Sensor network-based spinning workshop temperature and humidity control method

Abstract

The application discloses a spinning workshop temperature and humidity control method based on a sensor network, which comprises the following steps: step one, constructing a spinning workshop sensor network for the spinning workshop and generating a workshop area node set; step two, collecting original spinning data sets and generating standardized time series data blocks; step three, calculating the spatial adjacency coefficient, air supply transmission coefficient, return air coupling coefficient and thermal disturbance influence coefficient between the area nodes, constructing an area correlation matrix and a control object matrix; step four, time domain unfolding the area correlation matrix and the control object matrix, and generating an input sample tensor; step five, inputting the input sample tensor into an improved StemGNN model, and outputting a prediction result matrix; and step six, calculating a target control correction matrix of each area node according to the prediction result matrix, and generating corresponding execution instructions. The improved StemGNN model is used to realize accurate adjustment of the temperature and humidity of the spinning workshop.

Description

Technical Field

[0001] This invention relates to the field of environmental control technology, and in particular to a method for controlling temperature and humidity in a spinning workshop based on sensor networks. Background Technology

[0002] Spinning workshops are highly dependent on temperature and humidity conditions during production. The stability of environmental parameters within the workshop directly affects fiber opening, drafting, twisting, and yarn quality, further influencing breakage rate, fly waste, and equipment operational stability. Existing spinning workshops typically use air conditioning units, supply and return air systems, and humidification devices to regulate the workshop environment, and monitor environmental conditions using temperature and humidity sensors. Current technologies mostly rely on local or limited fixed-point data collection, combined with manual experience or simple threshold control strategies to adjust the supply, return, and humidification equipment, thereby achieving overall temperature and humidity control within the workshop.

[0003] However, as the density of equipment layout in spinning workshops continues to increase, the differences in temperature and humidity distribution across different workstations, machines, and supply and return air paths are becoming increasingly apparent. Traditional methods, which use a small number of measuring points to reflect the overall environmental state, are insufficient to accurately characterize the actual environmental changes in different areas within the workshop. Furthermore, existing technologies mostly only perform simple summaries or static comparisons of the collected temperature, humidity, wind speed, and equipment operation data, lacking comprehensive modeling of spatial adjacency relationships, supply air transmission relationships, return air coupling relationships, and thermal disturbance effects between regional nodes. This makes it difficult for the control process to reflect the interconnected effects between different areas. Moreover, traditional control methods generally focus on whether environmental values ​​exceed limits at the current moment, lacking a predictive mechanism for environmental change trends over multiple future control cycles. Therefore, when equipment load changes, heat source fluctuations, or supply and return air disturbances occur, problems such as response lag, insufficient control accuracy, and localized imbalances often arise.

[0004] Therefore, how to provide a sensor network-based method for temperature and humidity control in spinning workshops is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a sensor network-based method for temperature and humidity control in spinning workshops. This invention constructs a sensor network in the spinning workshop, standardizes the environmental data of regional nodes, and establishes a regional correlation model by combining spatial adjacency, air supply transmission, return air coupling, and thermal disturbance relationships. Then, it uses an improved StemGNN model to achieve multi-control cycle prediction, thereby generating target control correction quantities and execution instructions, and realizing precise regional adjustment and forward-looking control of temperature and humidity in the spinning workshop.

[0006] The temperature and humidity control method for a spinning workshop based on a sensor network according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Define the different distribution boundaries of the spinning workshop, construct a sensor network for the spinning workshop, and generate a set of nodes for the workshop area;

[0008] Step 2: Collect the original spinning dataset of each region node in the workshop area node set and generate standardized time-series data blocks;

[0009] Step 3: Based on the standardized time-series data blocks, calculate the spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient between regional nodes, and construct the regional correlation matrix and the control object matrix;

[0010] Step 4: Perform temporal expansion on the regional correlation matrix and the control object matrix, and generate the input sample tensor;

[0011] Step 5: Input the input sample tensor into the improved StemGNN model, and output the prediction result matrix for multiple future control cycles through the graph structure generation module, frequency domain decomposition module, time-frequency coupling propagation module and multi-step prediction output module.

[0012] Step 6: Calculate the target control correction matrix for each regional node based on the prediction result matrix, and generate the corresponding execution instructions.

[0013] Optionally, step one specifically includes:

[0014] Based on the different distribution boundaries within the spinning workshop, the planar area of ​​the spinning workshop is divided and labeled to generate different node boundaries;

[0015] Multiple sensor installation points are set up along the length, width, and height of the machine within the boundary of each partition node. Temperature sensor, humidity sensor, wind speed sensor, equipment condition acquisition node or air valve status acquisition node are installed at each sensor installation point to construct a sensor network for the spinning workshop.

[0016] Write a unique node number for each sensor node that has been installed in the construction of the sensor network in the spinning workshop, write the partition number according to the partition node boundary where the current sensor node is located, write the spatial coordinates and height coordinates according to the horizontal, vertical and vertical positions of each sensor node in the workshop coordinate system, and write the sampling period parameter according to the data acquisition frequency corresponding to the sensor node.

[0017] According to the machine layout direction, air supply and return direction and spatial location, multiple sensor nodes are grouped into corresponding area nodes to form a workshop area node set.

[0018] Optionally, step two specifically involves:

[0019] In the workshop area node set, control each sensor node to synchronously collect temperature value, relative humidity value, wind speed value, machine operating frequency value, motor load value, air supply outlet opening value, return air outlet opening value and humidification device operating status value according to a unified sampling period to obtain the original spinning dataset;

[0020] The original spinning dataset is resampled, missing data is filled in, anomalies are removed and units are standardized according to a unified timestamp. Data from multiple sampling points in the same region are aggregated in units of regional nodes to generate standardized time-series data blocks corresponding to each regional node.

[0021] Optionally, step three specifically includes:

[0022] Based on the spatial and height coordinates of each region node in the standardized time series data block, the planar distance and vertical height difference between any two region nodes are calculated one by one, and compared with the preset distance threshold and height difference threshold respectively. Region node pairs that are both less than the preset threshold are selected as adjacent region node pairs.

[0023] The spatial adjacency coefficient between nodes in each region is obtained by taking the weighted average of the inverse of the planar distance and the vertical height difference between each pair of adjacent region nodes.

[0024] The air supply outlet opening difference, the distance between nodes along the air supply direction, and the length of the air supply path are weighted according to preset weights to obtain the air supply transmission coefficient between nodes in each region; and the return air outlet opening difference, the distance between nodes along the return air direction, and the length of the return air path overlap are weighted according to preset weights to obtain the return air coupling coefficient between nodes in each region.

[0025] The spatial distance between the heat source region and the target region, the load variation of the heat source side equipment, and the temperature variation within adjacent sampling periods are weighted according to preset weights to obtain the thermal disturbance influence coefficient of the heat source region node on the target region node.

[0026] Based on the spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient, a regional correlation matrix and a control object matrix are generated.

[0027] Optionally, the step of generating the regional correlation matrix and the control object matrix based on the spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient specifically involves:

[0028] The spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient between any two regional nodes are located according to the row and column order of the regional nodes. The spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient of the corresponding regional node pair are filled in the column positions of the same row in sequence to generate a regional association matrix arranged in the order of regional node numbers.

