Air compressor circulation control system based on LSTM deep learning network

Abstract

The invention relates to the technical field of industrial automatic control, in particular to an air compressor circulation control system based on an LSTM deep learning network, which comprises a data acquisition unit, an LSTM prediction and decision unit, an execution control unit and an upper monitoring and interaction unit. According to the method, through multi-source data fusion, attention mechanism feature screening, LSTM time sequence modeling and reinforcement learning optimization, accurate prediction of the operation state of the air compressor and generation of the optimal control strategy giving consideration to energy efficiency and stability are achieved, control safety is guaranteed by means of digital twin virtual verification, the execution dynamic state is optimized in combination with an MPC adaptive algorithm, and the optimal control strategy is obtained. A physical execution mechanism is precisely driven through a PLC or a frequency converter, it is ensured that a control strategy is reliable, full-process supervision and knowledge graph diagnosis are achieved through visual display, an intelligent maintenance scheme is generated, man-machine interaction supports flexible operation, data storage backtracking assists analysis and tracing, and the system is high in practicability. And the intelligent level, the control precision and the operation and maintenance efficiency of operation of the air compressor are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of industrial automatic control technology, specifically to an air compressor cyclic control system based on an LSTM deep learning network. Background Technology

[0002] An air compressor, also known as an air handling unit, is a mechanical device that draws in air at normal pressure, increases the gas pressure through mechanical compression, and outputs high-pressure air. Its core function is to convert the potential energy (pressure) of air into usable energy. It is widely used in industrial pneumatic tool driving, mechanical manufacturing gas transportation, and construction equipment air supply, and is one of the key power sources in industrial production. With the increasing demands for stable air supply, energy consumption control, and operational efficiency in industrial production, the performance of the air compressor control system has become a core factor affecting its operational efficiency. Existing technologies, such as traditional air compressor control systems, suffer from problems such as insufficient utilization of multi-source data, low accuracy in predicting operating conditions, difficulty in balancing energy efficiency and stability in control strategies, lack of safety verification during execution, and reliance on manual experience for inefficient operation and maintenance.

[0003] Based on this, the present invention provides an air compressor cyclic control system based on an LSTM deep learning network to solve the aforementioned technical problems. Summary of the Invention

[0004] The purpose of this invention is to provide a cyclic control system for air compressors based on an LSTM deep learning network. This invention achieves accurate prediction of the air compressor's operating status and generation of an optimal control strategy that balances energy efficiency and stability through multi-source data fusion, attention mechanism feature selection, LSTM time series modeling, and reinforcement learning optimization. It also ensures control safety by leveraging digital twin virtual verification, optimizes execution dynamics by combining MPC adaptive algorithms, and ensures reliable implementation of the control strategy by precisely driving physical actuators through PLC or frequency converters. Furthermore, it enables full-process monitoring through visualization, generates intelligent maintenance solutions through knowledge graph diagnosis, supports flexible operation through human-machine interaction, and facilitates analysis and traceability through data storage and backtracking. This significantly improves the intelligence level, control accuracy, and maintenance efficiency of air compressor operation.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a cyclic control system for an air compressor based on an LSTM deep learning network, comprising a data acquisition unit, an LSTM prediction and decision unit, an execution control unit, and a supervisory control and interaction unit, wherein:

[0007] The data acquisition unit is responsible for collecting key parameters such as pressure, temperature, and flow rate during the operation of the air compressor.

[0008] The LSTM prediction and decision unit is used to fuse multi-source heterogeneous time series data, introduce an attention mechanism to enhance the perception of key features, and build a long-term and short-term collaborative prediction framework based on the LSTM model. Under the premise of ensuring prediction accuracy and response speed, it predicts the changing trend of the air compressor's operating status and generates the optimal control strategy.

[0009] The execution control unit is used to perform virtual verification and risk assessment of the execution process based on the control strategy output by the LSTM prediction and decision unit and in conjunction with the digital twin system. After confirming safety, it adjusts the response dynamics through an adaptive control algorithm to drive the valves and motor actuators of the air compressor.

[0010] The upper-level monitoring and interaction unit is used to display the system's operating status, energy consumption analysis, and AI decision-making basis in real time through a visual interface. It integrates knowledge graphs to generate intelligent fault diagnosis and maintenance suggestions, and supports parameter setting, command issuance, and human-computer interaction functions.

