A priori active explosion-proof closed-loop control method and system

Abstract

The application discloses a kind of priori active explosion-proof closed-loop control method and system, it is related to industrial safety monitoring technical field.System includes multimodal perception layer, intrinsically safe transmission layer, AI Agent decision layer and active execution layer;Multimodal perception layer includes integrated encapsulation multimodal perception terminal, integrates four kinds of sensing modules of heat, light, gas, pressure and is equipped with synchronous sampling mechanism;Intrinsic safety transmission layer transmits data under the condition of limiting transmission loop energy;AI Agent decision layer includes data enhancement module using generative adversarial network and risk early warning model using Bi-LSTM and attention mechanism, for generating extreme condition simulation data and carrying out explosion risk pre-judgment to generate hierarchical decision instruction;Active execution layer executes instruction and returns confirmation signal and constitutes closed loop.The application realizes the improvement from passive explosion-proof to active priori explosion-proof, with the advantages of high identification accuracy, fast response speed, high intrinsic safety level.

Description

Technical Field

[0001] This invention relates to the field of industrial safety monitoring technology, and more specifically, to a priori active explosion-proof closed-loop control method and system. Background Technology

[0002] Industrial explosion protection is a core requirement for safe production in industries such as petrochemicals, mining, and metallurgy. Existing explosion protection technologies are mainly based on passive explosion isolation, the core principle of which is to allow an explosion to occur inside the equipment and rely on the mechanical strength of the explosion-proof enclosure to isolate the propagation of flames and blast shock waves. This is a post-explosion blocking protection method.

[0003] The passive explosion-proof technology has the following shortcomings in practical applications: First, the explosion-proof surface of the explosion-proof enclosure is prone to failure due to corrosion, deformation, or improper maintenance during long-term operation. Once the integrity of the enclosure is damaged, the explosion-proof function is lost. Second, this technology only blocks the explosion after it occurs and cannot intervene in the precursor stage before the formation of the three elements of an explosion, making it difficult to avoid accident losses from the source. Third, its protective effect is significantly reduced when facing extreme conditions such as instantaneous strong electric arcs, rapid penetration of high-concentration flammable gases, and rapid temperature increases.

[0004] In terms of active monitoring and early warning, existing explosion-proof monitoring systems mostly adopt a discrete sensor layout, which has the following problems: the data acquisition of each sensing dimension is difficult to synchronize in time, resulting in deviations in the extraction of pre-explosion features; the split installation requires multiple holes to be opened on the explosion-proof cabinet, affecting the cabinet's sealing and structural strength, and may create monitoring blind spots.

[0005] In terms of data analysis and decision-making, extreme condition data of pre-explosion precursors are sporadic, sudden, and highly dangerous, making it difficult to collect large quantities in real-world scenarios. This limits the generalization ability of traditional supervised learning-based risk warning models due to insufficient training samples. Traditional data augmentation methods, such as interpolation and synthetic minority class oversampling, only perform simple transformations on existing samples, failing to generate new samples that conform to the physical evolution of pre-explosion precursors. Furthermore, traditional time-series prediction models have limited ability to capture forward and backward dependencies in pre-explosion precursor time-series data and exhibit lag in response to highly sensitive warning features such as micro-arcs and temperature rise rates, making it difficult to meet the requirements for sub-second real-time prediction.

[0006] In terms of transmission and execution, existing explosion-proof systems lack inherently safe design in their data transmission links, and the circuit energy control in the transmission stage is insufficient, posing a risk of becoming a potential ignition source. Simultaneously, the interface protocols and data formats between different levels of the system are inconsistent, and the coordination among the sensing, transmission, decision-making, and execution modules is insufficient, making it difficult to form an efficient, end-to-end closed-loop control. The execution layer generally adopts a single power-off protection method, lacking graded intervention measures for different risk levels, and there is a lack of feedback verification after execution, making it difficult to ensure the reliable execution of the blocking action.

[0007] However, simply combining multimodal sensors with general AI models still fails to solve the following key problems: First, the microsecond-level sampling asynchrony of the various sensors destroys the cross-dimensional temporal characteristics of pre-explosion precursors at the source; second, extreme condition data exhibits long-tailed sparsity, and conventional interpolation methods disrupt the coupling phase relationships between multidimensional physical quantities; third, traditional threshold determination lags behind the physical processes of heat conduction and gas diffusion, making it impossible to implement blocking during the latency period. These problems constitute physical barriers that have not yet been overcome in existing technologies.

[0008] In summary, existing industrial explosion-proof technologies still have many areas for improvement in terms of early warning sensing, extreme data enhancement, real-time risk prediction, intrinsically safe transmission, graded execution, and closed-loop feedback. Summary of the Invention

[0009] This invention aims to address the following problems existing in current industrial explosion-proof technologies: passive explosion-proof modes have a delayed response, making intervention impossible during the pre-explosion phase; asynchronous data acquisition from discrete sensors leads to deviations in the extraction of pre-explosion features; real data for extreme conditions is scarce, making it difficult for traditional data augmentation methods to generate samples that conform to the laws of physical evolution; traditional time-series prediction models have low sensitivity to key pre-explosion features and high prediction delays; transmission links lack inherently safe design, posing a risk of becoming potential ignition sources; and the system lacks sufficient coordination at various levels, lacking a hierarchical decision-making and execution feedback verification mechanism.

[0010] To solve the above-mentioned technical problems, the present invention provides a priori active explosion-proof closed-loop control system, comprising: The multimodal sensing layer is used to simultaneously collect four-dimensional sensing data of heat, light, air, and pressure from the industrial site. The intrinsically safe transmission layer includes a signal conditioning module, an electrical isolation module, and an energy limiting module. The intrinsically safe transmission layer transmits the data collected by the multimodal sensing layer to the AIAgent decision layer under the condition of limiting the electrical energy of the transmission circuit. The AI ​​Agent decision layer includes a data augmentation module and a risk warning model. The data augmentation module uses a generative adversarial network to generate extreme working condition simulation data to expand the training dataset. The risk warning model uses a hybrid model that combines Bi-LSTM and an attention mechanism. The risk warning model is trained based on the expanded training dataset and predicts the probability of explosion risk from real-time perceived data and generates hierarchical decision instructions. The active execution layer includes multiple execution mechanisms. The active execution layer receives and executes the hierarchical decision instructions issued by the AI ​​Agent decision layer, and returns an execution confirmation signal to the AI ​​Agent decision layer to form an instruction-execution-feedback closed loop.

[0011] Furthermore, the multimodal sensing layer includes an FPGA chip, which synchronously triggers the acquisition of thermal sensing modules, optical sensing modules, gas sensing modules, and pressure sensing modules, with a sampling frequency of 1kHz and a sampling time deviation of ≤1μs.

