Energy consumption equipment adaptive management and control platform based on ring spinning industry big data

Abstract

The application belongs to the technical field of energy consumption management, and provides an energy consumption equipment adaptive management and control platform based on ring spinning industrial big data, which comprises a model construction module, a first transfer learning module, a second transfer learning module, and a fusion and management and control module. The adaptive management and control platform can effectively eliminate the data empty window period of new equipment access, solve the contradiction between local optimization and global energy consumption optimization, significantly improve the energy consumption utilization rate of the ring spinning production line, and greatly reduce the dependence of model training on labeled data and the implementation maintenance cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy consumption management, in particular to an energy consumption device adaptive management and control platform based on ring spinning industrial big data. BACKGROUND

[0002] At present, the energy consumption management and control platform based on big data has become a hot spot of industry research and application. Such a platform usually deploys a sensor network to collect the power consumption data of spinning devices (such as spinning machines, bobbin winder, air conditioner fan, air compressor, etc.) in real time, and obtains process parameters such as yield, speed, variety, etc. in combination with a production execution system (MES). At the same time, relying on data-driven models (such as deep learning, support vector machine, random forest, etc. machine learning algorithm), through learning of a large amount of historical data, a nonlinear mapping relationship between energy consumption and device operating parameters, environmental parameters, and production parameters is constructed, and then real-time monitoring, abnormal alarm, energy efficiency analysis and optimization control of energy consumption are realized.

[0003] When a new model or new manufacturer's device is connected to the industrial Internet of Things platform, since the new device lacks historical operation data, an accurate data-driven energy consumption model cannot be trained for it immediately, resulting in that the platform cannot effectively monitor and optimize the control of the energy consumption of the new device in a period after the new device is connected, and there is a management blind area.

[0004] Therefore, how to further improve the effective energy consumption management of the energy consumption device adaptive management and control platform for newly connected devices is a technical problem to be solved at present. SUMMARY

[0005] To solve the above technical problems, the present application provides an energy consumption device adaptive management and control platform based on ring spinning industrial big data.

[0006] The present application provides an energy consumption device adaptive management and control platform based on ring spinning industrial big data, comprising a model construction module, a first transfer learning module, a second transfer learning module, and a fusion and control module.

[0007] The model construction module is used to analyze and obtain the key physical characteristic parameters of the connected new device, and to construct a physical mechanism-based energy consumption benchmark model for the new device based on the key physical characteristic parameters.

[0008] The first transfer learning module is used to extract small sample features from the real-time data stream generated by the new device using the general feature extractor in the first adaptive control submodule of similar energy consumption devices, use the parameters of the energy consumption benchmark model as the initial weights of the terminal output layer or a few top layers including the output layer of the general feature extractor, freeze the weight parameters of the remaining layers, and use the small sample features to perform supervised fine-tuning training on the unfrozen terminal output layer or a few top layers including the terminal output layer to obtain the second adaptive control submodule.

[0009] The second transfer learning module is used to test and retrain the second adaptive control submodule based on the energy consumption coupling knowledge graph and typical operating condition dataset corresponding to the region to which the new device belongs, so as to obtain the third adaptive control submodule; wherein, the energy consumption coupling knowledge graph is established based on the energy consumption device association topology network relationship corresponding to the region.

[0010] The fusion and control module is used to fuse the third adaptive control submodule with other first adaptive control submodules to obtain a new adaptive control module, and to adaptively control all energy-consuming devices in the region, including new devices, based on the new adaptive control module.

[0011] This invention initializes network weights using equipment physical characteristic parameters, ensuring both model convergence speed and physical rationality. Furthermore, by combining a two-stage transfer learning mechanism, it achieves rapid adaptation to individual equipment features and completes system-level multi-equipment collaborative optimization. The adaptive control platform of this invention effectively eliminates the data gap period when new equipment is connected, resolves the contradiction between local optimization and global energy efficiency, significantly improves the energy utilization rate of the entire ring spinning production line, and substantially reduces the model's dependence on labeled data and implementation and maintenance costs. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the structure of an adaptive control platform for energy consumption equipment based on big data in the ring spinning industry, as disclosed in an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of the system architecture disclosed in the embodiments of the present invention.

[0014] Figure 3 This is a schematic diagram of the structure of the general feature extractor disclosed in the embodiments of the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0016] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0017] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0018] As shown in Figure 1 , the embodiments of the present application disclose an energy-consuming equipment adaptive control platform 100 based on ring spinning industrial big data, comprising a model construction module 1001, a first transfer learning module 1002, a second transfer learning module 1003, and a fusion and control module 1004.

[0019] The model construction module 1001 is configured to analyze and obtain the key physical characteristic parameters of the accessed new equipment, and to construct a physical mechanism-based energy consumption benchmark model for the new equipment based on the key physical characteristic parameters.

