Textile workshop air conditioner cooperative control and real-time monitoring system based on IoT and PPO algorithms

Abstract

The invention discloses a textile workshop air conditioner cooperative control and real-time monitoring system based on IoT and PPO algorithms, and the system comprises a monitoring data collection module which is used for collecting original monitoring data in a textile workshop; the data preprocessing module is used for executing data preprocessing on the original monitoring data to generate a standard monitoring tensor; the state modeling module is used for generating a monitoring state feature vector sequence from the standard monitoring tensor through the improved ConvGRU network; the control environment construction module is used for constructing a reinforcement learning control environment; the air conditioner control strategy generation module is used for generating an air conditioner control strategy from the monitoring state feature vector sequence through an improved PPO algorithm; and the control instruction execution module is used for executing control operation on the air conditioning equipment in the textile workshop through the control keeping mechanism. According to the invention, through the improved ConvGRU network and the improved PPO algorithm, the cooperative control capability of the air conditioner in the textile workshop is improved.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent control technology, and in particular to a collaborative control and real-time monitoring system for air conditioning in textile workshops based on IoT and PPO algorithms. Background Technology

[0002] With the increasing demand for intelligent manufacturing and adaptive control of industrial environments, intelligent regulation and energy consumption optimization technologies for air conditioning systems in complex production spaces such as textile workshops have received widespread attention. Existing industrial air conditioning control systems mainly rely on fixed threshold settings or simple temperature and humidity feedback strategies for adjustment, but these still present the following problems in practical applications: Environmental monitoring data is diverse in dimensions and highly heterogeneous in time and space. There is a highly nonlinear coupling relationship between the air conditioning operation status in different areas and environmental disturbances, making it difficult for traditional methods to accurately model dynamic trends. There is a problem of linkage imbalance in comfort regulation between monitoring areas, which can easily cause some monitoring areas to be too cold or too hot, affecting production efficiency and worker health. The update of control strategies usually relies on manual setting or fixed parameter models, which is difficult to adapt to the dynamic adjustment needs under different seasons, different workshop loads and production conditions, resulting in high energy consumption and large fluctuations in comfort. Existing strategy networks based on single features or linear control logic lack robustness to sudden changes in environmental disturbances, and are prone to control lag or frequent oscillations in actual deployment, affecting the stability of the air conditioning system.

[0003] Therefore, how to provide a collaborative control and real-time monitoring system for air conditioning in textile workshops based on IoT and PPO algorithms is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a collaborative control and real-time monitoring system for air conditioning in textile workshops based on IoT and the PPO algorithm. This invention fully utilizes temporal modeling neural networks and reinforcement learning strategy optimization methods. It details how a modified ConvGRU network is used to perform temporal state modeling on monitoring data from the textile workshop, and how an improved PPO algorithm is used to generate an air conditioning control strategy. This invention possesses advantages such as high state modeling accuracy, strong intelligent control response, and strong collaborative optimization capabilities for comfort and energy consumption.

[0005] The textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm according to an embodiment of the present invention includes the following modules: The monitoring data acquisition module is used to collect raw monitoring data within the textile workshop; The data preprocessing module is used to perform data preprocessing on the raw monitoring data and generate standard monitoring tensors; The state modeling module is used to perform channel attention enhancement, temporal attention weighting, dilated convolutional gating modeling and state feature extraction on the standard monitoring tensor through an improved ConvGRU network to generate a monitoring state feature vector sequence. The improved ConvGRU network introduces attention guidance and multi-scale dilated convolutional gating structure. The control environment construction module is used to build a reinforcement learning control environment based on the monitoring state feature vector sequence; An air conditioning control strategy generation module is used to generate an air conditioning control strategy from a sequence of monitored state feature vectors using an improved PPO algorithm in a reinforcement learning control environment. The improved PPO algorithm introduces a multi-branch policy network and a state perturbation pruning mechanism. The control command execution module is used to execute control operations on the air conditioning equipment in the textile workshop through the control hold mechanism, which implements the air conditioning control strategy.

[0006] Optionally, the modules can be connected through the following steps: Step 1: Collect raw monitoring data within the textile workshop; Step 2: Perform data preprocessing on the raw monitoring data to generate a standard monitoring tensor; Step 3: Construct an improved ConvGRU network. The improved ConvGRU network introduces attention guidance and multi-scale dilated convolutional gating structure to enhance the attention of the standard monitoring tensor execution channel, perform temporal attention weighting, dilated convolutional gating modeling and state feature extraction, and generate a monitoring state feature vector sequence. Step 4: Construct a reinforcement learning control environment based on the monitoring state feature vector sequence; Step 5: Based on the reinforcement learning control environment, the monitoring state feature vector sequence is used to generate an air conditioning control strategy through an improved PPO algorithm; the improved PPO algorithm introduces a multi-branch policy network and a state perturbation pruning mechanism. Step Six: Based on the air conditioning control strategy, control operations are performed on the air conditioning equipment in the textile workshop through a control hold mechanism.

[0007] Optionally, step one specifically includes: The textile workshop is divided into monitoring areas, and IoT sensing nodes are deployed in each monitoring area to collect raw monitoring data. The IoT sensing nodes include temperature and humidity sensors, dust concentration sensors, carbon dioxide concentration sensors, air conditioning status acquisition modules, and power acquisition modules. The original monitoring data includes air temperature, air humidity, dust concentration, carbon dioxide concentration, air conditioning operating status parameters, set temperature, air supply volume, and energy consumption data.

[0008] Optionally, step two specifically includes: The data preprocessing includes data cleaning, time step alignment, missing value imputation, and normalization. The original monitoring data is labeled with time steps to generate the original monitoring data sequence; The original monitoring data sequence was cleaned by physical screening and the Z-Score method to obtain a cleaned monitoring data sequence. The cleaning monitoring data sequence is aligned by time steps according to a set time interval to obtain an aligned monitoring data sequence; Missing values ​​in the aligned monitoring data sequence are filled using linear interpolation to obtain the filled monitoring data sequence; Obtain the maximum and minimum values ​​of each type of monitoring data, and transform each monitoring data in the monitoring data sequence to the [0,1] interval using the minimum-maximum normalization method to obtain the normalized monitoring data sequence; The normalized monitoring data sequence is restructured according to the time step dimension × monitoring area dimension × monitoring feature dimension to obtain the standard monitoring tensor.