[0029] The current temperature, humidity, wind speed, equipment load, and actuator status values ​​of each regional node are concatenated according to the fixed field order within the same node to form a state feature vector of a single regional node. Then, all state feature vectors are arranged row by row according to the node order of the regional nodes to generate a control object matrix.

[0030] Optionally, step four specifically includes:

[0031] Following the time-incrementing direction, the control object matrices from multiple consecutive sampling periods are stacked layer by layer along the time dimension to generate a node time series matrix that reflects the continuous change process of nodes in each region.

[0032] Set the sliding time window length and sliding step size for the node time sequence matrix, and start from the position corresponding to the earliest sampling period. Move the sliding time window along the time dimension one by one according to the preset sliding step size. At each sliding position, extract the continuous node time sequence sub-matrix within the corresponding time range to generate multiple time window candidate segments arranged in time order.

[0033] For each time window candidate segment, extract all regional node temporal features contained between the start and end times, while keeping the order of regional nodes, time order, and internal field order of nodes unchanged, and organize each time window candidate segment into a corresponding time window sample feature block;

[0034] For each time window sample feature block, read the region association matrix corresponding to the unified timestamp of each sampling period within the current time window, and extract the region association matrix sequence of each sampling period according to the time order within the time window to form a node association relationship subsequence that matches the time window sample feature block;

[0035] The region node feature sequence in each time window sample feature block is combined with the corresponding node association subsequence and written to generate a single time window sample;

[0036] All samples from individual time windows are collected in the order in which they were generated, and then uniformly arranged according to sample number, time window length, number of regional nodes, and node feature dimension to form the input sample tensor.

[0037] Optionally, the improved StemGNN model is specifically as follows:

[0038] The input sample tensor is input into the graph structure generation module, and the temperature change sequence, humidity change sequence, wind speed change sequence, equipment load change sequence, and execution status change sequence of each region node in the input sample tensor within the continuous time window are used as node features at the current time step.

[0039] Extract the node features of each regional node at each time step according to the order of regional nodes, and splice the node features of the same regional node in a continuous time window to form the historical representation of the corresponding regional node.

[0040] The difference between the historical representations of any two regional nodes is taken as the node association strength between the corresponding regional nodes, and a dynamic association graph matrix reflecting the dynamic association relationship of regional nodes under the current time window is generated.

[0041] The dynamic correlation graph matrix is ​​input into the frequency domain decomposition module. The temperature change sequence, humidity change sequence and wind speed change sequence of each regional node corresponding to the dynamic correlation graph matrix within the continuous time window are decomposed in the frequency domain by the fast Fourier transform. The time domain change sequence corresponding to each regional node is converted into the frequency domain response sequence. The trend component, periodic component and disturbance component are extracted according to the preset frequency range to generate the frequency domain feature tensor of the regional node.

[0042] The frequency domain feature tensor of the regional nodes is input into the time-frequency coupled propagation module. The dynamic correlation graph matrix is ​​used as the information propagation constraint relationship between regional nodes. The change trend part, periodic fluctuation part and instantaneous disturbance part of each regional node are propagated along the dynamic correlation path between regional nodes, forming a joint state representation that can reflect the current state of the regional node itself and the propagation influence of adjacent regional nodes.

[0043] Arrange the joint state representations of all region nodes in the order of region nodes to generate a joint state representation matrix;

[0044] The joint state representation matrix is ​​input to the multi-step prediction output module, and the joint state representation of each region node is decoded step by step with a preset length. The step by step decoding process is as follows:

[0045] Temperature prediction features, humidity prediction features, and wind speed prediction features are extracted from the joint state representation of each regional node, and temperature prediction sequences, humidity prediction sequences, and wind speed prediction sequences corresponding to each regional node in multiple control cycles are generated.

[0046] According to the order of regional nodes and the order of prediction time, the temperature prediction sequence, humidity prediction sequence and wind speed prediction sequence of all regional nodes in multiple control cycles are arranged to form a prediction result matrix.

[0047] Optionally, step six specifically includes:

[0048] The temperature prediction sequence, humidity prediction sequence and wind speed prediction sequence of each region node in the prediction result matrix are compared with the preset temperature control interval, preset humidity control interval and preset wind speed control interval at each time step.

[0049] For prediction times when the predicted temperature is higher than the upper limit of the temperature control range, determine the cooling correction requirement for the corresponding regional nodes; for prediction times when the predicted temperature is lower than the lower limit of the temperature control range, determine the heating correction requirement for the corresponding regional nodes; and generate the temperature correction for each regional node according to the direction and magnitude of the temperature deviation from the control range.

[0050] For prediction times when the humidity forecast value is higher than the upper limit of the humidity control range, determine the dehumidification correction requirement for the corresponding regional nodes; for prediction times when the humidity forecast value is lower than the lower limit of the humidity control range, determine the humidification correction requirement for the corresponding regional nodes; and generate the humidity correction amount for each regional node according to the direction and magnitude of the humidity deviation from the control range.

[0051] For forecast times when the predicted wind speed is higher than the upper limit of the wind speed control range, determine the wind reduction correction requirement for the corresponding regional nodes; for forecast times when the predicted wind speed is lower than the lower limit of the wind speed control range, determine the wind increase correction requirement for the corresponding regional nodes; and generate the wind speed correction amount for each regional node according to the direction and magnitude of the deviation of the wind speed from the control range.

[0052] The temperature correction sequence, humidity correction sequence and wind speed correction sequence corresponding to each regional node are merged according to the prediction time order to form the target control correction matrix of each regional node in multiple control cycles.

[0053] Execution instructions are generated based on the target control correction matrix, and partition adjustments are performed.

[0054] The beneficial effects of this invention are:

[0055] Compared to existing control methods that rely on a small number of fixed measuring points, static threshold judgment, and lag-based adjustment, this invention can achieve finer-grained regional division based on different distribution boundaries in the spinning workshop. By deploying temperature sensors, humidity sensors, wind speed sensors, equipment condition acquisition nodes, and valve status acquisition nodes within the boundaries of each zone node, a sensor network covering multiple locations, heights, and operating states within the workshop is formed, resulting in more comprehensive and accurate acquisition of workshop environmental conditions. By performing resampling, missing data completion, anomaly removal, and dimensional unification processing on the original spinning dataset under a unified timestamp, the consistency and usability of subsequent analysis data can be improved, avoiding control judgment distortion caused by asynchronous sampling, missing data, or abnormal fluctuations. Furthermore, this invention not only utilizes standardized time-series data blocks to acquire the environmental status of each regional node but also combines spatial coordinates, height coordinates, air supply vent opening, return air vent opening, and equipment load changes to construct a regional correlation matrix and a control object matrix. This allows for a unified expression of spatial adjacency relationships, air supply transmission relationships, return air coupling relationships, and thermal disturbance influence relationships between regional nodes, thereby overcoming the problem of inaccurate representation of inter-regional linkage relationships in existing technologies.

[0056] By processing the input sample tensor using an improved StemGNN model, and utilizing graph structure generation, frequency domain decomposition, time-frequency coupled propagation, and multi-step prediction output, the method enables advance prediction of temperature, humidity, and wind speed trends over multiple control cycles. It can pre-generate the target control correction matrix and execution commands before environmental parameters actually exceed limits, thereby improving control response speed and adjustment foresight. This method transforms temperature and humidity control in spinning workshops from traditional passive response regulation to proactive predictive regulation oriented towards regional nodes. This improves the accuracy of temperature and humidity control in the workshop, reduces local environmental fluctuations, enhances the consistency of regulation between different areas, and provides a reliable basis for zoned coordinated control of air supply, return air, and humidification devices. Ultimately, it helps improve the stability of the spinning production environment and the overall operational reliability. Attached Figure Description

[0057] 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:

[0058] Figure 1 This is an overall flowchart of the temperature and humidity control method for a spinning workshop based on a sensor network proposed in this invention.