[0012] Furthermore, in the intrinsically safe transmission layer, the electrical isolation module uses an optocoupler isolation chip with an isolation voltage ≥2500VAC; the energy limiting module includes a current-limiting resistor and a Zener diode to limit the maximum current of the transmission circuit to within 20mA.

[0013] Furthermore, the data augmentation module employs a deep convolutional generative adversarial network, which includes a generator and a discriminator. The generator learns the data distribution of real extreme conditions through a transposed convolutional structure to generate simulated time-series data.

[0014] Furthermore, the risk warning model includes a bidirectional long short-term memory network layer and an additive attention layer connected in sequence. The bidirectional long short-term memory network layer is used to capture the forward and backward dependencies of time-series data, and the additive attention layer is used to assign attention weights to high-weight features of explosion precursors.

[0015] Furthermore, the AI ​​Agent decision layer also generates graded decision instructions based on preset three-level graded decision rules. The three-level graded decision rules include: generating a safety monitoring instruction when the explosion risk probability value is in the first interval, generating a first-level blocking instruction when it is in the second interval, and generating a second-level blocking instruction when it is in the third interval.

[0016] Furthermore, the active execution layer includes a gas replacement device, a power cut-off device, and a locking device; when a first-level blocking command is generated, the gas replacement device is triggered to perform inert gas replacement; when a second-level blocking command is generated, the power cut-off device and / or the locking device are triggered, and the gas replacement device is also triggered to perform inert gas replacement.

[0017] This invention also provides a priori active explosion-proof closed-loop control method, comprising the following steps: S1: Simultaneously collect four-dimensional sensing data of heat, light, air, and pressure from the industrial site through a multimodal sensing layer; S2: The collected four-dimensional perception data is processed with electrical isolation and energy limitation, and then transmitted to the AI ​​Agent decision layer under the condition of limiting the electrical energy of the transmission circuit; S3: Offline construction of explosion-proof dataset, which includes real industrial working condition data and extreme working condition simulation data generated by generative adversarial networks; S4: Use the explosion-proof dataset to train a hybrid early warning model that combines Bi-LSTM and attention mechanism, input the real-time received four-dimensional perception data into the trained hybrid early warning model, and output the explosion risk probability value; S5: Based on the explosion risk probability value, generate a graded decision instruction according to the preset grading rules; S6: Execute the hierarchical decision instruction, drive the corresponding execution mechanism to perform risk blocking actions, and return an execution confirmation signal to form an instruction-execution-feedback closed loop.

[0018] Furthermore, the preset grading rules in step S5 include: when the explosion risk probability value is in the first interval, a safety monitoring command is generated; when the explosion risk probability value is in the second interval, a first-level blocking command is generated to trigger the gas replacement device to perform inert gas replacement; when the explosion risk probability value is in the third interval, a second-level blocking command is generated to trigger the power cut-off device and / or the interlocking device, and to trigger the gas replacement device to perform inert gas replacement.

[0019] Furthermore, the instruction-execution-feedback closed loop in step S6 includes: After the action is completed, a confirmation signal is returned to the AI ​​Agent decision-making layer; If no confirmation signal is received within the preset time, a second command will be sent. If there is still no response after the second transmission, an execution error alarm will be output.

[0020] Compared with the prior art, the present invention has the following beneficial effects: This invention adopts a four-layer full-link closed-loop architecture, which overturns the traditional passive explosion-proof post-disaster remediation mode and realizes the pre-disaster prevention from the perception of explosion precursors to the active blocking of risks, with a system end-to-end response time of ≤0.28s.

[0021] This invention integrates four-dimensional sensing modules (thermal, optical, gas, and pressure) through a multimodal sensing layer and combines them with an FPGA synchronous sampling mechanism to achieve microsecond-level synchronous acquisition with a sampling time deviation of ≤1μs. This solves the problem of data timing misalignment in traditional discrete sensors and provides an accurate data foundation for subsequent feature extraction and risk prediction.

[0022] This invention employs an improved DCGAN deep convolutional generative adversarial network, which learns the data distribution of real explosion precursors through adversarial training, generating highly physically consistent extreme condition simulation data. The KL divergence between the generated data and the real data reaches 0.08, effectively making up for the shortcoming of the scarcity of real data for extreme conditions and improving the extreme condition recognition rate to over 98.7%.

[0023] This invention employs a hybrid early warning model that combines Bi-LSTM with an attention mechanism. By capturing the forward and backward dependencies of time-series data through Bi-LSTM and dynamically focusing on high-weight early warning features through additive attention, it achieves a dual improvement in recognition accuracy ≥99.5% and single-sample inference latency ≤32ms.

[0024] This invention constructs an intrinsically safe transmission link through an optocoupler isolation circuit and a current-limiting and voltage-regulating module, limiting the maximum current of the transmission circuit to within 20mA. While ensuring low-latency data transmission, it eliminates the risk of the transmission link becoming a potential ignition source, and the entire link complies with the ExiaIICT4Ga intrinsically safe standard.

[0025] This invention establishes a three-level intelligent hierarchical decision-making rule, which transforms risk probability values ​​into hierarchical execution instructions and sets up an instruction-execution-feedback closed-loop verification mechanism to ensure the reliability of the execution actions.

[0026] This invention is not a simple combination of multiple sensors and general AI. FPGA microsecond-level synchronous perception (≤1μs) ensures the physical integrity of the joint features of pre-explosion precursors in terms of time sequence; the improved DCGAN preserves the phase relationship of cross-channel data distribution, filling the cognitive blind spot of extreme working conditions; Bi-LSTM and attention mechanism focus on sub-macroscopic features such as micro-arc pulse width, making the decision window lead the macroscopic physical threshold. The three work together to achieve the improvement from "post-explosion isolation" to "pre-explosion active blocking". Attached Figure Description

[0027] Figure 1 This is a diagram of the four-layer closed-loop overall architecture of the priori active explosion-proof closed-loop control system of the present invention.

[0028] Figure 2A This is a schematic diagram of the overall structure of the multimodal integrated sensing terminal of the present invention.

[0029] Figure 2B This is a schematic diagram of the structure of the multimodal sensing module of the present invention.

[0030] Figure 2C This is a schematic diagram of the asymmetric airflow guide channel of the present invention.

[0031] Figure 3 This is a circuit diagram of the intrinsically secure transmission link of the present invention.

[0032] Figure 4 This is a schematic diagram of the structure of the improved DCGAN network of the present invention.

[0033] Figure 5 This is a schematic diagram of the structure of the Bi-LSTM+Attention hybrid early warning model of the present invention.