[0011] The data acquisition unit includes a sensor module, a signal preprocessing module, and a data transmission module, wherein:

[0012] The sensor module is used to collect physical signals of pressure, temperature, and flow rate in real time during the operation of the air compressor.

[0013] The signal preprocessing module is used to filter, reduce noise, and convert the format of the raw acquired data.

[0014] The data transmission module is used to stably transmit the processed data to the LSTM prediction and decision unit.

[0015] The LSTM prediction and decision unit includes a data fusion module, an attention weighting module, an LSTM prediction module, and a policy optimization module, wherein:

[0016] The data fusion module is used to perform time alignment and feature standardization on multi-source heterogeneous time-series data.

[0017] The attention weighting module is used to dynamically allocate feature weights through an attention mechanism.

[0018] The LSTM prediction module: models temporal dependencies based on a long short-term memory network and outputs prediction results of state change trends;

[0019] The strategy optimization module is used to generate control strategies that balance energy efficiency and stability through reinforcement learning algorithms.

[0020] The attention weighting module dynamically allocates feature weights through an attention mechanism, as follows:

[0021] A1: Receives multi-source temporal feature vectors from the data fusion module;

[0022] A2: Calculate the relevance score of each feature at the current time step, and generate dynamic weights through normalization;

[0023] A3: The dynamic weights are weighted and fused with the original features to output an enhanced feature representation, thereby improving the LSTM model's ability to identify key operating states.

[0024] The LSTM prediction module models temporal dependencies based on a long short-term memory network and outputs prediction results of state change trends. The specific operation is as follows:

[0025] B1: Divide the input weighted feature sequence into time steps to obtain the input sequence x. t , t=1,2,…,T;

[0026] B2: Calculate the forget gate, input gate, cell state update value and output gate for each time step in sequence, and update the hidden state to capture long-term dependencies in the input sequence;

[0027] B3: Map the hidden state sequence [h1,h2,…,h] through a fully connected layer. T Output the predicted state change trend for future time steps.

[0028] The strategy optimization module generates a control strategy that balances energy efficiency and stability using a reinforcement learning algorithm. The specific operation is as follows:

[0029] C1: Define the state space S as a fusion vector of the air compressor's current pressure, temperature, flow rate, and historical operating characteristics, and the action space A as a combination of valve opening adjustment and motor speed setpoint.

[0030] C2: Constructing a multi-objective reward function:

[0031] r(s,a)=ω1·r energy (s,a)+ω2·r stability (s,a)

[0032] Where, r energy (s,a) represents the energy consumption optimization reward, r stability (s,a) represents the stability reward, and ω1 and ω2 represent the dynamic weights.

[0033] C3: The policy network π(a|s;θ) is iteratively optimized using a deep deterministic policy gradient algorithm. Samples (s,a,r,s′) are stored in an empirical replay pool, and the loss function L(θ)=E[(Q(s,;θ)] is minimized. v )-(r+γQ′(s′,a′;θ v′)))2];

[0034] C4: When the policy network converges to a preset threshold, output the optimal action that matches the current state, i.e., the specific parameter value of the control policy.

[0035] The execution control unit includes a digital twin verification module, an adaptive control module, and an execution drive module, wherein:

[0036] The digital twin verification module is used to simulate control strategies in a virtual model and assess execution risks.

[0037] The adaptive control module is used to dynamically adjust the valve opening or motor speed using an MPC algorithm to optimize the response speed.

[0038] The execution drive module is used to convert control signals into physical actions via a PLC or frequency converter.

[0039] The digital twin verification module simulates control strategies in a virtual model and assesses execution risks. The specific operations are as follows:

[0040] D1: Construct a digital twin model of the physical entity of the air compressor, which includes multi-domain mappings of geometric, physical, and behavioral dimensions;

[0041] D2: Input the control strategy parameters output by the LSTM prediction and decision unit, perform dynamic simulation in the virtual model with a time step of Δt = 0.1s, and generate pressure fluctuation curve, temperature change curve and energy consumption curve;

[0042] D3: Establish a risk assessment indicator system, including the overpressure risk value R. p =∫(P(t)-P max ) + dt, temperature exceedance risk value R T =∑I(T(t)>T max ), Actuator load risk value R L =max(α) t / α max ,n t / n max );

[0043] D4: When the overall risk value R = ω p R p +ω T R T +ω L R L If the risk threshold R0 is less than or equal to the risk threshold, the control strategy is deemed safe to execute; otherwise, a risk warning and parameter correction suggestions are output.

