the input buffer of the local inference engine;

[0027] calling the inference engine to execute the shared backbone network and double-task output head of the lightweight deep neural network model M to obtain an efficiency logarithm value z E and a leakage logarithm value z R ;

[0028] calculating a pumping efficiency score E t = 100 x σ(z E ); where σ(·) is a Logistid Sigmoid function, and E t is linearly distributed between 0-100;

[0029] calculating a leakage risk R t = σ(z R / T s ); where T s is a temperature calibration coefficient determined through offline verification to improve the probability calibration degree;

[0030] writing {E t , R t}, together with a timestamp and a model version number, into an edge event queue for subsequent linkage control and uploading.

[0031] Preferably, the lightweight deep neural network model M is composed of an input layer, a batch normalization layer, a channel attention compression-excitation bottleneck block, a one-dimensional deep separable convolution layer, a global average pooling layer, and an efficiency output fully connected layer and a risk output fully connected layer arranged in parallel after the global average pooling layer, which are connected in series; wherein the input layer is used to receive the six-dimensional multi-parameter feature vector and reformat it into a 1 x 6 x 1 shaped tensor; the batch normalization layer is used to perform zero-mean unit-variance normalization on the tensor to suppress feature distribution drift; the channel attention compression-excitation bottleneck block performs adaptive re-labeling on channel weights through a global average pooling and a two-level fully connected-ReLU-Sigmoid structure to highlight monitoring parameters; the one-dimensional deep separable convolution layer adopts a 3-sized channel-by-channel convolution followed by a 1 x 1 point-by-point convolution to extract nonlinear coupling features between each monitoring parameter and compress the parameter amount; the global average pooling layer is used to reduce the dimension of the convolution output to a vector of length 64; the efficiency output fully connected layer receives the vector of length 64 and generates the pumping efficiency logarithm value z E , and the risk output fully connected layer synchronously receives the vector of length 64 and generates the leakage risk logarithm value z R .

[0032] Preferably, in the edge computing unit, the multi-parameter feature vector, the pumping efficiency score, and the leakage risk are compressed and encrypted using an elliptic curve encryption algorithm, and are sent to the central cloud platform through a wireless network according to a set uploading period, including:

[0033] After completing an inference, the edge computing unit packs the multi-parameter feature vector, the extraction efficiency score, the leakage risk, and the device identification code and millisecond-level timestamp obtained at the same sampling time into a structured message format to form an original upload message;

[0034] A lightweight lossless compression algorithm is called on the original upload message to compress the message byte stream to one-third of the original length to obtain a compressed message;

[0035] A pre-burned center cloud platform public key is used to perform an elliptic curve integrated encryption scheme on a national standard 256-bit elliptic curve to perform symmetric key packaging and data encryption on the compressed message, and then an elliptic curve digital signature algorithm is used to generate a digital signature for the ciphertext;

[0036] The encrypted and signed ciphertext is loaded into a message queue telemetry transport protocol message body, a 32-bit cyclic redundancy check code is attached as a link-level error detection field, and the quality of service level is marked as "first class" in the message header to ensure at least one reliable delivery;

[0037] When the local timer reaches the set upload period, the message is published to the center cloud platform through the wireless network; if a platform confirmation response is not received within the preset waiting time, an exponential backoff retransmission mechanism is automatically entered until a confirmation is received or the maximum number of retransmissions is reached to complete reliable encrypted uplink transmission.

[0038] Preferably, the wireless network is any of 5G NR or sixth-generation wireless local area network link.

[0039] A gas extraction multi-parameter monitoring system based on edge computing includes:

[0040] A multi-source monitoring and collecting unit is used to arrange a plurality of sensor nodes in the target area of the gas extraction borehole, pipeline and gas collection station, and to collect original monitoring parameters composed of multiple parameters using the sensor nodes; each sensor node is connected to an edge computing unit arranged around the sensor node through an industrial field bus;

[0041] An original data preprocessing unit is used to perform time synchronization, zero drift correction, outlier rejection and dimensionless normalization on the received original monitoring parameters in the edge computing unit to form a multi-parameter feature vector;

[0042] An edge inference analysis unit is used to call a pre-deployed lightweight deep neural network model to infer the multi-parameter feature vector and output an extraction efficiency score and a leakage risk;

[0043] A local risk response control unit is configured to trigger a local audible and visual alarm and issue a speed reduction instruction to a nearby extraction pump inverter by using the edge computing unit when the leakage risk degree is greater than a first threshold value; and adjust an extraction negative pressure set value by using the edge computing unit when the extraction efficiency score is lower than a second threshold value.

[0044] A secure communication transmission unit is configured to compress and encrypt the multi-parameter feature vector, the extraction efficiency score and the leakage risk degree by using an elliptic curve encryption algorithm in the edge computing unit, and send them to a central cloud platform through a wireless network according to a set upload period.

[0045] A model incremental learning and hot updating unit is configured to perform incremental training on the lightweight deep neural network model based on historical monitoring data and downhole feedback results in the central cloud platform, generate an updated model, and issue the updated model to each edge computing unit through a secure channel to complete hot replacement.

