Centralized control method and remote control system for evaporative cooling equipment

Abstract

The present application provides a centralized control method and remote control system for evaporative cooling equipment, which is based on an embedded branch network and a state mapping branch network to form an equipment status monitoring network. The normalization component in each branch network is deployed before the branch component. At the same time, the entry node and the exit node of the normalization component in the branch network are jump-connected. After completing the pre-normalization, the post-normalization jump connection is added. This connection mechanism ensures that the gradient can be stably transmitted in the deep network, so that even with the increase in the number of network layers, the debugging and stabilization of the network can be guaranteed, and the network has higher robustness. In addition, the design of post-normalization in the equipment status monitoring network can also ensure the universality and quality of the equipment status monitoring network, so as to increase the performance of the equipment status monitoring network, and the reliability of status monitoring is higher, so as to generate accurate control strategies.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a centralized control method and a remote control system for evaporative cooling equipment. Background Art

[0002] Evaporative cooling equipment plays an important role in many fields such as industrial production and environmental temperature regulation. With the development of technology, the demand for efficient control and accurate monitoring of evaporative cooling equipment is growing. Traditional evaporative cooling equipment control is mainly based on simple sensor feedback. For example, the operating parameters of the equipment, such as fan speed or cooling medium flow, are adjusted only based on the temperature value detected by the temperature sensor. This control method lacks comprehensive consideration of the overall operating status of the equipment and can only respond to a single factor, which can easily lead to low operating efficiency of the equipment under complex working conditions or failure to detect in time. In terms of equipment monitoring, traditional methods mostly set fixed threshold ranges for individual parameters to determine whether the equipment is operating normally. For example, when the outlet water temperature of the cooling equipment exceeds a preset maximum temperature, it is determined that there may be a problem with the equipment. However, this threshold-based monitoring method ignores the complex interrelationships between various parameters in the operation of the equipment and cannot accurately reflect the true operating status of the equipment.

[0003] With the development of artificial intelligence technology, deep neural networks have been tried to be applied to the monitoring of evaporative cooling equipment. However, there are many challenges in applying deep neural networks to such equipment. Deep neural networks are prone to gradient vanishing or gradient exploding problems when dealing with multi-layer structures. This is because as the number of network layers increases, the gradient value may become extremely small (gradient vanishing) or extremely large (gradient exploding) during the back propagation gradient calculation process, making the network difficult to train or the training results unstable. Summary of the invention

[0004] The object of the present invention is to provide a centralized control method and a remote control system for evaporative cooling equipment. This application is implemented as follows:

[0005] In a first aspect, an embodiment of the present application provides a centralized control method for evaporative cooling equipment, the method being applied to a remote control system, the remote control system being deployed with an equipment status monitoring network, the equipment status monitoring network comprising x embedded branch networks connected in sequence and y state mapping branch networks connected in sequence, each of the embedded branch networks and each of the state mapping branch networks comprising a normalization component and a branch component connected in sequence, x>1, y>2, the method comprising: obtaining an equipment monitoring parameter sequence of a target evaporative cooling equipment; loading the equipment monitoring parameter sequence into the x embedded branch networks, performing implicit representation extraction one by one based on each of the x embedded branch networks, and obtaining an embedded implicit representation of the output of the x embedded branch networks; each of the embedded branch networks is used to perform implicit representation extraction based on the normalization component and the branch component one by one , and based on the jump connections of the entry nodes and the exit nodes of the normalization component, the implicit representation extraction result of the embedded branch network is processed; the embedded implicit representation is loaded into the y state mapping branch networks, and the implicit representation extraction is performed one by one based on each of the y state mapping branch networks to obtain the state mapping implicit representation output by the y state mapping branch networks; each of the state mapping branch networks is used to perform implicit representation extraction based on the normalization component and the branch component one by one, and the implicit representation extraction result of the state mapping branch network is processed based on the jump connection; the cooling equipment operating state of the equipment monitoring parameter sequence is inferred based on the state mapping implicit representation; the corresponding equipment control strategy is generated based on the cooling equipment operating state, so as to control the target evaporative cooling equipment based on the equipment control strategy.

[0006] On the other hand, the present application provides a remote control system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method as described above is implemented.