[0034] Figure 6 This is a schematic diagram of the three-level intelligent decision-making and execution feedback process of the present invention.

[0035] Figure 7 This is a schematic diagram illustrating the implementation principle of the present invention.

[0036] The attached diagram is labeled as follows: 1. Integrated explosion-proof base; 2. Ultraviolet detector; 3. Thermal sensor array; 4. Laser gas detection module; 5. Piezoresistive pressure sensor; 6. Asymmetric airflow guide channel. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Example 1: Priority-based active explosion-proof closed-loop control system This embodiment provides a priori active explosion-proof closed-loop control system. (Refer to...) Figure 1 and Figure 7 The system comprises a four-layer, fully closed-loop architecture: a multimodal perception layer, an intrinsically safe transmission layer, an AI Agent decision-making layer, and an active execution layer. Each layer interacts with the others via standardized interfaces, forming a complete closed loop from data acquisition, secure transmission, intelligent decision-making to active execution and feedback verification. The system's end-to-end response time is ≤0.28s, and it fully complies with the ExiaIICT4Ga intrinsically safe standard.

[0039] I. Multimodal Sensing Layer The multimodal sensing layer is used to simultaneously collect four-dimensional sensing data of heat, light, gas, and pressure inside the industrial explosion-proof enclosure. This layer adopts an integrated multimodal sensing terminal, which integrates the heat sensing module, light sensing module, gas sensing module, and pressure sensing module into the same enclosure. Installation can be completed with only a single mounting hole on the explosion-proof cabinet, effectively maintaining the overall sealing and pressure resistance of the explosion-proof cabinet.

[0040] Reference Figures 2A to 2CThe multimodal sensing terminal includes an integrated explosion-proof base 1, within which a composite sensing cavity is located. The composite sensing cavity employs a honeycomb-shaped independent sensing compartment design, with four functional areas accommodating a thermal sensing module, a light sensing module, a gas sensing module, and a pressure sensing module, respectively. Each functional area is isolated by a metallic electromagnetic shielding layer with a shielding effectiveness ≥40dB (meeting the commercial-grade shielding requirements of IEEE Std 299-2006), ensuring that the sensing modules do not interfere with each other during operation. The metallic electromagnetic shielding layer can be made of copper, aluminum, or steel, with a thickness between 0.1mm and 1mm.

[0041] Within the composite sensing cavity, an asymmetric airflow guide channel 6 is also provided. The air inlet of the asymmetric airflow guide channel 6 is located on the side edge of the integrated explosion-proof base 1, and the air outlet faces the detection area of ​​the gas sensing module (laser gas detection module 4). Simultaneously, the overall spatial trajectory of the asymmetric airflow guide channel 6 is outside the conical field of view of the optical sensing module (ultraviolet detector 2). This structural design allows for the use of micro-pressure differences in the air to guide external gas quickly to the detection area of ​​the gas sensing module, achieving rapid response; at the same time, it avoids the airflow directly blowing against the optical window of the optical sensing module, preventing window contamination.

[0042] The integrated explosion-proof base 1 also features a porous flame-arresting and pressure-relief structure at its bottom. This structure consists of a stainless steel microporous plate with a pore size of 0.2mm-0.5mm and a thickness of 1.5mm-2.0mm. When the internal pressure of the explosion-proof housing abnormally increases, pressure can be relieved through the microporous plate. Simultaneously, because the pore size of the microporous plate is much smaller than the propagation diameter of an explosive gas flame (i.e., smaller than the maximum test safe gap MESG), even if internal ignition occurs in extreme circumstances, the flame cannot propagate outward through the microporous plate, thus achieving a flame-arresting function.

[0043] The detailed configuration and installation methods for each sensing module are as follows: (1) Thermal sensing module The thermal sensing module employs an 8-channel NTC thermistor array. The array chips are directly mounted to the inner wall of an independent sensing chamber using flip-chip bonding technology. Lead-free solder paste (Sn96.5Ag3Cu0.5) is used for the solder joints, and the reflow soldering temperature is 245℃±5℃. A 0.1mm thick transparent silicone coating covers the chip surface, protecting the solder joints while allowing infrared heat radiation to penetrate. The coating edges are precisely controlled using a dispensing machine to prevent overflow into the electrical connection area.

[0044] The thermal sensing module has a temperature measurement range of -40℃ to 125℃, a measurement accuracy of ±0.3℃, and a response time of ≤50ms. This module is used to monitor the temperature changes and temperature rise rate of key components inside the explosion-proof enclosure (such as wiring terminals and hot spots on the circuit board), and to detect early signs of overheating caused by poor contact or overload.

[0045] (2) Light sensing module The optical sensing module employs an ultraviolet detector 2. The ultraviolet detector 2 is mounted on an optical window mount at the center of the composite sensing cavity, with a response wavelength range of 185nm to 260nm, a detection sensitivity of 0.1mJ, and a pulse response time ≤10μs. This module is used to capture weak arc signals generated due to insulation aging, poor circuit contact, etc., and to identify precursors to ignition sources.

[0046] (3) Gas sensing module The gas sensing module uses a laser gas detection module 4, compatible with combustible gases such as methane and ethane. After the module is embedded in an independent sensing chamber, it is fixed to the chamber body via M4 stainless steel threads. Loctite 577 sealant is applied to the threads, and the torque is controlled at 2.5 N·m ± 0.2 N·m. The laser transmitting / receiving window is made of sapphire glass (Mohs hardness 9) and sealed with a metal pressure ring and silicone gasket. The clamping force is controlled by four M3 screws, each with a torque of 0.8 N·m.

[0047] The laser gas detection module 4 has a detection limit of ≤1ppm, a response speed of ≤200ms, and a measurement range of 0 to 100ppm. This module is used to monitor trace amounts of combustible gas seepage caused by seal failure and to detect early signs of combustible material accumulation.

[0048] (4) Pressure sensing module The pressure sensing module uses a piezoresistive pressure sensor. The bottom of the sensor and the contact surface with the independent sensing chamber are sealed with a flat surface, with a corrugated metal gasket (0.3mm thick, 316L stainless steel) sandwiched in between. A constant pressure of 500N is applied via a hydraulic pre-tightening device. The lead interface uses a ceramic sealed socket (model HERMETIC-12P). After the socket is soldered to the sensing chamber, it is filled with thermally conductive silicone grease (such as TG-300) to balance thermal stress.

[0049] The piezoresistive pressure sensor 5 has a measurement range of 0 to 1 MPa, a measurement accuracy of ±0.5%FS, and an anti-interference capability of ≥60dB. This module is used to sense abnormal pressure pulsations inside the explosion-proof enclosure and assist in judging changes in the sealing status.

