Novel VPSA method biogas decarbonization purification equipment intelligent control joint control system

Abstract

The invention discloses a novel VPSA method biogas decarbonization purification equipment intelligent control joint control system, belongs to the industrial automation and artificial intelligence fusion field, and comprises a global perception and data acquisition layer, an edge intelligent decision and control layer, an accurate execution and linkage control layer, and a cloud edge collaborative management and digital twinborn layer four-layer architecture. The global sensing layer is provided with a 48-node sensor network and a 10kHz acquisition card to complete data acquisition preprocessing; the edge decision-making layer carries a reinforcement learning algorithm and an LSTM-CNN model through an ARM + FPGA architecture; the precise execution layer realizes millisecond-level response through a high-speed execution mechanism and a Petri net linkage matrix; and the cloud side management layer realizes remote management and virtual debugging by relying on a cloud platform and a digital twin system, solves the problems of rigid control, weak fault early warning and the like of traditional equipment, and is suitable for the fields of biomass energy utilization and environmental protection engineering.

Description

Technical Field

[0001] This invention relates to the field of industrial automation and artificial intelligence technology integration, specifically to a novel intelligent control system for VPSA biogas decarbonization and purification equipment. Background Technology

[0002] Currently, the control system of VPSA biogas decarbonization equipment has the following technical bottlenecks: First, the control strategy is fixed, and it is impossible to dynamically optimize the operating parameters according to the real-time changes in biogas composition and flow, resulting in low equipment operating efficiency; Second, fault diagnosis relies on manual inspection and threshold alarms, lacking the ability to provide early warning of potential faults, and the equipment sudden shutdown rate is as high as 10%-12%; Third, the data utilization rate is low, and a large amount of operating data has not been effectively analyzed and mined, making it difficult to support process optimization and decision-making; Fourth, there is a lack of collaborative linkage mechanism between equipment, making it impossible to achieve optimal control of overall performance. Summary of the Invention

[0003] The purpose of this invention is to provide a novel intelligent control system for VPSA biogas decarbonization and purification equipment to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a novel intelligent control system for VPSA biogas decarbonization and purification equipment, comprising a global perception and data acquisition layer, an edge intelligent decision-making and control layer, a precise execution and linkage control layer, and a cloud-edge collaborative management and digital twin layer; the global perception and data acquisition layer includes a multi-parameter sensor network and a data acquisition and preprocessing module; the multi-parameter sensor network sets 48 sensor nodes for the adsorption tower, compressor, vacuum pump, and preprocessing unit of the VPSA biogas decarbonization and purification system; the data acquisition and preprocessing module uses a high-precision data acquisition card with a sampling frequency of 10kHz to obtain the data collected in the multi-parameter sensor network; the collected data is combined with edge computing nodes to complete data filtering, outlier removal, and normalization processing to obtain a standardized dataset; The edge intelligent decision-making and control layer includes an edge computing server, a fault prediction and health management model, and an intelligent controller. The edge computing server adopts a heterogeneous computing architecture of ARM processor and FPGA acceleration chip, and integrates a dynamic optimization algorithm based on reinforcement learning to optimize the operating efficiency, energy consumption index, and adsorbent lifespan of VPSA equipment. The fault prediction and health management model is an LSTM-CNN composite model that analyzes time-series data collected by the global perception and data acquisition layer to provide early warning of faults and establish an equipment health assessment system. The intelligent controller adopts a hybrid control architecture of PLC and IPC, and has built-in Modbus TCP, Profinet, and EtherCAT protocol interfaces. The precise execution and linkage control layer includes high-speed actuators and an equipment linkage control matrix. In the high-speed actuators, the adsorption tower switching valve is driven by an electro-hydraulic servo (response time < 300ms), paired with a flow meter with an accuracy of ±0.3%. The compressor and vacuum pump are equipped with servo motor drivers, supporting speed adjustment at 0.01Hz levels. The equipment linkage control matrix is ​​built on a Petri net. When the biogas flow rate suddenly increases by 20%, it triggers a linkage process of "standby adsorption tower startup - compressor speed increase - pretreatment flow increase." In case of equipment failure, it activates an emergency mechanism of "faulty equipment shutdown - standby equipment switching - dynamic adjustment of operating parameters." The cloud-edge collaborative management and digital twin layer includes a cloud platform and a digital twin system; the cloud platform receives data from the edge intelligent decision and control layer and generates process optimization reports and decision information, supporting remote monitoring and parameter adjustment; the digital twin system, based on 3D modeling and real-time data-driven operation, realizes equipment physical state mapping, operating condition simulation, virtual debugging, and fault diagnosis.

[0005] Preferably, the specific implementation steps of the reinforcement learning-based dynamic optimization algorithm for optimizing the operating efficiency, energy consumption index, and adsorbent lifespan of the VPSA equipment are as follows: Step 1: Data Preprocessing and Objective Function Definition First, at the overall perception and data acquisition layer, ≥200,000 sets of historical operating data for VPSA equipment are collected. The data dimensions cover two categories: status parameters and process parameters. It contains 28 real-time monitoring data points, all collected by sensors in the overall perception and data acquisition layer, defined as... ,in This refers to the CO2 concentration in biogas. This refers to the CH4 concentration in biogas. Biogas flow rate; The average temperature inside the adsorption tower; This refers to the vibration amplitude of the compressor. Other preset status parameters to be collected; For the set of state parameters Normalization is performed using the Min-Max standardization formula: ,in For the first The normalized value of the term parameter, , These are the minimum and maximum values ​​of the parameter in historical data, respectively. Construction of multi-objective optimization functions: Improving equipment operating efficiency Energy consumption indicators Adsorbent lifespan To optimize the objective, a total reward function is constructed by weighted summation. The formula is: Among them, equipment operating efficiency The methane purification efficiency, i.e., the ratio of the purified CH4 purity to the CH4 purity of the feed gas, is calculated in real time by the gas concentration sensor in the adsorption tower, with weights... Energy consumption indicators Energy consumption per unit of biogas processing, calculated by compressor / vacuum pump current and voltage sensors, weighted. Adsorbent lifespan This represents the amount of adsorbent activity decay within a single process cycle, and is dimensionless, ranging from [0,1]. It is derived from changes in adsorption tower temperature and CO2 concentration, with weights... ; Process parameter action space definition: Action space A is constructed based on 18 preset optimized key process parameters, and is defined as follows: ,in Adsorption time; The number of equalization cycles. Vacuum desorption pressure; This refers to the compressor speed; Other preset parameters; Step 2, DQN model structure design: The DQN model consists of a three-layer neural network: an input layer, a hidden layer, and an output layer. The structure is as follows: Input layer: 28 neurons, corresponding to the normalized set of state parameters. Dimension, reception As model input; Hidden layers: Two fully connected layers, the first with 64 neurons and the second with 32 neurons, the activation function is... This is used to extract nonlinear features of state parameters; the output layer has 18 neurons, corresponding to the dimension of the action space A, and outputs the "value" of each process parameter. value" —Indicates the current state Select action The long-term cumulative rewards are obtained, and the output layer activation function uses Linear; Model training process: Initialize the experience replay pool The capacity is set to 100,000 records to store historical training samples. Sample format: ,in For time steps, for Current state for Actions performed at all times for Rewards earned at any time for Next state at time 1; Initialize the main network With the target network The main network is used for real-time computing. Value, the target network is used to calculate the target Values, initial values ​​of both parameters and same; Iterative training: from the experience replay pool The current state can be obtained by randomly sampling a sample of size 32 from the data, or by obtaining the current state from real-time data. One of them, through - Greedy strategy for choosing actions : Execute action This means adjusting the corresponding process parameters and collecting data through the global sensing and data acquisition layer. state of time Substitute into the reward function to calculate ;Will Store in the experience replay pool ; Target value The time-difference formula is used for calculation, where let The discount factor is set to 0.9, representing the weight of future rewards. Network prediction Value Mean squared error is used as the loss function. : ,in , The first The target of each sample Values, states, and actions; learning rate through the Adam optimizer. Minimize loss function Backpropagation updates the main network parameters When the loss function For 1000 consecutive steps, the value is less than 0.01, and the average reward is... Once the value stabilizes above 0.8, the model training is complete, and the final model is deployed to the edge computing server. Step 3: Real-time parameter optimization: The edge computing server receives real-time data from the global perception and data acquisition layer through the intelligent controller, and updates the state parameter set every 100ms. Then, normalization is performed according to the Min-Max formula in step 1 to obtain... ;Will Input the trained DQN main network Output 18 process parameters corresponding to value ,choose The combination of actions with the highest value is taken as the optimal process parameter. The solution formula is: ,in , For the first The range of values ​​for each process parameter.