[0020] As shown in Figure 2 , when a new model or new manufacturer's energy-consuming equipment (such as an intelligent spinning frame, an efficient air conditioner main unit, a new type of air compressor, etc., refer to the energy-consuming equipment 1... energy-consuming equipment n in Figure 2 , the module automatically identifies the equipment identity through the embedded protocol analysis engine and accesses the built-in or cloud-based equipment metadata database, for example, obtains the digital nameplate information of the equipment through OPC UA, Modbus, etc. It can be understood that these key physical characteristic parameters include but are not limited to rated power, rated efficiency, rated speed, design operating point, fan / pump characteristic curve (PQ curve), motor load-efficiency curve, etc., which are determined according to the characteristics of the new equipment and will not be described here.

[0021] Subsequently, the module uses the above parameters to quickly construct the energy consumption benchmark model of the new equipment according to the principles of energy conservation law, similarity law, motor science, fluid mechanics, etc. For example, for fan and pump devices, the similarity law model (i.e. P / P0=(n / n0)³) that the power consumption is proportional to the cube of the speed is loaded; for motor-driven devices, the theoretical copper loss and iron loss are calculated based on the equivalent circuit model.

[0022] The first transfer learning module 1002 is configured to extract small sample features from real-time data streams generated by the new device using a general feature extractor in the first adaptive control sub-module of the similar energy consumption device, use parameters of the energy consumption benchmark model as initial weights of an end output layer or a few top layers including the end output layer of the general feature extractor, freeze weight parameters of the remaining layers, perform supervised fine-tuning training on the end output layer or the few top layers including the end output layer that are not frozen using the small sample features, and obtain a second adaptive control sub-module.

[0023] The platform pre-stores an optimized control model of a device similar in type or function to the current new device, i.e., the first adaptive control sub-module, which has been trained. The deep network (such as a one-dimensional CNN or LSTM) in the sub-module serves as a general feature extractor and has strong feature abstraction capability. After the first adaptive control sub-module is invoked, the general feature extractor in the first adaptive control sub-module is used to extract small sample features from real-time data streams of the new device.

[0024] Then, the parameters of the energy consumption benchmark model calculated by the model construction module are used as initial weights of an end output layer or a few top layers including the end output layer of the general feature extractor to be trained, instead of random initialization. The weight parameters of all layers of the general feature extractor except the few top layers including the end output layer are frozen, so that they remain unchanged in subsequent training. Subsequently, only the small sample features generated by the new device are used to perform supervised fine-tuning training on the end output layer or the few top layers including the end output layer that are not frozen.

[0025] The module starts learning from a highly reasonable initial state conforming to physical laws, instead of a random state, which can greatly accelerate the training convergence speed and ensure that the learning process is always constrained within a reasonable range defined by physical principles, avoiding the generation of incorrect outputs that violate physical laws in the early stage of the model. The finally obtained second adaptive control sub-module not only inherits the general feature extraction capability but also quickly adapts to the individual characteristics of the new device, thereby significantly improving the accuracy of subsequent adaptive control.

[0026] The second transfer learning module 1003 is configured to test and retrain the second adaptive control sub-module based on an energy consumption coupling knowledge graph corresponding to a region to which the new device belongs and a typical working condition data set, and obtain a third adaptive control sub-module. The energy consumption coupling knowledge graph is established based on an energy consumption device correlation topological network relationship corresponding to the region.

[0027] The module is used to ensure that the second adaptive control sub-module optimized by the first migration learning module can not only reduce its own energy consumption, but also work collaboratively with other devices in the region to improve the probability of achieving global energy efficiency optimization and reduce the occurrence of local optimization and global deterioration.

[0028] Specifically, the module relies on a pre-constructed energy consumption coupling knowledge graph. The graph digitally defines the complex topological relationship and energy flow logic among all energy-consuming devices in the region (such as a workshop or a production line). Its nodes include various energy-using devices, energy storage units, and environmental parameters; edges represent the energy supply, energy consumption, and influence between them, such as air compressors providing air sources for all spinning machines, and workshop air conditioning loads being affected by the heat generated by spinning machines.

[0029] The module places the second adaptive control sub-module in a virtual system collaborative environment and uses a typical working condition data set (such as historical system data of high load, low load, and product variety switching scenarios) to simulate and retrain it. It can be understood that the objective function of this training is to optimize the total energy consumption of the region, not just to minimize the energy consumption of the new device.

[0030] Through the above testing and retraining process, the second adaptive control sub-module of the new device can consider the energy consumption impact on upstream and downstream and parallel devices when making decisions, thereby evolving into a third adaptive control sub-module with a system-wide view.

[0031] The fusion and control module 1004 is configured to fuse the third adaptive control sub-module with other first adaptive control sub-modules to obtain a new adaptive control module, and to perform adaptive control on all energy-consuming devices in the region including the new device based on the new adaptive control module.