[0046] According to the specific embodiments of the present application, the following technical effects are disclosed:

[0047] The present application proposes a multi-parameter monitoring method based on edge computing to solve the problems of strong data upload dependency, large response time delay, delayed abnormal early warning, and untimely model updating in the existing gas extraction monitoring system. By deploying edge computing units on site, local synchronous processing and feature fusion of multiple parameters such as gas concentration, negative pressure, and flow rate are realized, replacing the traditional single-point upload method that relies on cloud computing. The lightweight neural network model is used to complete real-time inference of extraction efficiency and leakage risk on the edge side, enabling early warning and control to be executed immediately underground, significantly improving the system response speed and independent operation capability. The use of elliptic curve encryption ensures the confidentiality and integrity of data transmission, enhancing the security of the system in complex downhole communication environments. Combined with the incremental training and model hot replacement mechanism of the central cloud platform, the model can be continuously optimized based on historical data and downhole feedback, overcoming the defects of model update lag and inability to adapt to dynamic changes in extraction in traditional systems, and overall improving the intelligent level, robustness, and practical value of the system. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0049] Figure 1 The method flowchart provided for the embodiments of the present application;

[0050] Figure 2 A multi-parameter feature vector construction flowchart is provided for the embodiments of the present application.

[0051] Figure 3 A system structure schematic diagram is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0053] The purpose of the present application is to provide a gas extraction multi-parameter monitoring method and system based on edge computing, which overcomes the defects of dependence on the cloud, response lag and untimely model updating in the prior art, and realizes local intelligent analysis, rapid early warning and dynamic self-adaptive optimization of multi-parameter monitoring in the gas extraction process.

[0054] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0055] Figure 1 A method flowchart is provided for the embodiments of the present application, as shown in Figure 1 The present application provides a gas extraction multi-parameter monitoring method based on edge computing, which comprises:

[0056] Step 100: A plurality of sensor nodes are arranged in the gas extraction boreholes, pipelines and gas collection stations of a target area, and the original monitoring parameters composed of multiple parameters are collected by the sensor nodes; each sensor node is connected to an edge computing unit arranged around the sensor node through an industrial field bus;

[0057] Step 200: In the edge computing unit, the received original monitoring parameters are subjected to time synchronization, zero drift correction, outlier elimination and dimensionless normalization to form a multi-parameter feature vector;

[0058] Step 300: A pre-deployed lightweight deep neural network model is called to infer the multi-parameter feature vector, and the extraction efficiency score and the leakage risk degree are outputted;

[0059] Step 400: When the leakage risk degree is greater than a first threshold value, the edge computing unit triggers a local audible and visual alarm and issues a speed reduction instruction to the nearest extraction pump inverter; when the extraction efficiency score is lower than a second threshold value, the edge computing unit adjusts the extraction negative pressure set value;

[0060] Step 500: In the edge computing unit, the multi-parameter feature vector, the extraction efficiency score and the leakage danger degree are compressed and encrypted by using an elliptic curve encryption algorithm, and are sent to the central cloud platform through a wireless network according to a set upload period;

[0061] Step 600: In the central cloud platform, the lightweight deep neural network model is incrementally trained based on historical monitoring data and downhole feedback results, an updated model is generated, and the updated model is distributed to each edge computing unit through a secure channel to complete hot replacement.

[0062] Preferably, the original monitoring parameters include: gas concentration, extraction negative pressure, gas flow, temperature, humidity and hydrogen sulfide concentration.

[0063] In step 100 of the embodiment, first, a representative borehole outlet, negative pressure pipeline intersection point and gas collection station main pipe in the gas extraction area are selected as the layout positions of the sensor nodes; based on the relevant specifications of the Coal Mine Safety Regulations and the General Technical Conditions for Electrical Equipment in Coal Mines, gas concentration sensors, negative pressure transmitters, thermal gas flow meters, temperature and humidity integrated probes and electrochemical hydrogen sulfide detectors with intrinsic safety certification are installed respectively, and at least three types of the above six types of sensors are installed at each node to ensure the temporal and spatial representativeness and safety coverage of the parameters collected. Each type of sensor is configured with a unique device identifier and a communication address, and is connected in communication with the adjacent edge computing unit through an RS-485 industrial field bus and using a Modbus RTU protocol, and the number of devices mounted on a single bus does not exceed 32 to meet the stable operation requirements of anti-interference and bus bandwidth. The sensor sampling period is set to 5 seconds by default, and can be configured to be in the interval of 1 to 10 seconds according to the extraction disturbance frequency.

[0064] To achieve accurate acquisition and unified management of raw monitoring parameters, this embodiment uses the edge computing unit to call the device identification service to automatically identify and register the category, unit, range, and physical interface type of each access sensor. Simultaneously, a time synchronization mechanism is used to obtain a unified millisecond-level timestamp provided by the edge time synchronization module, ensuring that all parameters are comparable and have a basis for fusion within the same time window. The raw monitoring parameters specifically include: gas concentration (unit: volume fraction percentage), extraction negative pressure (unit: Pascals), gas flow rate (unit: standard cubic meters per hour), ambient temperature (unit: degrees Celsius), ambient humidity (unit: percentage), and hydrogen sulfide concentration (unit: parts per million by volume). The acquired data is organized into a time series format and stored in the buffer of the edge computing unit, serving as the input basis for subsequent feature construction and model inference. Since traditional systems often only achieve single-channel acquisition of gas concentration and negative pressure, lacking a unified time base and fusion capability, they struggle to support subsequent intelligent analysis and response control. This embodiment, however, integrates multi-source parameters through a fieldbus and forms a structured input, significantly improving the completeness and practicality of monitoring.

[0065] Preferably, the historical monitoring data includes a time series constructed from the multi-parameter feature vector, the extraction efficiency score, and the leakage risk; the downhole feedback results include alarm response records and the efficiency change trend after extraction adjustment.

