Air separation plant energy efficiency intelligent management system based on data analysis

Abstract

The application relates to the technical field of air separation equipment management, and discloses an air separation equipment energy efficiency intelligent management system based on data analysis, which comprises data acquisition, energy efficiency analysis, optimization strategy generation, real-time adjustment and dynamic feedback modules. The data acquisition module collects and stores data in real time in a multi-source heterogeneous manner. The energy efficiency analysis module extracts features by using a hybrid convolutional neural network and an adaptive wavelet packet decomposition algorithm. The optimization strategy generation module formulates strategies by combining deep reinforcement learning and a multi-objective particle swarm optimization algorithm. The real-time adjustment module dynamically adjusts equipment parameters based on fuzzy logic control. The dynamic feedback module forms a closed-loop optimization. The system solves the problem of traditional air separation equipment energy efficiency management, can accurately collect and process data, deeply analyze energy efficiency, intelligently generate optimization strategies, accurately adjust parameters in real time and continuously optimize, improves the energy efficiency of air separation equipment, reduces energy consumption, enhances the operation stability and reliability of the equipment, and helps sustainable industrial development.

Description

Technical Field

[0001] This invention relates to the field of air separation equipment management technology, specifically to an intelligent energy efficiency management system for air separation equipment based on data analysis. Background Technology

[0002] Air separation equipment, as a key piece of equipment in the industrial sector, is widely used in numerous industries such as chemical, steel, and electronics. It is primarily used to separate and purify gases such as oxygen, nitrogen, and argon from the air to meet the gas requirements of various production processes. Against the backdrop of global advocacy for energy conservation, emission reduction, and green development, the energy efficiency of air separation equipment has attracted significant attention. This is not only related to enterprises' production cost control but also crucial for the sustainable development of the industry. However, air separation equipment currently faces many challenges in energy efficiency management, urgently requiring effective solutions.

[0003] Air separation equipment operation involves various types of data, such as compressor power data, distillation column temperature gradient data, gas purity monitoring data, and cooling water flow data. This data is characterized by its multi-source heterogeneity. Traditional data acquisition methods often rely on manual periodic inspection records or simple sensor data collection, which suffers from problems such as untimely data acquisition, low accuracy, and the inability to achieve simultaneous acquisition at multiple time granularities. This makes the acquired data unable to comprehensively and accurately reflect the real-time operating status of the equipment, failing to meet the needs for in-depth analysis and optimization. Furthermore, due to the lack of spatiotemporal correlation in the data, storage and processing face numerous difficulties, hindering effective support for subsequent energy efficiency analysis and decision-making. Existing energy efficiency analysis methods for air separation equipment are mostly based on simple empirical formulas or traditional statistical analysis, making it difficult to deeply explore the complex energy efficiency characteristics and correlations behind the data. For example, calculating energy efficiency solely by monitoring the overall energy consumption and gas production of the equipment cannot accurately pinpoint the specific links and causes of excessive energy consumption. Faced with the complex operating conditions and large amounts of operational data of air separation equipment, traditional methods cannot fully utilize the value of the data, nor can they analyze equipment energy efficiency from multiple scales and dimensions, resulting in a lack of strong technical support for optimized equipment operation.

[0004] Developing energy efficiency optimization strategies for air separation equipment is a complex multi-objective optimization problem, requiring comprehensive consideration of factors such as energy consumption reduction, gas purity improvement, equipment stability, and adjustment frequency. Traditional optimization algorithms struggle to simultaneously address these objectives and suffer from low computational efficiency when processing large-scale, high-dimensional equipment operating data. While deep reinforcement learning and multi-objective particle swarm optimization algorithms have theoretical advantages, their application in the air separation equipment field is limited. They lack effective integration and optimization tailored to the specific characteristics of air separation equipment, making it difficult to generate practical energy efficiency optimization strategies and equipment operating parameter adjustment schemes.

[0005] During the operation of air separation equipment, operating conditions change constantly, requiring timely adjustments to equipment operating parameters to maintain optimal energy efficiency. However, current control methods for most air separation equipment are relatively crude, lacking precise, real-time adjustment mechanisms. While fuzzy logic control shows promise, in the field of air separation equipment, its fuzzy rule base construction is often inadequate, and membership functions cannot be dynamically adjusted in real time according to the equipment's operating status, resulting in poor control performance. Furthermore, existing systems generally lack effective dynamic feedback mechanisms, failing to monitor optimization effects in real time and promptly transmit feedback information to the system for adjustments, making it difficult to form closed-loop optimization and limiting the continuous improvement of air separation equipment's energy efficiency. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent energy efficiency management system for air separation equipment based on data analysis, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an intelligent energy efficiency management system for air separation equipment based on data analysis, the system comprising:

[0008] It includes a data acquisition module, an energy efficiency analysis module, an optimization strategy generation module, a real-time adjustment module, and a dynamic feedback module;

[0009] The data acquisition module is used to collect multi-source heterogeneous data during the operation of the air separation equipment. Specifically, it collects data in real time by deploying temperature sensors, pressure sensors, flow sensors and energy consumption monitoring devices to obtain the original dataset of the air separation equipment, and sends the original dataset of the air separation equipment to the energy efficiency analysis module.

[0010] The real-time data acquisition specifically refers to synchronous acquisition based on multiple time granularities, including acquiring high-precision instantaneous data every 0.5 seconds and steady-state operating data every 30 seconds. Through the synchronous acquisition of multiple time granularities, a spatiotemporally correlated original data sequence is generated and stored in a distributed time series database.