[0009] Optionally, step three specifically includes: The improved ConvGRU network includes an attention guidance module, a dilated convolutional gating module, and a state feature output module; In the attention guidance module, channel attention and temporal attention are applied to the standard monitoring tensor to generate an attention-enhanced monitoring matrix sequence; In the dilated convolutional gating module, the hidden state matrix is ​​initialized, and two dilation rates are set; the dilated convolutional gating module includes an update gate, a reset gate, and a candidate state gate. Perform convolution operations on the attention-enhanced monitoring matrix of the current time step with two different dilation rates, and then perform weighted fusion to obtain the dilated fused monitoring matrix. Perform convolution operations on the hidden state matrix of the previous time step with two different dilation rates, and then perform weighted fusion to obtain the dilated fused state matrix. At the update gate and reset gate, the extended fusion monitoring matrix and the extended fusion state matrix are linearly added together and then normalized by the Sigmoid activation function to obtain the update gate weight matrix and the reset gate weight matrix, respectively. At the candidate state gate, the extended fusion state matrix is ​​multiplied element-wise with the reset gate weight matrix and linearly added to the extended fusion monitoring matrix, and the candidate hidden state matrix is ​​generated by the Tanh activation function. Based on the updated gate weight matrix, the hidden state matrix of the previous time step and the candidate hidden state matrix are gatedly fused to obtain the hidden state matrix of the current time step. In the state feature output module, the hidden state matrix of each time step is averaged along the monitoring region dimension to obtain the monitoring state feature vector. The monitoring state feature vectors of all time steps are then arranged in the order of the time steps to form a monitoring state feature vector sequence.

[0010] Optionally, in the attention guidance module, channel attention and temporal attention are applied to the standard monitoring tensor to generate an attention-enhanced monitoring matrix sequence, specifically including: The standard monitoring tensor is subjected to two-dimensional average pooling in the time step dimension and the monitoring area dimension to obtain the global monitoring feature vector. The global monitoring feature vector is input into a two-layer fully connected network for compression and expansion to obtain the channel attention weight vector. The first fully connected network is followed by a ReLU activation function, and the second fully connected network is followed by a Sigmoid activation function. The channel attention weight vector is multiplied element-wise with the standard monitoring tensor along the monitoring feature dimension to obtain the channel weighted monitoring tensor; the channel weighted monitoring tensor is represented as a sequence of channel weighted monitoring matrices at the time step; At each time step, the channel-weighted monitoring matrix is ​​averaged along the monitoring region dimension to obtain the channel-weighted monitoring vector sequence; Based on the channel-weighted monitoring vector sequence, the channel-weighted monitoring vector of the current time step is mapped to a query vector through a linear transformation, and each channel-weighted monitoring vector of the historical time step is mapped to a key vector through a linear transformation. The historical time step refers to all time steps before the current time step. Based on the query vector and key vector, the relevance score between the current time step and each historical time step is calculated through attention operation, and all relevance scores are normalized to obtain the time attention weight vector of the current time step to the historical time step. The time attention weight vector is weighted and fused with the weighted monitoring matrix of all channels in the historical time step in the time step dimension to obtain the attention enhancement monitoring matrix. By concatenating the attention enhancement monitoring matrices at all time steps, a sequence of attention enhancement monitoring matrices is obtained.

[0011] Optionally, step four specifically includes: The reinforcement learning control environment includes a state space, an action space, and a reward function; The state space is composed of a sequence of monitoring state feature vectors, and the action space is defined as a set of control actions. The set of control actions includes start / stop control quantity, air volume adjustment quantity, and set temperature adjustment quantity for each monitoring area, wherein the start / stop control quantity takes the value of 0 or 1. The reward function is a weighted average of comfort index and energy consumption index.

[0012] Optionally, step five specifically includes: The improved PPO algorithm includes a multi-branch policy network, a state value network, and a pruning and update module; the multi-branch policy network includes a start-stop control sub-network, an air volume regulation sub-network, and a temperature regulation sub-network. The monitoring status feature vector of the current time step is input into the start-stop control sub-network. The monitoring status feature vector is subjected to feature mapping and nonlinear transformation through a two-layer fully connected network, and the start-stop control probability distribution is output through Softmax activation. The monitoring status feature vector of the current time step is input into the air volume regulation sub-network and the temperature regulation sub-network respectively. The air volume regulation sub-network and the temperature regulation sub-network have the same structure. They are linearly mapped through a fully connected network to generate the air volume regulation mean and the temperature regulation mean. They are also generated through another fully connected network and the Softplus activation function to generate the air volume regulation standard deviation and the temperature regulation standard deviation. Based on the mean and standard deviation of air volume adjustment, a probability distribution of air volume adjustment is generated using a Gaussian distribution; based on the mean and standard deviation of temperature adjustment, a probability distribution of temperature adjustment is generated using a Gaussian distribution. The probability distributions of start / stop control, air volume regulation, and temperature regulation are multiplied together to obtain the probability distribution of joint action. Based on the joint action probability distribution, control actions are sampled from the action space to generate start / stop control quantities, air volume adjustment quantities, and set temperature adjustment quantities for each monitoring area, and these are combined to form an air conditioning control strategy. Obtain the comfort index and energy consumption index corresponding to the air conditioning control strategy, and substitute them into the reward function to obtain the actual reward value; The monitoring state feature vector at the current time step is input into the state value network, and a value estimate is generated through a two-layer fully connected network. Calculate the difference between the actual return and the estimated value to obtain the advantage function value; Calculate the ratio of the joint action probability distribution of the current multi-branch policy network to the joint action probability distribution before the multi-branch policy network was updated, and obtain the policy probability ratio. In the pruning update module, the state perturbation amplitude is constructed to calculate the pruning threshold, and the pruning objective function value is calculated based on the state perturbation pruning mechanism; Based on the pruning objective function value, the multi-branch policy network is updated through backpropagation; the advantage function value is used as a supervision signal to synchronously update the state value network.