[0059] Figure 2 This is a schematic diagram illustrating the steps of generating the regional correlation matrix in the sensor network-based temperature and humidity control method for spinning workshops proposed in this invention. Detailed Implementation

[0060] 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.

[0061] refer to Figures 1-2 A sensor network-based method for temperature and humidity control in a spinning workshop includes the following steps:

[0062] Step 1: Define the different distribution boundaries of the spinning workshop, construct a sensor network for the spinning workshop, and generate a set of nodes for the workshop area;

[0063] Step 2: Collect the original spinning dataset of each region node in the workshop area node set and generate standardized time-series data blocks;

[0064] Step 3: Based on the standardized time-series data blocks, calculate the spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient between regional nodes, and construct the regional correlation matrix and the control object matrix;

[0065] Step 4: Perform temporal expansion on the regional correlation matrix and the control object matrix, and generate the input sample tensor;

[0066] Step 5: Input the input sample tensor into the improved StemGNN model, and output the prediction result matrix for multiple future control cycles through the graph structure generation module, frequency domain decomposition module, time-frequency coupling propagation module and multi-step prediction output module.

[0067] Step 6: Calculate the target control correction matrix for each regional node based on the prediction result matrix, and generate the corresponding execution instructions.

[0068] This step addresses the challenges of uneven environmental conditions, complex local disturbance transmission, and strong lag in traditional control methods within spinning workshops, enabling more precise and stable temperature and humidity regulation. By constructing a set of workshop area nodes and standardizing the time-series processing of multi-source raw spinning data, the consistency and effectiveness of environmental data acquisition are improved, reducing interference from abnormal data and discrete sampling on control decisions. Calculating the spatial adjacency coefficient, supply air transmission coefficient, return air coupling coefficient, and thermal disturbance influence coefficient between area nodes effectively characterizes the environmental coupling relationships between different areas, enabling the control process to move beyond static adjustment of a single node and possess holistic analytical capabilities oriented towards regional linkages. Furthermore, combining an improved StemGNN model to predict multiple future control cycles allows for the early identification of temperature, humidity, and wind speed trends, generating target control correction matrices and corresponding execution commands. This enhances the foresight, accuracy, and dynamic response capabilities of temperature and humidity control in spinning workshops, contributing to improved environmental stability and control reliability.

[0069] In this embodiment, step one specifically includes:

[0070] Based on the actual distribution boundaries of the opening equipment, carding equipment, drawing equipment, roving equipment, spinning equipment, winding equipment, main air supply duct, main return air duct, and raw material temporary storage location in the spinning workshop, the workshop's planar area is divided into zones and marked, generating node boundaries for the opening zone, carding zone, drawing zone, roving zone, spinning zone, winding zone, air supply duct zone, return air duct zone, and raw material temporary storage area.

[0071] Multiple sensor installation points are set up along the length, width and height of the machine within the boundary of each partition node. Temperature sensor, humidity sensor, wind speed sensor, equipment condition acquisition node or air valve status acquisition node are installed at each sensor installation point.

[0072] Among them, temperature sensors, humidity sensors, and wind speed sensors are used to collect environmental status data at the corresponding installation points; equipment operating condition acquisition nodes are used to collect operating frequency and load status data of equipment in the corresponding zone; and air valve status acquisition nodes are used to collect opening status data of supply air valves and return air valves in the corresponding zone.

[0073] Write a unique node number for each sensor node that has been installed, write the partition number according to the partition node boundary where the sensor node is located, write the spatial coordinates and height coordinates according to the horizontal, vertical and longitudinal positions of the sensor node in the workshop coordinate system, and write the sampling period parameter according to the data acquisition frequency corresponding to the sensor node to form a set of nodes in the workshop area.

[0074] This step involves zoning and calibrating the actual distribution boundaries of the spinning workshop's cleaning equipment, carding equipment, drawing equipment, roving equipment, spinning equipment, winding equipment, main air supply ducts, main return air ducts, and raw material storage locations. This refines the traditional, generalized, and extensive environmental monitoring method into a node-based monitoring method targeting different functional areas and spatial levels, thereby improving the targeting and accuracy of environmental perception in the spinning workshop. By setting multiple sensor installation points along the length, width, and height of the machine within each zone node boundary, and deploying temperature sensors, humidity sensors, wind speed sensors, equipment operating condition acquisition nodes, and valve status acquisition nodes, multiple sources of information can be simultaneously acquired, including environmental conditions, equipment operating conditions, and air supply and return air execution status. This ensures that the collected results not only reflect local temperature and humidity changes but also the impact of equipment load changes and valve adjustments on the regional environment, thus enhancing the data completeness of subsequent control and analysis. By writing a unique node number, zone number, spatial coordinates, height coordinates, and sampling period parameters to each sensor node, a unified correspondence can be established between the collected information from different nodes in terms of spatial location, region, and time frequency. This facilitates the subsequent orderly collection, comparison, and correlation processing of multi-node data. Furthermore, after forming a set of nodes in the workshop area, a clear data organization foundation can be provided for subsequent raw spinning data collection, standardized time-series processing, regional correlation modeling, and predictive control. This effectively overcomes the problems of insufficient environmental monitoring accuracy caused by scattered measurement point layout, unclear regional boundaries, inaccurate spatial positioning, and single data sources in existing technologies. It is beneficial to improve the zoning precision level and overall regulation reliability of temperature and humidity control in the spinning workshop.

[0075] In this embodiment, step two specifically includes:

[0076] In the workshop area node set, control each sensor node to synchronously collect temperature value, relative humidity value, wind speed value, machine operating frequency value, motor load value, air supply outlet opening value, return air outlet opening value and humidification device operating status value according to a unified sampling period to obtain the original spinning dataset;

[0077] The original spinning dataset is resampled, missing data is filled in, anomalies are removed and units are standardized according to a unified timestamp. Data from multiple sampling points in the same region are aggregated in units of regional nodes to generate standardized time-series data blocks corresponding to each regional node.

[0078] This step involves controlling each sensor node in the workshop area to synchronously collect temperature, relative humidity, wind speed, machine operating frequency, motor load, air supply vent opening, air return vent opening, and humidification device operating status values ​​at a unified sampling period. This ensures that the environmental conditions, equipment operating conditions, and air supply and return adjustment conditions of the spinning workshop are uniformly acquired under the same time reference, thus avoiding data misalignment caused by different sampling frequencies and times, and improving the temporal consistency and comparability of the original spinning dataset. By performing resampling, missing data completion, anomaly removal, and dimensional unification processing on the original spinning dataset according to a unified timestamp, interference factors such as data dispersion, missing items, abnormal fluctuations, and dimensional inconsistencies during the collection process can be effectively eliminated, improving the standardization and usability of the original data and making the data foundation used in subsequent analysis processes more stable and reliable. Furthermore, by aggregating data from multiple sampling points within the same region at the regional node level, standardized time-series data blocks corresponding to each regional node are generated. This transforms dispersed multi-point collected information into a continuous time-series data representation with regional representativeness. This not only helps reflect the dynamic changes in the environmental and equipment states of each regional node but also provides a unified data structure to support subsequent calculations of regional relationships, construction of control objects, and input for prediction models. Therefore, this invention effectively overcomes the problems of insufficient analytical accuracy caused by asynchronous multi-source sampling data, unstable data quality, and inconsistent regional data organization methods in existing technologies. It is beneficial for improving the data foundation quality of temperature and humidity control in spinning workshops and the reliability of subsequent prediction and adjustment.