[0006] Preferably, the fault prediction and health management model analyzes the time-series data collected by the full-domain perception and data acquisition layer to provide early warnings of faults and establish an equipment health assessment system. The specific implementation logic is as follows: A. Data Acquisition and Preprocessing: For the pre-defined core fault types of VPSA equipment (10 categories including bearing wear, valve seal failure, adsorbent poisoning, and motor phase loss), two types of key data are collected: (1) Time-series monitoring dataset It includes four types of core monitoring data, all of which are time-series data (sampling frequency 10kHz, synchronously acquired by a data acquisition card), defined as: Vibration signal time-series data (unit: mm / s), from a vibration acceleration sensor (frequency response 0.1-10kHz) of a compressor / vacuum pump. Each sample is a 10-second vibration sequence, with dimensions of [missing information]. ( (i.e., 10kHz × 10s). Temperature field time series data (unit: °C), from a distributed fiber optic temperature sensor (accuracy ±0.1 °C) of the adsorption tower. Each sample is a 5-minute temperature sequence. ( That is, 1 time / second × 300 seconds; 6 refers to the 6 monitoring points of the fiber optic sensor. Oil wear elemental time-series data (unit: ppm), from an oil spectrometer for compressors / vacuum pumps (supports detection of 12 elements). Each sample is a 1-hour elemental concentration sequence. ( (i.e., 1 time / hour; 12 refers to 12 wear elements). Current and voltage time-series data (unit: A / V), from current and voltage sensors of compressors / vacuum pumps (accuracy ±0.2%), each sample is a 1-minute electrical parameter sequence, with dimensions of [missing information]. ( That is, 1 time / second × 60 seconds; 2 represents the two parameters of current and voltage. The overall dimension of the time series monitoring dataset is ,in F represents the sample size (50,000 samples were collected in this invention, including 12,000 fault samples), and F represents the feature dimension. .

[0007] (2) Fault label set For each time series sample, the fault type and fault occurrence time are labeled. The fault type is represented by one-hot encoding, defined as: ,in, , This indicates that the sample corresponds to the k-th type of fault (e.g. Due to bearing wear, (Valve seal failure) Indicates that the fault does not fall into this category; also indicates the time the fault occurred. (Unit: h), used for subsequent early warning time calculation (e.g., if there are 96 hours remaining before the fault occurs when the sample is collected, then...). ); B. Construction and training of CNN-LSTM composite network: Network architecture design: CNN module: Extracting spatial features: for time-series data To analyze local features (such as peak values ​​of vibration signals and hotspot regions in the temperature field), a 1D-CNN (one-dimensional convolutional) module is designed, with the following structure: Convolutional layer 1: Input dimension ( ), number of kernels 32, kernel size 3, stride 1, activation function Output dimension Pooling layer 1: Max pooling, pooling kernel size 2, stride 2, output dimension Convolutional layer 2: 64 kernels, kernel size 3, stride 1, activation function... Output dimension Pooling layer 2: Max pooling, pooling kernel size 2, stride 2, output dimension Flattening layer: Transforms the two-dimensional features output by the pooling layer into a one-dimensional feature vector, with dimensions of... (like (At that time, the dimension is 1,600,000). LSTM module: Capturing time series trends: First, process the time series data... Preprocessing is performed on the preprocessed time series data. The LSTM module is designed with the following structure: LSTM layer 1: Input dimension is the same as the feature vector dimension output by the CNN module (e.g., 1,600,000), number of hidden units is 128, activation function is tanh, and output dimension is... LSTM layer 2: 64 hidden units, tanh activation function, output dimension... Dropout layer: Dropout probability 0.5 to prevent model overfitting, output dimension The LSTM module captures timing trends through a gating mechanism. Fully Connected Layer: Fault Classification and Early Warning Fully connected layer 1: Input dimension Output dimension Activation function ; Fully Connected Layer 2 (Classification Output): Output Dimension (Corresponding to 10 types of faults), activation function Output the probability of various faults. The formula is: ,in, For the 2nd pair of fully connected layers Linear output for fault types, Indicates that the sample belongs to the first The probability of such failures (e.g.) (This indicates a 92% probability of bearing wear). Fully Connected Layer 3 (Early Warning Time Prediction Output): Output Dimension The activation function is Linear, and the fault warning time is output. (Unit: h), the formula is: in, The output features of LSTM layer 2 This is the weight matrix. For bias terms (such as output) (This indicates that a failure is expected to occur in 96 hours). Model training process: The Adam optimizer is used for dual-objective training, employing both classification and regression loss. The steps are as follows: Dataset partitioning: 50,000 preprocessed time series data sets The samples were divided into a training set (35,000 sets), a validation set (10,000 sets), and a test set (5,000 sets) in a 7:2:1 ratio. Initialize network parameters: Use He normal distribution to initialize the weights of CNN and fully connected layers, and use orthogonal initialization for LSTM layer weights; Iterative training (100 epochs in total): Input the training set samples into the CNN-LSTM network, and output the fault probability. With warning time Classification loss Cross-entropy loss is used to calculate the relationship between the predicted fault type and the true label. The difference is expressed by the formula: .in, Set the batch size (to 32). for The first sample The true label of the type of fault, For the corresponding predicted probability; regression loss The mean squared error (MSE) is used to calculate the predicted warning time and the actual failure time. The difference is expressed by the formula: Total loss The weighted summation (classification loss weight 0.6, regression loss weight 0.4, because fault type identification has higher priority) is calculated using the following formula: ; Backpropagation and parameter update: Minimizing the total loss using gradient descent (Adam optimizer) Reverse update the weights and biases of CNN, LSTM, and fully connected layers; Validation and early stopping: After each epoch, the loss and prediction accuracy are calculated using the validation set. If the loss on the validation set does not decrease for 10 consecutive epochs, training is stopped and the optimal model is saved. Model Evaluation: The model performance was evaluated using a test set, requiring a fault prediction accuracy of ≥96% and an early warning time error of ≤±6h. ; C. Construction of Equipment Health Assessment System: Determining the weights of health assessment indicators: The weights of the four monitoring indicators were determined using the Analytic Hierarchy Process (AHP). The steps are as follows: Construct a hierarchical structure: The target layer is "device health". The criteria layer consists of four types of monitoring indicators (vibration). ,temperature Oil Current The scheme layer consists of the specific monitoring values ​​of each indicator; a judgment matrix is ​​constructed: five VPSA equipment experts are invited to compare the criteria layer indicators pairwise (using the 1-9 scaling method, e.g.) Compare (Important, scale 3), resulting in the judgment matrix. ; Weight Calculation and Consistency Test: Calculate the weights of each indicator using the eigenvalue method. ; Health Score: The fuzzy comprehensive evaluation method is used to calculate the equipment health index in three levels. (Values ​​range from 0 to 100, with higher scores indicating better health), the formula is: ;in, For the first Health scores of various monitoring indicators; Maintenance process triggered: When health index Below the preset threshold, the threshold is specifically divided into: Normal: Sub-health: Failure risk: When this happens, the system will automatically trigger the maintenance process.