[0013] Optionally, in the pruning update module, the state perturbation amplitude is constructed to calculate the pruning threshold, and the pruning objective function value is calculated based on the state perturbation pruning mechanism, specifically as follows: The amplitude of the state disturbance is calculated based on the Euclidean distance between the monitoring state feature vectors of two consecutive time steps. Set a basic pruning threshold, map the state perturbation amplitude through a negative exponential function to obtain the perturbation adjustment factor, and multiply the basic pruning threshold by the perturbation adjustment factor to obtain the pruning threshold; With 1 as the center, construct the clipping interval based on the clipping threshold; Based on the pruning interval, the policy probability ratio is pruned to obtain the pruning policy probability ratio. The pruning policy probability ratio is then multiplied by the dominance function value to obtain the pruning update term. The updated term is obtained by multiplying the policy probability ratio by the advantage function value; The minimum value of the pruning update item and the update item is used as the pruning target value at the current time step, and the expected value of the pruning target value at all time steps is calculated to obtain the pruning objective function value.

[0014] Optionally, step six specifically includes: The control holding mechanism is specifically as follows: a control cycle is set, which includes several time steps. If the modulo operation between the current time step and the control cycle is equal to 0, the air conditioning control strategy of the current time step is used as the execution strategy within the control cycle; otherwise, the air conditioning control strategy of the previous control cycle is maintained. Based on the air conditioning control strategy, the start / stop control quantity of each monitoring area is converted into a start / stop control command. If the start / stop control command is equal to 1, an air conditioning start command is generated; if the start / stop control command is equal to 0, an air conditioning stop command is generated. Set the minimum air volume, maximum air volume, minimum set temperature, and maximum set temperature for each monitoring area; calculate the difference between the maximum and minimum air volume to obtain the air volume control range; calculate the difference between the maximum and minimum set temperatures to obtain the set temperature range. Multiply the air supply volume adjustment amount by the air supply volume control range and add it to the minimum air supply volume to obtain the target air supply volume control value; multiply the set temperature adjustment amount by the set temperature range and add it to the minimum set temperature to obtain the target set temperature control value. Based on the start / stop control command, the target air volume control value, and the target set temperature control value, control commands are issued to the air conditioning execution units in each monitoring area to control the air conditioning to complete start / stop control, air volume adjustment, and set temperature operation.

[0015] The beneficial effects of this invention are: First, this invention deploys IoT sensing nodes in the textile workshop according to the monitoring area through a monitoring data acquisition module to collect air temperature, humidity, dust concentration, carbon dioxide concentration, and air conditioning operating status parameters, ensuring the spatiotemporal distribution integrity and data source diversity of the monitoring data. Combined with a data preprocessing module, a standard monitoring tensor is constructed using physical filtering, the Z-Score method, linear interpolation, and normalization, effectively improving the consistency of the data temporal structure and the stability of the model input.

[0016] Secondly, the state modeling module introduces an improved ConvGRU network, combining an attention-guided module and a dilated convolutional gating module to enhance channel attention and weight temporal attention in the standard monitoring tensor, preserving the correlation between multi-scale information and key historical states in time-series modeling. The dilated convolutional gating module dynamically fuses environmental states through multi-scale convolution operations and a gating structure, achieving robust modeling of dynamically changing operating conditions and outputting a sequence of monitoring state feature vectors.

[0017] Furthermore, the control environment construction module and the air conditioning control strategy generation module proposed in this invention together constitute the core of reinforcement learning optimization. The reinforcement learning control environment is constructed based on the state space, action space, and reward function. The improved PPO algorithm introduces a multi-branch policy network and a state perturbation pruning mechanism, enhancing the expressive power of multivariable control output and the stability of policy updates, while maintaining excellent convergence performance and response speed even in the face of sudden environmental changes or load fluctuations.

[0018] In summary, this invention enables dynamic coordinated adjustment and real-time monitoring of the air conditioning system in textile workshops, improving the accuracy and efficiency of environmental comfort control in textile workshops and significantly reducing the energy consumption of the air conditioning system. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the module of the textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm proposed in this invention. Figure 2 This is a flowchart of the improved ConvGRU network structure in the textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm proposed in this invention. Figure 3 This is a flowchart of the improved PPO algorithm in the collaborative control and real-time monitoring system for air conditioning in a textile workshop based on IoT and PPO algorithm proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A collaborative control and real-time monitoring system for air conditioning in textile workshops based on IoT and PPO algorithms includes the following modules: The monitoring data acquisition module is used to collect raw monitoring data within the textile workshop; The data preprocessing module is used to perform data preprocessing on the raw monitoring data and generate standard monitoring tensors; The state modeling module is used to perform channel attention enhancement, temporal attention weighting, dilated convolutional gating modeling and state feature extraction on the standard monitoring tensor through an improved ConvGRU network to generate a monitoring state feature vector sequence. The improved ConvGRU network introduces attention guidance and multi-scale dilated convolutional gating structure. The control environment construction module is used to build a reinforcement learning control environment based on the monitoring state feature vector sequence; An air conditioning control strategy generation module is used to generate an air conditioning control strategy from a sequence of monitored state feature vectors using an improved PPO algorithm in a reinforcement learning control environment. The improved PPO algorithm introduces a multi-branch policy network and a state perturbation pruning mechanism. The control command execution module is used to execute control operations on the air conditioning equipment in the textile workshop through the control hold mechanism, which implements the air conditioning control strategy.