[0066] Preferably, such as Figure 2 As shown, within the edge computing unit, the received raw monitoring parameters are time-synchronized, zero-drift corrected, outlier removed, and dimensionally normalized to form a multi-parameter feature vector, including:

[0067] Based on a unified time synchronization mechanism, a parameter sequence with a consistent timestamp is generated for each type of original monitoring parameter;

[0068] The zero-point offset values ​​of the parameter sequences of gas concentration C, extraction negative pressure P and gas flow rate Q are obtained by looking up the table respectively, and the parameter values ​​are corrected item by item by combining the ambient temperature compensation factor to obtain the corrected parameter sequences C′, P′ and Q′.

[0069] For the corrected C′, P′, Q′ and the original temperature T, humidity H, and hydrogen sulfide concentration S parameter sequences, outliers are detected and removed using the sliding window median standard deviation method to obtain six types of parameter sequences. Each sliding window contains N time points. If the parameter value deviates from the median of the sliding window by more than a preset multiple standard deviation threshold, it is repaired by linear interpolation of adjacent valid data.

[0070] The six parameter sequences after processing are uniformly normalized and mapped to the interval [-1,1]. The six-dimensional data at time point t constitute the multi-parameter feature vector at time t.

[0071] In this embodiment, after receiving the original monitoring parameters from each sensor node, the edge computing unit first calls the local time service to uniformly timestamp the data from different sources. Specifically, the edge computing unit uses the embedded high-precision crystal oscillator module as the main clock source, and refers to the external GPS synchronization signal or the industrial Ethernet PTP protocol to mark the original parameters (including gas concentration, suction negative pressure, gas flow, temperature, humidity and hydrogen sulfide concentration) with a uniform timestamp at the millisecond level, thereby constructing a time-aligned multi-class parameter sequence. Subsequently, for the three types of physical quantities of gas concentration, suction negative pressure and flow, the calibration value is obtained according to the zero point offset lookup table of the sensor at the current working temperature, and the zero drift correction is performed according to the following formula: corrected value = measured value - offset value under temperature compensation; wherein the compensation factor is fixed in the local edge database at the factory calibration of the device. This embodiment obtains the corrected gas concentration sequence, negative pressure sequence and flow sequence, respectively denoted as C', P' and Q', which lays an accurate foundation for subsequent anomaly processing and modeling analysis.

[0072] After completing the preliminary correction, this embodiment further performs outlier rejection processing on the six types of parameter sequences of C', P', Q', and original temperature T, humidity H and hydrogen sulfide concentration S using the sliding window median standard deviation method. The specific operation is: a sliding window containing N = 11 consecutive sampling points is constructed with each time point as the center, the median and standard deviation of all valid values in the window are calculated, and an adaptive threshold is set. If a value deviates from the median by more than three times the standard deviation, it is determined to be an outlier, and the linear interpolation result of the adjacent two valid points is used to replace it. The above algorithm runs on the low-power edge processor in this embodiment, has low computational complexity and strong noise resistance, and is more suitable for the drift and interference scene of the downhole sensor than the static threshold method in the prior art. After completing the outlier rejection, the edge computing unit performs minimum-maximum normalization on the six types of processed parameter sequences, so that each type of value is compressed to the [-1, 1] interval. Finally, the six types of normalized values at a time point t are arranged in order to construct a six-dimensional multi-parameter feature vector as the input of this time point for the subsequent neural network inference process. The feature vector has time sequence synchronization, physical consistency and scale uniformity, providing high-quality input for downhole edge intelligent reasoning and overcoming the problems of input dimension imbalance, error accumulation and response lag in existing methods.

[0073] Preferably, the calculation formula of the sliding window median standard deviation method is:

[0074]

[0075] wherein x t is the monitoring value at the current time; μ t = median(Wt Let W be a sliding window centered at t with a length of N. t Median of the data; σ t =1.4826×median{|w-μ t |:w∈W t} is a robust standard deviation estimate of the absolute deviation of the relative median of the peer data after normalization by a coefficient of 1.4826; N is the number of sample points in the sliding window (N takes an odd number, typically 11 or 21); The adaptive multiplication factor is derived from the universal threshold of Gaussian noise and increases monotonically with the window size.

[0076] Exemplary, {|w-μ t |:w∈W t} represents the set of absolute differences between each sample value w and the midpoint μ of the sliding window W.

[0077] In this embodiment, to improve the accuracy of outlier identification in underground gas monitoring data and avoid misjudgments caused by the dependence on fixed thresholds in traditional methods, this embodiment adopts a median standard deviation method based on a sliding window, and designs the following adaptive outlier identification criterion: The monitoring value at a certain moment is denoted as x. t Centered on this moment, construct a sliding window W of length N. t Where N is an odd number, typically 11 or 21. First, calculate the median μ of the data within this window. t =median(W t This is used as a local baseline value at that time point. Next, the set of absolute deviations between each data point within the window and the median is calculated, and the median MAD is determined. t =median(|x-μ) t To obtain a scaling estimate consistent with the normal standard deviation, an empirical correction factor of 1.4826 is used to scale this value, resulting in a robust standard deviation estimate σ. t =1.4826×MAD t Finally, determine whether the current value satisfies the following inequality:

[0078]

[0079] If the inequality holds, then we consider x to be true. tThe outlier is replaced by a linear interpolation value constructed with its two nearest non-outlier points. If not, the original value is kept. The multiplication factor of "square root of two logarithmic N" is derived from the concentration probability boundary estimation under Gaussian distribution, which can adaptively adjust the sensitivity according to the window length, without manually setting specific threshold values, enhancing the universality and stability of the algorithm. The method effectively filters out isolated abnormal readings caused by electromagnetic interference, sudden changes in downhole environment or short-term drift of sensors, significantly improves the robustness of monitoring parameter input and the credibility of neural network inference results, and is different from the existing fixed multiple threshold method which is easily misled by noise.