[0079] In this embodiment, step three specifically includes:

[0080] Based on the spatial and height coordinates of each region node in the standardized time series data block, the planar distance and vertical height difference between any two region nodes are calculated one by one, and compared with the preset distance threshold and height difference threshold respectively. Region node pairs that are both less than the preset threshold are selected as adjacent region node pairs.

[0081] The spatial adjacency coefficient between nodes in each region is obtained by taking the weighted average of the inverse of the planar distance and the vertical height difference between each pair of adjacent region nodes.

[0082] The air supply outlet opening difference, the distance between nodes along the air supply direction, and the length of the air supply path are weighted according to preset weights to obtain the air supply transmission coefficient between nodes in each region; and the return air outlet opening difference, the distance between nodes along the return air direction, and the length of the return air path overlap are weighted according to preset weights to obtain the return air coupling coefficient between nodes in each region.

[0083] The spatial distance between the heat source region and the target region, the load variation of the heat source side equipment, and the temperature variation within adjacent sampling periods are weighted according to preset weights to obtain the thermal disturbance influence coefficient of the heat source region node on the target region node.

[0084] The spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient between any two regional nodes are located according to the row and column order of the regional nodes. The spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient of the corresponding regional node pair are filled in the column positions of the same row in sequence to generate a regional association matrix arranged in the order of regional node numbers.

[0085] The current temperature, humidity, wind speed, equipment load, and actuator status values ​​of each regional node are concatenated according to the fixed field order within the same node to form a state feature vector of a single regional node. Then, all state feature vectors are arranged row by row according to the node order of the regional nodes to generate a control object matrix.

[0086] This step calculates the planar distance and vertical height difference between any two regional nodes based on the spatial and height coordinates of each regional node in the standardized time-series data block, and selects adjacent regional node pairs. This establishes a more accurate basis for node association within the spinning workshop, reflecting the actual spatial distribution characteristics, thereby improving the accuracy of identifying environmental impact relationships between different regions. By weighted averaging the reciprocal of the planar distance and the vertical height difference between adjacent regional node pairs, a spatial adjacency coefficient is obtained, which accurately characterizes the spatial proximity of regional nodes, providing a reliable basis for subsequent regional linkage analysis. Further weighting of the air supply outlet opening difference, the distance between nodes along the air supply direction, and the length of the air supply path yields the air supply transmission coefficient. Similarly, weighting of the return air outlet opening difference, the distance between nodes along the return air direction, and the overlap length of the return air path yields the return air coupling coefficient. This quantifies the impact of air supply and return processes on the regional environmental state, enhancing the ability to characterize the influence of airflow organization within the workshop. By combining the spatial distance between the heat source area and the target area, the load variation of the equipment on the heat source side, and the temperature variation within adjacent sampling periods, a thermal disturbance influence coefficient is obtained, which can further reflect the propagation characteristics of thermal disturbances caused by equipment operation. Based on the above multi-type coefficients, a regional correlation matrix is ​​generated, and a control object matrix is ​​constructed by combining the current temperature, humidity, wind speed, equipment load, and actuator status values ​​of each regional node. This enables a unified expression of the regional correlation and current control status of the spinning workshop, thereby effectively overcoming the analytical bias caused by relying only on single-point environmental parameters and lacking inter-regional coupling modeling in existing technologies. This is beneficial to improving the accuracy and overall coordination of subsequent prediction results and control decisions.

[0087] In this embodiment, step four specifically includes:

[0088] Following the time-incrementing direction, the control object matrices from multiple consecutive sampling periods are stacked layer by layer along the time dimension to generate a node time series matrix that reflects the continuous change process of nodes in each region.

[0089] Set the sliding time window length and sliding step size for the node time sequence matrix, and start from the position corresponding to the earliest sampling period. Move the sliding time window along the time dimension one by one according to the preset sliding step size. At each sliding position, extract the continuous node time sequence sub-matrix within the corresponding time range to generate multiple time window candidate segments arranged in time order.

[0090] For each time window candidate segment, extract all regional node temporal features contained between the start and end times, while keeping the order of regional nodes, time order, and internal field order of nodes unchanged, and organize each time window candidate segment into a corresponding time window sample feature block;

[0091] For each time window sample feature block, read the region association matrix corresponding to the unified timestamp of each sampling period within the current time window, and extract the region association matrix sequence of each sampling period according to the time order within the time window to form a node association relationship subsequence that matches the time window sample feature block;

[0092] The region node feature sequence in each time window sample feature block is combined with the corresponding node association subsequence and written to generate a single time window sample;

[0093] All samples from individual time windows are collected in the order in which they were generated, and then uniformly arranged according to sample number, time window length, number of regional nodes, and node feature dimension to form the input sample tensor.

[0094] This step generates a node time-series matrix that reflects the continuous changes of nodes in each region by stacking the control object matrices from multiple consecutive sampling periods layer by layer along the time dimension in an increasing direction. This transforms the regional state information, originally discretely distributed at different sampling times, into a continuous time-series expression, thus more completely reflecting the dynamic evolution of each region node in the spinning workshop in terms of temperature, humidity, wind speed, equipment load, and actuator status. By setting the sliding time window length and sliding step size for the node time-series matrix and successively extracting continuous node time-series sub-matrices along the time dimension, multiple time window candidate segments with a unified structure can be constructed while preserving temporal continuity, thus standardizing the expression of regional state change patterns in different time segments. Furthermore, by extracting all regional node time-series features from each time window candidate segment while maintaining the order of regional nodes, time order, and the order of internal fields within nodes, the structural consistency and information integrity during sample construction can be guaranteed, avoiding the impact of chaotic feature arrangement on subsequent model recognition performance. By combining the regional association matrix corresponding to the unified timestamps of each sampling period within the current time window, a node association subsequence matching the feature blocks of the time window samples is formed. Combining the regional node feature sequence with the node association subsequence allows the input data to simultaneously include the temporal variation information of the regional nodes themselves and the association influence information between regional nodes. The resulting input sample tensor not only provides a unified, standardized, and structurally complete input foundation for the improved StemGNN model but also effectively overcomes the problems of loose temporal data organization, separation of regional association information and state features, and insufficient sample representation ability in existing technologies. This is beneficial for improving the temporal sensitivity, regional association modeling ability, and overall prediction accuracy of subsequent temperature and humidity predictions.

[0095] In this embodiment, the improved StemGNN model is specifically as follows:

[0096] The input sample tensor is input into the graph structure generation module, and the temperature change sequence, humidity change sequence, wind speed change sequence, equipment load change sequence, and execution status change sequence of each region node in the input sample tensor within the continuous time window are used as node features at the current time step.

[0097] Extract the node features of each regional node at each time step according to the order of regional nodes, and splice the node features of the same regional node in a continuous time window to form the historical representation of the corresponding regional node.