[0022] In this embodiment, the modules are connected through the following steps: Step 1: Collect raw monitoring data within the textile workshop; Step 2: Perform data preprocessing on the raw monitoring data to generate a standard monitoring tensor; Step 3: Construct an improved ConvGRU network. The improved ConvGRU network introduces attention guidance and multi-scale dilated convolutional gating structure to enhance the attention of the standard monitoring tensor execution channel, perform temporal attention weighting, dilated convolutional gating modeling and state feature extraction, and generate a monitoring state feature vector sequence. Step 4: Construct a reinforcement learning control environment based on the monitoring state feature vector sequence; Step 5: Based on the reinforcement learning control environment, the monitoring state feature vector sequence is used to generate an air conditioning control strategy through an improved PPO algorithm; the improved PPO algorithm introduces a multi-branch policy network and a state perturbation pruning mechanism. Step Six: Based on the air conditioning control strategy, control operations are performed on the air conditioning equipment in the textile workshop through a control hold mechanism.

[0023] In this embodiment, step one specifically includes: The textile workshop is divided into monitoring areas, and IoT sensing nodes are deployed in each monitoring area to collect raw monitoring data. The IoT sensing nodes include temperature and humidity sensors, dust concentration sensors, carbon dioxide concentration sensors, air conditioning status acquisition modules, and power acquisition modules. The original monitoring data includes air temperature, air humidity, dust concentration, carbon dioxide concentration, air conditioning operating status parameters, set temperature, air supply volume, and energy consumption data.

[0024] In this embodiment, step two specifically includes: The data preprocessing includes data cleaning, time step alignment, missing value imputation, and normalization. The original monitoring data is labeled with time steps to generate the original monitoring data sequence; The original monitoring data sequence was cleaned using physical screening and the Z-Score method to obtain a cleaned monitoring data sequence, specifically: Set the upper and lower limits of physical values ​​for each type of monitoring data, and remove monitoring data that are higher than the upper limit and lower than the lower limit of physical values ​​from the original monitoring data sequence to obtain the initial screening monitoring data sequence; Set the sliding window width, use the sliding window statistical method to calculate the local mean and local standard deviation of each type of initial screening monitoring data within the sliding window width, and use the Z-Score method to identify and remove anomalies in the initial screening monitoring data at each time step to obtain the cleaned monitoring data sequence. The cleaning monitoring data sequence is aligned by time steps according to a set time interval to obtain an aligned monitoring data sequence; Missing values ​​in the aligned monitoring data sequence are filled using linear interpolation to obtain the filled monitoring data sequence; Obtain the maximum and minimum values ​​of each type of monitoring data, and transform each monitoring data in the monitoring data sequence to the [0,1] interval using the minimum-maximum normalization method to obtain the normalized monitoring data sequence; The normalized monitoring data sequence is restructured according to the time step dimension × monitoring area dimension × monitoring feature dimension to obtain the standard monitoring tensor.

[0025] In this embodiment, step three specifically includes: The improved ConvGRU network includes an attention guidance module, a dilated convolutional gating module, and a state feature output module; In the attention guidance module, channel attention and temporal attention are applied to the standard monitoring tensor to generate an attention-enhanced monitoring matrix sequence; In the dilated convolutional gating module, the hidden state matrix is ​​initialized, and two dilation rates are set; the dilated convolutional gating module includes an update gate, a reset gate, and a candidate state gate. Perform convolution operations on the attention-enhanced monitoring matrix of the current time step with two different dilation rates, and then perform weighted fusion to obtain the dilated fused monitoring matrix. Perform convolution operations on the hidden state matrix of the previous time step with two different dilation rates, and then perform weighted fusion to obtain the dilated fused state matrix. At the update gate and reset gate, the extended fusion monitoring matrix and the extended fusion state matrix are linearly added together and then normalized by the Sigmoid activation function to obtain the update gate weight matrix and the reset gate weight matrix, respectively. At the candidate state gate, the extended fusion state matrix is ​​multiplied element-wise with the reset gate weight matrix and linearly added to the extended fusion monitoring matrix, and the candidate hidden state matrix is ​​generated by the Tanh activation function. Based on the updated gate weight matrix, the hidden state matrix of the previous time step and the candidate hidden state matrix are gatedly fused to obtain the hidden state matrix of the current time step. This invention introduces two convolution operations with different dilation rates and performs multi-scale dilated convolution modeling on the attention-enhanced monitoring matrix at the current time step and the hidden state matrix at the previous time step, achieving simultaneous capture of short-term local changes and medium-to-long-term trend information. This effectively overcomes the problem of insufficient temporal dependency modeling caused by traditional ConvGRU networks relying solely on a single receptive field. Simultaneously, the update gate, reset gate, and candidate state gate perform gating calculations based on the dilated fusion monitoring matrix, enabling the improved ConvGRU network to adaptively adjust the historical state retention ratio and new state injection intensity according to the magnitude of environmental state changes. This maintains stable and continuous state evolution under both abrupt environmental changes and stable operating scenarios, significantly improving the expressive power and robustness of the hidden state matrix for complex conditions.

[0026] In the state feature output module, the hidden state matrix of each time step is averaged along the monitoring region dimension to obtain the monitoring state feature vector. The monitoring state feature vectors of all time steps are then arranged in the order of the time steps to form a monitoring state feature vector sequence.

[0027] In this embodiment, the step of generating an attention-enhanced monitoring matrix sequence by performing channel attention and temporal attention on the standard monitoring tensor in the attention guidance module specifically includes: The standard monitoring tensor is subjected to two-dimensional average pooling in the time step dimension and the monitoring area dimension to obtain the global monitoring feature vector. The global monitoring feature vector is input into a two-layer fully connected network for compression and expansion to obtain the channel attention weight vector. The first fully connected network is followed by a ReLU activation function, and the second fully connected network is followed by a Sigmoid activation function. The channel attention weight vector is multiplied element-wise with the standard monitoring tensor along the monitoring feature dimension to obtain the channel weighted monitoring tensor; the channel weighted monitoring tensor is represented as a sequence of channel weighted monitoring matrices at the time step; At each time step, the channel-weighted monitoring matrix is ​​averaged along the monitoring region dimension to obtain the channel-weighted monitoring vector sequence; Based on the channel-weighted monitoring vector sequence, the channel-weighted monitoring vector of the current time step is mapped to a query vector through a linear transformation, and each channel-weighted monitoring vector of the historical time step is mapped to a key vector through a linear transformation. The historical time step refers to all time steps before the current time step. Based on the query vector and key vector, the relevance score between the current time step and each historical time step is calculated through attention operation, and all relevance scores are normalized to obtain the time attention weight vector of the current time step to the historical time step. The time attention weight vector is weighted and fused with the weighted monitoring matrix of all channels in the historical time step in the time step dimension to obtain the attention enhancement monitoring matrix. By concatenating the attention enhancement monitoring matrices at all time steps, a sequence of attention enhancement monitoring matrices is obtained.