[0080] Preferably, a pre-deployed lightweight deep neural network model is called to infer the multi-parameter feature vector, output the extraction efficiency score and leakage risk degree, including:

[0081] Before loading the lightweight deep neural network model M, the edge computing unit calculates the SHA-256 hash value of the model file and compares it with the pre-stored reference value, and performs the inference process after verification;

[0082] The multi-parameter feature vector obtained by normalization is packaged as a 1x6 floating point tensor and written into the input buffer of the local inference engine;

[0083] The inference engine is called to execute the shared backbone network and double-task output head of the lightweight deep neural network model M, and the efficiency logarithmic value z E and the leakage logarithmic value z R are obtained.

[0084] The extraction efficiency score E t = 100xσ(z E ); where σ(·) is the Logistid Sigmoid function, making E t linearly distributed between 0 and 100.

[0085] The leakage risk degree R t = σ(z R / T s ); where T s is the temperature calibration coefficient determined by offline verification, used to improve the probability calibration degree.

[0086] {E t , R t} is written into the edge event queue together with the timestamp and model version number for subsequent linkage control and uploading.

[0087] In this embodiment, in order to guarantee the integrity and credibility of the inference model running on the edge device, the model file is first checked for integrity before the model is called. Specifically, before starting the inference engine, the edge computing unit calculates the hash value of the model binary file stored locally using the SHA-256 digest algorithm, and compares it character by character with the standard reference value previously burned in the trusted storage. Only when the hash values are completely consistent, the model can be loaded into memory, otherwise the inference execution is prevented and a security alarm log is generated. Compared with the "no check direct loading" method in the prior art, this processing method significantly enhances the tamper resistance of the model running, and is especially suitable for downhole industrial application scenarios with high requirements for model security.

[0088] After completing the model loading, the normalized six-dimensional multi-parameter feature vector is packaged as a single batch, single channel one-dimensional floating point tensor, and written to the input buffer area of the edge inference engine (such as TensorRT or ONNX Runtime). The inference engine is called to execute a lightweight deep neural network model, which is composed of a shared backbone network composed of a group of convolution-activation structures and a parallelly arranged double-output task head: the first output head is used to generate the logarithmic value of the sampling efficiency, and the second output head is used to generate the logarithmic value of the leakage danger degree. After the inference is completed, the Sigmoid function is used to map the efficiency logarithmic value, and multiplied by one hundred, so that the sampling efficiency score is linearly distributed in the interval of 0 to 100; at the same time, the leakage logarithmic value is divided by the temperature calibration coefficient obtained by offline verification and then input into the Sigmoid function to generate a leakage danger degree value between 0 and 1, thereby improving the probability expression accuracy of this output value under different working conditions. Finally, the score result and the corresponding timestamp and current model version number are packaged as structured event data and written to the edge event queue for subsequent real-time reading by the linkage control module or reporting to the cloud according to the set period. This process breaks through the technical bottlenecks of existing methods such as large model size, inability to sink, and output without probability significance, through lightweight model compression, calibration enhancement, and structured cache management, and realizes stable, verifiable and credible intelligent inference in a resource-limited environment.

[0089] Preferably, the lightweight deep neural network model M is composed of an input layer, a batch normalization layer, a channel attention compression-excitation bottleneck block, a one-dimensional deep separable convolution layer, a global average pooling layer, and an efficiency output fully connected layer and a risk output fully connected layer arranged in parallel after the global average pooling layer in series; wherein the input layer is used to receive the six-dimensional multi-parameter feature vector and reformat it into a 1x6x1 shaped tensor; the batch normalization layer is used to perform zero mean unit variance normalization on the tensor to suppress feature distribution drift; the channel attention compression-excitation bottleneck block performs adaptive re-labeling on channel weights through global average pooling and a two-level fully connected- ReLU-Sigmoid structure to highlight monitoring parameters; the one-dimensional deep separable convolution layer adopts a 3-size convolution kernel for channel-by-channel convolution followed by a 1x1 point-by-point convolution to extract nonlinear coupling features between each monitoring parameter and compress parameter quantity; the global average pooling layer is used to reduce the dimension of the convolution output to a vector of length 64; the efficiency output fully connected layer receives the vector of length 64 and generates a decimation efficiency logarithmic value z E , and the risk output fully connected layer synchronously receives the vector of length 64 and generates a leakage risk logarithmic value z R .

[0090] In this embodiment, the called lightweight deep neural network model M is designed specifically for edge computing unit deployment scenarios, with the characteristics of compact structure and efficient inference. The model includes an input layer, a batch normalization layer, a channel attention compression-excitation bottleneck block, a one-dimensional deep separable convolution layer, a global average pooling layer, and an efficiency output fully connected layer and a risk output fully connected layer arranged in parallel after the global average pooling layer from input to output. The input layer is used to receive the normalized six-dimensional multi-parameter feature vector and reconstruct it into a 1x6x1 tensor to adapt to the subsequent convolution calculation structure. The batch normalization layer performs zero mean, unit variance standardization processing on the channel dimension of the tensor, effectively suppressing the feature distribution drift that may occur during training and inference, and improving the stability of the model. Then, the model learns the weights of the six channels through a channel attention compression-excitation bottleneck block. This module first performs global average pooling on each channel to form a one-dimensional channel description vector, and then passes through two fully connected layers with a rectified linear unit (ReLU) as the activation function and a Sigmoid function at the end. The output of the six channel weights is re-labeled with the original channel by multiplication, thereby adaptively improving the important parameter response capability under the current working condition.