[0007] The beneficial effects of the present application include at least: the present application is based on an embedded branch network and a state mapping branch network to form a device state monitoring network, the normalization component in each branch network is deployed before the branch component, and at the same time, the entry node and the exit node of the normalization component in the branch network are jump-connected, and the post-normalization jump connection is added on the premise of completing the pre-normalization. This connection mechanism ensures that the gradient can be stably transmitted in the deep network, so that even with the increase in the number of network layers, the debugging and stabilization of the network can be guaranteed, and the network has higher robustness. In addition, the design of post-normalization in the device state monitoring network can also ensure the universality and quality of the device state monitoring network, so as to increase the performance of the device state monitoring network, and the reliability of state monitoring is higher, so as to generate accurate control strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a flow chart of a centralized control method of evaporative cooling equipment provided in an embodiment of the present application.

[0009] Figure 2 It is a schematic diagram of the composition of a remote control system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] The execution subject of the centralized control method of the evaporative cooling device in the embodiment of the present application is a remote control system, including but not limited to a server, a personal computer, a laptop, a tablet computer, a smart phone, etc. Figure 1 As shown, the method is applied to a remote control system, which is deployed with a device status monitoring network. The device status monitoring network includes x embedded branch networks connected in sequence and y state mapping branch networks connected in sequence. Each embedded branch network and each state mapping branch network include a normalized component and a branch component connected in sequence, x>1, y>2.

[0011] The equipment state monitoring network is mainly composed of x embedded branch networks and y state mapping branch networks, where x>1, y>2. Each embedded branch network contains a normalization component and a branch component connected in sequence. The normalization component is of great significance in the entire network structure. For example, in data processing, the normalization component may use batch normalization technology. Suppose there is a parameter value in a set of equipment monitoring parameter sequences, such as the inlet water temperature monitoring value of the evaporative cooling equipment, and its original value range may be large and unstable. Batch normalization is processed by the following formula: , where x is the original input value (inlet water temperature monitoring value), μ is the mean in a batch of data, and σ is the standard deviation. is a very small number (for example, 10 -8 ), is the result of normalization. This normalization operation helps map the data to a more stable range, which is convenient for subsequent branch component processing.

[0012] The branch components can be different types of network structures or algorithms. Taking the parallel weight focusing component (such as the Multi-Head Self-Attention Network) as an example, when processing the implicit representation extraction embedded in the branch network, it can focus on the weights of different parts in the equipment monitoring parameter sequence. Assuming that the equipment monitoring parameter sequence contains multiple parameters, such as inlet water temperature, outlet water temperature, air flow rate, etc., the parallel weight focusing component can assign different weights according to the importance of these parameters in the evaluation of the equipment operation status. If at a certain moment, the outlet water temperature is more critical to judging the equipment operation status, then it will give a higher weight to the outlet water temperature parameter in order to extract the implicit representation more accurately.

[0013] When implicit representation extraction is performed in the embedding branch network, it is performed one by one based on the normalization component and the branch component. For example, for the first embedding branch network, the device monitoring parameter sequence is first input into the normalization component for data normalization, and then the normalized data is input into the branch component (such as the parallel weight focusing component or the multi-layer perceptron component) for processing to obtain the preliminary implicit representation result.

[0014] At the same time, the jump connections of the entry and exit nodes of the normalization component are used to process the implicit representation extraction results. Assume that in a simple example of an embedded branch network containing three nodes, node 1 is the input node (before entering the normalization component), node 2 is the output node of the normalization component, and node 3 is the output node of the branch component. The jump connection is a direct connection from node 1 to node 3. When calculating the final implicit representation, the results of the direct path from node 1 to node 3 and the regular path through node 2 are comprehensively considered. This jump connection mechanism ensures that the gradient can be stably transmitted in the deep network. For example, when the back-propagation algorithm calculates the gradient, the gradient can be directly transmitted from node 3 to node 1 through the jump connection, avoiding the problem of gradient vanishing or gradient exploding in the deep network, making the network easier to converge during the debugging process, and as the number of network layers increases, the network can still remain stable.

[0015] When there are x embedded branch networks (x>1), they are connected in sequence. For example, the output implicit representation of the first embedded branch network is used as the input of the second embedded branch network, and so on. This cascaded approach can gradually and deeply extract the implicit representation of the equipment monitoring parameter sequence. Each embedded branch network further refines and optimizes the implicit representation based on the extraction results of the previous network, like a multi-layer information extraction process. For example, the first embedded branch network may preliminarily extract the implicit representation of temperature-related parameters in the equipment operating status, and the second embedded branch network further combines flow-related parameters on this basis for a more comprehensive implicit representation extraction.