[0028] In this invention, the channel attention mechanism generates channel attention weights based on the global monitoring feature vector, dynamically adjusting the importance of different monitoring feature dimensions. This allows the improved ConvGRU network to focus on more discriminative feature channels, improving the accuracy of environmental state representation. The temporal attention mechanism constructs the attention weight distribution between the current time step and historical time steps based on the channel-weighted monitoring vector sequence. By weightedly fusing historical information, it effectively enhances the improved ConvGRU network's ability to perceive key temporal dependencies and its long-term information memory capacity. Through these improvements, the improved ConvGRU network exhibits stronger adaptability and robustness in dealing with environmental disturbances and state abrupt changes.

[0029] In this embodiment, step four specifically includes: The reinforcement learning control environment includes a state space, an action space, and a reward function; The state space is composed of a sequence of monitoring state feature vectors, and the action space is defined as a set of control actions. The set of control actions includes start / stop control quantity, air volume adjustment quantity, and set temperature adjustment quantity for each monitoring area, wherein the start / stop control quantity takes the value of 0 or 1. The reward function is a weighted sum of comfort index and energy consumption index. The comfort index represents the environmental comfort score of the textile workshop based on the current set of control actions, and the energy consumption index represents the air conditioning energy consumption value corresponding to the current set of control actions. The environmental comfort score is generated by combining the current monitoring state feature vector and the set of control actions through a multilayer perceptron regression network. The air conditioning energy consumption value is obtained in real time through the power acquisition module deployed in each monitoring area. In this embodiment, step five specifically includes: The improved PPO algorithm includes a multi-branch policy network, a state value network, and a pruning and update module; the multi-branch policy network includes a start-stop control sub-network, an air volume regulation sub-network, and a temperature regulation sub-network. The monitoring status feature vector of the current time step is input into the start-stop control sub-network. The monitoring status feature vector is subjected to feature mapping and nonlinear transformation through a two-layer fully connected network, and the start-stop control probability distribution is output through Softmax activation. The monitoring status feature vector of the current time step is input into the air volume regulation sub-network and the temperature regulation sub-network respectively. The air volume regulation sub-network and the temperature regulation sub-network have the same structure. They are linearly mapped through a fully connected network to generate the air volume regulation mean and the temperature regulation mean. They are also generated through another fully connected network and the Softplus activation function to generate the air volume regulation standard deviation and the temperature regulation standard deviation. Based on the mean and standard deviation of air volume adjustment, a probability distribution of air volume adjustment is generated using a Gaussian distribution; based on the mean and standard deviation of temperature adjustment, a probability distribution of temperature adjustment is generated using a Gaussian distribution. The probability distributions of start / stop control, air volume regulation, and temperature regulation are multiplied together to obtain the probability distribution of joint action. In this invention, the multi-branch policy network is divided into a start-stop control sub-network, an air volume regulation sub-network, and a temperature regulation sub-network. This enables decoupled modeling of different control dimensions of the air conditioning system, effectively improving the policy network's expressive power and training efficiency in the multi-dimensional action space. The start-stop control sub-network outputs discrete binary classification actions through Softmax activation, which is beneficial for learning the policy boundaries of air conditioning on and off. The air volume regulation and temperature regulation sub-networks adopt Gaussian distribution structures modeled with mean and standard deviation, which are more suitable for handling continuous control variables. Through structural separation, different sub-networks can be independently optimized according to their respective task characteristics, thereby avoiding action conflicts and gradient interference problems. The probability distributions output by each sub-network are jointly modeled to construct a complete joint action probability distribution, enabling the system to achieve collaborative optimization among multiple control variables, improving policy accuracy and control robustness.

[0030] Based on the joint action probability distribution, control actions are sampled from the action space to generate start / stop control quantities, air volume adjustment quantities, and set temperature adjustment quantities for each monitoring area, and these are combined to form an air conditioning control strategy. Obtain the comfort index and energy consumption index corresponding to the air conditioning control strategy, and substitute them into the reward function to obtain the actual reward value; The monitoring state feature vector at the current time step is input into the state value network, and a value estimate is generated through a two-layer fully connected network. Calculate the difference between the actual return and the estimated value to obtain the advantage function value; Calculate the ratio of the joint action probability distribution of the current multi-branch policy network to the joint action probability distribution before the multi-branch policy network was updated, and obtain the policy probability ratio. In the pruning update module, the state perturbation amplitude is constructed to calculate the pruning threshold, and the pruning objective function value is calculated based on the state perturbation pruning mechanism; Based on the pruning objective function value, the multi-branch policy network is updated through backpropagation; the advantage function value is used as a supervision signal to synchronously update the state value network.

[0031] In this embodiment, the step of constructing the state perturbation amplitude to calculate the pruning threshold in the pruning update module, and calculating the pruning objective function value based on the state perturbation pruning mechanism, specifically involves: The amplitude of the state disturbance is calculated based on the Euclidean distance between the monitoring state feature vectors of two consecutive time steps. Set a basic pruning threshold, map the state perturbation amplitude through a negative exponential function to obtain the perturbation adjustment factor, and multiply the basic pruning threshold by the perturbation adjustment factor to obtain the pruning threshold; With 1 as the center, construct the clipping interval [1-e, 1+e] based on the clipping threshold e; Based on the pruning interval, the policy probability ratio is pruned to obtain the pruning policy probability ratio. The pruning policy probability ratio is then multiplied by the dominance function value to obtain the pruning update term. The updated term is obtained by multiplying the policy probability ratio by the advantage function value; The minimum value of the pruning update item and the update item is used as the pruning target value at the current time step, and the expected value of the pruning target value at all time steps is calculated to obtain the pruning objective function value.