[0044] The upper-level monitoring and interaction unit includes a visualization module, a knowledge graph diagnostic module, a human-computer interaction module, and a data storage and backtracking module, wherein:

[0045] The visualization module is used to present operational data, energy consumption analysis, and AI decision-making processes in real time through the interface.

[0046] The knowledge graph diagnostic module generates diagnostic results and repair plans based on fault feature correlation analysis.

[0047] The human-computer interaction module is used to support operators in configuring parameters and issuing commands.

[0048] The data storage and backtracking module is used to record historical system operation data for analysis and tracing.

[0049] The knowledge graph diagnostic module generates diagnostic results and repair plans based on fault feature correlation analysis. The specific operation is as follows:

[0050] E1: Real-time acquisition of air compressor operating parameters and alarm signals, and extraction of fault feature vectors;

[0051] E2: Match predefined fault mode association rules in the knowledge graph;

[0052] E3: Calculate the propagation probability between faulty nodes using a graph neural network;

[0053] E4: Generate a solution that includes root cause identification, scope of impact, and repair steps;

[0054] E5: Compare the repair plan with the historical work order database and recommend the optimal handling strategy.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] This invention achieves accurate prediction of air compressor operating status and generation of optimal control strategies that balance energy efficiency and stability through multi-source data fusion, attention mechanism feature selection, LSTM time series modeling, and reinforcement learning optimization. It also ensures control safety through digital twin virtual verification, optimizes execution dynamics by combining MPC adaptive algorithms, and ensures reliable implementation of control strategies by precisely driving physical actuators through PLC or frequency converters. Furthermore, it enables full-process monitoring through visualization, generates intelligent maintenance solutions through knowledge graph diagnosis, supports flexible operation through human-machine interaction, and facilitates analysis and traceability through data storage and backtracking. This significantly improves the intelligence level, control accuracy, and maintenance efficiency of air compressor operation. Attached Figure Description

[0057] Figure 1 This is a system diagram of an air compressor cyclic control system based on an LSTM deep learning network according to the present invention.

[0058] Figure 2 This is a system architecture diagram of an air compressor cyclic control system based on an LSTM deep learning network according to the present invention.

[0059] Figure 3 This is a flowchart illustrating the digital twin verification process in an air compressor cyclic control system based on an LSTM deep learning network, as described in this invention.

[0060] Explanation of icon numbers:

[0061] 100. Data Acquisition Unit; 101. Sensor Module; 102. Signal Preprocessing Module; 103. Data Transmission Module; 200. LSTM Prediction and Decision Unit; 201. Data Fusion Module; 202. Attention Weighting Module; 203. LSTM Prediction Module; 204. Policy Optimization Module; 300. Execution Control Unit; 301. Digital Twin Verification Module; 302. Adaptive Control Module; 303. Execution Drive Module; 400. Upper-Level Monitoring and Interaction Unit; 401. Visualization Display Module; 402. Knowledge Graph Diagnostic Module; 403. Human-Computer Interaction Module; 404. Data Storage and Backtracking Module. Detailed Implementation

[0062] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0063] Example:

[0064] like Figures 1-3As shown, this embodiment provides a cyclic control system for an air compressor based on an LSTM deep learning network, including a data acquisition unit 100, an LSTM prediction and decision unit 200, an execution control unit 300, and a supervisory control and interaction unit 400. Specifically: the data acquisition unit 100 is responsible for collecting key parameters such as pressure, temperature, and flow rate during the operation of the air compressor; the LSTM prediction and decision unit 200 is used to fuse multi-source heterogeneous time-series data, introduce an attention mechanism to enhance the perception of key features, and construct a long-term and short-term collaborative prediction framework based on the LSTM model, ensuring both prediction accuracy and response speed. The system predicts the changing trends of the air compressor's operating status and generates the optimal control strategy. The execution control unit 300, based on the control strategy output by the LSTM prediction and decision unit 200, combines a digital twin system to perform virtual verification and risk assessment of the execution process. After confirming safety, it adjusts the response dynamics through an adaptive control algorithm to drive the air compressor's valves and motor actuators. The upper-level monitoring and interaction unit 400 displays the system's operating status, energy consumption analysis, and AI decision-making basis in real time through a visual interface. It integrates a knowledge graph for intelligent fault diagnosis and maintenance suggestion generation, and supports parameter setting, command issuance, and human-machine interaction functions.