[0011] The energy efficiency analysis module is used to extract energy efficiency features and perform correlation analysis on the raw data. Specifically, it uses a hybrid convolutional neural network and an adaptive wavelet packet decomposition algorithm to extract multi-scale features from the raw dataset of the space division equipment to obtain an energy efficiency feature matrix, and then sends the energy efficiency feature matrix to the optimization strategy generation module and the real-time adjustment module.

[0012] The optimization strategy generation module is used to generate energy efficiency optimization strategies and equipment operating parameter adjustment schemes. Specifically, based on the energy efficiency feature matrix, it uses deep reinforcement learning combined with multi-objective particle swarm optimization algorithm to generate energy efficiency optimization strategies, obtains the space separation equipment optimization strategy set, and sends the optimization strategy set to the real-time adjustment module.

[0013] The real-time adjustment module is used to execute optimization strategies and dynamically adjust the equipment operating parameters. Specifically, it constructs a dynamic regulator based on fuzzy logic control, generates real-time control instructions in combination with the optimization strategy set, and transmits the control instructions to the space separation equipment execution unit.

[0014] The dynamic feedback module is used to monitor the optimization effect in real time and feed it back to the system. Specifically, it constructs an incremental online learning model to dynamically update the features of the adjusted equipment operation data and feeds the updated feature data back to the energy efficiency analysis module, forming a closed-loop optimization link.

[0015] Preferably, in the data acquisition module, the original dataset of the air separation equipment specifically includes compressor unit power data, distillation column temperature gradient data, gas purity monitoring data, cooling water flow data, and equipment historical maintenance record data; the equipment historical maintenance record data includes equipment start-up and shutdown logs, component replacement records, and abnormal event marking data.

[0016] Preferably, in the energy efficiency analysis module, the processing steps of the hybrid convolutional neural network and adaptive wavelet packet decomposition algorithm include:

[0017] Step S1: Perform time-frequency domain decomposition on the original data using an adaptive wavelet packet decomposition algorithm to generate multi-resolution sub-band signals;

[0018] Step S2: Construct a three-dimensional convolution kernel for each sub-band signal, and extract joint spatial-frequency features through parallel convolutional layers;

[0019] Step S3: Use a channel attention mechanism to assign weights to the multi-channel features and generate an optimized energy efficiency feature matrix;

[0020] Step S4: Nonlinearly superimpose the features of different sub-bands using a cross-channel feature fusion algorithm to form a global energy efficiency characterization vector.

[0021] Preferably, the specific implementation steps of the deep reinforcement learning combined with the multi-objective particle swarm optimization algorithm include:

[0022] A Markov decision process model is constructed, defining the state space as the energy efficiency feature matrix and the action space as the set of equipment parameter adjustment instructions.

[0023] A dual-depth Q-network is used for policy exploration, and a Pareto front screening mechanism is introduced to reduce the dimensionality of the solution set for the multi-objective optimization problem;

[0024] The optimal energy efficiency strategy combination is generated by globally optimizing the solution set using a dynamic inertial weighted particle swarm algorithm.

[0025] Preferably, the dynamic regulator of the fuzzy logic control specifically includes a fuzzy rule base construction module, a membership function dynamic adjustment submodule, and a defuzzification processor;

[0026] The fuzzy rule base is constructed by generating fuzzy control rules based on expert experience and historical optimization data, and using a genetic algorithm to iteratively optimize the rule weights.

[0027] The membership function dynamic adjustment submodule corrects the width and center point of the triangle membership function in real time by learning the rate of change of the device's operating status online.

[0028] The deblurring processor uses the centroid method to convert fuzzy output into precise control commands.

[0029] Preferably, the incremental online learning model is specifically a variational autoencoder driven by streaming data, and its processing flow includes:

[0030] Step P1: Perform sliding window sampling on the real-time input device operation data to generate dynamic data blocks;

[0031] Step P2: Reconstruct features and map the latent space of the data block using a variational autoencoder to generate incremental feature vectors;

[0032] Step P3: Use KL divergence to measure the difference between the old and new feature distributions and dynamically update the encoder parameters to adapt to changes in device status.

[0033] Preferably, in the multi-objective particle swarm optimization algorithm, the objective function is defined as:

[0034] f1 = α * energy consumption reduction rate + β * gas purity deviation

[0035] f2 = γ * Equipment stability index - δ * Adjustment frequency

[0036] Among them, α, β, γ, and δ are dynamic weighting coefficients that are adaptively adjusted during the equipment operation phase.

[0037] Preferably, the method for constructing the three-dimensional convolutional kernel is as follows: separable convolutional kernels are set in the time dimension, spatial dimension and frequency domain dimension respectively, and a composite convolutional weight matrix is ​​generated by tensor multiplication; the update of the composite convolutional weight matrix adopts a gradient descent combined with Nesterov acceleration strategy.

[0038] Preferably, the sliding window sampling strategy for the dynamic data block includes:

[0039] The adaptive window size adjustment based on equipment operation mode switching detection reduces the window size to improve response speed when a sudden change in operating conditions is detected; and expands the window size to enhance the robustness of feature extraction during steady-state operation.

[0040] Preferably, the iterative optimization step of the genetic algorithm for the weights of fuzzy rules includes:

[0041] Encoding stage: Converting rule weights into binary gene sequences;

[0042] Selection phase: A tournament selection mechanism is used to screen for highly fit individuals;

[0043] Crossover and mutation phase: A new generation of population is generated using a uniform crossover operator and a dynamic mutation probability strategy;

[0044] The fitness function is defined as the inverse ratio of the fitting error of the rule base to historical optimization data to the weighted sum of the rule complexity.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] The system's data acquisition module utilizes multiple sensors to achieve real-time acquisition of multi-source heterogeneous data. It employs a multi-time-granularity synchronous acquisition method, collecting high-precision instantaneous data every 0.5 seconds and steady-state operating data every 30 seconds, generating a spatiotemporally correlated raw data sequence and storing it in a distributed time-series database. Compared to traditional acquisition methods, this approach not only significantly increases the acquisition frequency but also obtains more comprehensive and detailed operational information, accurately reflecting the real-time status of the equipment from multiple dimensions. For example, when monitoring compressor unit power changes, high-precision instantaneous data can capture power fluctuations at startup and loading moments, while steady-state operating data helps analyze energy consumption patterns during normal operation, providing a solid and reliable data foundation for subsequent energy efficiency analysis and optimization strategy development.