[0032] In this invention, the state perturbation pruning mechanism introduces the Euclidean distance between the monitored state feature vectors of consecutive time steps as the state perturbation amplitude, adaptively adjusting the pruning threshold for policy updates. This effectively solves the problem of excessively fast or slow policy updates in traditional PPO algorithms under scenarios of drastic state changes or sudden environmental changes. The state perturbation pruning mechanism uses a negative exponential function to map the state perturbation amplitude to a perturbation adjustment factor, dynamically scaling or widening the pruning interval. This enhances policy convergence speed when the state is stable and suppresses drastic updates when the state is unstable, improving the stability and generalization ability of policy learning. Furthermore, this mechanism effectively alleviates the overfitting problem and enhances the adaptability and robustness of the policy network to complex temporal environments.

[0033] In this embodiment, step six specifically includes: The control holding mechanism is specifically as follows: a control cycle is set, which includes several time steps. If the modulo operation between the current time step and the control cycle is equal to 0, the air conditioning control strategy of the current time step is used as the execution strategy within the control cycle; otherwise, the air conditioning control strategy of the previous control cycle is maintained. Based on the air conditioning control strategy, the start / stop control quantity of each monitoring area is converted into a start / stop control command. If the start / stop control command is equal to 1, an air conditioning start command is generated; if the start / stop control command is equal to 0, an air conditioning stop command is generated. Set the minimum air volume, maximum air volume, minimum set temperature, and maximum set temperature for each monitoring area; calculate the difference between the maximum and minimum air volume to obtain the air volume control range; calculate the difference between the maximum and minimum set temperatures to obtain the set temperature range. Multiply the air supply volume adjustment amount by the air supply volume control range and add it to the minimum air supply volume to obtain the target air supply volume control value; multiply the set temperature adjustment amount by the set temperature range and add it to the minimum set temperature to obtain the target set temperature control value. Based on the start / stop control command, the target air volume control value, and the target set temperature control value, control commands are issued to the air conditioning execution units in each monitoring area to control the air conditioning to complete start / stop control, air volume adjustment, and set temperature operation.

[0034] Example 1 To verify the feasibility of this invention in practice, it was applied to a large textile production workshop for practical verification. The workshop has a floor area of ​​approximately 8,000 square meters and is divided into 12 monitoring zones. The workshop is equipped with a ceiling-mounted air conditioning system. Traditional control methods mainly rely on temperature threshold settings and timer switching, which makes it difficult to achieve dynamic adjustment and energy efficiency optimization.

[0035] During implementation, temperature and humidity sensors, dust concentration sensors, carbon dioxide sensors, power metering modules, and air conditioning operation status acquisition devices are deployed in each monitoring area. Data is collected every 30 seconds, and the raw monitoring data is preprocessed to generate a standard monitoring tensor. In the state modeling phase, an improved ConvGRU network is used to transform the standard monitoring tensor into a sequence of monitoring state feature vectors. In the control core, an improved PPO algorithm is used to generate the air conditioning control strategy. During control execution, a control hold mechanism is used to avoid frequent control switching due to fluctuations. The system evaluates the control strategy every 10 minutes and issues control commands uniformly.

[0036] To verify the control and energy-saving effects of the system of the present invention, the system of the present invention was compared with the PPO algorithm control system and the ConvGRU+PPO algorithm control system in experiments, with an experimental period of 30 consecutive days. The comparison indicators included average energy consumption, PM2.5 control qualification rate, CO2 control qualification rate, temperature standard deviation, humidity standard deviation, strategy convergence time, and comfort score. The experimental results are shown in Table 1.

[0037] Table 1. Performance Comparison of Different Invented Systems in Collaborative Control of Air Conditioning in Textile Workshops

[0038] As shown in Table 1, the system of this invention significantly outperforms the comparative system in multiple comparative indicators. Regarding energy consumption, the average energy consumption of the system of this invention is 4365.2 kWh, which is 555.6 kWh less than the PPO algorithm control system and 338.4 kWh less than the ConvGRU+PPO algorithm control system, representing reductions of approximately 11.3% and 7.2% respectively, significantly improving the operating energy efficiency of the air conditioning system. In terms of air quality control, the PM2.5 compliance rate of the system of this invention reaches 96.3%, and the CO2 compliance rate reaches 94.7%, representing improvements of 7.9 and 7.8 percentage points respectively compared to the PPO algorithm control system, and improvements of 5.1 and 4.9 percentage points respectively compared to the ConvGRU+PPO control system, demonstrating that the system of this invention has higher precision in pollutant control.