[0065] It should be noted that the data acquisition unit 100 acquires the air compressor's operating parameters in real time and transmits them to the LSTM prediction and decision unit 200 for multimodal time series analysis and strategy generation. After digital twin safety verification by the execution control unit 300, the physical equipment is driven, and finally, the upper-level monitoring and interaction unit 400 realizes full-process visual supervision and intelligent diagnosis.

[0066] In this embodiment, it should also be noted that the data acquisition unit 100 includes a sensor module 101, a signal preprocessing module 102, and a data transmission module 103, wherein: the sensor module 101 is used to acquire physical signals of pressure, temperature, and flow rate during the operation of the air compressor in real time; the signal preprocessing module 102 is used to filter, reduce noise, and convert the format of the raw acquired data; and the data transmission module 103 is used to stably transmit the processed data to the LSTM prediction and decision unit 200.

[0067] It should be noted that the sensor module 101 senses the operating parameters of the air compressor in real time, and after the signal preprocessing module 102 performs filtering and standardization processing, the high-quality data is finally stably transmitted to the LSTM prediction and decision unit 200 by the data transmission module 103.

[0068] Furthermore, it should be noted that the pressure sensor uses a diffused silicon pressure transmitter with a measurement range of 0-1.6 MPa, an accuracy of ±0.1% FS, and a sampling frequency of 10 Hz; the temperature sensor uses a PT100 resistance temperature detector (RTD) with a measurement range of -50 to 200℃, a resolution of 0.1℃, and a response time ≤1 s; and the flow sensor uses a vortex flow meter with a range of 0-100 m³ / h. 3 / h, accuracy ±0.5%, supports pulse / 4-20mA output.

[0069] In this embodiment, it should also be noted that the LSTM prediction and decision unit 200 includes a data fusion module 201, an attention weighting module 202, an LSTM prediction module 203, and a policy optimization module 204. Specifically: the data fusion module 201 is used for time alignment and feature standardization of multi-source heterogeneous time-series data; the attention weighting module 202 is used for dynamically allocating feature weights through an attention mechanism; the specific operations are as follows: A1: Receive multi-source time-series feature vectors from the data fusion module 201; A2: Calculate the correlation score of each feature at the current time step and generate dynamic weights through normalization; A3: Perform weighted fusion of the dynamic weights and the original features, outputting the enhanced feature representation to enhance the LSTM model's ability to identify key operating states. The LSTM prediction module 203 models temporal dependencies based on a long short-term memory network and outputs the prediction result of state change trends; the specific operations are as follows: B1: Divide the input weighted feature sequence into time steps to obtain the input sequence x. t B2: Calculate the forget gate, input gate, cell state update value, and output gate at each time step, and update the hidden state to capture long-term dependencies in the input sequence; B3: Map the hidden state sequence [h1, h2, ..., h...] through a fully connected layer. T Output the predicted state change trend for future time steps. Strategy optimization module 204: Used to generate control strategies that balance energy efficiency and stability through reinforcement learning algorithms. Specific operations are as follows: C1: Define the state space S as a fusion vector of the air compressor's current pressure, temperature, flow rate, and historical operating characteristics; the action space A is a combination of valve opening adjustment and motor speed setpoint; C2: Construct a multi-objective reward function:

[0070] r(s,a)=ω1·r energy (s,a)+ω2·r stability (s,a)

[0071] Where, r energy (s,a) represents the energy consumption optimization reward, r stability(s,a) represents the stability reward, and ω1 and ω2 represent the dynamic weights; C3: The policy network π(a|s;θ) is iteratively optimized using a deep deterministic policy gradient algorithm, and (s,a,r,s′) samples are stored in an experience replay pool to minimize the loss function L(θ)=E[(Q(s,;θ)] v )-(r+γQ′(s′,a′;θ v C4: When the policy network converges to a preset threshold, output the optimal action that matches the current state, i.e., the specific parameter value of the control policy.