[0047] By employing a hybrid convolutional neural network and an adaptive wavelet packet decomposition algorithm, the system can extract multi-scale features from the original data, uncovering deep-seated energy efficiency characteristics and complex correlations. The adaptive wavelet packet decomposition algorithm performs fine-grained decomposition of the original data in the time-frequency domain, and multi-resolution sub-band signals can reveal the data's characteristics at different frequencies and time scales. Three-dimensional convolutional kernels and parallel convolutional layers work together to effectively extract joint spatial-frequency features. Channel attention mechanisms and cross-channel feature fusion algorithms further optimize and integrate features, forming a global energy efficiency representation vector. This deep analysis approach can accurately pinpoint the links and causes of excessive energy consumption in space separation equipment, providing strong support for the formulation of optimization strategies. Compared to traditional simple energy efficiency analysis methods, the analysis results are more accurate and comprehensive, contributing to more efficient energy utilization. The optimization strategy generation module utilizes deep reinforcement learning combined with a multi-objective particle swarm optimization algorithm, effectively solving the complex multi-objective optimization problem of energy efficiency optimization for space separation equipment. The constructed Markov decision process model clearly defines the state space and action space. The strategy exploration using a dual-depth Q-network combined with a Pareto front screening mechanism efficiently handles multi-objective optimization problems and reduces solution set redundancy. A dynamic inertial weighted particle swarm optimization algorithm performs global optimization, generating the optimal energy efficiency strategy combination that comprehensively considers energy consumption reduction rate, gas purity deviation, equipment stability index, and adjustment frequency. Compared with traditional optimization algorithms, this method more comprehensively balances multiple optimization objectives, and the generated optimization strategy better fits the actual operating needs of air separation equipment. This helps to reduce energy consumption, improve equipment stability, and extend equipment lifespan while ensuring gas product quality.

[0048] The real-time adjustment module, built upon a fuzzy logic control-based dynamic regulator, enables real-time and precise adjustment of equipment operating parameters. The fuzzy rule base is generated based on expert experience and historical optimization data, and its rule weights are iteratively optimized using a genetic algorithm to make the control rules more consistent with the actual operating conditions of the equipment. The membership function dynamic adjustment submodule corrects the changes in equipment operating status in real time, enhancing the adaptability of fuzzy logic control. The defuzzification processor uses the centroid method to convert fuzzy outputs into precise control commands, driving the air separation unit's actions. Taking the adjustment of the reflux ratio in a distillation column as an example, the system can quickly and accurately adjust the reflux ratio based on real-time monitored temperature, pressure, and other data, ensuring efficient and stable operation of the distillation process, improving gas separation efficiency, and further enhancing equipment energy efficiency. The dynamic feedback module constructs an incremental online learning model using a variational autoencoder driven by streaming data. This model dynamically updates the adjusted equipment operating data and feeds the updated feature data back to the energy efficiency analysis module, forming a closed-loop optimization chain. This allows the system to monitor the optimization effect in real time, promptly detect changes in equipment operating status, and adjust optimization strategies and operating parameters based on new data. As time goes by and data accumulates, the system continuously optimizes its performance, improves the energy efficiency management level of the air separation equipment, and ensures that the equipment can operate efficiently under different working conditions to adapt to complex and ever-changing production needs. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the working principle of the intelligent energy efficiency management system for air separation equipment described in this invention.

[0050] Figure 2 A flowchart illustrating the workflow of the energy efficiency analysis algorithm for air separation equipment;

[0051] Figure 3 A flowchart of the energy efficiency optimization cycle for air separation equipment. Detailed Implementation

[0052] 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.

[0053] Please see Figure 1-3 This invention provides a technical solution: an intelligent energy efficiency management system for air separation equipment based on data analysis. The system consists of a data acquisition module, an energy efficiency analysis module, an optimization strategy generation module, a real-time adjustment module, and a dynamic feedback module working together to achieve intelligent management and optimization of the energy efficiency of air separation equipment.

[0054] The data acquisition module plays a crucial role in data acquisition during the operation of the air separation unit. This module collects operational data in real time by deploying temperature sensors, pressure sensors, flow sensors, and energy consumption monitoring devices. This acquisition is based on multi-time-granularity synchronous acquisition: high-precision instantaneous data is collected every 0.5 seconds to capture instantaneous changes during equipment operation; steady-state operating data is collected every 30 seconds to obtain data characteristics under stable operating conditions. This multi-time-granularity synchronous acquisition method generates spatiotemporally correlated raw data sequences and stores them in a distributed time-series database, forming the raw dataset of the air separation unit. This data comprehensively covers compressor unit power data, distillation column temperature gradient data, gas purity monitoring data, cooling water flow data, and historical equipment maintenance records (including equipment start-up and shutdown logs, component replacement records, and abnormal event marking data), providing rich and accurate basic information for subsequent analysis and decision-making.