[0091] After the channel weight adjustment, the model inputs the tensor into a one-dimensional depth separable convolution layer for feature extraction. First, a channel-by-channel convolution operation with a kernel size of 3 is performed independently on each channel to extract the time correlation pattern of each type of monitoring parameter. Then, a point-by-point convolution of 1x1 is performed to realize the interaction and compression between channels, greatly reducing the parameter quantity and improving the non-linear cross modeling capability. The output tensor after convolution is sent to the global average pooling layer, which is averaged in the channel dimension, and finally compressed into a fixed-length feature vector of length 64. This vector is then sent to the efficiency output fully connected layer and the risk output fully connected layer, respectively. The former outputs the logarithm of the extraction efficiency, and the latter outputs the logarithm of the leakage risk. The output values of the two are input into the Sigmoid activation function for standardization mapping to form the subsequent score values. The overall structure of the model is compact, with a parameter quantity controlled within 50,000 and an average inference delay of less than 50 milliseconds. It is suitable for edge processors with ARM Cortex-A55 architecture and above, meeting the real-time operation requirements of low power consumption and low delay in mine sites. Unlike traditional deep neural networks, which cannot be directly deployed on edge devices due to technical limitations, this model has the significant advantage of balancing operability and deployment efficiency.

[0092] Preferably, in the edge computing unit, the multi-parameter feature vector, the extraction efficiency score and the leakage risk degree are compressed and encrypted using an elliptic curve encryption algorithm, and sent to the central cloud platform through a wireless network according to a set upload period, including:

[0093] After completing an inference, the edge computing unit packs the multi-parameter feature vector, the extraction efficiency score, the leakage risk degree, the device identification code and the millisecond-level timestamp obtained at the same sampling time into a structured message format to form an original upload message;

[0094] A lightweight lossless compression algorithm is called for the original upload message to compress the message byte stream to one-third of the original length, obtaining a compressed message;

[0095] A pre-burned public key of the central cloud platform is used to perform an elliptic curve integrated encryption scheme on a national standard 256-bit elliptic curve to perform symmetric key packaging and data encryption on the compressed message, and then an elliptic curve digital signature algorithm is used to generate a digital signature for the ciphertext;

[0096] The encrypted and signed ciphertext is loaded into the message queue telemetry transport protocol message body, a 32-bit cyclic redundancy check code is attached as a link-level error detection field, and the quality of service level is marked as "first class" in the message header to ensure at least one reliable delivery;

[0097] When the local timer reaches the set upload period, the message is sent to the central cloud platform through the wireless network; if no platform confirmation response is received within the preset waiting time, the exponential backoff retransmission mechanism is automatically entered, and the message is retransmitted until the maximum number of retransmissions is reached or a confirmation is received, to complete the reliable encrypted uplink transmission.

[0098] Preferably, the wireless network is any of 5G NR or sixth generation wireless local area network link.

[0099] In this embodiment, to achieve high security, low redundancy and high reliability of the edge side inference result uplink transmission, after completing a model inference, the multi-parameter feature vector, the sampling efficiency score and the leakage danger degree generated by this inference are immediately packaged as a structured key-value pair format, together with the unique identification code of the sampling device and the millisecond level timestamp, and a structure definition library is called to generate an original upload message conforming to the general protocol. The message is compressed using a lightweight lossless compression algorithm, and the bit length after compression is reduced to about one third of the original, significantly reducing the bandwidth load of subsequent encryption and transmission. To avoid the potential leakage risk caused by plaintext uplink, this embodiment uses the national standard 256-bit elliptic curve (conforming to SM2 or NIST P-256 standard) to perform encryption processing on the compressed message. The specific process is as follows: first, call the local embedded encryption module, perform the elliptic curve integrated encryption scheme based on the pre-recorded cloud public key, realize key negotiation and symmetric encryption double protection; then perform elliptic curve digital signature algorithm on the generated ciphertext, use the edge side private key to generate a signature fragment, and attach it to the end of the ciphertext for platform signature verification.

[0100] After completing the encryption and signature processing, this embodiment packs the ciphertext into a message body conforming to the Message Queue Telemetry Transport protocol format, adds a 32-bit cyclic redundancy check code field to the end of the message for link layer error detection, and marks the quality of service level as level one to ensure that the message is successfully delivered at least once. This embodiment uses any of 5G New Radio or sixth generation wireless local area network for data transmission, and can dynamically switch the link type according to the underground communication coverage. When the local timer in the edge computing unit reaches the set upload period, the encrypted message is immediately sent to the central cloud platform through the above wireless link; if no confirmation response is received from the cloud platform within the set response waiting time, the exponential backoff retransmission mechanism is automatically started, the initial retry interval is 100 milliseconds, the retry interval is doubled each time, and the maximum number of retries is five; if the message is not confirmed after five attempts, an error is written into the log and the offline buffer mode is entered, waiting for the next scheduling period to resend. Compared with the existing technology which simply relies on physically isolated networks or plaintext uplink, this embodiment integrates compression, encryption, signature and retransmission to ensure that the inference result can also achieve high security, high integrity and high fault tolerance in the data upload process in harsh underground environments.