[0072] It should be noted that the data fusion module 201 realizes the alignment and standardization of multi-source data. After the attention weighting module 202 dynamically selects key features, the LSTM prediction module 203 models the temporal dependency relationship to generate state prediction. Finally, the reinforcement learning algorithm of the policy optimization module 204 outputs the optimal control policy that takes into account both energy efficiency and stability.

[0073] Furthermore, it should be noted that the relevance score is calculated as follows: s i =MLP(h i =W2tanh(W1h) i +b1)+b2, where h i Let W1 and W2 be the weight matrices, and b1 and b2 be the bias terms, representing the i-th eigenvector. Dynamic weight normalization is achieved by generating weights using the softmax function. Where n is the feature dimension, and the sum of weights is ∑α i =1. Weighted fusion formula in A3: Enhanced features Weight α of key features (such as pressure fluctuations) i "Priority processing is triggered when the value is ≥0.5".

[0074] Specific calculation formula in B2: Forgetting gate: f t =σ(W f ·[h t-1 ,x t ]+b f ), where σ is the Sigmoid function, W f For the forget gate weight, b f For the forget gate bias term, h t-1 x is the hidden state from the previous moment. t This is the current input; input gate: i t =σ(W i ·[h t-1 ,x t ]+b i ), where W i b represents the input gate weights. i Input gate bias term; candidate cell state values: Where tanh is the hyperbolic tangent activation function, W c b represents the weight of the candidate cell state values. c Bias term for candidate cell state values; Cell state update: Output gate: o t =σ(W o ·[h t-1 ,x t ]+b o ), where W o b is the output gate weight. o Output gate bias term; hidden state: h t =o t ⊙tanh(C t In C4, the convergence threshold is determined when the mean of the loss function over 500 consecutive iterations is ≤0.01 and the standard deviation of the policy output actions is ≤5%.

[0075] In this embodiment, it should also be noted that the execution control unit 300 includes a digital twin verification module 301, an adaptive control module 302, and an execution drive module 303, wherein: the digital twin verification module 301 is used to simulate the control strategy in a virtual model and assess the execution risk; the specific operations are as follows: D1: Construct a digital twin model of the air compressor physical entity, which includes multi-domain mapping of geometric, physical, and behavioral dimensions; D2: Input the control strategy parameters output by the LSTM prediction and decision unit 200, perform dynamic simulation with a time step Δt = 0.1s in the virtual model, and generate pressure fluctuation curves, temperature change curves, and energy consumption curves; D3: Set a risk assessment index system, including an overpressure risk value R. p =∫(P(t)-P max ) + dt, temperature exceedance risk value R T =∑I(T(t)>T max ), Actuator load risk value R L =max(α) t / α max ,n t / n max D4: When the comprehensive risk value R = ω p R p +ω T R T +ω L R LWhen the risk threshold R0 is less than or equal to the risk level, the control strategy is deemed safe to execute; otherwise, a risk warning and parameter correction suggestions are output. Adaptive control module 302: Used to dynamically adjust valve opening or motor speed using an MPC algorithm to optimize response speed; Execution drive module 303: Used to convert control signals into physical actions via a PLC or frequency converter.

[0076] It should be noted that the digital twin verification module 301 performs virtual simulation and risk assessment on the control strategy generated by LSTM. After the execution parameters are dynamically optimized by the MPC algorithm of the adaptive control module 302, they are finally converted into physical actions of the valve / motor by the execution drive module 303.

[0077] Furthermore, it should be noted that the specific formula for the MPC algorithm is as follows: Where ΔP is the pressure deviation, Δu is the rate of change of the control quantity, and λ = 0.2 is the weighting coefficient to suppress drastic fluctuations in the control quantity.

[0078] In this embodiment, it should also be noted that the upper-level monitoring and interaction unit 400 includes a visualization module 401, a knowledge graph diagnosis module 402, a human-computer interaction module 403, and a data storage and backtracking module 404. Specifically: the visualization module 401 is used to present operational data, energy consumption analysis, and AI decision-making processes in real time through an interface; the knowledge graph diagnosis module 402 generates diagnostic results and maintenance plans based on fault feature association analysis; the specific operations are as follows: E1: Real-time acquisition of air compressor operating parameters and alarm signals, extracting fault feature vectors; E2: Matching predefined fault mode association rules in the knowledge graph; E3: Calculating the propagation probability between fault nodes through a graph neural network; E4: Generating a solution including root cause location, impact range, and maintenance steps; E5: Comparing the maintenance plan with the historical work order database and recommending the optimal handling strategy. The human-computer interaction module 403 supports operators in configuring parameters and issuing commands; the data storage and backtracking module 404 records historical system operation data for analysis and traceability.