[0055] The energy efficiency analysis module receives the raw dataset from the space division equipment from the data acquisition module and performs in-depth energy efficiency feature extraction and correlation analysis on the raw data using a hybrid convolutional neural network and an adaptive wavelet packet decomposition algorithm. First, the adaptive wavelet packet decomposition algorithm is used to decompose the raw data in the time-frequency domain, breaking down the complex raw signal into multi-resolution sub-frequency band signals to allow for more detailed analysis of signal characteristics at different frequencies and time scales. Next, a three-dimensional convolutional kernel is constructed for each sub-frequency band signal, and spatial-frequency joint features are extracted through parallel convolutional layers to fully mine the feature information of the data in different dimensions. Then, a channel attention mechanism is used to assign weights to the multi-channel features, highlighting features that have a significant impact on energy efficiency analysis and generating an optimized energy efficiency feature matrix. Finally, a cross-channel feature fusion algorithm is used to nonlinearly superimpose the features of different sub-frequency bands to form a global energy efficiency representation vector that comprehensively characterizes the energy efficiency status of the space division equipment. This vector serves as an important basis for subsequent module decisions and is sent to the optimization strategy generation module and the real-time adjustment module.

[0056] The optimization strategy generation module, based on the energy efficiency feature matrix provided by the energy efficiency analysis module, uses deep reinforcement learning combined with a multi-objective particle swarm optimization algorithm to generate energy efficiency optimization strategies and equipment operating parameter adjustment schemes. Specifically, a Markov decision process model is constructed, defining the energy efficiency feature matrix as the state space and the set of equipment parameter adjustment instructions as the action space. A dual-deep Q-network is used for strategy exploration, and a Pareto front screening mechanism is introduced during the exploration process to reduce the dimensionality of the solution set of the multi-objective optimization problem, thereby reducing computational complexity and selecting better solutions. Then, a dynamic inertial weighted particle swarm optimization algorithm is used to globally optimize the solution set, comprehensively considering factors such as energy consumption reduction rate, gas purity deviation, equipment stability index, and adjustment frequency, to generate the optimal energy efficiency strategy combination, i.e., the air separation equipment optimization strategy set, which is then sent to the real-time adjustment module.

[0057] The real-time adjustment module is responsible for executing optimization strategies and dynamically adjusting equipment operating parameters. This module achieves this function by constructing a dynamic regulator based on fuzzy logic control. The fuzzy logic control dynamic regulator includes a fuzzy rule base construction module, a membership function dynamic adjustment submodule, and a defuzzification processor. The fuzzy rule base generates fuzzy control rules based on expert experience and historical optimization data, and uses a genetic algorithm to iteratively optimize the rule weights, making the rules more consistent with actual operating conditions. The membership function dynamic adjustment submodule learns the rate of change of equipment operating status online and corrects the width and center point of the triangular membership function in real time to adapt to changes in equipment operating status. The defuzzification processor uses the centroid method to convert fuzzy outputs into precise control commands and transmits these commands to the air separation unit's execution unit, achieving precise adjustment of the air separation unit's operating parameters.

[0058] The dynamic feedback module constructs an incremental online learning model to dynamically update the features of adjusted equipment operating data, enabling real-time monitoring and feedback of optimization effects to the system. Based on a streaming data-driven variational autoencoder as the incremental online learning model, it first performs sliding window sampling on the real-time input equipment operating data to generate dynamic data blocks. Then, the variational autoencoder reconstructs features and maps the latent space of the data blocks, generating incremental feature vectors. Finally, the KL divergence measure is used to measure the difference between the old and new feature distributions, dynamically updating the encoder parameters to adapt to changes in equipment status. The updated feature data is fed back to the energy efficiency analysis module, forming a closed-loop optimization chain that continuously optimizes the system's management and control of the air separation equipment's energy efficiency.

[0059] The present invention will be further described below with reference to Examples 1 to 5:

[0060] Example 1:

[0061] During the operation of air separation equipment, the accuracy and completeness of the data acquisition module play a decisive role in the performance of the entire system.

[0062] To collect power data from the compressor unit, a high-precision power sensor is installed on the power supply line of the compressor unit. This sensor, based on the principle of electromagnetic induction, can monitor changes in current and voltage in real time, and calculates the data using an internal calculation chip according to a formula. (Where P is power, U is voltage, and I is current) The real-time power of the compressor unit is calculated using a power factor. The sensor collects high-precision instantaneous power data at 0.5-second intervals, capturing instantaneous power fluctuations under different operating conditions such as startup, loading, and unloading. Simultaneously, steady-state operating power data is collected every 30 seconds to record the power consumption of the compressor unit under stable operating conditions, providing fundamental data for subsequent analysis of the compressor unit's energy efficiency.

[0063] The acquisition of temperature gradient data in the distillation column relies on multiple temperature sensors installed at different heights within the column. These sensors are evenly distributed across the column, forming a temperature measurement point array from top to bottom. The sensors utilize high-precision thermistors, whose resistance changes precisely with temperature. By measuring the thermistor resistance and converting it to a corresponding temperature value based on a pre-calibrated temperature-resistance curve, each temperature sensor synchronously acquires data at multiple time granularities: instantaneous temperature data is collected every 0.5 seconds to monitor rapid temperature changes within the column, such as those caused by feed fluctuations or reflux variations; steady-state temperature data is collected every 30 seconds to analyze the temperature distribution under stable operating conditions. Based on the collected temperature data from different heights, the temperature gradient of the distillation column can be calculated using the following formula: (where ΔT is the temperature difference between adjacent measurement points and Δh is the height difference between adjacent measurement points). Temperature gradient data is of great significance for judging the distillation effect and energy utilization efficiency of the distillation column.