[0039] Furthermore, regarding environmental stability, the temperature standard deviation of the system of this invention is only ±0.82°C, and the humidity standard deviation is ±2.8%RH, both significantly better than the comparative system. This indicates that the system of this invention can respond and adjust more quickly and smoothly when facing environmental disturbances, ensuring the thermal comfort of the workshop. In terms of strategy learning efficiency, the strategy convergence time of the system of this invention is 37.4 minutes, which is more than half shorter than the PPO algorithm control system and about 36% shorter than the ConvGRU+PPO algorithm control system, demonstrating the practical effectiveness of the introduced state disturbance pruning mechanism in improving learning efficiency and control response speed. In terms of user experience, the system of this invention achieved an average comfort score of 91.6 points during 30 days of operation, 13.3 points higher than the PPO algorithm control system and 7.9 points higher than the ConvGRU+PPO algorithm control system, fully demonstrating that the system of this invention can both ensure energy saving and provide a superior working environment experience.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm, characterized in that, include: The monitoring data acquisition module is used to collect raw monitoring data within the textile workshop; The data preprocessing module is used to perform data preprocessing on the raw monitoring data and generate standard monitoring tensors; The state modeling module is used to perform channel attention enhancement, temporal attention weighting, dilated convolutional gating modeling and state feature extraction on the standard monitoring tensor through an improved ConvGRU network to generate a monitoring state feature vector sequence. The improved ConvGRU network introduces attention guidance and multi-scale dilated convolutional gating structure. The control environment construction module is used to build a reinforcement learning control environment based on the monitoring state feature vector sequence; An air conditioning control strategy generation module is used to generate an air conditioning control strategy from a sequence of monitored state feature vectors using an improved PPO algorithm in a reinforcement learning control environment. The improved PPO algorithm introduces a multi-branch policy network and a state perturbation pruning mechanism. The control command execution module is used to execute control operations on the air conditioning equipment in the textile workshop through the control hold mechanism, which implements the air conditioning control strategy.

2. The textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm according to claim 1, characterized in that, The modules are connected through the following steps: Step 1: Collect raw monitoring data within the textile workshop; Step 2: Perform data preprocessing on the raw monitoring data to generate a standard monitoring tensor; Step 3: Construct an improved ConvGRU network. The improved ConvGRU network introduces attention guidance and multi-scale dilated convolutional gating structure to enhance the attention of the standard monitoring tensor execution channel, perform temporal attention weighting, dilated convolutional gating modeling and state feature extraction, and generate a monitoring state feature vector sequence. Step 4: Construct a reinforcement learning control environment based on the monitoring state feature vector sequence; Step 5: Based on the reinforcement learning control environment, the monitoring state feature vector sequence is used to generate an air conditioning control strategy through an improved PPO algorithm; the improved PPO algorithm introduces a multi-branch policy network and a state perturbation pruning mechanism. Step Six: Based on the air conditioning control strategy, control operations are performed on the air conditioning equipment in the textile workshop through a control hold mechanism.

3. The textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm according to claim 2, characterized in that, Step one specifically includes: The textile workshop is divided into monitoring areas, and IoT sensing nodes are deployed in each monitoring area to collect raw monitoring data. The IoT sensing nodes include temperature and humidity sensors, dust concentration sensors, carbon dioxide concentration sensors, air conditioning status acquisition modules, and power acquisition modules. The original monitoring data includes air temperature, air humidity, dust concentration, carbon dioxide concentration, air conditioning operating status parameters, set temperature, air supply volume, and energy consumption data.

4. The textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm according to claim 2, characterized in that, Step two specifically includes: The data preprocessing includes data cleaning, time step alignment, missing value imputation, and normalization. The original monitoring data is labeled with time steps to generate the original monitoring data sequence; The original monitoring data sequence was cleaned by physical screening and the Z-Score method to obtain a cleaned monitoring data sequence. The cleaning monitoring data sequence is aligned by time steps according to a set time interval to obtain an aligned monitoring data sequence; Missing values ​​in the aligned monitoring data sequence are filled using linear interpolation to obtain the filled monitoring data sequence; Obtain the maximum and minimum values ​​of each type of monitoring data, and transform each monitoring data in the monitoring data sequence to the [0,1] interval using the minimum-maximum normalization method to obtain the normalized monitoring data sequence; The normalized monitoring data sequence is restructured according to the time step dimension × monitoring area dimension × monitoring feature dimension to obtain the standard monitoring tensor.

5. The textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm according to claim 2, characterized in that, Step three specifically includes: The improved ConvGRU network includes an attention guidance module, a dilated convolutional gating module, and a state feature output module; In the attention guidance module, channel attention and temporal attention are applied to the standard monitoring tensor to generate an attention-enhanced monitoring matrix sequence; In the dilated convolutional gating module, the hidden state matrix is ​​initialized, and two dilation rates are set; the dilated convolutional gating module includes an update gate, a reset gate, and a candidate state gate. Perform convolution operations on the attention-enhanced monitoring matrix of the current time step with two different dilation rates, and then perform weighted fusion to obtain the dilated fused monitoring matrix. Perform convolution operations on the hidden state matrix of the previous time step with two different dilation rates, and then perform weighted fusion to obtain the dilated fused state matrix. At the update gate and reset gate, the extended fusion monitoring matrix and the extended fusion state matrix are linearly added together and then normalized by the Sigmoid activation function to obtain the update gate weight matrix and the reset gate weight matrix, respectively. At the candidate state gate, the extended fusion state matrix is ​​multiplied element-wise with the reset gate weight matrix and linearly added to the extended fusion monitoring matrix, and the candidate hidden state matrix is ​​generated by the Tanh activation function. Based on the updated gate weight matrix, the hidden state matrix of the previous time step and the candidate hidden state matrix are gatedly fused to obtain the hidden state matrix of the current time step. In the state feature output module, the hidden state matrix of each time step is averaged along the monitoring region dimension to obtain the monitoring state feature vector. The monitoring state feature vectors of all time steps are then arranged in the order of the time steps to form a monitoring state feature vector sequence.

6. The textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm according to claim 5, characterized in that, In the attention guidance module, channel attention and temporal attention are applied to the standard monitoring tensor to generate an attention-enhanced monitoring matrix sequence, specifically including: The standard monitoring tensor is subjected to two-dimensional average pooling in the time step dimension and the monitoring area dimension to obtain the global monitoring feature vector. The global monitoring feature vector is input into a two-layer fully connected network for compression and expansion to obtain the channel attention weight vector. The first fully connected network is followed by a ReLU activation function, and the second fully connected network is followed by a Sigmoid activation function. The channel attention weight vector is multiplied element-wise with the standard monitoring tensor along the monitoring feature dimension to obtain the channel weighted monitoring tensor; the channel weighted monitoring tensor is represented as a sequence of channel weighted monitoring matrices at the time step; At each time step, the channel-weighted monitoring matrix is ​​averaged along the monitoring region dimension to obtain the channel-weighted monitoring vector sequence; Based on the channel-weighted monitoring vector sequence, the channel-weighted monitoring vector of the current time step is mapped to a query vector through a linear transformation, and each channel-weighted monitoring vector of the historical time step is mapped to a key vector through a linear transformation. The historical time step refers to all time steps before the current time step. Based on the query vector and key vector, the relevance score between the current time step and each historical time step is calculated through attention operation, and all relevance scores are normalized to obtain the time attention weight vector of the current time step to the historical time step. The time attention weight vector is weighted and fused with the weighted monitoring matrix of all channels in the historical time step in the time step dimension to obtain the attention enhancement monitoring matrix. By concatenating the attention enhancement monitoring matrices at all time steps, a sequence of attention enhancement monitoring matrices is obtained.