[0008] Preferably, in the health score The calculation method is as follows: Vibration index health score Based on the peak vibration characteristics extracted by CNN, five health levels are defined: Normal: Peak value < 0.5 mm / s → Mild abnormality: 0.5-1.0 mm / s → Moderate abnormality: 1.0-1.5 mm / s → Severe abnormality: 1.5-2.0 mm / s → Fault risk: >2.0mm / s → ; Temperature index health score Based on the maximum deviation of the temperature field, i.e., the actual temperature minus the normal temperature, it is divided into 5 levels, specifically: Normal: Deviation < 2℃ → Mild abnormality: 2-5℃ → Moderate abnormality: 5-8℃ → Severe abnormality: 8-12℃ → Failure risk: >12℃ → ; Oil quality index health score Based on the total concentration of wear elements, it is divided into 5 levels, specifically: Normal: <10ppm → Mild abnormality: 10-20 ppm → Moderate abnormality: 20-30 ppm → Severe abnormality: 30-40 ppm → Failure risk: >40ppm → ; Current index health score Based on the current fluctuation range, calculated as (actual current - rated current) / rated current, it is divided into 5 levels: Normal: fluctuation < 5% → Mild abnormalities: 5%-10% → Moderate abnormality: 10%-15% → Severe abnormalities: 15%-20% → Failure risk: >20% → .

[0009] Preferably, the specific maintenance process triggered by the maintenance process is as follows: Sub-health state ( The edge decision layer generates "preventive maintenance recommendations," which include fault risk points (such as "slight abnormal compressor vibration, suspected early bearing wear") and recommended inspection cycles (such as "re-inspect every 24 hours"), and uploads them to the cloud platform. Fault risk status ( The edge decision layer automatically generates maintenance work orders. The work order information includes the fault location (e.g., "2# compressor"), suggested maintenance parts (e.g., "replace bearing"), required tools (e.g., "bearing disassembly kit"), and operation steps (e.g., "1. Shut down and disconnect power → 2. Remove end cover → 3. Replace bearing → 4. No-load test"). The work orders are dispatched to maintenance personnel through the cloud platform, triggering fault alarm push notifications.

[0010] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves a comprehensive breakthrough by deeply integrating a four-layer architecture of "full-domain perception, edge decision-making, precise execution, and cloud-edge collaboration," addressing the pain points of traditional VPSA biogas decarbonization equipment and yielding significant beneficial effects.

[0011] In terms of operational efficiency and energy consumption optimization, relying on the reinforcement learning algorithm of the ARM+FPGA heterogeneous architecture, 18 process parameters can be dynamically optimized within 1 second. Combined with the real-time data support of 48 high-precision sensors, the purification efficiency is improved, the methane purity is stabilized at over 98%, the energy consumption per unit of biogas treatment is reduced, the adsorbent service life is extended, and the economic benefits and resource utilization of the equipment are greatly improved.

[0012] In terms of fault early warning and maintenance cost control, the LSTM-CNN composite model can provide early warnings 72-120 hours in advance, with a fault prediction accuracy of ≥96%. Combined with a quantitative health assessment system, it can reduce the rate of sudden equipment downtime and reduce maintenance costs.

[0013] In terms of system collaboration and management efficiency, the Petri net-based linkage control matrix enables millisecond-level device collaboration. Faced with 20% traffic fluctuations or device failures, the response and switching times are <500ms and <100ms respectively, ensuring system stability. The combination of digital twins and cloud platforms not only supports virtual debugging, reducing on-site debugging time by 50%, but also promotes the transformation of management models from "experience-driven" to "data-driven", improving the efficiency of management personnel's inspections and enabling remote control to cover the entire process.

[0014] In addition, the system has high security and scalability, encrypted communication and dual data backup to ensure data security, and reserved interfaces to connect to the entire biogas treatment production line, providing a scalable technical solution for the intelligent upgrading of industries such as biomass energy utilization and environmental protection engineering. Attached Figure Description

[0015] Figure 1This is a schematic diagram of the overall system structure of the present invention; Figure 2 This is a schematic diagram of the global perception and data acquisition layer structure of the present invention; Figure 3 This is a schematic diagram of the edge intelligent decision-making and control layer structure of the present invention; Figure 4 This is a schematic diagram of the precise execution and linkage control layer structure of the present invention; Figure 5 This is a schematic diagram of the cloud-edge collaborative management and digital twin layer structure of the present invention. Detailed Implementation

[0016] 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.

[0017] Please see Figure 1-5 This invention provides a technical solution: a novel intelligent control system for VPSA biogas decarbonization and purification equipment, comprising a global perception and data acquisition layer, an edge intelligent decision-making and control layer, a precise execution and linkage control layer, and a cloud-edge collaborative management and digital twin layer; the global perception and data acquisition layer includes a multi-parameter sensor network and a data acquisition and preprocessing module; Multi-parameter sensor network: A distributed monitoring network of 48 sensor nodes is constructed for the core functional modules of the VPSA biogas decarbonization and purification system to achieve comprehensive perception of equipment status and process parameters. Adsorption Tower Monitoring: The system incorporates a distributed fiber optic temperature sensor and a miniature gas concentration sensor. The distributed fiber optic temperature sensor utilizes distributed temperature measurement technology, achieving a measurement accuracy of ±0.1℃. It can capture the temperature field distribution at different heights and in different areas within the adsorption tower in real time, reflecting the intensity of the adsorption reaction and the activity state of the adsorbent. The miniature gas concentration sensor employs infrared detection principles, achieving a detection accuracy of 0.05% for CH4 and 0.01% for CO2. It can accurately monitor the purity of methane and the residual amount of carbon dioxide within the tower, providing a basis for adjusting the adsorption time.