[0064] Gas purity monitoring data is acquired using a highly sensitive gas composition analyzer. This analyzer employs advanced technologies such as chromatography, mass spectrometry, or infrared absorption to accurately measure the content of various components in the gas produced by the air separation unit. Taking oxygen purity monitoring as an example, the analyzer extracts an oxygen sample from the air separation unit, performs a series of separation and detection processes internally, and accurately calculates the oxygen purity. Gas purity monitoring data is also collected at a frequency of 0.5 seconds for high-precision instantaneous data and 30 seconds for steady-state data to promptly detect trends in gas purity changes and provide a basis for adjusting equipment operating parameters.

[0065] Cooling water flow rate data is collected using electromagnetic flowmeters installed in the cooling water pipeline. Based on Faraday's law of electromagnetic induction, when conductive cooling water flows in a magnetic field, an induced electromotive force (EMF) is generated in a direction perpendicular to both the magnetic field and the water flow direction. The magnitude of this induced EMF is proportional to the cooling water flow velocity. By measuring the induced EMF and using the calibration coefficient of the electromagnetic flowmeter, the cooling water flow rate can be calculated. The calculation formula is Q = vA (where Q is the flow rate, v is the flow velocity, and A is the pipe cross-sectional area). The collected cooling water flow rate data is stored using a multi-time-granularity synchronous acquisition method, providing data support for analyzing the cooling effect and energy consumption relationship of the air separation unit.

[0066] The collection of historical maintenance records is achieved through data integration with the equipment maintenance management system. Equipment start-up and shutdown logs record the time of each start-up and shutdown of the air separation unit, the operator, and whether any abnormalities occurred during the start-up and shutdown process. Component replacement records detail the replacement time, model, and manufacturer of each component. Abnormal event tagging data marks faults, alarms, and other abnormal situations that occur during equipment operation, including the time of occurrence, type of abnormality, and handling measures. This data is transmitted to a distributed time-series database via a data interface and stored along with other operational data, providing historical data references for equipment fault diagnosis, performance evaluation, and energy efficiency optimization.

[0067] Example 2:

[0068] This embodiment describes the specific processing flow of the hybrid convolutional neural network and adaptive wavelet packet decomposition algorithm in the energy efficiency analysis module. This algorithm plays a crucial role in mining the energy efficiency characteristics of air separation unit operating data.

[0069] In step S1, the adaptive wavelet packet decomposition algorithm performs time-frequency domain decomposition on the original data. Taking a key operating parameter in an air separation unit (such as the vibration signal of a compressor unit) as an example, this signal is a complex signal that varies with time, containing different frequency components and time characteristics. The adaptive wavelet packet decomposition algorithm adaptively selects the wavelet basis function and the number of decomposition levels based on the energy distribution characteristics of the signal. Assuming the original signal is x(t), it is decomposed by the wavelet packet decomposition function W... j,k (t) (where j is the decomposition level and k is the frequency sub-band index) is decomposed to obtain the multi-resolution sub-band signal d. j,k (t), its calculation formula is: This decomposition method allows for detailed analysis of the original signal at different frequencies and time scales, extracting more representative features.

[0070] In step S2, a three-dimensional convolution kernel is constructed for each sub-band signal. Taking one of the sub-band signals d... j,k Taking (t) as an example, in the time dimension, a convolution kernel of appropriate length is set according to the time series characteristics of the signal; assuming the time dimension convolution kernel length is T. In the spatial dimension, considering the spatial correlation between different components of the space division equipment, the size of the spatial dimension convolution kernel is set; assuming the spatial dimension convolution kernel size is S×S. In the frequency domain dimension, the frequency domain convolution kernel parameters are set according to the frequency range of the sub-band signal. A composite convolution weight matrix C is generated through tensor multiplication. (where C) T C is the convolution kernel matrix in the time dimension. S C is the spatial dimension convolution kernel matrix. FThe frequency domain dimension convolution kernel matrix, (This refers to the tensor multiplication operator). Parallel convolutional layers are used to perform convolution operations on the sub-band signal to extract joint spatial-frequency features. Parallel convolutional layers can simultaneously perform convolution operations on sub-band signals from different channels, improving feature extraction efficiency.

[0071] In step S3, a channel attention mechanism is used to assign weights to the multi-channel features. Assume the multi-channel features extracted by the parallel convolutional layers are F = [f1, f2, ..., f...]. n ], where n is the number of channels. First, a global average pooling operation is performed on the features of each channel to obtain global statistical information of the channel features. (Where H, W, and D represent the dimensions of the feature map in the height, width, and depth directions, respectively). Then, the global statistics are input into a network containing a multilayer perceptron (MLP), the structure of which is MLP(g) i )=W2σ(W1g i (W1 and W2 are the weight matrices of the MLP, and σ is the activation function, such as the RelU function), to obtain the weight coefficient w for each channel. i Finally, the weighting coefficients are multiplied by the original channel features to generate the optimized energy efficiency feature matrix F. optimized =[w1f1,w2f2,…,w n f n This highlights the characteristic channels that are important for energy efficiency analysis and suppresses irrelevant or secondary characteristic channels.

[0072] In step S4, the features of different sub-bands are nonlinearly superimposed using a cross-channel feature fusion algorithm. Assume the feature matrices of the different sub-bands after optimization by the channel attention mechanism are F1, F2, ..., F... m (Where m is the number of sub-bands). First, the feature dimension of each sub-band feature matrix is ​​transformed to make them have the same dimension. Then, a nonlinear fusion function is used, such as... (where a) i The fusion coefficients are learned through training (tanh is the hyperbolic tangent activation function). Features from different sub-bands are nonlinearly superimposed to form a global energy efficiency representation vector G. This global energy efficiency representation vector integrates the feature information from different sub-bands, and can more comprehensively reflect the energy efficiency status of the space separation equipment.