7. The textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm according to claim 2, characterized in that, Step four specifically includes: The reinforcement learning control environment includes a state space, an action space, and a reward function; The state space is composed of a sequence of monitoring state feature vectors, and the action space is defined as a set of control actions. The set of control actions includes start / stop control quantity, air volume adjustment quantity, and set temperature adjustment quantity for each monitoring area, wherein the start / stop control quantity takes the value of 0 or 1. The reward function is a weighted average of comfort index and energy consumption index.

8. The textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm according to claim 2, characterized in that, Step five specifically includes: The improved PPO algorithm includes a multi-branch policy network, a state value network, and a pruning and update module; the multi-branch policy network includes a start-stop control sub-network, an air volume regulation sub-network, and a temperature regulation sub-network. The monitoring status feature vector of the current time step is input into the start-stop control sub-network. The monitoring status feature vector is subjected to feature mapping and nonlinear transformation through a two-layer fully connected network, and the start-stop control probability distribution is output through Softmax activation. The monitoring status feature vector of the current time step is input into the air volume regulation sub-network and the temperature regulation sub-network respectively. The air volume regulation sub-network and the temperature regulation sub-network have the same structure. They are linearly mapped through a fully connected network to generate the air volume regulation mean and the temperature regulation mean. They are also generated through another fully connected network and the Softplus activation function to generate the air volume regulation standard deviation and the temperature regulation standard deviation. Based on the mean and standard deviation of air volume adjustment, a probability distribution of air volume adjustment is generated using a Gaussian distribution; based on the mean and standard deviation of temperature adjustment, a probability distribution of temperature adjustment is generated using a Gaussian distribution. The probability distributions of start / stop control, air volume regulation, and temperature regulation are multiplied together to obtain the probability distribution of joint action. Based on the joint action probability distribution, control actions are sampled from the action space to generate start / stop control quantities, air volume adjustment quantities, and set temperature adjustment quantities for each monitoring area, and these are combined to form an air conditioning control strategy. Obtain the comfort index and energy consumption index corresponding to the air conditioning control strategy, and substitute them into the reward function to obtain the actual reward value; The monitoring state feature vector at the current time step is input into the state value network, and a value estimate is generated through a two-layer fully connected network. Calculate the difference between the actual return and the estimated value to obtain the advantage function value; Calculate the ratio of the joint action probability distribution of the current multi-branch policy network to the joint action probability distribution before the multi-branch policy network was updated, and obtain the policy probability ratio. In the pruning update module, the state perturbation amplitude is constructed to calculate the pruning threshold, and the pruning objective function value is calculated based on the state perturbation pruning mechanism; Based on the pruning objective function value, the multi-branch policy network is updated through backpropagation; the advantage function value is used as a supervision signal to synchronously update the state value network.

9. The textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm according to claim 8, characterized in that, In the pruning update module, the state perturbation amplitude is constructed to calculate the pruning threshold, and the pruning objective function value is calculated based on the state perturbation pruning mechanism, specifically as follows: The amplitude of the state disturbance is calculated based on the Euclidean distance between the monitoring state feature vectors of two consecutive time steps. Set a basic pruning threshold, map the state perturbation amplitude through a negative exponential function to obtain the perturbation adjustment factor, and multiply the basic pruning threshold by the perturbation adjustment factor to obtain the pruning threshold; With 1 as the center, construct the clipping interval based on the clipping threshold; Based on the pruning interval, the policy probability ratio is pruned to obtain the pruning policy probability ratio. The pruning policy probability ratio is then multiplied by the dominance function value to obtain the pruning update term. The updated term is obtained by multiplying the policy probability ratio by the advantage function value; The minimum value of the pruning update item and the update item is used as the pruning target value at the current time step, and the expected value of the pruning target value at all time steps is calculated to obtain the pruning objective function value.

10. The textile workshop air conditioning collaborative control and real-time monitoring system based on IoT and PPO algorithm according to claim 2, characterized in that, Step six specifically includes: The control and hold mechanism is as follows: a control cycle is set, which includes several time steps. If the modulo operation between the current time step and the control cycle is equal to 0, the air conditioning control strategy of the current time step is used as the execution strategy within the control cycle; otherwise, the air conditioning control strategy of the previous control cycle is maintained. Based on the air conditioning control strategy, the start / stop control quantity of each monitoring area is converted into a start / stop control command. If the start / stop control command is equal to 1, an air conditioning start command is generated; if the start / stop control command is equal to 0, an air conditioning stop command is generated. Set the minimum air volume, maximum air volume, minimum set temperature, and maximum set temperature for each monitoring area; calculate the difference between the maximum and minimum air volume to obtain the air volume control range; calculate the difference between the maximum and minimum set temperatures to obtain the set temperature range. Multiply the air volume adjustment amount by the air volume control range and add it to the minimum air volume to obtain the target air volume control value; multiply the set temperature adjustment amount by the set temperature range and add it to the minimum set temperature to obtain the target set temperature control value. Based on the start / stop control command, the target air supply volume control value, and the target set temperature control value, control commands are issued to the air conditioning execution units in each monitoring area to control the air conditioning to complete start / stop control, air supply volume adjustment, and set temperature operation.

Images