[0032] The module integrates the third adaptive control sub-module of the new device trained by system-level collaborative optimization with the latest control sub-modules of other existing devices in the region, i.e., other first adaptive control sub-modules, to form a unified and upgraded new adaptive control module. Obviously, the new adaptive control module is a multi-agent collaborative control system.

[0033] Finally, the adaptive control platform generates and executes collaborative control instructions (such as uniformly adjusting the speed setting, coordinating the device start-stop sequence, optimizing pressure distribution, etc.) based on the new adaptive control module containing new members and internally coordinated, thereby achieving unified, adaptive, and globally optimal energy consumption control of all energy-consuming devices in the region including the new device, and maximizing the energy utilization efficiency of the entire region.

[0034] The application initializes network weights by using device physical characteristic parameters, ensures model convergence speed and physical rationality, and combines a two-stage transfer learning mechanism to realize rapid adaptation of device individual characteristics and complete system-level multi-device collaborative optimization. The adaptive management and control platform can effectively eliminate the data empty window period of new device access, solve the contradiction between local optimization and global energy consumption optimization, significantly improve the energy utilization rate of ring spinning yarn production line, and greatly reduce the dependence of model training on labeled data and implementation maintenance cost.

[0035] As an example, the first transfer learning module 1002 extracts small sample features from real-time data streams generated by the new device using the general feature extractor in the first adaptive management and control sub-module of the similar energy consumption device, including: determining the general feature extractor in the first adaptive management and control sub-module of the similar energy consumption device based on multi-dimensional attribute feature matching, and the general feature extractor is constructed by using one-dimensional convolutional neural network or long short-term memory network architecture.

[0036] Among them, the similar energy consumption device is determined based on multi-dimensional attribute feature matching, and the general feature extractor in the first adaptive management and control sub-module of the similar energy consumption device. Specifically, the adaptive management and control platform maintains a feature database of the accessed devices, which stores the multi-dimensional attribute features of each device, including device type, rated power range, typical operating condition, mechanical structure characteristics, etc. When a new device is accessed, the adaptive management and control platform extracts its device nameplate parameters and initial configuration information to generate a corresponding multi-dimensional attribute feature vector. By calculating the cosine similarity between the multi-dimensional attribute feature vector and the existing device feature vector in the database, the device with the highest similarity is selected as the similar energy consumption device.

[0037] Subsequently, the platform calls the first adaptive management and control sub-module corresponding to the similar energy consumption device, and the general feature extractor in the first adaptive management and control sub-module is constructed by using one-dimensional convolutional neural network (1D-CNN) or long short-term memory network (LSTM) architecture. These networks have been trained on historical data and have strong time series feature extraction capability.

[0038] The real-time multi-dimensional running data stream generated after the new device is accessed is input into the general feature extractor; the low-dimensional abstract feature vector representing the device running state is extracted from the real-time data stream through the forward propagation calculation of the general feature extractor, forming a small sample feature set for subsequent fine-tuning training.

[0039] Among them, as Figure 3As shown, the general feature extractor includes an input layer, a feature abstraction layer (including a time-series feature extraction path, a context feature extraction path), a feature fusion layer, a dimension compression layer, and a feature optimization layer. The specific processing process is as follows: the input layer receives the pre-processed real-time data stream and converts it into a three-dimensional tensor format suitable for network processing. The data is first copied to the double-path architecture of the feature abstraction layer and is subjected to feature extraction, specifically: in the time-series feature extraction path, a one-dimensional convolutional neural network gradually extracts the local feature patterns of the data through multiple convolution and pooling operations; the shallow network captures fine-grained short-term features (such as vibration impact), and the deep network aggregates these features to form higher-level abstract representations (such as running state patterns). In the context feature extraction path, the bidirectional LSTM network processes the time-series data from both forward and backward directions, learns long-term dependencies through its gating mechanism, and captures the gradual process and periodic characteristics of the device running state. Subsequently, the feature fusion layer splices and recalibrates the output features of the two paths. The SE attention mechanism automatically learns the importance weights of each feature channel, enhances features with high discrimination, and suppresses redundant features.

[0040] Then, the dimension compression layer compresses the time-series dimension through global average pooling operations to generate a fixed-length feature vector. After this process, the most important feature information is retained, and the data dimension is greatly reduced.

[0041] Finally, the feature optimization layer performs final projection and normalization on the feature vector, outputting a 512-dimensional dense low-dimensional feature vector. This low-dimensional feature vector has the following characteristics: high abstraction: retains the most representative and discriminative feature information in the original data.

[0042] Low dimensionality: the feature dimension is reduced by two orders of magnitude compared to the original data.

[0043] Separability: the feature vectors of different running states have good clustering characteristics in the feature space.