[0073] Example 3:

[0074] This embodiment focuses on the specific implementation steps of the deep reinforcement learning combined with multi-objective particle swarm optimization algorithm in claim 4. This algorithm is crucial for generating efficient energy efficiency optimization strategies for air separation equipment.

[0075] When constructing the Markov decision process model, the energy efficiency characteristic matrix output by the energy efficiency analysis module is used as the state space S. Assume the energy efficiency characteristic matrix is ​​E = [e1, e2, ..., e...]. n ], where e i Different energy efficiency characteristic indicators are used. The set of equipment parameter adjustment commands is defined as action space A, such as commands to adjust the compressor unit's intake air volume or the distillation column's reflux ratio. At each decision time t, the system adjusts the current state s... t ∈S selects an action a t ∈A, after performing this action, the system will transition to a new state s. t+1 and receive a reward value r t The reward value is set according to the system's optimization goal. For example, if energy consumption is reduced or gas purity is increased, a positive reward is given; otherwise, a negative reward is given.

[0076] A dual-deep Q-network (DDQN) is employed for policy exploration. DDQN introduces two neural networks: an evaluation network Q(s,a;θ) and a target network Q(s,a;θ). - This addresses the overestimation problem in traditional Deep Q-Networks (DQNs). During training, the current state s is first determined... t Select an action from the evaluation network. Then, the action is performed in the environment, and the new state s is observed. t+1 and reward r t Next, the target Q-value is calculated using the target network. (Where γ is the discount factor, used to balance the importance of current and future rewards). By minimizing the loss function L(θ) = E[(y t -Q(s t ,a t ;θ)) 2 The parameters θ of the evaluation network are updated using the stochastic gradient descent algorithm. After a certain number of training steps, the parameters of the evaluation network are copied to the target network, i.e., θ. - ←θ, to maintain the stability of the target network.

[0077] In multi-objective optimization problems, a Pareto front screening mechanism is introduced to reduce the dimensionality of the solution set. Multi-objective particle swarm optimization considers multiple objectives, such as energy consumption reduction rate, gas purity deviation, equipment stability index, and adjustment frequency. In each iteration, each particle in the swarm has a corresponding set of objective function values. The Pareto front refers to a set of non-dominated solutions in a multi-objective optimization problem; that is, none of these solutions can improve a particular objective without sacrificing other objectives. By comparing the objective function values ​​of the particles, particles at the Pareto front are selected as potential optimal solutions, reducing the number of solutions that need to be processed and lowering computational complexity.

[0078] The solution set is globally optimized using a dynamic inertial weighted particle swarm optimization (PSO) algorithm. In the PSO algorithm, each particle represents a possible solution, and the particle's position represents a combination of equipment operating parameters. By continuously updating the particle's velocity and position, the particle swarm gradually moves closer to the optimal solution, ultimately generating the optimal energy efficiency strategy combination, i.e., the space separation equipment optimization strategy set.

[0079] Example 4:

[0080] This embodiment describes the fuzzy logic control dynamic regulator in the real-time adjustment module. This regulator is a key component for achieving precise dynamic adjustment of the air separation unit's operating parameters. Its function is to transform the fuzzy control intent into executable, precise instructions based on the strategy output by the optimization strategy generation module, thereby driving the air separation unit to perform actions, adjusting the equipment's operating parameters towards improving energy efficiency, and ensuring the stable and efficient operation of the air separation unit.

[0081] Construction and Optimization of the Fuzzy Rule Base: The construction of the fuzzy rule base involves the deep integration of the rich experience of experts in the field of air separation equipment and a large amount of historical optimization data. Experts have accumulated strategies for adjusting equipment operating parameters under different conditions through long-term practice. For example, when observing a continuous rise in the top temperature of the distillation column while the pressure remains low, experience tells them that the reflux ratio should be appropriately increased to improve distillation efficiency and reduce energy consumption. Historical optimization data records the adjustment process of equipment operating parameters under various conditions and the corresponding changes in energy efficiency.

[0082] After organizing these experiences and data, they are transformed into specific fuzzy control rules. For example, rules can be set with distillation column temperature (T), pressure (P), and reflux ratio (R) as parameters: if T is "too high" and P is "too low", then R "increases"; if T is "moderate" and P is "moderate", then R "remains unchanged"; if T is "too low" and P is "too high", then R "decreases". Here, "too high", "moderate", "too low", "increase", "remain unchanged", and "decrease" are all fuzzy linguistic variables used to summarize the different states of equipment operating parameters and their corresponding adjustment directions.

[0083] To make these fuzzy rules more suitable for the actual operation of the air separation plant, the genetic algorithm is used to iteratively optimize the rule weights. First is the encoding step, where the weight of each rule is transformed into a binary gene sequence.

[0084] Then comes the selection stage, adopting the tournament selection mechanism. Randomly select k individuals from the population (k is the tournament size), compare their fitness values, and select the individual with the highest fitness to enter the next-generation population. The evaluation of fitness is based on the improvement effect of system energy efficiency. The more significant the energy efficiency improvement, the higher the individual fitness.

[0085] Next are the crossover and mutation stages. During crossover, the uniform crossover operator is used to randomly select gene positions of two individuals for exchange to generate new individuals. Mutation adopts the dynamic mutation probability strategy. At the initial stage of iteration, the mutation probability is set relatively low. As the number of iterations increases, the mutation probability is gradually increased to maintain the diversity of the population and prevent the algorithm from falling into a local optimal solution. Through continuous iteration, the rule weights are optimized so that the fuzzy rules can more accurately reflect the operation requirements of the air separation plant.