[0050] (5) Synchronous sampling mechanism The multimodal sensing layer also includes an FPGA chip for implementing a synchronous sampling mechanism. The FPGA chip sends a 1kHz synchronous trigger signal to the four sensing modules, and each module performs data acquisition on the same clock edge. Synchronous triggering of the four-dimensional sensing modules is implemented using Verilog programming, with a sampling time deviation of ≤1μs for the four sensing data streams.

[0051] The acquired data is temporarily stored in the sensor's built-in 2KB buffer and uploaded to the intrinsically safe transport layer in frames. Each frame contains four channels of sensing data and corresponding timestamp information. This synchronous sampling mechanism ensures strict temporal correspondence between the multi-dimensional sensing data collected at the same time, providing an accurate data foundation for subsequent feature extraction and risk prediction.

[0052] (6) Power supply and intrinsic safety design The multimodal sensing terminal operates at 3.3V DC, with an operating current ≤20mA and a total power consumption ≤0.066W. The internal circuitry is filled with epoxy resin to eliminate the possibility of exposed circuitry generating electrical sparks. The output interface incorporates a current-limiting resistor and an overvoltage protection diode to ensure inherent safety at the output.

[0053] II. Intrinsically Secure Transport Layer Reference Figure 3 The intrinsically safe transmission layer includes a signal conditioning module, an electrical isolation module, and an energy limiting module. The design goal of this layer is to limit the electrical energy of the transmission link to levels below which an explosive atmosphere cannot be ignited, ensuring low-latency and distortion-free transmission of sensed data, thus achieving the effect of data transmission with immediate spark interruption.

[0054] (1) Signal conditioning module The signal conditioning module receives various types of signals (including analog voltage signals, digital pulse signals, I2C digital signals, etc.) from the multimodal sensing layer and converts them into 3.3VTTL level digital signals. The module has a sampling rate of 1kHz, a quantization precision of 16 bits, and a built-in RC filter circuit (composed of a 1kΩ resistor and a 0.1μF capacitor) to filter out electromagnetic interference in industrial environments, achieving a signal-to-noise ratio (SNR) of ≥60dB for the output signal.

[0055] (2) Electrical isolation module The electrical isolation module employs an optocoupler isolation circuit. Its input is connected to the output of the signal conditioning module, and its output is connected to the input of the energy limiting module. In this embodiment, the optocoupler isolation circuit uses a TLP521-4 four-channel optocoupler isolation chip, with an isolation voltage ≥2500VAC and an isolation delay ≤10μs between its input and output sides. Through optocoupler isolation, complete electrical isolation between the sensing and transmission sides is achieved, effectively preventing energy crosstalk under fault conditions.

[0056] (3) Energy limiting module The energy limiting module includes a 100Ω current-limiting resistor in series and a 5V Zener diode in parallel. This module limits the maximum current at the output of the transmission circuit to within 20mA and the maximum voltage to within 5V. With these parameters, even if a short-circuit fault occurs in the transmission circuit, the released electrical energy will be insufficient to ignite an explosive gas atmosphere.

[0057] (4) Transmission protocol and data encapsulation The energy-limited digital signals are encapsulated according to the Modbus-RTU industrial bus protocol. The protocol parameters are configured as follows: baud rate 9600bps, data bits 8 bits, stop bits 1 bit, and checksum CRC-16. The four sensing data streams are encapsulated into a 20-byte single frame and transmitted to the AI ​​Agent decision layer via the bus, with a bus transmission delay of ≤50μs.

[0058] (5) Data integrity assurance To ensure data transmission integrity, the intrinsically safe transport layer is also equipped with a retransmission mechanism. When the AI ​​Agent decision layer fails to receive a data frame within a specified time, or if the CRC-16 check fails, the sender automatically triggers a data retransmission, with a maximum of three retransmissions. If the retransmission fails after reaching the maximum number of retransmissions, the system outputs a "transmission anomaly" alarm signal to the monitoring platform. Through this design, the intrinsically safe transport layer ensures real-time and reliable data transmission while fully complying with the ExiaIICT4Ga intrinsically safe standard.

[0059] III. AI Agent Decision-Making Layer The AI ​​Agent decision-making layer is deployed on the STM32F401RCT6 core control board or edge computing device, and is responsible for the construction and enhancement of the dataset, the training and inference of the risk warning model, and the generation of hierarchical decision instructions. This layer mainly includes three parts: a data enhancement module, a risk warning model, and three-level intelligent hierarchical decision rules.

[0060] (I) Construction of Explosion-proof Dedicated Dataset Before training the model, a dedicated explosion-proof dataset is first constructed. The dataset consists of two parts: The first part is a dataset of real-world industrial operating conditions. This dataset is derived from long-term operating records of explosion-proof equipment, covering four typical operating conditions: normal operation, terminal overheating, insulation aging and micro-arc, and gas infiltration due to seal failure. It contains over 100,000 time-series data points. Each data point contains sampled values ​​from 50 consecutive time steps, with each time step including sensor data in four dimensions: heat, light, gas, and pressure. The sampling frequency is 1 kHz. The data has been anonymized, retaining feature values ​​and risk level labels.

[0061] The second part is the extreme condition simulation dataset generated by the data augmentation module.

[0062] (ii) Data Augmentation Module To address the challenges of sporadic, sudden, and highly dangerous pre-explosion conditions in real-world industrial scenarios (such as rapid temperature rise, sudden infiltration of high-concentration combustible gases, strong electric arc energy, and mixed risks), and the difficulty of on-site data collection and the limited sample size, this embodiment incorporates a data augmentation module. (Refer to...) Figure 4 The data augmentation module uses an improved DCGAN (Deep Convolutional Generative Adversarial Network).

[0063] The improved DCGAN consists of two neural networks: a generator and a discriminator.

[0064] The generator employs a 4-layer transposed convolutional structure: the first layer receives a 100-dimensional random noise vector, which is transposed and then convolved to output 64 channels; the second layer outputs 32 channels; the third layer continues upsampling; and the fourth layer outputs 4 channels of time-series data in a 50×4-dimensional format, consistent with the format of real-world operating data. The first three layers use LeakyReLU activation (negative slope 0.2), and the output layer uses Tanh activation, constraining the output values ​​to the interval [-1, 1].

[0065] The discriminator employs a 4-layer convolutional structure: the first layer receives 4-channel temporal data and outputs 32 channels after convolution; the second layer outputs 64 channels, followed by a batch normalization layer for stable training; the third layer outputs 64 channels; and the fourth layer outputs binary classification probability values, followed by a sigmoid activation function. Each layer uses LeakyReLU (negative slope 0.2) as the activation function.