[0101] In step 600 of the embodiment, in order to adapt to the data distribution drift caused by the dynamic change of downhole working conditions, the fluctuation of extraction load and environmental disturbance, the embodiment constructs a periodic incremental training mechanism of a neural network model based on a central cloud platform. Specifically, first, historical monitoring data uploaded from an edge computing unit is received, the historical data is organized in a time series structure, and contains a multi-parameter feature vector, a corresponding timestamp, an extraction efficiency score and a leakage risk prediction value; at the same time, the central platform extracts downhole feedback results from an artificial watch system, a sound and light alarm log and an extraction pump frequency adjustment record, mainly including alarm response state, linkage action execution time, change trend of extraction efficiency within 5 minutes after control, etc., for labeling the correctness and timeliness of the model output. The embodiment adopts a multi-task learning framework with an attention calibration mechanism, and only locally fine-tunes the output layer and batch normalization layer parameters on the basis of freezing the main network parameters, so as to realize adaptive update of the latest data features and avoid model overall drift or catastrophic forgetting.

[0102] After training, the generated updated model is saved as a binary model file with a unified platform structure, and version control and digest signature processing are performed. In order to protect the transmission security and consistency of the model, the embodiment uses an end-to-end secure channel for distribution operation. The secure channel is based on Transport Layer Security Protocol version 1.3 (TLS1.3) and combines a two-way identity authentication mechanism to ensure that the model can only be received by the authorized edge computing unit. The model file is attached with a digest signature and a model version number before sending; the edge computing unit immediately checks the signature consistency after receiving, and on the premise of not interrupting the edge process, hot loads the updated model file to the inference engine cache area, replaces the original model parameters through memory mapping, and does not need to restart the device or interrupt the task. This hot replacement method effectively guarantees the continuous running ability of edge intelligence, which is different from the traditional model update method that needs to shut down or manually load, and realizes the self-evolution of the model in the field environment and the edge autonomy ability.

[0103] Corresponding to the above method, as shown in Figure 2 The embodiment also provides a gas extraction multi-parameter monitoring system based on edge computing, which comprises:

[0104] A multi-source monitoring and collecting unit is used for arranging a plurality of sensor nodes in a gas extraction borehole, a pipeline and a gas collection station of a target area, and collecting original monitoring parameters composed of multiple parameters by using the sensor nodes; each sensor node is connected with an edge computing unit arranged around the sensor node through an industrial field bus;

[0105] An original data preprocessing unit is used for performing time synchronization, zero drift correction, outlier elimination and dimensionless normalization on the received original monitoring parameters in the edge computing unit to form a multi-parameter feature vector;

[0106] An edge inference analysis unit is configured to call a pre-deployed lightweight deep neural network model to perform inference on the multi-parameter feature vector, and output a pumping efficiency score and a leakage risk degree;

[0107] A local risk response control unit is configured to trigger a local audible and visual alarm and issue a speed reduction instruction to a nearby pumping pump inverter by using the edge computing unit when the leakage risk degree is greater than a first threshold value, and adjust a pumping negative pressure set value by using the edge computing unit when the pumping efficiency score is lower than a second threshold value.

[0108] A secure communication transmission unit is configured to compress and encrypt the multi-parameter feature vector, the pumping efficiency score and the leakage risk degree by using an elliptic curve encryption algorithm in the edge computing unit, and send the data to a central cloud platform through a wireless network according to a set upload period.

[0109] A model incremental learning and hot update unit is configured to perform incremental training on the lightweight deep neural network model based on historical monitoring data and downhole feedback results in the central cloud platform, generate an updated model, and issue the updated model to each edge computing unit through a secure channel for hot replacement.

[0110] The present application has the following advantages:

[0111] (1) The present application realizes local processing and real-time inference of multiple types of monitoring parameters by deploying an edge computing unit at the gas pumping site, avoiding the high latency and communication interruption risks caused by the dependence of traditional methods on cloud computing. The edge side can independently complete data cleaning, feature extraction, model inference and early warning response, and can still maintain continuous operation in the event of an underground environment, effectively improving the local intelligent response capability of the pumping system.

[0112] (2) The present application proposes a complete preprocessing procedure for multiple source monitoring parameters such as gas concentration, negative pressure, flow rate, temperature and humidity, and harmful gases, including time synchronization, zero drift correction, sliding window robust outlier elimination and normalization processing, forming a structured multi-parameter feature vector. By introducing the median absolute deviation and adaptive threshold mechanism, the fault tolerance of the system to interference, drift and sudden change data in the complex underground environment is enhanced.

[0113] (3) The present application uses an elliptic curve encryption algorithm and a digital signature mechanism to double-protect the data after compression, combines a signed structured message with a retransmission confirmation mechanism to ensure the confidentiality, integrity and reliability of the uploaded data in public networks or underground wireless networks. Compared with the traditional plaintext upload method, the data security level of the system in complex mine communication conditions is significantly improved, and the risk of data leakage and tampering is reduced.

[0114] (4) The application constructs a cloud incremental training and edge model hot replacement mechanism, which can combine historical monitoring data and downhole feedback results to continuously optimize the neural network model structure and weight parameters, and complete model version replacement through a secure channel without interrupting task operation. This mechanism overcomes the problems of model update lag and poor adaptability in existing solutions, enabling the gas extraction monitoring system to have dynamic learning and continuous evolution capabilities, making it suitable for long-term operation scenarios.