[0044] These low-dimensional abstract feature vectors are organized in time series to form a small sample feature set data dimension for subsequent fine-tuning training. In addition, the entire above forward propagation process does not require human intervention, and the feature extraction and transformation are automatically completed by the pre-trained network parameters, ensuring the consistency and efficiency of feature extraction.

[0045] As an example, using the small sample features to supervise the fine-tuning training of the unfrozen end output layer or a small number of top layers including the end output layer, a second adaptive control sub-module is obtained, including: dividing the small sample feature set into a training set and a validation set, using mean square error or mean absolute error as the loss function, and using the small batch gradient descent algorithm for optimization.

[0046] Specifically, the adaptive management and control platform adopts a hierarchical sampling strategy to ensure that the proportion of feature samples of different operating states in the training set and the validation set remains consistent. The training set is used for model parameter updating, and the validation set is used for training process monitoring and hyperparameter tuning. For the energy consumption prediction task of the regression type, the mean square error (MSE) is selected as the loss function, which has a stronger punishment effect on large prediction errors and is beneficial to rapid convergence. At the same time, the adaptive moment estimation (Adam) optimizer is used, which combines the advantages of the momentum method and the RMSProp algorithm, and can maintain stable convergence performance in the case of sparse gradient and small batch data.

[0047] During the training process, the weight parameters of the frozen layers in the general feature extractor are kept unchanged, and only the weight parameters of the unfrozen end output layer or a small number of top layers including the end output layer are updated by back propagation.

[0048] Specifically, the adaptive management and control platform sets a gradient calculation flag to stop gradient calculation and back propagation for frozen layers, which can significantly reduce the consumption of computing resources in the training process. Only the parameters of the unfrozen fully connected layer are updated, and the learning rate is set to 1 / 10 of that in the pre-training stage to ensure the stability of parameter updating. This partial fine-tuning strategy in this embodiment can not only retain the feature extraction capability of the pre-trained model, but also enable the model to quickly adapt to the operating characteristics of new devices.

[0049] The training process is monitored through the validation set, and the training is terminated in advance when the validation set loss function value no longer decreases or reaches the preset training round, and finally the second adaptive management and control sub-module trained by fine-tuning is obtained.

[0050] Specifically, the training process is monitored through the validation set, and the early stopping mechanism is used to prevent overfitting. The adaptive management and control platform calculates the loss function value of the validation subset in real time, and when the validation loss no longer decreases for a plurality of consecutive training periods (such as 10 periods), the current optimal model parameters are automatically saved and the training process is terminated. At the same time, the maximum number of training periods (such as 100 periods) is set as a safeguard condition for terminating training to avoid indefinite training. Finally, the second adaptive management and control sub-module trained by fine-tuning is obtained, which retains the original general feature extraction capability, and its output layer parameters have been optimized and adjusted to accurately reflect the individual characteristic differences of new devices, such as mechanical transmission efficiency deviation, motor characteristic curve difference, etc.

[0051] As an example, the second adaptive control submodule is tested and retrained based on the energy consumption coupling knowledge graph corresponding to the region to which the new device belongs and the typical working condition data set, to obtain a third adaptive control submodule, including: extracting a typical working condition data set, including multi-device collaborative operation data sequences under normal operation working conditions, extreme production load working conditions and device start-stop transition working conditions of the region; wherein the data sequence contains energy consumption data, operating state parameters and environmental parameters of each device.

[0052] Specifically, the adaptive control platform selects representative working condition periods from the historical database, extracts energy consumption data (such as power, current values) of each device, operating state parameters (such as speed, pressure set value) and environmental parameters (such as temperature and humidity monitoring values), and performs timestamp alignment, missing value filling and standardization preprocessing on these multi-source heterogeneous data to form a high-quality typical working condition data set that can be used for reinforcement learning training.

[0053] An energy consumption coupling knowledge graph is constructed, the nodes of which include energy consumption devices, energy conversion units and environmental parameter monitoring points in the region, the edges represent the energy supply relationship, process coupling relationship and physical connection relationship between devices, and the edge weights are calculated by analyzing the energy flow data in the typical working condition data set.

[0054] Specifically, the adaptive control platform constructs an energy consumption coupling knowledge graph with energy consumption devices, energy conversion units and environmental parameter monitoring points as nodes, and energy supply relationship, process coupling relationship and physical connection relationship between devices as edges based on device topology relationship and process flow diagram. The edge weights are calculated by analyzing the energy flow data in the typical working condition data set, specifically by using the transfer entropy method to quantify the energy consumption influence strength between devices. For example, the weight of the air supply relationship between the air compressor and the spinning machine is determined by analyzing the correlation between the change of the air compressor output pressure and the energy consumption fluctuation of the spinning machine in the historical data.

[0055] A virtual simulation environment is constructed based on the typical working condition data set, and the second adaptive control submodule is placed in the environment as an agent, and a reward function is defined with the regional comprehensive energy efficiency index as the target.