[0018] Compressor and Vacuum Pump Monitoring: Both types of equipment are equipped with vibration acceleration sensors, oil spectrometers, and current and voltage sensors. The vibration acceleration sensors have a frequency response range of 0.1-10kHz and can collect vibration signals during equipment operation. By analyzing the vibration spectrum, mechanical faults such as bearing wear and rotor imbalance can be identified. The oil spectrometer uses inductively coupled plasma atomic emission spectrometry (ICP-AES) technology, supporting quantitative detection of 12 wear elements (such as iron, copper, and chromium) with a detection limit down to the ppm level. This reflects the wear degree of internal components and the lubrication system status. The current and voltage sensors utilize the Hall effect principle, with a measurement accuracy of ±0.2%, monitoring the equipment's input current, voltage, and power in real time to determine if there are electrical faults such as overload or phase loss.

[0019] Pretreatment Unit Monitoring: Pressure sensors, flow sensors, level sensors, and an online gas analyzer are deployed. Pressure sensors monitor the inlet and outlet pressures of the pretreatment unit to prevent excessive pressure from damaging the equipment or excessive pressure from affecting subsequent processes. The flow sensor uses a turbine flow meter with an accuracy of ±0.5%, monitoring the biogas processing flow rate in real time and providing a reference for system load adjustment. The level sensor monitors the levels of desulfurization and dehydration liquids in the pretreatment unit to prevent substandard treatment due to low levels or overflow due to high levels. The online gas analyzer uses gas chromatography to detect the content of components such as H2S and CO2 in the pretreated biogas in real time, ensuring that the biogas entering the VPSA system meets process requirements.

[0020] Data acquisition and preprocessing module: Employs a high-precision data acquisition card with a sampling frequency of 10kHz to ensure rapid capture of transient data (such as current fluctuations during equipment start-up and shutdown, and pressure changes during adsorption tower switching); after the acquired data is transmitted to the edge computing node, the following preprocessing operations are performed sequentially: Data filtering: Kalman filtering algorithm is used to eliminate sensor noise (such as environmental interference in vibration signals) and improve data accuracy; Outlier removal: Based on the 3σ criterion, abnormal data caused by sensor failure or transmission interference is identified and removed to avoid erroneous data affecting decision-making; Data normalization: Transform monitoring data of different dimensions (such as temperature, pressure, concentration) into the [0,1] interval to provide standardized data for subsequent algorithm modeling; The edge intelligent decision-making and control layer includes an edge computing server, a fault prediction and health management model, and an intelligent controller. The edge computing server employs a heterogeneous computing architecture combining an ARM processor and an FPGA acceleration chip, integrating a reinforcement learning-based dynamic optimization algorithm to optimize the operating efficiency, energy consumption, and adsorbent lifespan of the VPSA equipment. The fault prediction and health management model is an LSTM-CNN composite model that analyzes time-series data collected by the full-domain perception and data acquisition layer to provide early warnings of faults and establish an equipment health assessment system. The intelligent controller adopts a hybrid PLC and IPC control architecture. The PLC is responsible for real-time execution of digital and analog control (such as valve switching, motor start / stop, and speed regulation), while the IPC handles complex logic operations and parameter settings. The controller has a built-in multi-protocol communication interface, supporting industrial bus protocols such as Modbus TCP, Profinet, and EtherCAT. The protocol's communication cycle is ≤1ms, ensuring real-time transmission of control commands. The controller receives optimized parameters and control commands generated by the edge computing server, converts them into signals recognizable by the actuators (such as 4-20mA analog signals and switching signals), and achieves high-precision closed-loop control of the adsorption tower switching valve group, compressor frequency converter, vacuum pump speed control device, and pretreatment unit valves, with a control accuracy of ±0.5%. The specific implementation steps for optimizing the operating efficiency, energy consumption, and adsorbent lifespan of VPSA equipment based on a reinforcement learning-based dynamic optimization algorithm are as follows: Step 1: Data Preprocessing and Objective Function Definition First, at the overall perception and data acquisition layer, ≥200,000 sets of historical operating data for VPSA equipment are collected. The data dimensions cover two categories: status parameters and process parameters. It contains 28 real-time monitoring data points, all collected by sensors in the overall perception and data acquisition layer, defined as... ,in This refers to the CO2 concentration in biogas. This refers to the CH4 concentration in biogas. Biogas flow rate; The average temperature inside the adsorption tower; This refers to the vibration amplitude of the compressor. Other preset status parameters to be collected; For the set of state parameters Normalization is performed using the Min-Max standardization formula: ,in For the first The normalized value of the term parameter, , These are the minimum and maximum values ​​of the parameter in historical data, respectively. Construction of multi-objective optimization functions: Improving equipment operating efficiency Energy consumption indicators Adsorbent lifespan To optimize the objective, a total reward function is constructed by weighted summation. The formula is: Among them, equipment operating efficiency The methane purification efficiency, i.e., the ratio of the purified CH4 purity to the CH4 purity of the feed gas, is calculated in real time by the gas concentration sensor in the adsorption tower, with weights... Energy consumption indicators Energy consumption per unit of biogas processing, calculated by compressor / vacuum pump current and voltage sensors, weighted. Adsorbent lifespan This represents the amount of adsorbent activity decay within a single process cycle, and is dimensionless, ranging from [0,1]. It is derived from changes in adsorption tower temperature and CO2 concentration, with weights... ; Process parameter action space definition: Action space A is constructed based on 18 preset optimized key process parameters, and is defined as follows: ,in Adsorption time; The number of equalization cycles. Vacuum desorption pressure; This refers to the compressor speed; Other preset parameters; Step 2, DQN model structure design: The DQN model consists of a three-layer neural network: an input layer, a hidden layer, and an output layer. The structure is as follows: Input layer: 28 neurons, corresponding to the normalized set of state parameters. Dimension, reception As model input; Hidden layers: Two fully connected layers, the first with 64 neurons and the second with 32 neurons, the activation function is... This is used to extract nonlinear features of state parameters; the output layer has 18 neurons, corresponding to the dimension of the action space A, and outputs the "value" of each process parameter. value" —Indicates the current state Select action The long-term cumulative rewards are obtained, and the output layer activation function uses Linear; Model training process: Initialize the experience replay pool The capacity is set to 100,000 records to store historical training samples. Sample format: ,in For time steps, for Current state for Actions performed at all times for Rewards earned at any time for Next state at time 1; Initialize the main network With the target network The main network is used for real-time computing. Value, the target network is used to calculate the target Values, initial values ​​of both parameters and same; Iterative training: from the experience replay pool The current state can be obtained by randomly sampling a sample of size 32 from the data, or by obtaining the current state from real-time data. One of them, through - Greedy strategy for choosing actions : Execute action This means adjusting the corresponding process parameters and collecting data through the global sensing and data acquisition layer. state of time Substitute into the reward function to calculate ;Will Store in the experience replay pool ; Target value The time-difference formula is used for calculation, where let The discount factor is set to 0.9, representing the weight of future rewards. Network prediction Value Mean squared error is used as the loss function. : ,in , The first The target of each sample Values, states, and actions; learning rate through the Adam optimizer. Minimize loss function Backpropagation updates the main network parameters When the loss function The value is less than 0.01 for 1000 consecutive steps, and the average reward on the test set is... Once the value stabilizes above 0.8, the model training is complete, and the final model is deployed to the edge computing server. Step 3: Real-time parameter optimization: The edge computing server receives real-time data from the global perception and data acquisition layer through the intelligent controller, and updates the state parameter set every 100ms. Then, normalization is performed according to the Min-Max formula in step 1 to obtain... ;Will Input the trained DQN main network Output 18 process parameters corresponding to value ,choose The combination of actions with the highest value is taken as the optimal process parameter. The solution formula is: ,in , For the first The range of values ​​for each process parameter; 1. By analyzing the time-series data collected by the comprehensive perception and data acquisition layer through the aforementioned fault prediction and health management model, early warning of faults is provided, and an equipment health assessment system is established. The specific implementation logic is as follows: A. Data Acquisition and Preprocessing: For the pre-defined core fault types of VPSA equipment (10 categories including bearing wear, valve seal failure, adsorbent poisoning, and motor phase loss), two types of key data are collected: (1) Time-series monitoring dataset It includes four types of core monitoring data, all of which are time-series data (sampling frequency 10kHz, synchronously acquired by a data acquisition card), defined as: Vibration signal time-series data (unit: mm / s), from a vibration acceleration sensor (frequency response 0.1-10kHz) of a compressor / vacuum pump. Each sample is a 10-second vibration sequence, with dimensions of [missing information]. ( (i.e., 10kHz × 10s). Temperature field time series data (unit: °C), from a distributed fiber optic temperature sensor (accuracy ±0.1 °C) of the adsorption tower. Each sample is a 5-minute temperature sequence. ( That is, 1 time / second × 300 seconds; 6 refers to the 6 monitoring points of the fiber optic sensor. Oil wear elemental time-series data (unit: ppm), from an oil spectrometer for compressors / vacuum pumps (supports detection of 12 elements). Each sample is a 1-hour elemental concentration sequence. ( (i.e., 1 time / hour; 12 refers to 12 wear elements). Current and voltage time-series data (unit: A / V), from current and voltage sensors of compressors / vacuum pumps (accuracy ±0.2%), each sample is a 1-minute electrical parameter sequence, with dimensions of [missing information]. ( That is, 1 time / second × 60 seconds; 2 represents the two parameters of current and voltage. The overall dimension of the time series monitoring dataset is ,in F represents the sample size (50,000 samples were collected in this invention, including 12,000 fault samples), and F represents the feature dimension. .