[0098] The difference between the historical representations of any two regional nodes is taken as the node association strength between the corresponding regional nodes, and a dynamic association graph matrix reflecting the dynamic association relationship of regional nodes under the current time window is generated.

[0099] The dynamic correlation graph matrix is ​​input into the frequency domain decomposition module. The temperature change sequence, humidity change sequence and wind speed change sequence of each regional node corresponding to the dynamic correlation graph matrix within the continuous time window are decomposed in the frequency domain by the fast Fourier transform. The time domain change sequence corresponding to each regional node is converted into the frequency domain response sequence. The trend component, periodic component and disturbance component are extracted according to the preset frequency range to generate the frequency domain feature tensor of the regional node.

[0100] The frequency domain feature tensor of the regional nodes is input into the time-frequency coupled propagation module. The dynamic correlation graph matrix is ​​used as the information propagation constraint relationship between regional nodes. The change trend part, periodic fluctuation part and instantaneous disturbance part of each regional node are propagated along the dynamic correlation path between regional nodes, forming a joint state representation that can reflect the current state of the regional node itself and the propagation influence of adjacent regional nodes.

[0101] Arrange the joint state representations of all region nodes in the order of region nodes to generate a joint state representation matrix;

[0102] The joint state representation matrix is ​​input to the multi-step prediction output module, and the joint state representation of each region node is decoded step by step with a preset length. The step by step decoding process is as follows:

[0103] Temperature prediction features, humidity prediction features, and wind speed prediction features are extracted from the joint state representation of each regional node, and temperature prediction sequences, humidity prediction sequences, and wind speed prediction sequences corresponding to each regional node in multiple control cycles are generated.

[0104] According to the order of regional nodes and the order of prediction time, the temperature prediction sequence, humidity prediction sequence and wind speed prediction sequence of all regional nodes in multiple control cycles are arranged to form a prediction result matrix.

[0105] The improved StemGNN model proposed in this step shares similarities with the traditional StemGNN model in that both are designed for multivariate time series forecasting tasks, both employ a core processing approach combining graph structure modeling and time series change analysis, and neither treats each region node as an isolated object for independent prediction. Instead, they characterize the mutual influence between different nodes by establishing relationships between them, and then predict the changing trends in future time periods based on this. Both models demonstrate the combined utilization of node characteristics, node relationships, and temporal change patterns in their overall structure. Specifically, both require first organizing the historical change information of each region node within a continuous time window into the model input, then using graph structure-related processing units to characterize the coupling relationships between nodes, and finally combining frequency domain analysis or time series evolution analysis to extract trend and volatility information in node changes, ultimately outputting prediction results for multiple future time periods. Meanwhile, both the traditional StemGNN model and the improved StemGNN model proposed in this step serve the goal of multi-step prediction. They both emphasize continuous prediction of multiple future control cycles within the same model framework, rather than just outputting the prediction value at a single moment. Therefore, in terms of model function positioning, they both belong to the multivariate, multi-node, and multi-step joint prediction model based on graph time series fusion mechanism.

[0106] The difference lies in the fact that the improved StemGNN model proposed in this step does not directly adopt the general processing method of the traditional StemGNN model for general time series data. Instead, it combines the temperature and humidity control scenario of a spinning workshop and reconstructs the input object, graph structure generation method, frequency domain processing object, and propagation constraint mechanism for the regional node control task. Traditional StemGNN models usually focus on uniformly learning existing graph structures or general node relationships. However, the improved model in this step first extracts the temperature change sequence, humidity change sequence, wind speed change sequence, equipment load change sequence, and execution state change sequence from the input sample tensor, and uses them together as the node features at the current time step. Then, it generates a dynamic association graph matrix by the difference between the historical representations of regional nodes, so that the node relationships can adaptively adjust with the state changes within the current time window. Traditional StemGNN models typically focus on the original time-series signals themselves in their frequency domain analysis. However, the improved model in this step, under the constraints of a dynamic correlation graph matrix, performs Fast Fourier Transform decomposition on the temperature, humidity, and wind speed change sequences of each regional node within a continuous time window, extracting trend, periodic, and disturbance components. This allows the frequency domain features to correspond to the dynamic correlation relationships within the region. Furthermore, the improved model in this step incorporates a time-frequency coupling propagation module, propagating the trend, periodic fluctuation, and instantaneous disturbance components along the dynamic correlation path to form a joint state representation matrix. This matrix is ​​then decoded in multiple steps to output the prediction result matrix, thus enhancing its overall adaptability to regional linkage control scenarios in spinning workshops.

[0107] The beneficial effect of the improvements lies in the fact that by incorporating temperature change sequences, humidity change sequences, wind speed change sequences, equipment load change sequences, and execution status change sequences into the node feature construction process, the improved model proposed in this step can more comprehensively characterize the combined change features of each area node in the spinning workshop in terms of environmental conditions, equipment operating conditions, and execution adjustment conditions compared to traditional models, thereby improving the completeness of the prediction basis information. By constructing a dynamic correlation graph matrix using the differences between the historical representations of regional nodes, the model can dynamically identify the changes in the strength of correlations between regional nodes within different time windows, avoiding the problem that a fixed graph structure cannot reflect the actual changes in regional influence relationships. By performing fast Fourier transform decomposition under the constraints of the dynamic correlation graph matrix and extracting trend components, periodic components, and disturbance components, the long-term evolution characteristics, periodic fluctuation characteristics, and sudden disturbance characteristics of temperature and humidity changes in the spinning workshop can be distinguished and expressed, which is beneficial to improving the model's ability to identify complex environmental change patterns. Furthermore, by propagating various frequency domain components along the dynamic correlation path through the time-frequency coupling propagation module, the propagation influence of adjacent regional nodes can be uniformly coupled with the current regional node's own state. This makes the generated joint state representation more consistent with the actual linkage relationship under the combined effects of air supply, return air, and thermal disturbances in the spinning workshop. The resulting prediction matrix has stronger foresight, specificity, and regional consistency, which is conducive to improving the accuracy of subsequent target control correction calculations and execution command generation, thereby enhancing the stability and reliability of temperature and humidity control in the spinning workshop.

[0108] In this embodiment, step six specifically includes:

[0109] The temperature prediction sequence, humidity prediction sequence and wind speed prediction sequence of each region node in the prediction result matrix are compared with the preset temperature control interval, preset humidity control interval and preset wind speed control interval at each time step.

[0110] For prediction times when the predicted temperature is higher than the upper limit of the temperature control range, determine the cooling correction requirement for the corresponding regional nodes; for prediction times when the predicted temperature is lower than the lower limit of the temperature control range, determine the heating correction requirement for the corresponding regional nodes; and generate the temperature correction for each regional node according to the direction and magnitude of the temperature deviation from the control range.

[0111] For prediction times when the humidity forecast value is higher than the upper limit of the humidity control range, determine the dehumidification correction requirement for the corresponding regional nodes; for prediction times when the humidity forecast value is lower than the lower limit of the humidity control range, determine the humidification correction requirement for the corresponding regional nodes; and generate the humidity correction amount for each regional node according to the direction and magnitude of the humidity deviation from the control range.

[0112] For forecast times when the predicted wind speed is higher than the upper limit of the wind speed control range, determine the wind reduction correction requirement for the corresponding regional nodes; for forecast times when the predicted wind speed is lower than the lower limit of the wind speed control range, determine the wind increase correction requirement for the corresponding regional nodes; and generate the wind speed correction amount for each regional node according to the direction and magnitude of the deviation of the wind speed from the control range.