[0079] It should be noted that the visualization module 401 presents the system's operating status and AI decision-making process in real time, uses the knowledge graph diagnosis module 402 to realize intelligent fault analysis and maintenance suggestions, receives operation instructions and parameter adjustments through the human-computer interaction module 403, and uses the data storage and backtracking module 404 to complete the persistent storage of operating data.

[0080] Furthermore, it should be noted that the graph neural network calculation formula is as follows: in, It is an adjacency matrix. Given a degree matrix, output the fault propagation probability P(G). k|F)≥0.9 is considered the root cause.

[0081] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0082] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A cyclic control system for an air compressor based on an LSTM deep learning network, characterized in that, It includes a data acquisition unit (100), an LSTM prediction and decision unit (200), an execution control unit (300), and a supervisory control and interaction unit (400), wherein: The data acquisition unit (100) is responsible for collecting key parameters such as pressure, temperature, and flow rate during the operation of the air compressor. The LSTM prediction and decision unit (200) is used to fuse multi-source heterogeneous time series data, introduce an attention mechanism to enhance the perception of key features, build a long-term and short-term collaborative prediction framework based on the LSTM model, and predict the change trend of the air compressor's operating status and generate the optimal control strategy under the premise of ensuring prediction accuracy and response speed. The execution control unit (300) is used to perform virtual verification and risk assessment of the execution process based on the control strategy output by the LSTM prediction and decision unit (200) and in conjunction with the digital twin system. After confirming safety, it adjusts the response dynamics through an adaptive control algorithm to drive the valves and motor actuators of the air compressor. The upper-level monitoring and interaction unit (400) is used to display the system's operating status, energy consumption analysis, and AI decision-making basis in real time through a visual interface, integrate knowledge graphs to generate intelligent fault diagnosis and maintenance suggestions, and support parameter setting, command issuance, and human-computer interaction functions.

2. The air compressor cyclic control system based on LSTM deep learning network according to claim 1, characterized in that, The data acquisition unit (100) includes a sensor module (101), a signal preprocessing module (102), and a data transmission module (103), wherein: The sensor module (101) is used to collect physical signals of pressure, temperature and flow rate in real time during the operation of the air compressor. The signal preprocessing module (102) is used to filter, reduce noise, and convert the format of the raw acquired data; The data transmission module (103) is used to stably transmit the processed data to the LSTM prediction and decision unit (200).

3. The air compressor cyclic control system based on an LSTM deep learning network according to claim 1, characterized in that, The LSTM prediction and decision unit (200) includes a data fusion module (201), an attention weighting module (202), an LSTM prediction module (203), and a policy optimization module (204), wherein: The data fusion module (201) is used to perform time alignment and feature standardization on multi-source heterogeneous time-series data; The attention weighting module (202) is used to dynamically allocate feature weights through an attention mechanism; The LSTM prediction module (203) models temporal dependencies based on long short-term memory networks and outputs prediction results of state change trends. The strategy optimization module (204) is used to generate a control strategy that balances energy efficiency and stability through a reinforcement learning algorithm.

4. The air compressor cyclic control system based on an LSTM deep learning network according to claim 3, characterized in that, The attention weighting module (202) dynamically allocates feature weights through an attention mechanism, as follows: A1: Receives multi-source temporal feature vectors from the data fusion module (201); A2: Calculate the relevance score of each feature at the current time step, and generate dynamic weights through normalization; A3: The dynamic weights are weighted and fused with the original features to output an enhanced feature representation, thereby improving the LSTM model's ability to identify key operating states.

5. The air compressor cyclic control system based on an LSTM deep learning network according to claim 3, characterized in that, The LSTM prediction module (203) models temporal dependencies based on a long short-term memory network and outputs the prediction results of state change trends. The specific operation is as follows: B1: Divide the input weighted feature sequence into time steps to obtain the input sequence x. t , t=1,2,…,T; B2: Calculate the forget gate, input gate, cell state update value and output gate for each time step in sequence, and update the hidden state to capture long-term dependencies in the input sequence; B3: Map the hidden state sequence [h1,h2,…,h] through a fully connected layer. T Output the predicted state change trend for future time steps.