[0056] Specifically, the adaptive control platform uses digital twinning technology to establish a virtual simulation environment containing device dynamic models and environmental interaction models. The second adaptive control submodule is placed in the virtual simulation environment as an agent, and a reward function is defined with the regional comprehensive energy efficiency index as the target. The reward function considers multiple indicators such as unit energy consumption, device load balancing degree and energy utilization efficiency, and calculates the immediate reward value by weighted summation.

[0057] In the virtual simulation environment, the adjacent node state information is aggregated using the energy consumption coupling knowledge graph, and the second adaptive control sub-module is collaboratively optimized and trained through a multi-agent reinforcement learning algorithm; meanwhile, the feasibility of the generated control strategy is verified through the energy consumption coupling knowledge graph, and a penalty is given to the strategy that violates the physical constraints or may cause system oscillation, and finally the third adaptive control sub-module considering system-level collaborative optimization is obtained.

[0058] Specifically, the adaptive control platform uses a graph attention network (GAT) to aggregate the adjacent node state information of the energy consumption coupling knowledge graph, so that the agent can perceive the global state of the system. The second adaptive control sub-module is trained using a multi-agent proximal policy optimization (MAPPO) algorithm, and the feasibility of the generated control strategy is verified simultaneously in the training process. A negative reward penalty is given to the strategy that violates the device operation constraints (such as over-limit operation) or may cause system oscillation.

[0059] Finally, the training difficulty is gradually increased through a curriculum learning strategy. In the initial stage, basic training is carried out under steady-state conditions, and then gradually transitions to dynamic and extreme condition training. The convergence of the reward function is monitored during the training process, and when the regional comprehensive energy efficiency index reaches the preset threshold or the training round reaches the upper limit, the training is terminated, and the third adaptive control sub-module with system-level collaborative optimization capability is output. This sub-module not only considers the individual characteristics of the device, but also has the ability of global optimization of the system, which helps to achieve the overall optimization of regional energy consumption.

[0060] As an example, in the collaborative optimization training process, a curriculum learning strategy is adopted, and the training scene is gradually switched from simple steady-state conditions to complex dynamic conditions according to the complexity of the working conditions, and the proximal policy optimization algorithm is used to update the model parameters.

[0061] In this embodiment, a three-stage training plan for curriculum learning is developed, namely the initial stage, the intermediate stage, and the advanced stage.

[0062] In the initial stage, training is carried out under steady-state conditions, and the device operating parameter fluctuation range is controlled within ±5% of the rated value. The system is in a quasi-equilibrium state, and the agent is mainly trained to master basic energy consumption adjustment strategies, such as maintaining the device in the best efficiency interval.

[0063] In the intermediate stage, normal production fluctuations are introduced, and the device load variation range is expanded to ±15% of the rated value. Common working conditions such as variety switching and raw material changes are simulated, and the agent is trained to adapt to dynamic adjustment and learn predictive adjustment strategies.

[0064] In the advanced stage, extreme load conditions are used, with load fluctuations reaching ±30% of the rated value, and including emergency start and stop of the device and other sudden conditions. The agent is trained to have the ability to handle complex scenarios and system coordination ability.

[0065] Secondly, an automatic course scheduling mechanism based on performance is established. The adaptive control platform monitors the reward function value in real time during training, and calculates the average reward value using a sliding window of 10 training periods. When the average reward value exceeds the performance threshold of the current stage for 5 consecutive training periods (the steady-state stage threshold is set to 0.8, the dynamic stage threshold is set to 0.6, and the extreme stage threshold is set to 0.4), it is automatically promoted to the next difficulty stage. At the same time, the maximum number of training periods is limited in each stage (200 periods in the initial stage, 300 periods in the intermediate stage, and 500 periods in the advanced stage), to ensure that each stage can be fully trained.

[0066] Then, the proximal policy optimization (PPO-Clip) algorithm is used for model parameter update. The PPO-Clip algorithm uses generalized advantage estimation (GAE) to calculate the advantage function, where the discount factor γ is set to 0.99 and the GAE parameter λ is set to 0.95. When updating the policy, the clipping range ε is set to 0.2 to prevent the policy update step from being too large. The learning rate of the value function is set to 0.5 times the learning rate of the policy, and the policy network and value network are updated every 2048 time steps of sample data. At the same time, 4 environment replicas are used to collect training data in parallel to improve training efficiency.

[0067] Finally, training stability safeguards are implemented. A parameter soft update mechanism is used, with the new policy parameters mixed with the old policy parameters at a ratio of 0.95 to maintain the stability of policy updates. The gradient clipping threshold is set to 0.5 to prevent gradient explosion. Model checkpoints are saved every 50 training periods to support training interruption recovery. Key indicators such as policy loss function value, value function loss value, and reward function convergence are monitored in real time, and training is terminated when the reward value improvement is less than 1% for 20 consecutive training periods or the maximum number of training periods is reached.