[0113] The temperature correction sequence, humidity correction sequence and wind speed correction sequence corresponding to each regional node are merged according to the prediction time order to form the target control correction matrix of each regional node in multiple control cycles.

[0114] Execution instructions are generated based on the target control correction matrix, and partition adjustments are performed.

[0115] This step compares the predicted temperature, humidity, and wind speed sequences for each region node in the prediction result matrix with the preset control intervals time-by-time. This allows for the identification of deviation trends for each region node in subsequent control cycles before environmental parameters actually exceed limits. This transforms temperature and humidity control in the spinning workshop from traditional reactive adjustment to proactive, future-oriented adjustment. By determining the correction requirements for cooling, heating, dehumidification, humidification, wind reduction, and wind increase for each region node, and generating corresponding temperature, humidity, and wind speed corrections according to the direction and magnitude of the deviation, the control results better match the actual adjustment needs of different region nodes, avoiding the problems of insufficient or excessive adjustment in local areas under traditional uniform adjustment methods. Furthermore, merging the temperature, humidity, and wind speed correction sequences according to the prediction time sequence forms a target control correction matrix for each region node over multiple control cycles. This unifies the adjustment needs of multiple variables and multiple time points, providing a clear basis for subsequent execution command generation. Then, by generating execution instructions based on the target control correction matrix and performing zonal adjustments, targeted control of nodes in different regions can be achieved, making the environmental adjustment process of each node more timely, precise, and coordinated. Therefore, this invention effectively overcomes the problems of existing technologies where control adjustment relies on the current state, lacks utilization of future trends, and has insufficient zonal correction capabilities. It is beneficial for improving the response speed, adjustment accuracy, and regional stability of temperature and humidity control in spinning workshops.

[0116] Example 1: To verify the feasibility of this invention in practice, it was applied to a medium-to-large ring spinning workshop in a city of a certain province. The workshop is responsible for the continuous production of combed cotton yarn and carded blended yarn. The workshop is equipped with opening equipment, carding equipment, drawing equipment, roving equipment, spinning equipment, and winding equipment arranged sequentially. A main air supply duct is installed above the production area, and a main return air duct is installed along the upper side of the workshop. Localized thermal and humidity disturbances occur in the raw material storage area due to unpacking, transfer, and short-term stacking. The original environmental control method in this workshop mainly relied on a small number of fixed temperature and humidity measuring points, combined with the experience of on-duty personnel to adjust the operating intensity of the air supply valves, return air valves, and humidification devices. Although this could maintain the overall average environmental parameters of the workshop within a generally acceptable range, there were always significant localized environmental fluctuations during actual production, especially at the boundary between the spinning and winding areas, at the end of the air supply duct, and near the raw material storage area, where localized high temperatures, low humidity, or large fluctuations in wind speed frequently occurred. These problems directly lead to unstable fiber conditions. The workshop has repeatedly experienced issues such as increased yarn breakage, increased fly waste, fluctuations in winding tension, and unstable operation of local machines. Although on-duty personnel can manually adjust environmental parameters after they have significantly exceeded limits, the existing methods are not capable of identifying regional differences and cannot predict future trends. Therefore, repeated fluctuations often occur, with "one area being adjusted, and then adjacent areas deviating again." As a result, the stability of the workshop environment has always been difficult to improve.

[0117] In this scenario, the sensor network-based temperature and humidity control method for spinning workshops described in this invention is applied. First, based on the actual distribution boundaries of the opening, carding, drawing, roving, spinning, winding, main air supply duct, main return air duct, and raw material storage areas within the workshop, the workshop's planar area is recalibrated, creating clear node boundaries for the opening, carding, drawing, roving, spinning, winding, air supply duct, return air duct, and raw material storage areas. Subsequently, temperature sensors, humidity sensors, wind speed sensors, equipment operating condition data acquisition nodes, and valve status data acquisition nodes are deployed along the length, width, and height of each machine within these node boundaries, ensuring that data from different locations, heights, and operating states within the workshop can be collected synchronously. During on-site implementation, the entire workshop was divided into 36 regional nodes: 3 nodes in the opening and cleaning area, 5 nodes in the carding area, 4 nodes in the drawing area, 5 nodes in the roving area, 10 nodes in the spinning area, 5 nodes in the winding area, 2 nodes in the air supply duct area, 1 node in the return air duct area, and 1 node in the raw material storage area. Each regional node had 4 to 9 sampling points, depending on its area and equipment density. A total of 198 temperature sensors, 198 humidity sensors, 124 wind speed sensors, 36 equipment status acquisition nodes, and 22 valve status acquisition nodes were deployed throughout the workshop. All sensor nodes collected temperature, relative humidity, wind speed, machine operating frequency, motor load, air supply vent opening, return air vent opening, and humidifier operating status values ​​according to a unified sampling period, forming the original spinning dataset in the edge computing terminal.

[0118] Before the data enters the control process, the system performs resampling, missing data completion, anomaly removal, and dimensional unification on the original spinning dataset according to a unified timestamp. Then, it aggregates the data by regional nodes to form standardized time-series data blocks. Initial statistical results from the field showed that the proportion of timestamp misalignment caused by network latency, local interference, and momentary occlusion in the original collected data was 3.8%, the proportion of short-term missing data was 1.6%, and the proportion of abnormal spike data was 0.9%. After standardized time-series processing, the data availability increased from 93.7% in the original state to 99.2%, and the data alignment error between different regional nodes decreased from an average of 1.7 seconds to less than 0.2 seconds. This provides a stable data foundation for subsequent predictive analysis based on time-series and regional correlation relationships. Subsequently, the system calculates the spatial adjacency coefficient, air supply transmission coefficient, return air coupling coefficient, and thermal disturbance influence coefficient between regional nodes based on the spatial and height coordinates of each regional node, the opening of the air supply vent, the opening of the return air vent, equipment load changes, and temperature changes within adjacent sampling periods. Based on these calculations, it constructs a regional correlation matrix and a control object matrix. During this process, the system found that there was a high air transfer coefficient between the nodes in the middle area of ​​the spinning zone and the nodes at the end of the air supply duct; a high thermal disturbance influence coefficient between the raw material storage area and the adjacent drawing zone; and a high return air coupling coefficient among a group of nodes in the winding zone near the main return air duct. This result is basically consistent with the long-term experience and judgment of the field engineers, but in the past, these relationships could only be perceived through experience and could not be quantified and directly used for control.

[0119] During the control operation phase, the system stacks the control object matrices from multiple consecutive sampling periods layer by layer along the time dimension in an ascending direction to generate a node time series matrix. An input sample tensor is constructed using a sliding time window approach, and this tensor is then input into an improved StemGNN model. In this implementation scenario, this model dynamically generates a correlation graph between regions based on the temperature, humidity, wind speed, equipment load, and execution status change sequences of each region's nodes within a continuous time window. It then extracts trend, periodic, and disturbance components using a Fast Fourier Transform, and completes time-frequency coupling propagation along the dynamic correlation path. Finally, it outputs the predicted temperature, humidity, and wind speed sequences for each region's nodes in the next multiple control periods. During continuous operation in the field, the system performs rolling predictions for the next six control periods and compares the prediction results with the preset control intervals time-by-time. It automatically generates temperature, humidity, and wind speed corrections, forming a target control correction matrix and corresponding execution commands to drive the supply air valve, return air valve, and humidification device for zoned adjustment. One of the most typical scenarios in the application occurs in the eastern node of the spinning zone. When the load of adjacent equipment increases and the air supply opening decreases, the traditional method usually requires the on-duty personnel to make adjustments only after the on-site temperature has increased significantly and the humidity has decreased significantly. However, the method of this invention can identify the trend of temperature rise and humidity drop in the area in the subsequent multiple control cycles before the environmental parameters exceed the upper limit of the process. The system issues small correction instructions in advance to coordinate the increase of humidity and the increase of air, which ultimately suppresses the environmental fluctuations in the area to a small range and prevents them from spreading to the adjacent spinning units.