[0115] The various embodiments are described in a progressive manner in the specification, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be mutually referred to. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0116] The principles and implementation modes of the application are described by applying specific examples in this paper, and the above embodiment description is only used to help understand the method and core idea of the application; at the same time, for those skilled in the art, according to the idea of the application, the specific implementation mode and application range will be changed. In view of the above, the content of the specification should not be understood as a limitation of the application.

Claims

1. A multi-parameter monitoring method for gas extraction based on edge computing, characterized in that, include: Several sensor nodes are deployed in the gas extraction boreholes, pipelines and gas gathering stations in the target area, and the original monitoring parameters composed of multiple parameters are collected by the sensor nodes respectively. Each of the sensor nodes is connected to an edge computing unit located around the sensor node via an industrial fieldbus; Within the edge computing unit, the received raw monitoring parameters are time-synchronized, zero-drift corrected, outlier removed, and dimension-normalized to form a multi-parameter feature vector. The pre-deployed lightweight deep neural network model is invoked to infer the multi-parameter feature vector and output the extraction efficiency score and leakage risk. When the leakage risk level is greater than the first threshold, the edge computing unit triggers a local audible and visual alarm and sends a speed reduction command to the nearest extraction pump inverter; when the extraction efficiency score is lower than the second threshold, the edge computing unit adjusts the extraction negative pressure setting value. Within the edge computing unit, an elliptic curve cryptography algorithm is used to compress and encrypt the multi-parameter feature vector, the sampling efficiency score, and the leakage risk, and then transmits them to the central cloud platform via a wireless network according to a set upload cycle. Within the central cloud platform, incremental training is performed on the lightweight deep neural network model based on historical monitoring data and downhole feedback results to generate an updated model. The updated model is then distributed to each edge computing unit via a secure channel to complete hot replacement. The pre-deployed lightweight deep neural network model is invoked to infer the multi-parameter feature vector, outputting an extraction efficiency score and leakage risk level, including: Before loading a lightweight deep neural network model, the edge computing unit calculates the SHA-256 hash value of the model file and compares it with a pre-stored benchmark value. If the verification passes, the inference process is performed. The normalized multi-parameter feature vector is encapsulated into a 1×6 floating-point tensor and written into the input cache of the local inference engine. The inference engine is invoked to execute the shared backbone network and dual-task output head of the lightweight deep neural network model, and the efficiency logarithm z is obtained. E And the logarithm of leakage z R ; Calculate the extraction efficiency score E t =100×σ(z) E ); where σ(·) is the Logistic Sigmoid function, such that E t Linear distribution between 0 and 100; calculate the leakage hazard R. t =σ(z) R / T s ); where T s These are temperature calibration coefficients determined through offline verification, used to improve probabilistic calibration accuracy; {E t ,R t The timestamp and model version number are written to the edge event queue for subsequent linkage control and uploading.

2. The multi-parameter monitoring method for gas extraction based on edge computing according to claim 1, characterized in that, The original monitoring parameters include: gas concentration, extraction negative pressure, gas flow rate, temperature, humidity, and hydrogen sulfide concentration.

3. The multi-parameter monitoring method for gas extraction based on edge computing according to claim 1, characterized in that, The historical monitoring data includes a time series constructed from the multi-parameter feature vector, the extraction efficiency score, and the leakage risk; the downhole feedback results include alarm response records and the efficiency change trend after extraction adjustment.

4. The multi-parameter monitoring method for gas extraction based on edge computing according to claim 2, characterized in that, Within the edge computing unit, the received raw monitoring parameters are time-synchronized, zero-drift corrected, outlier removed, and dimensionally normalized to form a multi-parameter feature vector, including: Based on a unified time synchronization mechanism, a parameter sequence with a consistent timestamp is generated for each type of original monitoring parameter; The zero-point offset values ​​of the parameter sequences of gas concentration C, extraction negative pressure P and gas flow rate Q are obtained by looking up the table respectively, and the parameter values ​​are corrected item by item by combining the ambient temperature compensation factor to obtain the corrected parameter sequences C′, P′ and Q′. For the corrected C′, P′, Q′ and the original temperature T, humidity H, and hydrogen sulfide concentration S parameter sequences, outliers are detected and removed using the sliding window median standard deviation method to obtain six types of parameter sequences. Each sliding window contains N time points. If the parameter value deviates from the median of the sliding window by more than a preset multiple standard deviation threshold, it is repaired by linear interpolation of adjacent valid data. The six parameter sequences after processing are uniformly normalized and mapped to the interval [-1,1]. The six-dimensional data at time point t constitute the multi-parameter feature vector at time t.

5. The multi-parameter monitoring method for gas extraction based on edge computing according to claim 4, characterized in that, The formula for calculating the standard deviation of the median in the sliding window is as follows: Where, x t The current monitored value; μ t =median(W t Let W be a sliding window centered at t with a length of N. t Median of the data; σ t =1.4826×median{|w-μ t |:w∈W t } is a robust standard deviation estimate of the absolute deviation of the relative median of the peer data after normalization by a coefficient of 1.4826; N is the number of sample points in the sliding window, N is an odd number, typically 11 or 21; The adaptive multiplication factor is derived from the universal threshold of Gaussian noise and increases monotonically with the window size.