[0021] (2) Fault label set For each time series sample, the fault type and fault occurrence time are labeled. The fault type is represented by one-hot encoding, defined as: ,in, , This indicates that the sample corresponds to the k-th type of fault (e.g. Due to bearing wear, (Valve seal failure) Indicates that the fault does not fall into this category; also indicates the time the fault occurred. (Unit: h), used for subsequent early warning time calculation (e.g., if there are 96 hours remaining before the fault occurs when the sample is collected, then...). ); B. Construction and training of CNN-LSTM composite network: Network architecture design: CNN module: Extracting spatial features: for time-series data To analyze local features (such as peak values ​​of vibration signals and hotspot regions in the temperature field), a 1D-CNN (one-dimensional convolutional) module is designed, with the following structure: Convolutional layer 1: Input dimension ( ), number of kernels 32, kernel size 3, stride 1, activation function Output dimension Pooling layer 1: Max pooling, pooling kernel size 2, stride 2, output dimension Convolutional layer 2: 64 kernels, kernel size 3, stride 1, activation function... Output dimension Pooling layer 2: Max pooling, pooling kernel size 2, stride 2, output dimension Flattening layer: Transforms the two-dimensional features output by the pooling layer into a one-dimensional feature vector, with dimensions of... (like (At that time, the dimension is 1,600,000). LSTM module: Capturing time series trends: First, process the time series data... Preprocessing is performed on the preprocessed time series data. The LSTM module is designed with the following structure: LSTM layer 1: Input dimension is the same as the feature vector dimension output by the CNN module (e.g., 1,600,000), number of hidden units is 128, activation function is tanh, and output dimension is... LSTM layer 2: 64 hidden units, tanh activation function, output dimension... Dropout layer: Dropout probability 0.5 to prevent model overfitting, output dimension The LSTM module captures timing trends through a gating mechanism. Fully Connected Layer: Fault Classification and Early Warning Fully connected layer 1: Input dimension Output dimension Activation function ; Fully Connected Layer 2 (Classification Output): Output Dimension (Corresponding to 10 types of faults), activation function Output the probability of various faults. The formula is: ,in, For the 2nd pair of fully connected layers Linear output for fault types, Indicates that the sample belongs to the first The probability of such failures (e.g.) (This indicates a 92% probability of bearing wear). Fully Connected Layer 3 (Early Warning Time Prediction Output): Output Dimension The activation function is Linear, and the fault warning time is output. (Unit: h), the formula is: in, The output features of LSTM layer 2 This is the weight matrix. For bias terms (such as output) (This indicates that a failure is expected to occur in 96 hours). Model training process: The Adam optimizer is used for dual-objective training, employing both classification and regression loss. The steps are as follows: Dataset partitioning: 50,000 preprocessed time series data sets The samples were divided into a training set (35,000 sets), a validation set (10,000 sets), and a test set (5,000 sets) in a 7:2:1 ratio. Initialize network parameters: Use He normal distribution to initialize the weights of CNN and fully connected layers, and use orthogonal initialization for LSTM layer weights; Iterative training (100 epochs in total): Input the training set samples into the CNN-LSTM network, and output the fault probability. With warning time Classification loss Cross-entropy loss is used to calculate the relationship between the predicted fault type and the true label. The difference is expressed by the formula: .in, Set the batch size (to 32). for The first sample The true label of the type of fault, For the corresponding predicted probability; regression loss The mean squared error (MSE) is used to calculate the predicted warning time and the actual failure time. The difference is expressed by the formula: Total loss The weighted summation (classification loss weight 0.6, regression loss weight 0.4, because fault type identification has higher priority) is calculated using the following formula: ; Backpropagation and parameter update: Minimizing the total loss using gradient descent (Adam optimizer) Reverse update the weights and biases of CNN, LSTM, and fully connected layers; Validation and early stopping: After each epoch, the loss and prediction accuracy are calculated using the validation set. If the loss on the validation set does not decrease for 10 consecutive epochs, training is stopped and the optimal model is saved. Model Evaluation: The model performance was evaluated using a test set, requiring a fault prediction accuracy of ≥96% and an early warning time error of ≤±6h. ; C. Construction of Equipment Health Assessment System: Determining the weights of health assessment indicators: The weights of the four monitoring indicators were determined using the Analytic Hierarchy Process (AHP). The steps are as follows: Construct a hierarchical structure: The target layer is "device health". The criteria layer consists of four types of monitoring indicators (vibration). ,temperature Oil Current The scheme layer consists of the specific monitoring values ​​of each indicator; a judgment matrix is ​​constructed: five VPSA equipment experts are invited to compare the criteria layer indicators pairwise (using the 1-9 scaling method, e.g.) Compare (Important, scale 3), resulting in the judgment matrix. ; Weight Calculation and Consistency Test: Calculate the weights of each indicator using the eigenvalue method. ; Health Score: The fuzzy comprehensive evaluation method is used to calculate the equipment health index in three levels. (Values ​​range from 0 to 100, with higher scores indicating better health), the formula is: ;in, For the first Health scores of various monitoring indicators; The calculation method is as follows: Vibration index health score Based on the peak vibration characteristics extracted by CNN, five health levels are defined: Normal: Peak value < 0.5 mm / s → Mild abnormality: 0.5-1.0 mm / s → Moderate abnormality: 1.0-1.5 mm / s → Severe abnormality: 1.5-2.0 mm / s → Fault risk: >2.0mm / s → ; Temperature index health score Based on the maximum deviation of the temperature field, i.e., the actual temperature minus the normal temperature, it is divided into 5 levels, specifically: Normal: Deviation < 2℃ → Mild abnormality: 2-5℃ → Moderate abnormality: 5-8℃ → Severe abnormality: 8-12℃ → Failure risk: >12℃ → ; Oil quality index health score Based on the total concentration of wear elements, it is divided into 5 levels, specifically: Normal: <10ppm → Mild abnormality: 10-20 ppm → Moderate abnormality: 20-30 ppm → Severe abnormality: 30-40 ppm → Failure risk: >40ppm → ; Current index health score Based on the current fluctuation range, calculated as (actual current - rated current) / rated current, it is divided into 5 levels: Normal: fluctuation < 5% → Mild abnormalities: 5%-10% → Moderate abnormality: 10%-15% → Severe abnormalities: 15%-20% → Failure risk: >20% → Maintenance process triggered: When health index Below the preset threshold, the threshold is specifically divided into: Normal: Sub-health: Failure risk: When this happens, the system will automatically trigger the maintenance process. The specific maintenance process is as follows: Sub-health state ( The edge decision layer generates "preventive maintenance recommendations," which include fault risk points (such as "slight abnormal compressor vibration, suspected early bearing wear") and recommended inspection cycles (such as "re-inspect every 24 hours"), and uploads them to the cloud platform. Fault risk status ( The edge decision layer automatically generates maintenance work orders. Work order information includes the fault location (e.g., "compressor #2"), recommended maintenance components (e.g., "replace bearing"), required tools (e.g., "bearing removal kit"), and operation steps (e.g., "1. Shut down and disconnect power → 2. Remove end cover → 3. Replace bearing → 4. No-load test"). These work orders are dispatched to maintenance personnel via the cloud platform, triggering fault alarm push notifications. The precise execution and linkage control layer includes high-speed actuators and an equipment linkage control matrix; high-speed actuators: employing high-precision, fast-response execution components to ensure the precise implementation of control commands. Adsorption tower switching valve: It adopts an electro-hydraulic servo drive method with a drive voltage of 24V and a response time of <300ms, which can quickly complete the opening and closing of the valve and the adjustment of the opening degree; it is equipped with an electromagnetic flow meter with an accuracy of ±0.3% to provide real-time feedback on gas flow, forming a closed-loop flow control to ensure the stability of the gas inlet and outlet of the adsorption tower. Compressor and vacuum pump: Equipped with servo motor driver, adopting vector control technology, supporting 0.01Hz level speed adjustment, with speed control accuracy of ±0.1%; the driver has overload protection, overvoltage protection, and overcurrent protection functions. When the equipment malfunctions, it will automatically stop and feed back the fault signal to the edge computing server. Equipment Interlocking Control Matrix: Based on existing Petri nets, an interlocking control model is constructed, defining the logical relationships and triggering conditions between devices to achieve millisecond-level coordinated response. Normal operating condition linkage: When the system detects a sudden increase of 20% in biogas flow, the linkage model immediately triggers the linkage process of "starting up the standby adsorption tower → increasing the compressor speed by 15%-20% → increasing the opening of the pretreatment unit flow valve by 20%-25%". The entire process response time is <500ms, ensuring that the system load quickly matches the flow change and avoiding adsorption tower overload. Fault condition linkage: When the fault prediction model detects that a compressor has a risk of bearing wear, the linkage model activates the backup compressor in advance to complete preheating and parameter calibration; when the faulty compressor stops, it immediately switches to the backup compressor, and at the same time adjusts the adsorption tower adsorption time and vacuum pump speed to ensure stable system operating parameters and downtime <100ms, minimizing production interruption. The cloud-edge collaborative management and digital twin layer includes a cloud platform and a digital twin system. The cloud platform receives data from the edge intelligent decision-making and control layer and generates process optimization reports and decision information. It supports remote monitoring and parameter adjustment. Remote monitoring and control: Managers can log in to the cloud platform through a computer client to view the real-time operating status of the equipment (such as data from various sensors, equipment health index, and purification efficiency). They can remotely issue parameter adjustment commands (which require authorization verification). The commands are transmitted to the edge computing server through encrypted communication to ensure operational security.