[0120] To verify the practical effectiveness of this invention, comparative tests were conducted in the same workshop under the same shift, production line load level, and similar external climatic conditions. The comparison was made with the existing manual experience-based control method using fixed thresholds, while this invention employs a regional node sensor network combined with an improved StemGNN predictive control method. The test location was the aforementioned spinning workshop, and the test duration consisted of 72 consecutive hours of data collection. Particular attention was paid to recording the environmental stability and production operation indicators of typical areas such as the spinning zone, winding zone, drawing zone, and raw material storage zone.

[0121] Table 1. Comparison of the Application Effects of Predictive Control in Spinning Workshops

[0122]

[0123] As shown in Table 1, using the original method, among the 36 area nodes in the workshop, the average percentage of area nodes whose temperature exceeded the preset control range at any given time was 18.6%, whose humidity exceeded the preset control range was 22.4%, and whose wind speed deviated from the preset range was 16.9%. After adopting the method of the present invention, the above three percentages decreased to 6.3%, 7.8%, and 5.4%, respectively. Under the original method, the standard deviation of temperature for a group of area nodes on the east side of the spinning area was 1.28 degrees Celsius, the standard deviation of humidity was 4.7%RH, and the standard deviation of wind speed was 0.19 meters per second. After adopting the method of the present invention, the standard deviation of temperature decreased to 0.46 degrees Celsius, the standard deviation of humidity decreased to 1.8%RH, and the standard deviation of wind speed decreased to 0.07 meters per second. Under the original method, the average recognition lag time of the system for local environmental anomalies was 4.6 minutes, while the present invention, based on the prediction result matrix, can identify potential deviation trends approximately 3 control cycles in advance, resulting in an average early warning lead time of 2.9 minutes. Under the original method, the average number of invalid start-stop cycles of the humidifier was 37 times per day, and the number of frequent corrections of the air supply valve was 52 times. After adopting the method of the present invention, these numbers were reduced to 18 and 29 times respectively, indicating that the control action is more targeted and reduces unnecessary repeated adjustments.

[0124] The advantages of this invention are not merely reflected in the improvement of a single environmental parameter, but in the overall optimization of the entire environmental control chain in the workshop area. Because this invention transforms the previously coarse overall environmental monitoring into refined monitoring oriented towards multiple areas, heights, and states through zonal calibration and the construction of regional node sets, data efficiency and synchronization success rates are significantly improved, directly ensuring the reliability of subsequent modeling. Furthermore, because this invention further constructs a regional correlation matrix using spatial adjacency coefficients, supply air transmission coefficients, return air coupling coefficients, and thermal disturbance influence coefficients, the previously implicit regional linkages within the workshop are quantified. Therefore, in typical complex areas such as the spinning zone, winding zone, and raw material storage zone, temperature, humidity, and wind speed fluctuations are significantly reduced. In particular, the raw material storage zone, which was previously prone to environmental shifts in adjacent drawing zones due to short-term stacking and thermal disturbances, can now have such effects mitigated in advance through prediction and zonal adjustment after the implementation of this invention, with humidity fluctuations decreasing by more than 60%. From the production operation results, the number of yarn breakages and the frequency of fly waste cleaning have both decreased significantly, indicating that the improvement in environmental stability has translated into improved process operation quality.

[0125] During implementation, workshop operators experienced a significant improvement in their understanding of the system. Previously, operators frequently checked data from multiple dispersed measuring points and relied on experience to determine whether adjustments were needed, which often resulted in delayed judgments and excessive adjustments. With this invention, the system directly provides control corrections and execution commands at the regional node level. Operators are primarily responsible for confirming the execution status and handling special operating conditions, significantly reducing the burden of manual intervention. Especially during continuous production phases with rapidly changing equipment loads, this invention can predict and fine-tune environmental parameters before they become significantly unstable, preventing the spread of localized thermal and humidity imbalances to adjacent areas. This demonstrates that this invention not only solves the problems of incomplete environmental parameter collection, difficulty in quantifying regional relationships, delayed control actions, and insufficient zonal adjustment in existing technologies, but also possesses strong engineering feasibility and continuous operational stability in real spinning workshop scenarios, providing practical technical support for precise temperature and humidity control in spinning workshops.

[0126] 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 method for temperature and humidity control in a spinning workshop based on sensor networks, characterized in that, Includes the following steps: Step 1: Define the different distribution boundaries of the spinning workshop, construct a sensor network for the spinning workshop, and generate a set of nodes for the workshop area; Step 2: Collect the original spinning dataset of each region node in the workshop area node set and generate standardized time-series data blocks; Step 3: Based on the standardized time-series data blocks, calculate the spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient between regional nodes, and construct the regional correlation matrix and the control object matrix; Step 4: Perform temporal expansion on the regional correlation matrix and the control object matrix, and generate the input sample tensor; Step 5: Input the input sample tensor into the improved StemGNN model, and output the prediction result matrix for multiple future control cycles through the graph structure generation module, frequency domain decomposition module, time-frequency coupling propagation module and multi-step prediction output module. Step 6: Calculate the target control correction matrix for each regional node based on the prediction result matrix, and generate the corresponding execution instructions.

2. The method for temperature and humidity control in a spinning workshop based on a sensor network according to claim 1, characterized in that, Step one specifically involves: Based on the different distribution boundaries within the spinning workshop, the planar area of ​​the spinning workshop is divided and labeled to generate different node boundaries; Multiple sensor installation points are set up along the length, width, and height of the machine within the boundary of each partition node. Temperature sensor, humidity sensor, wind speed sensor, equipment condition acquisition node or air valve status acquisition node are installed at each sensor installation point to construct a sensor network for the spinning workshop. Write a unique node number for each sensor node that has been installed in the construction of the sensor network in the spinning workshop, write the partition number according to the partition node boundary where the current sensor node is located, write the spatial coordinates and height coordinates according to the horizontal, vertical and vertical positions of each sensor node in the workshop coordinate system, and write the sampling period parameter according to the data acquisition frequency corresponding to the sensor node. According to the machine layout direction, air supply and return direction and spatial location, multiple sensor nodes are grouped into corresponding area nodes to form a workshop area node set.

3. The temperature and humidity control method for a spinning workshop based on a sensor network according to claim 1, characterized in that, Step two specifically involves: In the workshop area node set, control each sensor node to synchronously collect temperature value, relative humidity value, wind speed value, machine operating frequency value, motor load value, air supply outlet opening value, return air outlet opening value and humidification device operating status value according to a unified sampling period to obtain the original spinning dataset; The original spinning dataset is resampled, missing data is filled in, anomalies are removed and units are standardized according to a unified timestamp. Data from multiple sampling points in the same region are aggregated in units of regional nodes to generate standardized time-series data blocks corresponding to each regional node.