[0086] Dynamic adjustment sub-module of membership function: Taking the temperature membership function as an example, it is initially set as a triangular membership function. During the operation of the air separation plant, temperature is a key parameter, and its change affects the overall performance and energy efficiency of the plant. For the fuzzy linguistic variable "high temperature", a temperature interval [a, b, c] is set, and the calculation rule of the corresponding triangular membership function μT(x) is as follows: when x ≤ a, μT(x) = 0; when a < x ≤ b, μT(x) = (x - a) / (b - a); when b < x ≤ c, μT(x) = (c - x) / (c - b); when x > c, μT(x) = 0. Here, a is a relatively high temperature threshold, b is the typical temperature value representing "high temperature", and c is a higher temperature threshold.

[0087] During the operation of the plant, the change rate of the plant operation state is obtained through online learning, and the values of a, b, and c are dynamically adjusted according to the temperature change trend. If the temperature rises rapidly and is close to the upper limit of the current membership function, it means that the plant operation state has changed and the membership function needs to be adjusted. At this time, appropriately increase the value of c and adjust the value of b to ensure that the membership function can accurately reflect the current temperature state. When adjusting, comprehensively consider the temperature change rate and historical data. For example, when the temperature change rate exceeds the preset threshold and continues to rise for a period of time, the adjustment amplitudes of c and b can be appropriately increased according to experience; if the temperature changes relatively smoothly, then fine-tune the values of a, b, and c. By continuously monitoring the change rate of the plant operation state, the width and center point of the triangular membership function are continuously corrected, so that the fuzzy logic control can more accurately adapt to the actual operation state of the plant.

[0088] Defuzzification Processor: The defuzzification processor uses the centroid method to convert fuzzy output into precise control commands. After fuzzy inference yields the output fuzzy set B, its membership function is μB(y), where y represents the value of the output variable.

[0089] After obtaining precise control commands in this way, they are transmitted to the air separation unit's execution unit. For example, if fuzzy inference indicates that the feed rate needs to be increased, the defuzzification processor will calculate the specific increase in feed rate, such as an increase of 5 cubic meters per hour. Then, this command is sent to the feed pump's control unit, which drives the feed pump to adjust the flow rate according to the command. This adjustment may be achieved by regulating the pump's motor speed or valve opening, thereby precisely adjusting the feed rate to the air separation unit, optimizing equipment operating parameters, and improving equipment energy efficiency.

[0090] Example 5:

[0091] This embodiment mainly focuses on the incremental online learning model of the dynamic feedback module, which enables real-time monitoring of optimization effects and dynamic updates of the system.

[0092] A variational autoencoder based on streaming data, used as an incremental online learning model, first performs sliding window sampling on the real-time input device operation data. Assume the real-time input device operation data sequence is X = [x1, x2, ..., x...]. N ], where x i The device runs a data vector at each time point. Setting the sliding window size to w, each sample yields a data block X of length w. window =[x t-w+1 ,x t-w+2 ,…,x t ], where t is the current time point. As time progresses, the sliding window moves forward continuously, generating a series of dynamic data blocks.

[0093] A variational autoencoder (VAE) consists of two parts: an encoder and a decoder. The encoder maps the input data block to a latent space. Encoders are typically implemented using neural networks, which learn the feature distribution of the input data to transform it into latent variables. The decoder's role is to reconstruct the latent variables as approximations of the original data block.

[0094] During training, the goal of the variational autoencoder is to maximize the evidence lower bound (ELBO). By maximizing the ELBO, the variational autoencoder can learn the latent feature distribution of the data and achieve effective reconstruction of the data.

[0095] The KL divergence is used to measure the difference between the old and new feature distributions. A large KL divergence value indicates a significant difference between the old and new feature distributions, suggesting a substantial change in the device's operating state. In this case, the encoder parameters need to be dynamically updated to adapt to the change in device state. The encoder and decoder parameters are updated based on the gradient information from ELBO using the backpropagation algorithm, enabling the variational autoencoder to continuously adapt to new data features.

[0096] The updated feature data is fed back to the energy efficiency analysis module. Based on the new feature data, the energy efficiency analysis module re-extracts and performs correlation analysis on energy efficiency features, adjusts the optimization strategy generated by the optimization strategy generation module, and the real-time adjustment module further adjusts the operating parameters of the air separation unit according to the new optimization strategy. Through this closed-loop optimization chain, the system can respond to changes in equipment operating status in real time and continuously optimize the energy efficiency management of the air separation unit.

[0097] For example, when an air separation unit undergoes maintenance or has some components replaced, its operating status may change. The dynamic feedback module can promptly capture this change through an incremental online learning model, using KL divergence to determine differences in feature distribution, updating encoder parameters, and feeding the new feature data back to the energy efficiency analysis module. The energy efficiency analysis module reassesses the unit's energy efficiency based on the new feature data, while the optimization strategy generation module generates a more suitable optimization strategy for the current unit's state. The module then adjusts its execution in real time to implement the new strategy, thereby achieving dynamic optimization of the air separation unit's energy efficiency.