[0022] The digital twin system, based on 3D modeling and real-time data-driven operation, enables physical state mapping, operational condition simulation, virtual debugging, and fault diagnosis of equipment. Specifically, it utilizes the existing Unity3D engine to construct a 3D digital model of the equipment, combined with real-time data-driven technology, to achieve digital management of the entire equipment lifecycle. Physical state mapping: The digital twin model receives real-time data such as temperature, pressure, and vibration transmitted from the edge intelligent decision-making and control layer. The operating status of the equipment is displayed intuitively through color changes (e.g., red for areas with excessively high temperatures) and animation effects (e.g., valve opening and closing actions). The mapping latency is <100ms. Operating condition simulation and optimization: Supports inputting different operating condition parameters (such as biogas CO2 concentration 30%-60%, flow rate 50-150m³ / h, ambient temperature -10℃-40℃) to simulate the equipment's operating status under extreme conditions and output parameter adjustment schemes; for example, simulating the equipment's operating status in low-temperature winter environments, it outputs an optimization scheme of "extending the preheating time by 10 minutes and increasing the temperature of the adsorbent insulation layer by 5℃", providing pre-validation for on-site process adjustments; Virtual commissioning and troubleshooting: During the equipment installation and commissioning phase, virtual commissioning can be performed through a digital twin system. Different parameter combinations can be set to verify the stability of equipment operation and reduce on-site commissioning time by more than 50%. When equipment malfunctions, technicians can reproduce and troubleshoot the fault in the digital twin model, simulate the operating status after component replacement, and improve maintenance efficiency.