[0068] In this embodiment, through the combination of the above course learning strategy and proximal policy optimization algorithm, the agent can gradually master the optimization strategy from simple to complex, avoiding the convergence difficulty problem caused by direct training in a complex environment.

[0069] As an example, the energy consumption coupling knowledge graph is constructed, specifically: based on the average mutual information intensity, the average correlation coefficient, and the energy transfer loss rate corresponding to the region, the overall energy consumption coupling degree index of the region is calculated.

[0070] Based on the overall energy consumption coupling degree index, the construction resolution is determined, and based on the determined construction resolution and the energy consumption equipment association topology network relationship corresponding to the region, the energy consumption coupling knowledge graph is established.

[0071] As an example, the average mutual information strength is calculated by analyzing the mutual information value of the energy consumption data between the devices in the region; the average correlation coefficient is calculated by counting the correlation coefficient matrix of the energy consumption fluctuation between the devices; and the energy transmission loss rate is calculated by evaluating the transmission efficiency of the energy flow between the devices.

[0072] In this embodiment, the adaptive management and control platform collects the historical energy consumption time series data of all devices in the region, for example, the data sampling interval is 1 minute and the time span is the last 3 months. Based on these data, the overall energy consumption coupling degree index is calculated by weighted calculation based on the average mutual information strength, the average correlation coefficient and the energy transmission loss rate after normalization processing.

[0073] Among them, the calculation method of each dimension index is as follows: (1) the average mutual information strength, which is used to quantify the mutual dependence of the energy consumption between the devices. The calculation process is as follows: the historical energy consumption time series data of all devices in the region is preprocessed, including missing value filling and outlier removal; the continuous energy consumption data is discretized into 10 intervals by equal frequency binning method to calculate the probability distribution. For any two devices, calculate the mutual information value; after calculating the mutual information value of all device pairs, take the arithmetic mean to get the average mutual information strength. The larger the value is, the stronger the statistical dependence of the energy consumption between the devices is. The specific calculation formula of mutual information is not described here.

[0074] (2) the average correlation coefficient, which is used to measure the synchronicity and linear correlation degree of the energy consumption change of the devices. The calculation process is as follows: the energy consumption time series of each device is standardized, then the Pearson correlation coefficient between each pair of devices is calculated, and finally the average value of the absolute value of all device pair correlation coefficients is calculated. The closer the value is to 1, the stronger the synchronicity of the energy consumption change of the devices is.

[0075] (3) the energy transmission loss rate, which is used to evaluate the efficiency of energy transmission between devices. Taking the compressed air system as an example: first, determine the energy supply device (air compressor) and the use device (spinning machine, bobbin winder, etc.), measure the output power of the air compressor and the input power of each gas using device through the installed flow meter and pressure sensor. The total loss rate of the system is calculated as follows: ; wherein, is the number of energy supply devices (such as air compressor, transformer, etc.); is the number of energy using devices; is the actual power obtained by the jth using device (unit: kW); is the input power of the ith supply device (unit: kW).

[0076] This value reflects the proportion of energy loss in the transmission process, and the smaller the value is, the higher the energy utilization efficiency is.

[0077] Then, the construction resolution of the knowledge graph is determined based on the overall energy consumption coupling degree index. Specifically, for example, the first threshold is set to 0.8, and the second threshold is set to 0.5. When the overall energy consumption coupling degree index is greater than 0.8, the device-level resolution is adopted, that is, the graph is constructed with single devices as nodes. When the overall energy consumption coupling degree index is between 0.5 and 0.8, the subsystem-level resolution is adopted, that is, the device cluster is taken as the node. When the overall energy consumption coupling degree index is less than 0.5, the system-level resolution is adopted, that is, the production area is taken as the node.

[0078] Finally, based on the determined resolution and device association topology network relationship, the energy consumption coupling knowledge graph is constructed. Taking the device-level resolution as an example, the adaptive control platform executes the following steps: creating device nodes, including spinning machines, bobbin winder machines and other production devices, and air compressors, air conditioning units and other auxiliary devices; establishing connection relationship edges, including energy supply relationship, process coupling relationship and physical connection relationship; using the transfer entropy method to calculate the edge weight to quantify the energy consumption influence strength between devices; using a graph database to store the graph, and adding attribute information to each node and edge; establishing a regular updating mechanism to recalculate the edge weight and node attribute every 24 hours.

[0079] As an example, the third adaptive control submodule is fused with other first adaptive control submodules to obtain a new adaptive control module, including: performing consistency detection on the third adaptive control submodule and other first adaptive control submodules, and performing integration processing after the consistency detection to obtain the new adaptive control module.