[0066] During training, samples labeled "red level (high risk)" are first selected from a real industrial working condition dataset as real samples. Then, real samples and simulated samples generated by the generator are alternately input into the discriminator for adversarial training. The generator's training goal is to make the distribution of the generated simulated data as close as possible to the distribution of real data, so that the discriminator cannot accurately distinguish between real and fake data; the discriminator's training goal is to distinguish as accurately as possible whether the input data is a real sample or a generated sample.

[0067] The training parameters were set as follows: batch size 64, number of iterations 10000, learning rate 0.0002, Adam optimizer first-order moment decay coefficient β1=0.5, second-order moment decay coefficient β2=0.999, and weight decay coefficient 1e-5. During training, the generator and discriminator alternately updated the network weights: first, the generator parameters were fixed, and the discriminator parameters were updated; then, the discriminator parameters were fixed, and the generator parameters were updated.

[0068] As training progresses, the quality of the simulated data generated by the generator gradually improves, while the discriminator's accuracy gradually decreases. When the discriminator's accuracy in recognizing both real and generated samples approaches 50%, it indicates that the generator has learned the data distribution characteristics of real extreme conditions, and the discriminator can no longer distinguish between real and fake data, at which point training terminates.

[0069] After training, the generator can be used to generate more than 30,000 extreme operating condition simulation time series data, including various scenarios such as rapid temperature rise, high concentration gas infiltration, strong electric arc energy, and mixed risks.

[0070] The KL divergence (KL divergence) calculation shows that the generated data and the real data have a KL divergence of 0.08, indicating that the two have a high degree of consistency in data distribution. The generated data can serve as an effective supplement to the real extreme working condition samples.

[0071] (III) Data Preprocessing Process After merging real-world industrial operating data with the generated extreme operating condition simulation data, the following preprocessing procedure is performed: (1) 3σ criterion outlier removal: Calculate the mean μ and standard deviation σ for each dimension of data. Data points that deviate from the mean by more than 3σ are identified as outliers and removed. The missing positions are filled with the median of the data in that dimension. For example, for the temperature dimension, if the temperature value of a sampling point deviates from the temperature mean by more than 3 times the standard deviation, the point is identified as an outlier and replaced with the median of the temperature of all normal sampling points in that dimension.

[0072] (2) Linear interpolation to supplement missing data: For individual sampling points that may be missing during the data acquisition process, linear interpolation is used to supplement them. That is, the value of the missing point is estimated according to the linear relationship based on the values ​​of the two valid sampling points before and after the missing point.

[0073] (3) Min-Max Normalization: The Min-Max normalization method is used to linearly map the data of each dimension to the interval [0, 1], eliminating the dimensional differences between different perceptual dimensions. The normalization formula is: x'=(x-x_min) / (x_max-x_min); Where x represents the original data, and x_min and x_max represent the minimum and maximum values ​​of the data in this dimension, respectively.

[0074] (4) Sliding window construction: Long time series data is divided into short sequences according to a fixed window size of 50 time steps (corresponding to a time length of 50ms), and the sliding step size is 10 time steps. Each window corresponds to a risk prediction label. For example, for a time series data with a length of 200 time steps, 16 training samples can be generated by sliding with a step size of 10.

[0075] (5) Data set partitioning: The preprocessed dataset is divided into training set (70%), validation set (15%) and test set (15%) in chronological order to simulate the characteristics of data generation over time in real industrial scenarios and avoid time travel between training data and test data.

[0076] (iv) Risk early warning model Reference Figure 5 The risk warning model employs a hybrid model combining Bi-LSTM and an attention mechanism to predict the probability of explosion risks from real-time perceived data. The model consists of an input layer, an embedding layer, a bidirectional long short-term memory network layer, an additive attention layer, and a fully connected layer connected sequentially.

[0077] The input layer receives a 50×4-dimensional temporal feature matrix, where 50 represents the time step and 4 represents the four sensing dimensions of heat, light, air, and pressure.

[0078] The embedding layer is a fully connected layer that maps 4-dimensional input features to a 64-dimensional feature space, with an output dimension of 50×64, to enhance the expressive power of the features.

[0079] The bidirectional Long Short-Term Memory (LSTM) network consists of a forward LSTM network and a backward LSTM network. Each hidden layer contains 128 neurons, and the activation function is ReLU. The forward LSTM processes the temporal data in ascending chronological order (from time step 1 to time step 50) to extract temporal dependencies from the past to the future; the backward LSTM processes the temporal data in descending chronological order (from time step 50 to time step 1) to extract temporal dependencies from the future to the past.

[0080] The outputs of the forward LSTM and the backward LSTM are concatenated at each time step to form a feature matrix with a dimension of 50×256. Through this bidirectional structure, the model can comprehensively utilize the forward and backward correlation information of the time series data to effectively distinguish between slow changes under normal operating conditions (such as natural fluctuations in ambient temperature) and abnormal abrupt changes under fault conditions (such as a sharp temperature jump caused by a short circuit).

[0081] The additive attention layer receives a 50×256-dimensional feature matrix output from the bidirectional long short-term memory network layer. This layer first performs a linear transformation on the feature vector at each time step using a trainable weight matrix W (256×64 dimension), then obtains the hidden representation through the tanh activation function. Next, it calculates the attention score at each time step using a trainable score vector v (64×1 dimension). The score is normalized using the Softmax function and used as the attention weight. Finally, the attention weights are weighted and summed with the original feature vector to obtain a 256-dimensional context vector. Through this attention mechanism, the model can automatically increase the attention given to features that contribute significantly to explosion risk prediction (such as micro-arc pulse width and temperature rise rate), while suppressing the influence of irrelevant features or noise.

[0082] The fully connected layer consists of two layers: the first fully connected layer has 64 neurons, uses the ReLU activation function, and is followed by a Dropout layer with a dropout rate of 0.3 to prevent overfitting; the second fully connected layer has 3 neurons, uses the Softmax activation function, and outputs probability values ​​corresponding to three risk levels: green (safe), yellow (critical), and red (high risk), with the sum of the three probability values ​​being 1. The probability value corresponding to the red level is the explosion risk probability value P, which ranges from 0 to 1.

[0083] During model training, the model is implemented using either PyTorch or TensorFlow 2.8 deep learning frameworks. The loss function is cross-entropy loss, the optimizer is Adam, the initial learning rate is 0.001, and the batch size is 32. The training epochs are set to 50 epochs. After each epoch, the loss value and classification accuracy are calculated on the validation set.

[0084] Early stopping mechanism is enabled during training: when the validation set loss does not decrease for 10 consecutive iterations, training is stopped and the model weights with the lowest loss on the validation set are saved to prevent overfitting.