[0023] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A novel intelligent control system for VPSA biogas decarbonization and purification equipment, characterized in that, include: The system comprises a global perception and data acquisition layer, an edge intelligent decision-making and control layer, a precise execution and linkage control layer, and a cloud-edge collaborative management and digital twin layer; the global perception and data acquisition layer includes a multi-parameter sensor network and a data acquisition and preprocessing module. The multi-parameter sensor network is set up with 48 sensor nodes for the adsorption tower, compressor, vacuum pump and pretreatment unit of the VPSA biogas decarbonization and purification system. The data acquisition and preprocessing module uses a high-precision data acquisition card with a sampling frequency of 10kHz to obtain the data collected in the multi-parameter sensor network. The collected data is combined with edge computing nodes to complete data filtering, outlier removal and normalization to obtain a standardized dataset. The edge intelligent decision-making and control layer includes an edge computing server, a fault prediction and health management model, and an intelligent controller. The edge computing server adopts a heterogeneous computing architecture of ARM processor and FPGA acceleration chip, and integrates a dynamic optimization algorithm based on reinforcement learning to optimize the operating efficiency, energy consumption index, and adsorbent lifespan of VPSA equipment. The fault prediction and health management model is an LSTM-CNN composite model that analyzes time-series data collected by the global perception and data acquisition layer to provide early warning of faults and establish an equipment health assessment system. The intelligent controller adopts a hybrid control architecture of PLC and IPC, and has built-in Modbus TCP, Profinet, and EtherCAT protocol interfaces. The precise execution and linkage control layer includes a high-speed actuator and an equipment linkage control matrix. In the high-speed actuator, the adsorption tower switching valve is driven by an electro-hydraulic servo and equipped with a flow meter with an accuracy of ±0.3%. The compressor and vacuum pump are equipped with servo motor drivers, supporting speed adjustment at 0.01Hz levels. The equipment linkage control matrix is ​​built on a Petri net. When the biogas flow rate suddenly increases by 20%, it triggers a linkage process of "standby adsorption tower startup - compressor speed increase - pretreatment flow increase." In case of equipment failure, it activates an emergency mechanism of "faulty equipment shutdown - standby equipment switching - dynamic adjustment of operating parameters." The cloud-edge collaborative management and digital twin layer includes a cloud platform and a digital twin system; the cloud platform receives data from the edge intelligent decision and control layer and generates process optimization reports and decision information, supporting remote monitoring and parameter adjustment; the digital twin system, based on 3D modeling and real-time data-driven operation, realizes equipment physical state mapping, operating condition simulation, virtual debugging, and fault diagnosis.

2. The intelligent control system for a novel VPSA biogas decarbonization and purification equipment according to claim 1, characterized in that: The specific implementation steps of the reinforcement learning-based dynamic optimization algorithm for optimizing the operating efficiency, energy consumption, and adsorbent lifespan of VPSA equipment are as follows: Step 1: Data Preprocessing and Objective Function Definition First, at the overall perception and data acquisition layer, ≥200,000 sets of historical operating data for VPSA equipment are collected. The data dimensions cover two categories: status parameters and process parameters. It contains 28 real-time monitoring data points, all collected by sensors in the overall perception and data acquisition layer, defined as... ,in This refers to the CO2 concentration in biogas. This refers to the CH4 concentration in biogas. Biogas flow rate; The average temperature inside the adsorption tower; This refers to the vibration amplitude of the compressor. Other preset status parameters to be collected; For the set of state parameters Normalization is performed using the Min-Max standardization formula: ,in For the first The normalized value of the term parameter, , These are the minimum and maximum values ​​of the parameter in historical data, respectively. Construction of multi-objective optimization functions: Improving equipment operating efficiency Energy consumption indicators Adsorbent lifespan To optimize the objective, a total reward function is constructed by weighted summation. The formula is: Among them, equipment operating efficiency The methane purification efficiency, i.e., the ratio of the purified CH4 purity to the CH4 purity of the feed gas, is calculated in real time by the gas concentration sensor in the adsorption tower, with weights... Energy consumption indicators Energy consumption per unit of biogas processing, calculated by compressor / vacuum pump current and voltage sensors, weighted. Adsorbent lifespan This represents the amount of adsorbent activity decay within a single process cycle, and is dimensionless, ranging from [0,1]. It is derived from changes in adsorption tower temperature and CO2 concentration, with weights... ; Process parameter action space definition: Action space A is constructed based on 18 preset optimized key process parameters, and is defined as follows: ,in Adsorption time; The number of equalization cycles. Vacuum desorption pressure; This refers to the compressor speed; Other preset parameters; Step 2, DQN model structure design: The DQN model consists of a three-layer neural network: an input layer, a hidden layer, and an output layer. The structure is as follows: Input layer: 28 neurons, corresponding to the normalized set of state parameters. Dimension, reception As input to the model; Hidden layers: Two fully connected layers are used, with 64 neurons in the first layer and 32 neurons in the second layer. The activation function is... This is used to extract the nonlinear features of state parameters; Output layer: Number of neurons = 18, corresponding to the dimension of action space A, outputting the value of each process parameter. value" —Indicates the current state Select action The long-term cumulative rewards are obtained, and the output layer activation function uses Linear; Model training process: Initialize the experience replay pool The capacity is set to 100,000 records to store historical training samples. Sample format: ,in For time steps, for Current state for Actions performed at all times for Rewards earned at any time for Next state at time 1; Initialize the main network With the target network The main network is used for real-time computing. Value, the target network is used to calculate the target Values, initial values ​​of both parameters and same; Iterative training: from the experience replay pool The current state can be obtained by randomly sampling a sample of size 32 from the data, or by obtaining the current state from real-time data. One of them, through - Greedy strategy for choosing actions : Execute action This means adjusting the corresponding process parameters and collecting data through the global sensing and data acquisition layer. state of time Substitute into the reward function to calculate ;Will Store in the experience replay pool ; Target value The time-difference formula is used for calculation, where let The discount factor is set to 0.9, representing the weight of future rewards. Network prediction Value Mean squared error is used as the loss function. : ,in , The first The target of each sample Values, states, and actions; learning rate through the Adam optimizer. Minimize loss function Backpropagation updates the main network parameters When the loss function For 1000 consecutive steps, the value is less than 0.01, and the average reward is... Once the value stabilizes above 0.8, the model training is complete, and the final model is deployed to the edge computing server. Step 3: Real-time parameter optimization: The edge computing server receives real-time data from the global perception and data acquisition layer through the intelligent controller, and updates the state parameter set every 100ms. Then, normalization is performed according to the Min-Max formula in step 1 to obtain... ;Will Input the trained DQN main network Output 18 process parameters corresponding to value ,choose The combination of actions with the highest value is taken as the optimal process parameter. The solution formula is: ,in , For the first The range of values ​​for each process parameter.