4. The method for temperature and humidity control in a spinning workshop based on a sensor network according to claim 1, characterized in that, Step three specifically involves: Based on the spatial and height coordinates of each region node in the standardized time series data block, the planar distance and vertical height difference between any two region nodes are calculated one by one, and compared with the preset distance threshold and height difference threshold respectively. Region node pairs that are both less than the preset threshold are selected as adjacent region node pairs. The spatial adjacency coefficient between nodes in each region is obtained by taking the weighted average of the inverse of the planar distance and the vertical height difference between each pair of adjacent region nodes. The air supply outlet opening difference, the distance between nodes along the air supply direction, and the length of the air supply path are weighted according to preset weights to obtain the air supply transmission coefficient between nodes in each region; and the return air outlet opening difference, the distance between nodes along the return air direction, and the length of the return air path overlap are weighted according to preset weights to obtain the return air coupling coefficient between nodes in each region. The spatial distance between the heat source region and the target region, the load variation of the heat source side equipment, and the temperature variation within adjacent sampling periods are weighted according to preset weights to obtain the thermal disturbance influence coefficient of the heat source region node on the target region node. Based on the spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient, a regional correlation matrix and a control object matrix are generated.

5. The temperature and humidity control method for a spinning workshop based on a sensor network according to claim 4, characterized in that, The specific steps for generating the regional correlation matrix and the control object matrix based on the spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient are as follows: The spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient between any two regional nodes are located according to the row and column order of the regional nodes. The spatial adjacency coefficient, supply air transfer coefficient, return air coupling coefficient, and thermal disturbance influence coefficient of the corresponding regional node pair are filled in the column positions of the same row in sequence to generate a regional association matrix arranged in the order of regional node numbers. The current temperature, humidity, wind speed, equipment load, and actuator status values ​​of each regional node are concatenated according to the fixed field order within the same node to form a state feature vector of a single regional node. Then, all state feature vectors are arranged row by row according to the node order of the regional nodes to generate a control object matrix.

6. The method for temperature and humidity control in a spinning workshop based on a sensor network according to claim 1, characterized in that, Step four specifically involves: Following the time-incrementing direction, the control object matrices from multiple consecutive sampling periods are stacked layer by layer along the time dimension to generate a node time series matrix that reflects the continuous change process of nodes in each region. Set the sliding time window length and sliding step size for the node time sequence matrix, and start from the position corresponding to the earliest sampling period. Move the sliding time window along the time dimension one by one according to the preset sliding step size. At each sliding position, extract the continuous node time sequence sub-matrix within the corresponding time range to generate multiple time window candidate segments arranged in time order. For each time window candidate segment, extract all regional node temporal features contained between the start and end times, while keeping the order of regional nodes, time order, and internal field order of nodes unchanged, and organize each time window candidate segment into a corresponding time window sample feature block; For each time window sample feature block, read the region association matrix corresponding to the unified timestamp of each sampling period within the current time window, and extract the region association matrix sequence of each sampling period according to the time order within the time window to form a node association relationship subsequence that matches the time window sample feature block; The region node feature sequence in each time window sample feature block is combined with the corresponding node association subsequence and written to generate a single time window sample; All samples from individual time windows are collected in the order in which they were generated, and then uniformly arranged according to sample number, time window length, number of regional nodes, and node feature dimension to form the input sample tensor.

7. The method for temperature and humidity control in a spinning workshop based on a sensor network according to claim 1, characterized in that, The improved StemGNN model is specifically as follows: The input sample tensor is input into the graph structure generation module, and the temperature change sequence, humidity change sequence, wind speed change sequence, equipment load change sequence, and execution status change sequence of each region node in the input sample tensor within the continuous time window are used as node features at the current time step. Extract the node features of each regional node at each time step according to the order of regional nodes, and splice the node features of the same regional node in a continuous time window to form the historical representation of the corresponding regional node. The difference between the historical representations of any two regional nodes is taken as the node association strength between the corresponding regional nodes, and a dynamic association graph matrix reflecting the dynamic association relationship of regional nodes under the current time window is generated. The dynamic correlation graph matrix is ​​input into the frequency domain decomposition module. The temperature change sequence, humidity change sequence and wind speed change sequence of each regional node corresponding to the dynamic correlation graph matrix within the continuous time window are decomposed in the frequency domain by the fast Fourier transform. The time domain change sequence corresponding to each regional node is converted into the frequency domain response sequence. The trend component, periodic component and disturbance component are extracted according to the preset frequency range to generate the frequency domain feature tensor of the regional node. The frequency domain feature tensor of the regional nodes is input into the time-frequency coupled propagation module. The dynamic correlation graph matrix is ​​used as the information propagation constraint relationship between regional nodes. The change trend part, periodic fluctuation part and instantaneous disturbance part of each regional node are propagated along the dynamic correlation path between regional nodes, forming a joint state representation that can reflect the current state of the regional node itself and the propagation influence of adjacent regional nodes. Arrange the joint state representations of all region nodes in the order of region nodes to generate a joint state representation matrix; The joint state representation matrix is ​​input to the multi-step prediction output module, and the joint state representation of each region node is decoded step by step with a preset length. The step by step decoding process is as follows: Temperature prediction features, humidity prediction features, and wind speed prediction features are extracted from the joint state representation of each regional node, and temperature prediction sequences, humidity prediction sequences, and wind speed prediction sequences corresponding to each regional node in multiple control cycles are generated. According to the order of regional nodes and the order of prediction time, the temperature prediction sequence, humidity prediction sequence and wind speed prediction sequence of all regional nodes in multiple control cycles are arranged to form a prediction result matrix.

8. The method for temperature and humidity control in a spinning workshop based on a sensor network according to claim 1, characterized in that, Step six specifically involves: The temperature prediction sequence, humidity prediction sequence and wind speed prediction sequence of each region node in the prediction result matrix are compared with the preset temperature control interval, preset humidity control interval and preset wind speed control interval at each time step. For prediction times when the predicted temperature is higher than the upper limit of the temperature control range, determine the cooling correction requirement for the corresponding regional nodes; for prediction times when the predicted temperature is lower than the lower limit of the temperature control range, determine the heating correction requirement for the corresponding regional nodes; and generate the temperature correction for each regional node according to the direction and magnitude of the temperature deviation from the control range. For prediction times when the humidity forecast value is higher than the upper limit of the humidity control range, determine the dehumidification correction requirement for the corresponding regional nodes; for prediction times when the humidity forecast value is lower than the lower limit of the humidity control range, determine the humidification correction requirement for the corresponding regional nodes; and generate the humidity correction amount for each regional node according to the direction and magnitude of the humidity deviation from the control range. For forecast times when the predicted wind speed is higher than the upper limit of the wind speed control range, determine the wind reduction correction requirement for the corresponding regional nodes; for forecast times when the predicted wind speed is lower than the lower limit of the wind speed control range, determine the wind increase correction requirement for the corresponding regional nodes; and generate the wind speed correction amount for each regional node according to the direction and magnitude of the deviation of the wind speed from the control range. The temperature correction sequence, humidity correction sequence and wind speed correction sequence corresponding to each regional node are merged according to the prediction time order to form the target control correction matrix of each regional node in multiple control cycles. Execution instructions are generated based on the target control correction matrix, and partition adjustments are performed.

Images