6. The air compressor cyclic control system based on an LSTM deep learning network according to claim 3, characterized in that, The strategy optimization module (204) generates a control strategy that balances energy efficiency and stability through a reinforcement learning algorithm. The specific operation is as follows: C1: Define the state space S as a fusion vector of the air compressor's current pressure, temperature, flow rate, and historical operating characteristics, and the action space A as a combination of valve opening adjustment and motor speed setpoint. C2: Constructing a multi-objective reward function: r(s,a)=ω1·r energy (s,a)+ω2·r stability (s,a) Where, r energy (s,a) represents the energy consumption optimization reward, r stability (s,a) represents the stability reward, and ω1 and ω2 represent the dynamic weights. C3: The policy network π(a|s;θ) is iteratively optimized using a deep deterministic policy gradient algorithm. Samples (s,a,r,s′) are stored in an empirical replay pool, and the loss function L(θ)=E[(Q(s,;θ)] is minimized. v )-(r+γQ′(s′,a′;θ v ′)))2]; C4: When the policy network converges to a preset threshold, output the optimal action that matches the current state, i.e., the specific parameter value of the control policy.

7. The air compressor cyclic control system based on an LSTM deep learning network according to claim 1, characterized in that, The execution control unit (300) includes a digital twin verification module (301), an adaptive control module (302), and an execution drive module (303), wherein: The digital twin verification module (301) is used to simulate control strategies in a virtual model and assess execution risks. The adaptive control module (302) is used to dynamically adjust the valve opening or motor speed using the MPC algorithm to optimize the response speed; The execution drive module (303) is used to convert control signals into physical actions via a PLC or frequency converter.

8. The air compressor cyclic control system based on an LSTM deep learning network according to claim 7, characterized in that, The digital twin verification module (301) simulates the control strategy in the virtual model and assesses the execution risk. The specific operation is as follows: D1: Construct a digital twin model of the physical entity of the air compressor, which includes multi-domain mappings of geometric, physical, and behavioral dimensions; D2: Input the control strategy parameters output by the LSTM prediction and decision unit (200), perform dynamic simulation with a time step of Δt = 0.1s in the virtual model, and generate pressure fluctuation curve, temperature change curve and energy consumption curve; D3: Establish a risk assessment indicator system, including the overpressure risk value R. p =∫(P(t)-P max ) + dt, temperature exceedance risk value R T =∑I(T(t)>T max ), Actuator load risk value R L =max(α) t / α max ,n t / n max ); D4: When the overall risk value R = ω p R p +ω T R T +ω L R L If the risk threshold R0 is less than or equal to the risk threshold, the control strategy is deemed safe to execute; otherwise, a risk warning and parameter correction suggestions are output.

9. The air compressor cyclic control system based on an LSTM deep learning network according to claim 1, characterized in that, The upper-level monitoring and interaction unit (400) includes a visualization display module (401), a knowledge graph diagnosis module (402), a human-computer interaction module (403), and a data storage and backtracking module (404), wherein: The visualization module (401) is used to present operational data, energy consumption analysis, and AI decision-making processes in real time through the interface. The knowledge graph diagnostic module (402) generates diagnostic results and maintenance plans based on fault feature correlation analysis; The human-computer interaction module (403) is used to support operators in configuring parameters and issuing commands; The data storage and backtracking module (404) is used to record historical data of system operation for analysis and backtracking.

10. The air compressor cyclic control system based on an LSTM deep learning network according to claim 9, characterized in that, The knowledge graph diagnostic module (402) generates diagnostic results and maintenance plans based on fault feature correlation analysis. The specific operation is as follows: E1: Real-time acquisition of air compressor operating parameters and alarm signals, and extraction of fault feature vectors; E2: Match predefined fault mode association rules in the knowledge graph; E3: Calculate the propagation probability between faulty nodes using a graph neural network; E4: Generate a solution that includes root cause identification, scope of impact, and repair steps; E5: Compare the repair plan with the historical work order database and recommend the optimal handling strategy.

Images