[0085] To further verify the model's generalization ability, blind testing was conducted on the model after training using a test set that was not involved in the training and validation process. The blind test results showed that the model had a 100% recall rate and a 0% false negative rate for red-level high-risk states. Comprehensive testing demonstrated that the model achieved an accuracy of ≥99.5% in identifying pre-explosion warning features on the test set, with a single-sample inference latency of ≤32ms.

[0086] (v) Three-level intelligent hierarchical decision-making rules The AI ​​Agent decision-making layer generates corresponding decision instructions based on the explosion risk probability value P output by the risk warning model, according to a preset three-level hierarchical decision-making rule. (Refer to...) Figure 6 The specific rules are shown in the table below: IV. Active Execution Layer The active execution layer includes a gas replacement device, a power cut-off device, and a locking device, which are used to receive and execute hierarchical decision instructions issued by the AIAgent decision-making layer.

[0087] (1) Gas replacement device The gas replacement device employs a miniature nitrogen replacement system, consisting of a 1L high-pressure nitrogen tank (operating pressure 0.8MPa), an intrinsically safe solenoid valve, and a flow control module. The intrinsically safe solenoid valve has a response time ≤20ms.

[0088] The flow control module adjusts the output flow based on the instruction type: When instruction 0x01 (Level 1 Blocking) is received, nitrogen gas is output at the first flow rate (1L / min); When instruction 0x02 (secondary blocking) is received, nitrogen gas is output at a second flow rate (2L / min), which is twice the first flow rate.

[0089] By injecting inert nitrogen into the explosion-proof enclosure, the oxygen concentration inside the enclosure can be diluted to below 12% and the flammable gas concentration can be reduced to below 25% of the lower explosive limit (LEL) within 300ms, thereby destroying the oxidizer and flammable material conditions required for an explosion at the source.

[0090] (2) Power cut-off device The power cut-off device uses an intrinsically safe solid-state relay connected in series in the main power supply circuit of the explosion-proof equipment. Its response time is ≤50μs, operating voltage is 12VDC, operating current is ≤50mA, it has no contact wear, and a service life of ≥100,000 cycles. When command 0x02 is received, the solid-state relay immediately disconnects, cutting off the equipment power supply and eliminating the electrical ignition source before ignition energy accumulates.

[0091] (3) Locking device The interlocking device is an electromagnetic interlocking device with an operating voltage of 3.3VDC, an operating current of ≤30mA, and a locking force of ≥500N. Upon receiving command 0x02, the electromagnetic interlocking device activates, locking the equipment hatch door to prevent accidental opening by personnel in hazardous conditions. The interlocking device can only be unlocked when the risk level drops to green and an unlocking command (0x03) is received.

[0092] V. Command-Execution-Feedback Closed-Loop Verification Mechanism To ensure that decision-making instructions are executed reliably, the system sets up a closed-loop verification mechanism between the AI ​​Agent decision-making layer and the active execution layer.

[0093] After completing the instruction execution, the active execution layer returns an acknowledgment signal to the AI ​​Agent decision-making layer. The correspondence between the acknowledgment signal and the decision instruction is as follows: instruction 0x00 corresponds to acknowledgment signal 0x0A, instruction 0x01 corresponds to acknowledgment signal 0x0B, and instruction 0x02 corresponds to acknowledgment signal 0x0C.

[0094] After issuing a decision command, the AI ​​Agent decision-making layer starts a timer with a duration of 500ms. If a corresponding confirmation signal is received within this time, the command is considered successfully executed, the execution status is recorded, and the next monitoring cycle continues. If no confirmation signal is received after 500ms, the AI ​​Agent decision-making layer immediately resends the same decision command to the active execution layer. If a confirmation signal is still not received within the specified time after the second resend, the system determines that an abnormality has occurred in the execution process, outputs an "execution abnormality" alarm signal, and pushes the alarm information to the monitoring platform.

[0095] VI. Data Visualization and Encrypted Evidence Storage The system is equipped with a visual monitoring interface that supports simultaneous access from three terminals: industrial SCADA monitoring platform, local touch screen, and mobile APP. The monitoring interface displays in real time the current values ​​of each sensing module (temperature, arc energy, combustible gas concentration, pressure), current risk level (green / yellow / red), the content of the most recent decision command, the working status of the actuators, and the system operating status, among other information.

[0096] Meanwhile, the system encrypts and stores all perception data, risk prediction results, decision command records, and execution feedback records generated during operation. The encryption method uses AES-256CBC mode with a key length of 32 bytes. Each record is encrypted with AES after generation, then Base64 encoded before storage. The stored data is tamper-proof and provides a reliable basis for subsequent fault analysis and accident tracing.

[0097] Example 2: A priori active explosion-proof closed-loop control method This embodiment provides a priori active explosion-proof closed-loop control method, which can be implemented based on the system described in Embodiment 1. The method includes the following steps: S1: Multimodal synchronous sensing The FPGA chip in the multimodal sensing layer sends a 1kHz synchronous trigger signal to the thermal sensing module, optical sensing module, gas sensing module, and pressure sensing module. The four sensing modules collect temperature, electric arc, combustible gas concentration, and pressure data respectively on the same clock edge, with a sampling time deviation ≤1μs. The collected four-channel sensing data, along with timestamp information, are temporarily stored in a 2KB buffer and uploaded to the intrinsically safe transport layer in frames, with each frame containing complete data from all four channels.

[0098] S2: Intrinsically safe data transmission The signal conditioning module of the intrinsically safe transmission layer receives four channels of sensing data and converts them into a standard digital signal at 3.3 VTL level.

[0099] The optocoupler isolation circuit (using TLP521-4 chip) provides electrical isolation between the signal input side and the output side, with an isolation voltage ≥2500VAC.

[0100] The current-limiting and voltage-regulating module limits the current in the transmission circuit to within 20mA and the voltage to within 5V by connecting a 100Ω current-limiting resistor in series and a 5V Zener diode in parallel.

[0101] After energy management, the data is encapsulated into a 20-byte single frame according to the Modbus-RTU protocol (9600bps baud rate, 8 data bits, 1 stop bit), and a checksum is added using the CRC-16 check algorithm before being transmitted to the AI ​​Agent decision layer. If the CRC check fails at the receiving end, data retransmission is triggered, with a maximum of 3 retransmissions; if retransmission still fails, a transmission abnormality alarm is output.

[0102] S3: Explosion-proof Dataset Construction and Data Augmentation The AI ​​Agent decision-making layer builds a dedicated explosion-proof dataset during the offline phase.

[0103] First, acquire real-world industrial operating condition data, including various conditions such as normal operation, terminal overheating, insulation aging micro-arc, and gas infiltration due to seal failure.