3. The intelligent control system for a novel VPSA biogas decarbonization and purification equipment according to claim 1, characterized in that: By analyzing the time-series data collected by the comprehensive perception and data acquisition layer through the fault prediction and health management model, early warning of faults is provided, and an equipment health assessment system is established. The specific implementation logic is as follows: A. Data Acquisition and Preprocessing: Based on the preset core fault types of VPSA equipment, two types of key data are collected: (1) Time-series monitoring dataset It includes four types of core monitoring data, all of which are time-series data, defined as follows: Vibration signal time-series data, from vibration acceleration sensors of compressors / vacuum pumps. Each sample is a 10-second vibration sequence, with dimensions of [missing information]. ; Temperature field time series data, from distributed fiber optic temperature sensors in the adsorption tower. Each sample is a 5-minute temperature sequence, with dimensions of [dimension missing]. ; : Time-series data of elemental composition in oil wear, from an oil spectrometer for compressors / vacuum pumps. Each sample is a 1-hour sequence of elemental concentrations. ; Current and voltage time-series data, from current and voltage sensors of compressors / vacuum pumps. Each sample is a 1-minute sequence of electrical parameters, with dimensions of [missing information]. ; The overall dimension of the time series monitoring dataset is ,in Where is the number of samples, and F is the feature dimension. (2) Fault label set For each time series sample, the fault type and fault occurrence time are labeled. The fault type is represented by one-hot encoding, defined as: ,in, , This indicates that the sample corresponds to the k-th type of fault. Indicates that the fault does not fall into this category; also indicates the time the fault occurred. (Unit: h), used for subsequent early warning time calculation (e.g., if there are 96 hours remaining before the fault occurs when the sample is collected, then...). ); B. Construction and training of CNN-LSTM composite network: Network architecture design: CNN module: Extracting spatial features: for time-series data Based on local features, a 1D-CNN module is designed with the following structure: Convolutional layer 1: Input dimension Number of convolution kernels: 32; kernel size: 3; stride: 1; activation function: Output dimension Pooling layer 1: Max pooling, pooling kernel size 2, stride 2, output dimension Convolutional layer 2: 64 kernels, kernel size 3, stride 1, activation function... Output dimension Pooling layer 2: Max pooling, pooling kernel size 2, stride 2, output dimension Flattening layer: Transforms the two-dimensional features output by the pooling layer into a one-dimensional feature vector, with dimensions of... ; LSTM module: Capturing time series trends: First, process the time series data... Preprocessing is performed on the preprocessed time series data. The LSTM module is designed with the following structure: LSTM layer 1: Input dimension is the same as the feature vector output from the CNN module, number of hidden units is 128, activation function is tanh, and output dimension is... LSTM layer 2: 64 hidden units, tanh activation function, output dimension... Dropout layer: Dropout probability 0.5 to prevent model overfitting, output dimension The LSTM module captures timing trends through a gating mechanism. Fully Connected Layer: Fault Classification and Early Warning Fully connected layer 1: Input dimension Output dimension Activation function ; Fully Connected Layer 2: Output Dimension Activation function Output the probability of various faults. The formula is: ,in, For the 2nd pair of fully connected layers Linear output for fault types, Indicates that the sample belongs to the first The probability of a type of failure; Fully Connected Layer 3: Output Dimension The activation function is Linear, and the fault warning time is output. The formula is: in, The output features of LSTM layer 2 This is the weight matrix. For bias terms; Model training process: The Adam optimizer is used for dual-objective training, employing both classification and regression loss. The steps are as follows: Dataset partitioning: 50,000 preprocessed time series data sets The samples were divided into training, validation, and test sets in a 7:2:1 ratio. Initialize network parameters: Use He normal distribution to initialize the weights of CNN and fully connected layers, and use orthogonal initialization for LSTM layer weights; Iterative training: Input the training set samples into the CNN-LSTM network, and output the fault probability. With warning time Classification loss Cross-entropy loss is used to calculate the relationship between the predicted fault type and the true label. The difference is expressed by the formula: .in, For batch size, for The first sample The true label of the type of fault, For the corresponding predicted probability; regression loss The mean squared error is used to calculate the predicted warning time and the actual failure time. The difference is expressed by the formula: Total loss The weighted summation formula is: ; Backpropagation and parameter update: Minimizing the total loss using gradient descent Reverse update the weights and biases of CNN, LSTM, and fully connected layers; Validation and early stopping: After each epoch, the loss and prediction accuracy are calculated using the validation set. If the loss on the validation set does not decrease for 10 consecutive epochs, training is stopped and the optimal model is saved. Model Evaluation: The model performance was evaluated using a test set, requiring a fault prediction accuracy of ≥96% and an early warning time error of ≤±6h. ; C. Construction of Equipment Health Assessment System: Determining the weights of health assessment indicators: The weights of the four monitoring indicators were determined using the analytic hierarchy process (AHP). The steps are as follows: Build a hierarchical structure: the target layer is "device health". The criteria layer consists of four categories of monitoring indicators, and the scheme layer consists of the specific monitoring values ​​of each indicator. A judgment matrix is ​​constructed by inviting five VPSA equipment experts to pairwise compare the criteria layer indicators to obtain the judgment matrix. ; Weight Calculation and Consistency Test: Calculate the weights of each indicator using the eigenvalue method. ; Health Score: The fuzzy comprehensive evaluation method is used to calculate the equipment health index in three levels. The formula is: ;in, For the first Health scores of various monitoring indicators; Maintenance process triggered: When health index Below the preset threshold, the threshold is specifically divided into: Normal: Sub-health: Failure risk: When this happens, the system will automatically trigger the maintenance process.

4. The intelligent control system for a novel VPSA biogas decarbonization and purification equipment according to claim 3, characterized in that: The health score The calculation method is as follows: Vibration index health score Based on the peak vibration characteristics extracted by CNN, five health levels are defined: Normal: Peak value < 0.5 mm / s → ; Mild abnormality: 0.5-1.0 mm / s → ; Moderate abnormality: 1.0-1.5 mm / s → ; Severe abnormality: 1.5-2.0 mm / s → Fault risk: >2.0mm / s → ; Temperature index health score Based on the maximum deviation of the temperature field, i.e., the actual temperature minus the normal temperature, it is divided into 5 levels, specifically: Normal: Deviation < 2℃ → ; Mild abnormality: 2-5℃ → Moderate abnormality: 5-8℃ → Severe abnormality: 8-12℃ → Failure risk: >12℃ → ; Oil quality index health score Based on the total concentration of wear elements, it is divided into 5 levels, specifically: Normal: <10ppm → Mild abnormality: 10-20 ppm → Moderate abnormality: 20-30 ppm → Severe abnormality: 30-40 ppm → Failure risk: >40ppm → ; Current index health score Based on the current fluctuation range, calculated as (actual current - rated current) / rated current, it is divided into 5 levels: Normal: fluctuation < 5% → Mild abnormalities: 5%-10% → Moderate abnormality: 10%-15% → Severe abnormalities: 15%-20% → Failure risk: >20% → .

5. The intelligent control system for a novel VPSA biogas decarbonization and purification equipment according to claim 3, characterized in that: The specific maintenance process triggered by the aforementioned maintenance process is as follows: Sub-health status: The edge decision layer generates "preventive maintenance suggestions", which include fault risk points and recommended inspection cycles, and uploads them to the cloud platform; Fault risk status: The edge decision layer automatically generates maintenance work orders. The work order information includes the fault location, recommended maintenance parts, required tools, and operation steps. The work orders are dispatched to maintenance personnel through the cloud platform, triggering fault alarm push notifications.

Images