[0098] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0099] 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. An air separation equipment energy efficiency intelligent management system based on data analysis, characterized by: comprising a data acquisition module, an energy efficiency analysis module, an optimization strategy generation module, a real-time adjustment module, and a dynamic feedback module; the data acquisition module is used to acquire multi-source heterogeneous data in the running process of the air separation equipment, specifically by deploying temperature sensors, pressure sensors, flow sensors, and energy consumption monitoring devices to perform real-time data acquisition, obtain an air separation equipment original data set, and send the air separation equipment original data set to the energy efficiency analysis module; the real-time data acquisition specifically refers to synchronous acquisition based on multiple time granularities, including collecting high-precision instantaneous data every 0.5 seconds and collecting steady-state operation data every 30 seconds, and generating a spatiotemporal correlated original data sequence through the multiple time granularities synchronous acquisition and storing it in a distributed time series database; the energy efficiency analysis module is used to perform energy efficiency feature extraction and correlation analysis on the original data, specifically by using a hybrid convolutional neural network and an adaptive wavelet packet decomposition algorithm to perform multi-scale feature extraction on the air separation equipment original data set, obtain an energy efficiency feature matrix, and send the energy efficiency feature matrix to the optimization strategy generation module and the real-time adjustment module; the optimization strategy generation module is used to generate energy efficiency optimization strategies and device operation parameter adjustment schemes, specifically by using a deep reinforcement learning combined with a multi-objective particle swarm optimization algorithm to generate energy efficiency optimization strategies based on the energy efficiency feature matrix, obtaining an air separation equipment optimization strategy set, and sending the optimization strategy set to the real-time adjustment module; the real-time adjustment module is used to execute the optimization strategies and dynamically adjust the device operation parameters, specifically by constructing a dynamic regulator based on fuzzy logic control, generating real-time control instructions in combination with the optimization strategy set, and transmitting the control instructions to an air separation equipment execution unit; the dynamic feedback module is used to monitor the optimization effect in real time and feed back to the system, specifically by constructing an incremental online learning model to update the dynamic features of the adjusted device operation data, and feeding back the updated feature data to the energy efficiency analysis module to form a closed-loop optimization link; the processing steps of the hybrid convolutional neural network and the adaptive wavelet packet decomposition algorithm include: Step S1: Time-frequency domain decomposition of the original data is performed by the adaptive wavelet packet decomposition algorithm to generate multi-resolution sub-band signals; Step S2: A three-dimensional convolution kernel is constructed for each sub-band signal, and spatial-frequency domain joint features are extracted through a parallel convolution layer; Step S3: A channel attention mechanism is used to allocate weights to multi-channel features to generate an optimized energy efficiency feature matrix; Step S4: Different sub-band features are nonlinearly superimposed by a cross-channel feature fusion algorithm to form a global energy efficiency representation vector; in the multi-objective particle swarm optimization algorithm, the objective function is defined as: ； ； wherein, , , , is a dynamic weight coefficient, which is self-adaptively adjusted by the device running stage.

2. The data analysis based air space equipment energy efficiency intelligent management system according to claim 1, characterized in that, In the data acquisition module, the air separation equipment original data set specifically includes compressor unit power data, rectifying column temperature gradient data, gas purity monitoring data, cooling water flow data and equipment historical maintenance record data; the equipment historical maintenance record data includes equipment start-stop log, component replacement record and abnormal event marker data.

3. The data analytics based air-spaced equipment energy efficiency intelligent management system of claim 1, wherein, The specific implementation steps of the deep reinforcement learning combined with the multi-objective particle swarm optimization algorithm include: A Markov decision process model is constructed, and a state space is defined as an energy efficiency feature matrix, and an action space is defined as a device parameter adjustment instruction set; A double deep Q network is used for policy exploration, and a Pareto front screening mechanism is introduced to reduce the dimension of the solution set of the multi-objective optimization problem; A dynamic inertia weight particle swarm optimization algorithm is used for global optimization of the solution set to generate an optimal energy efficiency strategy combination.

4. The data analytics based air-spaced equipment energy efficiency intelligent management system of claim 3, wherein, The dynamic regulator of the fuzzy logic control specifically includes a fuzzy rule base construction, a membership function dynamic adjustment submodule and a defuzzification processor; The fuzzy rule base construction generates fuzzy control rules according to expert experience and historical optimization data, and iteratively optimizes the rule weights using a genetic algorithm; The membership function dynamic adjustment submodule adjusts the width and center point of the triangular membership function in real time by learning the change rate of the device operating state online; The defuzzification processor converts the fuzzy output into accurate control instructions using the barycentric method.

5. The data analytics based air-spaced equipment energy efficiency intelligent management system of claim 1, wherein, The incremental online learning model is specifically a variational autoencoder driven by streaming data, and its processing procedure includes: Step P1: sliding window sampling of real-time input device operating data to generate dynamic data blocks; Step P2: feature reconstruction and latent space mapping of the data blocks by the variational autoencoder to generate incremental feature vectors; Step P3: dynamically updating the encoder parameters to adapt to the device state changes by measuring the difference between the new and old feature distributions using KL divergence.

6. The data analytics based air-spaced equipment energy efficiency intelligent management system of claim 1, wherein, The construction method of the three-dimensional convolution kernel is to set separable convolution kernels in the time dimension, the spatial dimension and the frequency domain, respectively, and generate a composite convolution weight matrix through tensor product; and the update of the composite convolution weight matrix adopts gradient descent combined with Nesterov acceleration strategy.

7. The data analytics based air-spaced equipment energy efficiency intelligent management system of claim 5, wherein, The sliding window sampling strategy of the dynamic data block includes: Based on the adaptive window size adjustment of the device operating mode switching detection, when a working condition mutation is detected, the window is reduced to improve the response speed; and when in steady state operation, the window is expanded to enhance the robustness of feature extraction.

8. The data analytics based air-spaced equipment energy efficiency intelligent management system of claim 4, wherein, The iterative optimization steps of the genetic algorithm for the fuzzy rule weight include: Encoding stage: converting the rule weight into a binary gene sequence; Selection stage: screening high fitness individuals using the tournament selection mechanism; Crossing and mutation stage: using the uniform crossover operator and dynamic mutation probability strategy to generate a new generation of population; The fitness function is defined as the inverse ratio of the weighted sum of the fitting error of the rule base to the historical optimization data and the rule complexity.

Images