[0104] Then, the improved DCGAN data augmentation module is used to generate extreme condition simulation data: (1) Extract samples labeled as red-level high-risk from real industrial data as real samples; (2) Construct an improved DCGAN network, with the generator being a 4-layer transposed convolutional structure and the discriminator being a 4-layer convolutional structure; (3) Input 100-dimensional random noise into the generator and output 50×4-dimensional simulated time series data; alternately input real samples and generated samples into the discriminator to distinguish between real and fake samples; (4) The generator and discriminator are optimized alternately. The training parameters are set as follows: batch size 64, number of iterations 10000, and learning rate 0.0002. (5) Training is stopped when the discriminator’s recognition accuracy for both real and generated samples is close to 50%. (6) Using the trained generator, over 30,000 extreme operating condition simulation data were generated, including scenarios such as rapid temperature rise, high-concentration gas infiltration, strong electric arc energy, and mixed risks. The KL divergence between the generated data and the real data was 0.08.

[0105] The real industrial operating condition data is merged with the generated extreme operating condition simulation data, and the following preprocessing is performed sequentially: (1) 3σ criterion outlier removal: Data points that deviate from the mean by more than 3σ are identified as outliers and filled with the median of the data in that dimension; (2) Linear interpolation to supplement missing data; (3) Min-Max normalization: Mapping the data of each dimension to the interval [0, 1]; (4) Sliding window construction: The window size is 50 time steps, and the sliding step size is 10 time steps; (5) Data set partitioning: The dataset is divided into training set, validation set and test set in a ratio of 7:1.5:1.5.

[0106] S4: Risk Warning Model Training and Risk Prediction The hybrid early warning model combining Bi-LSTM and attention mechanism is trained using the training set constructed in step S3.

[0107] The model structure is as follows: Input layer (receives a 50×4-dimensional temporal feature matrix) → Embedding layer (maps the 4-dimensional features to a 64-dimensional space) → Bidirectional Long Short-Term Memory network layer (128 neurons each for forward LSTM and backward LSTM, ReLU activation function, outputting a 50×256-dimensional feature matrix) → Additive attention layer (calculates attention weights through weight matrix and score vector, outputting a 256-dimensional context vector) → Fully connected layer (64 neurons for ReLU in the first layer, 3 neurons for Softmax in the second layer, outputting green, yellow, and red probability values).

[0108] Training configuration: loss function is cross-entropy loss, optimizer is Adam, initial learning rate is 0.001, batch size is 32, and number of iterations is 50. Early stopping mechanism is enabled during training: training stops when the validation set loss does not decrease for 10 consecutive iterations, and the model weights with the lowest validation set loss are saved.

[0109] After training, the real-time received four-dimensional perception data is preprocessed using the same procedures as the training data and then input into the model. After processing layer by layer, the model outputs probability values ​​for three risk levels: green, yellow, and red. The probability value corresponding to the red level is taken as the explosion risk probability value P.

[0110] S5: Three-level intelligent hierarchical decision-making The AI ​​Agent decision-making layer generates decision instructions based on the explosion risk probability value P, according to a preset three-level classification rule: When P < 30%, it is judged as green level (safe), and a safety monitoring command 0x00 is generated; When 30%≤P<70%, it is judged as yellow level (critical), and a level 1 blocking instruction 0x01 is generated; When P ≥ 70%, it is judged as red level (high risk), and a level 2 blocking instruction 0x02 is generated.

[0111] S6: Proactive Risk Intervention and Feedback Verification After receiving the decision instruction, the active execution layer executes the corresponding action: If the instruction is 0x00, the actuator will not move, and the system will continue to collect and upload data; If the instruction is 0x01, the intrinsically safe solenoid valve of the gas replacement device will open, and the flow control module will output nitrogen at a flow rate of 1L / min for replacement. Within 300ms, the oxygen concentration inside the shell will be diluted to below 12% and the flammable gas concentration will be reduced to below 25% of the lower explosive limit. If the command is 0x02, the power cut-off device (intrinsically safe solid-state relay) disconnects the main power supply circuit (response time ≤ 50μs), the locking device (electromagnetic locking device) locks the hatch (locking force ≥ 500N), the gas replacement device outputs nitrogen at a flow rate of 2L / min for replacement, and at the same time pushes an emergency alarm to the staff's mobile terminal.

[0112] After the action is completed, the active execution layer returns the corresponding confirmation signal to the AI ​​Agent decision layer: 0x00 corresponds to 0x0A, 0x01 corresponds to 0x0B, and 0x02 corresponds to 0x0C.

[0113] If the AI ​​Agent decision-making layer does not receive a confirmation signal within 500ms, it will trigger a second instruction to be sent; if no confirmation signal is received after the second sending, it will output an execution abnormality alarm and push it to the monitoring platform.

[0114] The above steps S1 to S6 are executed cyclically, forming a closed-loop control process for continuous monitoring, prior prediction, and proactive prevention of explosion risks at industrial sites.

[0115] Example 3: Performance Testing and Verification To verify the effectiveness of the technical solution of this invention, a test platform simulating an industrial explosion-proof scenario was built. The main equipment of the test platform includes: 0.5m... 3 Stainless steel explosion-proof housing, programmable heating module (temperature rise rate adjustable from 0 to 10℃ / s), arc generation module (adjustable from 0.1 to 5mJ), methane precision injection system (accuracy ±0.1ppm), high-speed data acquisition card (sampling rate 1MHz) and gas concentration analyzer.

[0116] The test was set up with three typical operating conditions: Operating Condition 1 (Insulation Aging Micro-Arc Condition): Arc energy 0.1–1 mJ, occurrence frequency 1–5 Hz, simulating intermittent weak arcs generated by line insulation aging. This condition is used to verify the system's ability to capture weak arc signals and its effectiveness in identifying ignition source precursors.

[0117] Operating Condition 2 (Overheating Condition): Temperature rise rate 1–8℃ / s, simulating rapid temperature rise caused by poor contact at the wiring terminals. This condition is used to verify the system's ability to identify temperature change trends and its effectiveness in distinguishing between normal and fault temperature rises.

[0118] Operating Condition 3 (Seal Failure and Combustible Gas Infiltration): Methane injection rate 1–15 ppm / s, simulating continuous infiltration of combustible gas due to seal aging. This condition is used to verify the system's ability to monitor changes in combustible gas concentration and its effectiveness in identifying precursors of combustible material accumulation.

[0119] Each type of operating condition was tested 50 times, and all performance indicators were recorded.