[0080] In this embodiment, the adaptive control platform first performs interface compatibility detection on each submodule, verifies whether the input and output data structures of the third adaptive control submodule and other first adaptive control submodules are consistent, including data dimensions, units, sampling frequencies and other parameters. The adaptive control platform adopts a standardized interface protocol to ensure that the data exchange formats of all submodules are uniform.

[0081] Then, the functional logic conflict detection is performed to analyze whether the control instructions generated by each submodule are contradictory. For example, it is checked whether the optimization instructions of the third adaptive control submodule conflict with the control targets of the existing device submodule, and whether there is a situation of mutual cancellation or negative superposition effect. The adaptive control platform automatically identifies potential functional conflicts by establishing a rule library and a constraint condition library.

[0082] Finally, the performance index consistency evaluation is performed to ensure that the performance levels of each submodule are equivalent. The adaptive control platform evaluates the control accuracy, response speed and stability indicators of each submodule, marks the submodules with large performance differences, and gives appropriate weight adjustment in subsequent fusion.

[0083] After passing the above tests, the new and old sub-modules can be integrated together. It can be understood that each sub-module can separately perform energy consumption optimization control on each energy consumption device, or perform energy consumption collaborative optimization based on an externally connected multi-agent collaborative decision mechanism, so as to achieve a more optimal overall energy consumption level.

[0084] As an example, based on the new adaptive control module, the adaptive control of all energy consumption devices in the region including the new device is carried out, including: generating a trigger signal in response to the completion of the fusion of the new adaptive control module, and sending the trigger signal to the new device; wherein the trigger signal is used to trigger the new device to switch to receiving the energy consumption control instruction of the adaptive control platform.

[0085] In this embodiment, when the new adaptive control module completes the fusion and passes the verification test, the adaptive control platform automatically generates a control right switching instruction. The instruction can be encapsulated by using a standard OPC UA protocol, and contains key information such as target device identification, effective timestamp and digital signature. It is transmitted to the local controller of the new device through the industrial ring network.

[0086] After receiving the instruction, the new device first verifies the validity of the digital signature to confirm the legality of the instruction source. After verification, the new device performs the control right switching operation: switches the running mode from local manual control to receiving platform optimization instruction, and at the same time, synchronizes the current running parameters and state data with the platform formally and continuously. After the switching is completed, the new device sends a confirmation signal to the platform, and the platform updates the device state table after receiving the confirmation signal, and marks the device as "accessed to control" state.

[0087] It can be understood that the adaptive control platform establishes a unified monitoring mechanism for all devices, and periodically collects real-time running data of all devices, including energy consumption data, device status and process parameters. Based on these data, the new adaptive control module generates an optimization control instruction corresponding to the period. The content of the control instruction includes device set value adjustment, running mode switching and energy efficiency optimization strategy. The control instruction is issued to each device controller through a distributed execution system. During the execution of the instruction, the adaptive control platform monitors the device response in real time, and when it detects that the control instruction deviates from the execution or the device is abnormal, a dynamic adjustment mechanism is started: first analyze the abnormal reason, then adjust the control strategy parameters, and if necessary, fall back to the safe running mode.

[0088] In addition, the adaptive control platform can further establish an efficiency evaluation system to continuously monitor the control effect. The evaluation indexes include system total energy consumption, device load balancing degree, energy efficiency index achievement rate, etc. For example, an efficiency evaluation report is generated every 24 hours, and the parameter settings of the control module are automatically optimized and adjusted to achieve continuous improvement.

[0089] While the application has been particularly shown and described with reference to preferred embodiments, it will be understood to those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application. Accordingly, the disclosed application is to be considered as illustrative and not restrictive, and the application is defined by the scope of the appended claims.

Claims

1. A self-adaptive management and control platform for energy-consuming equipment based on ring spinning industry big data, characterized in that, The model construction module is configured to parse key physical characteristic parameters of a newly connected device, and construct a physical mechanism-based energy consumption benchmark model for the newly connected device based on the key physical characteristic parameters. The second transfer learning module is configured to test and retrain the second adaptive control submodule based on an energy consumption coupling knowledge graph and a typical working condition data set corresponding to a region to which the newly connected device belongs, and obtain a third adaptive control submodule. 2.The energy consumption device adaptive management and control platform based on ring spinning industrial big data according to claim 1, characterized in that: The first transfer learning module extracts small sample features from real-time data streams generated by the newly connected device using a general feature extractor in a first adaptive control submodule of a similar energy consumption device, includes: based on multi-dimensional attribute feature matching, similar energy consumption devices are obtained, and the general feature extractor in the first adaptive control submodule of the similar energy consumption device is determined, and the general feature extractor is constructed using a one-dimensional convolutional neural network or a long short-term memory network architecture; the real-time multi-dimensional running data stream generated after the newly connected device is connected is input to the general feature extractor; low-dimensional abstract feature vectors representing the running state of the device are extracted from the real-time data stream through forward propagation calculation of the general feature extractor, and a small sample feature set for subsequent fine-tuning training is formed. 3.The energy consumption device adaptive management and control platform based on ring spinning industrial big data of claim 2, characterized in that: The general feature extractor includes an input layer, a feature abstraction layer, a feature fusion layer, a dimension compression layer, and a feature optimization layer.