6. The multi-parameter monitoring method for gas extraction based on edge computing according to claim 1, characterized in that, The lightweight deep neural network model consists of an input layer, a batch normalization layer, a channel attention compression-activation bottleneck block, a one-dimensional depthwise separable convolutional layer, a global average pooling layer, and an efficiency output fully connected layer and a risk output fully connected layer connected in parallel after the global average pooling layer. The input layer receives the six-dimensional multi-parameter feature vector and reshapes it into a 1×6×1 tensor. The batch normalization layer performs zero-mean unit variance normalization on the tensor to suppress feature distribution drift. The channel attention compression-activation bottleneck block... The attention compression-excitation bottleneck block adaptively recalibrates channel weights using global average pooling and a two-stage fully connected ReLU-Sigmoid structure to highlight the monitoring parameters. The one-dimensional depthwise separable convolutional layer employs channel-wise convolutions with a kernel size of 3 followed by 1×1 pointwise convolutions to extract nonlinear coupling features between monitoring parameters and compress the number of parameters. The global average pooling layer reduces the dimensionality of the convolutional output to a vector of length 64. The efficiency output fully connected layer receives the vector of length 64 and generates the logarithmic efficiency value z. E The risk output fully connected layer synchronously receives a vector of length 64 and generates a leakage logarithm value z. R .

7. The multi-parameter monitoring method for gas extraction based on edge computing according to claim 1, characterized in that, Within the edge computing unit, an elliptic curve cryptography algorithm is used to compress and encrypt the multi-parameter feature vector, the sampling efficiency score, and the leakage risk level, and then transmits them to the central cloud platform via a wireless network according to a set upload cycle, including: After completing one inference, the edge computing unit packages the multi-parameter feature vector, the sampling efficiency score, the leakage risk, the device identification code, and the millisecond-level timestamp obtained at the same sampling time into a structured message format to form the original upload message. A lightweight lossless compression algorithm is applied to the original uploaded message to compress the message byte stream to one-third of its original length, resulting in the compressed message. Using the pre-programmed public key of the central cloud platform, an elliptic curve integrated encryption scheme is executed on a 256-bit elliptic curve to encapsulate and encrypt the compressed message with a symmetric key, and then an elliptic curve digital signature algorithm is used to generate a digital signature for the ciphertext. The encrypted and signed ciphertext is loaded into the message queue telemetry transmission protocol message body, a 32-bit cyclic redundancy check code is added as a link-level error detection field, and the service quality level is marked as "Level 1" in the message header to ensure reliable delivery at least once. When the local timer reaches the set upload period, the message is published to the central cloud platform through the wireless network. If no confirmation response is received from the platform within the preset waiting time, the exponential backoff retransmission mechanism is automatically entered until confirmation is received or the maximum number of retransmissions is reached, so as to complete the reliable encrypted uplink transmission.

8. The multi-parameter monitoring method for gas extraction based on edge computing according to claim 1, characterized in that, The wireless network is either 5G NR or sixth-generation wireless local area network link.

9. A multi-parameter monitoring system for gas extraction based on edge computing, characterized in that, include: The multi-source monitoring and acquisition unit is used to deploy several sensor nodes in the gas extraction boreholes, pipelines and gas gathering stations in the target area, and to use the sensor nodes to collect raw monitoring parameters composed of multiple parameters; each sensor node is connected to the edge computing unit set around the sensor node through an industrial fieldbus. The raw data preprocessing unit is used within the edge computing unit to perform time synchronization, zero drift correction, outlier removal, and dimension normalization on the received raw monitoring parameters to form a multi-parameter feature vector. The edge inference analysis unit is used to invoke a pre-deployed lightweight deep neural network model to infer the multi-parameter feature vectors and output the extraction efficiency score and leakage risk. Invoking the pre-deployed lightweight deep neural network model to infer the multi-parameter feature vectors and output the extraction efficiency score and leakage risk includes: using the edge computing unit to calculate the SHA-256 hash value of the model file before loading the lightweight deep neural network model and comparing it with a pre-stored benchmark value, and performing the inference process after verification; encapsulating the normalized multi-parameter feature vectors into a 1×6 floating-point tensor and writing it into the input cache of the local inference engine; invoking the inference engine to execute the shared backbone network and dual-task output head of the lightweight deep neural network model to obtain the efficiency logarithm z. E And the logarithm of leakage z R ; Calculate the extraction efficiency score E t =100×σ(z) E ); where σ(·) is the Logistic Sigmoid function, such that E t Linear distribution between 0 and 100; calculate the leakage hazard R. t =σ(z) R / T s ); where T s These are temperature calibration coefficients determined through offline verification, used to improve probabilistic calibration accuracy; {E t ,R t The timestamp and model version number are written to the edge event queue for subsequent linkage control and uploading; The local risk response control unit is used to trigger a local audible and visual alarm and send a speed reduction command to the nearest extraction pump inverter when the leakage risk level is greater than a first threshold; and to adjust the extraction negative pressure setting value using the edge computing unit when the extraction efficiency score is lower than a second threshold. A secure communication transmission unit is used within the edge computing unit to compress and encrypt the multi-parameter feature vector, the sampling efficiency score, and the leakage risk using an elliptic curve encryption algorithm, and then transmit them to the central cloud platform via a wireless network according to a set upload cycle. The incremental learning and hot update unit is used to perform incremental training on the lightweight deep neural network model based on historical monitoring data and downhole feedback results within the central cloud platform, generate an updated model, and distribute the updated model to each edge computing unit through a secure channel to complete hot replacement.

Images