[0120] The test results are shown in the table below: The following are detailed records of two typical test cases: Typical Case 1 (Operating Condition 1): When the arc energy is 0.5mJ and the frequency is 3Hz, the optical sensing module of the multimodal sensing terminal captures the first micro-arc pulse signal within 0.12s. The intrinsically safe transmission layer takes 45μs to transmit data to the AI ​​Agent decision layer; the Bi-LSTM+Attention model takes 32ms to infer and outputs an explosion risk probability value P=0.85 (≥70%), which is judged as red-level high risk; The active actuator cut off the power supply and initiated nitrogen purging at a rate of 2 L / min within 48 μs. The entire end-to-end response time was 0.24 s, and no ignition occurred inside the casing, successfully preventing the risk of explosion.

[0121] Typical Case 2 (Operating Condition 3): When the methane injection rate is 12 ppm / s, the gas sensing module continuously monitors the increase in gas concentration. When the concentration reaches 30 ppm, the model outputs a risk probability value P=0.52 (between 30% and 70%), which is determined to be a yellow-level critical condition, triggering the first-level blocking command 0x01.

[0122] The active execution layer initiates nitrogen purging at a rate of 1 L / min. Within 300 ms, the concentration of combustible gas inside the inner shell drops from 30 ppm to 8 ppm, which is below 25% of the lower explosive limit (LEL) (the LEL for methane is approximately 4.4%, and 25% LEL is approximately 1.1% by volume). The risk level drops back to green, and the system resumes normal monitoring.

[0123] Test results show that the system and method provided by this invention meet the design specifications in terms of accuracy of explosion precursor identification, extreme condition identification rate, response time, and accident prevention success rate. Compared with traditional passive explosion-proof technology, this invention achieves pre-emptive prevention and has significant advantages in dimensions such as defense logic, response time, failure risk, adaptability to extreme conditions, and intrinsic safety level.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A priori active explosion-proof closed-loop control system, characterized in that, include: The multimodal sensing layer is used to simultaneously collect four-dimensional sensing data of heat, light, air, and pressure from the industrial site. The intrinsically safe transmission layer includes a signal conditioning module, an electrical isolation module, and an energy limiting module. The intrinsically safe transmission layer transmits the data collected by the multimodal sensing layer to the AI ​​Agent decision layer under the condition of limiting the electrical energy of the transmission circuit. The AI ​​Agent decision-making layer includes a data enhancement module and a risk warning model; The data augmentation module uses generative adversarial networks to generate extreme condition simulation data to expand the training dataset; The risk warning model adopts a hybrid model that combines Bi-LSTM and attention mechanism. The risk warning model is trained based on the expanded training dataset, and predicts the probability of explosion risk from real-time perceived data and generates hierarchical decision instructions. The active execution layer includes multiple execution mechanisms. The active execution layer receives and executes the hierarchical decision instructions issued by the AI ​​Agent decision layer, and returns an execution confirmation signal to the AI ​​Agent decision layer to form an instruction-execution-feedback closed loop.

2. The priori active explosion-proof closed-loop control system according to claim 1, characterized in that, The multimodal sensing layer includes an FPGA chip, which synchronously triggers the acquisition of thermal sensing modules, optical sensing modules, gas sensing modules, and pressure sensing modules. The sampling frequency is 1kHz, and the sampling time deviation is ≤1μs.

3. The priori active explosion-proof closed-loop control system according to claim 1, characterized in that, In the intrinsically safe transmission layer, the electrical isolation module uses an optocoupler isolation chip with an isolation voltage ≥2500VAC; the energy limiting module includes a current-limiting resistor and a Zener diode to limit the maximum current of the transmission circuit to within 20mA.

4. The priori active explosion-proof closed-loop control system according to claim 1, characterized in that, The data augmentation module employs a deep convolutional generative adversarial network, which includes a generator and a discriminator. The generator learns the data distribution of real extreme conditions through a transposed convolutional structure to generate simulated time-series data.

5. The priori active explosion-proof closed-loop control system according to claim 1, characterized in that, The risk warning model includes a bidirectional long short-term memory network layer and an additive attention layer connected in sequence. The bidirectional long short-term memory network layer is used to capture the forward and backward dependencies of time-series data, and the additive attention layer is used to assign attention weights to high-weight features of explosion precursors.

6. The priori active explosion-proof closed-loop control system according to claim 1, characterized in that, The AI ​​Agent decision layer also generates graded decision instructions based on preset three-level graded decision rules. The three-level graded decision rules include: generating a safety monitoring instruction when the explosion risk probability value is in the first interval, generating a first-level blocking instruction when it is in the second interval, and generating a second-level blocking instruction when it is in the third interval.

7. The priori active explosion-proof closed-loop control system according to claim 6, characterized in that, The active execution layer includes a gas replacement device, a power cut-off device, and a locking device; When a Level 1 blocking command is generated, the gas replacement device is triggered to perform inert gas replacement; When a secondary blocking command is generated, the power cut-off device and / or the interlocking device are triggered, and the gas replacement device is triggered to perform inert gas replacement.

8. A priori active explosion-proof closed-loop control method, characterized in that, Includes the following steps: S1: Simultaneously collect four-dimensional sensing data of heat, light, air, and pressure from the industrial site through a multimodal sensing layer; S2: The collected four-dimensional perception data is processed with electrical isolation and energy limitation, and then transmitted to the AI ​​Agent decision layer under the condition of limiting the electrical energy of the transmission circuit; S3: Offline construction of explosion-proof dataset, which includes real industrial working condition data and extreme working condition simulation data generated by generative adversarial networks; S4: Use the explosion-proof dataset to train a hybrid early warning model that combines Bi-LSTM and attention mechanism, input the real-time received four-dimensional perception data into the trained hybrid early warning model, and output the explosion risk probability value; S5: Based on the explosion risk probability value, generate a graded decision instruction according to the preset grading rules; S6: Execute the hierarchical decision instruction, drive the corresponding execution mechanism to perform risk blocking actions, and return an execution confirmation signal to form an instruction-execution-feedback closed loop.

9. The priori active explosion-proof closed-loop control method according to claim 8, characterized in that, The preset grading rules in step S5 include: When the explosion risk probability value is in the first range, a safety monitoring command is generated; When the explosion risk probability value is in the second range, a first-level blocking command is generated, triggering the gas replacement device to perform inert gas replacement. When the explosion risk probability value is in the third range, a secondary blocking command is generated, which triggers the power cut-off device and / or the interlocking device, and triggers the gas replacement device to perform inert gas replacement.

10. The priori active explosion-proof closed-loop control method according to claim 8, characterized in that, The instruction-execution-feedback closed loop in step S6 includes: After the action is completed, a confirmation signal is returned to the AI ​​Agent decision-making layer; If no confirmation signal is received within the preset time, a second command will be sent. If there is still no response after the second transmission, an execution error alarm will be output.

Images