4. The energy consumption device adaptive management and control platform based on ring spinning industry big data according to claim 3, characterized in that: The second adaptive control submodule is obtained by fine-tuning the training of the non-frozen end output layer using the small sample features, including: dividing the small sample feature set into a training set and a validation set, taking mean square error or mean absolute error as a loss function, and using a small batch gradient descent algorithm for optimization; the training process is monitored through the validation set, and the training is terminated in advance when the validation set loss function value no longer decreases or reaches a preset training round, and finally the fine-tuned second adaptive control submodule is obtained. The second adaptive control submodule is obtained by fine-tuning the training of the non-frozen end output layer using the small sample features, including: dividing the small sample feature set into a training set and a validation set, taking mean square error or mean absolute error as a loss function, and using a small batch gradient descent algorithm for optimization; the training process is monitored through the validation set, and the training is terminated in advance when the validation set loss function value no longer decreases or reaches a preset training round, and finally the fine-tuned second adaptive control submodule is obtained. 5.The energy consumption device adaptive management and control platform based on ring spinning industrial big data of claim 4, characterized in that: The second adaptive control submodule is tested and retrained based on an energy consumption coupling knowledge graph corresponding to a region to which the new device belongs and a typical working condition data set, to obtain a third adaptive control submodule, including: extracting a typical working condition data set, including multi-device collaborative operation data sequences in a region normal operation working condition, an extreme production load working condition, and a device start-stop transition working condition; wherein the data sequence contains energy consumption data, operating state parameters, and environmental parameters of each device; constructing an energy consumption coupling knowledge graph, the nodes of which include energy consumption devices, energy conversion units, and environmental parameter monitoring points in the region, the edges represent the energy supply relationship, process coupling relationship, and physical connection relationship between devices, and the edge weights are calculated by analyzing the energy flow data in the typical working condition data set; a virtual simulation environment is constructed based on the typical working condition data set, the second adaptive control submodule is placed in the environment as an intelligent agent, and a reward function is defined with the region comprehensive energy efficiency index as the target; in the virtual simulation environment, the adjacent node state information is aggregated using the energy consumption coupling knowledge graph, and the second adaptive control submodule is collaboratively optimized and trained through a multi-agent reinforcement learning algorithm; at the same time, the feasibility of the generated control strategy is verified through the energy consumption coupling knowledge graph, strategies that violate physical constraints or may cause system oscillation are punished, and finally a third adaptive control submodule considering system-level collaborative optimization is obtained. 6.The energy consumption device adaptive management and control platform based on ring spinning industrial big data of claim 5, characterized in that: During the collaborative optimization training process, a curriculum learning strategy is adopted, the training scene is gradually switched from a simple steady-state working condition to a complex dynamic working condition according to the working condition complexity, and a proximal policy optimization algorithm is used to update the model parameters. 7.The energy consumption device adaptive management and control platform based on ring spinning industrial big data of claim 5, wherein: The energy consumption coupling knowledge graph is constructed, specifically: based on the average mutual information intensity, the average correlation coefficient, and the energy transfer loss rate corresponding to the region to which the new device belongs, the overall energy consumption coupling degree index of the region is calculated; based on the overall energy consumption coupling degree index, the construction resolution is determined, and based on the determined construction resolution and the energy consumption device association topology network relationship corresponding to the region, the energy consumption coupling knowledge graph is established. 8.The energy consumption device adaptive management and control platform based on ring spinning industrial big data of claim 7, wherein: The average mutual information intensity is calculated by analyzing the mutual information value of the energy consumption data between the devices in the region; the average correlation coefficient is calculated by counting the correlation coefficient matrix of the energy consumption fluctuation between the devices; and the energy transfer loss rate is calculated by evaluating the transfer efficiency of the energy flow between the devices. 9.The energy consumption device adaptive management and control platform based on ring spinning industrial big data of claim 1, wherein: The third adaptive control submodule and other first adaptive control submodules are fused to obtain a new adaptive control module, including: performing consistency detection on the third adaptive control submodule and other first adaptive control submodules, and performing integration processing after consistency detection to obtain a new adaptive control module. 10.The energy consumption device adaptive management and control platform based on ring spinning industrial big data of claim 1, wherein: Based on the new adaptive control module, all energy consumption devices in the region including the new device are adaptively controlled, including: generating a trigger signal in response to the completion of the fusion of the new adaptive control module, and sending the trigger signal to the new device; wherein the trigger signal is used to trigger the new device to switch to receiving energy consumption control instructions from the adaptive control platform.

Images