[0016] Similar to the embedding branch network, each state mapping branch network also includes a normalization component and a branch component connected in sequence. The normalization component also plays a role in stabilizing the data range, providing more favorable conditions for the processing of the branch component. The branch component can adopt different structures according to different design requirements. For example, in some state mapping branch networks, the branch component can be a parallel weight focusing component or a multi-layer perceptron component.

[0017] Taking the multi-layer perceptron component as an example, when processing the implicit representation extraction in the state mapping branch network, it performs nonlinear transformation on the implicit representation of the input through multiple neuron layers. Assuming that the implicit representation of the input is a vector that contains various feature information about the operating status of the device previously embedded in the branch network, the multi-layer perceptron component processes this vector through a series of weight matrices and activation functions (such as the ReLU function: y=max(0,x)), and gradually maps it to a representation space that is more in line with the target state mapping.

[0018] The implicit representation extraction in the state mapping branch network is also performed one by one based on the normalization component and the branch component. For each state mapping branch network, the input embedded implicit representation (output from the embedded branch network) is first normalized, and then processed by the branch component to obtain the preliminary state mapping implicit representation result. Similarly, the implicit representation extraction result of the state mapping branch network is processed using jump connections. For example, assuming a simple state mapping branch network structure, the input is I, which becomes N after the normalization component and then becomes O after the branch component. The jump connection allows the information transfer path from I to O directly to exist. When calculating the final state mapping implicit representation, the influence of these two paths is comprehensively considered. This jump connection mechanism also ensures the stable transmission of gradients in the state mapping branch network, allowing the network to debug and converge more effectively when dealing with complex device state mapping relationships.

[0019] When there are y state mapping branch networks (y>2) connected in sequence, the way they collaborate is similar to that of an embedded branch network. The output state mapping implicit representation of the previous state mapping branch network serves as the input of the next state mapping branch network. For example, the first state mapping branch network may initially map the embedded implicit representation of the equipment monitoring parameter sequence to a representation space related to a certain feature of the equipment operating state, such as a representation space related to the cooling efficiency of the equipment. On this basis, subsequent state mapping branch networks are further mapped to a more comprehensive equipment operating state representation space, such as a representation space that takes into account other factors such as cooling efficiency and equipment energy consumption.

[0020] Through the collaboration of the above-mentioned embedded branch network and state mapping branch network, the equipment state monitoring network can implicitly represent the cooling equipment operating state of the inference equipment monitoring parameter sequence based on the state mapping. For example, the equipment state monitoring network monitors and processes the parameters such as the equipment inlet water temperature, outlet water temperature, air flow rate, humidity, etc., and finally determines whether the equipment is currently in normal operation, inefficient operation, or fault warning state through layer-by-layer mapping and extraction of the network. This reasoning process is based on the intrinsic relationship between the equipment operating parameters and the operating state learned by the network during the training (debugging) process.

[0021] Based on the inferred operating status of the cooling equipment, the remote control system can generate the corresponding equipment control strategy. If the equipment is in an inefficient operating state, the control strategy may be to adjust the fan speed, adjust the opening of the water inlet valve, etc. For example, if it is inferred that the low cooling efficiency of the equipment is due to insufficient air flow rate, then the control strategy may be to increase the fan speed to increase the air flow rate, thereby improving the cooling effect of the equipment. The generation of this equipment control strategy is a key link for the remote control system to achieve centralized control of evaporative cooling equipment. It relies on the accurate state monitoring and reasoning capabilities of the equipment state monitoring network.

[0022] The above-mentioned centralized control method of evaporative cooling equipment comprises the following steps:

[0023] Step S100: Acquire a device monitoring parameter sequence of a target evaporative cooling device.

[0024] The equipment monitoring parameter sequence is a set of data that can reflect the operating status of the evaporative cooling equipment. These parameters cover multiple aspects, such as temperature parameters, flow parameters, pressure parameters, etc. Taking the common evaporative cooling air conditioning equipment as an example, the temperature parameters include the inlet temperature, outlet temperature, cooling water inlet temperature and outlet temperature. The inlet temperature reflects the initial temperature of the air entering the equipment, while the outlet temperature reflects the change in air temperature after the evaporative cooling process. The difference between the two is one of the important indicators for measuring the cooling effect. In terms of flow parameters, there is the flow rate of cooling water, and its size directly affects the cooling efficiency. If the cooling water flow rate is too small, it may not be able to fully take away the heat, resulting in poor cooling effect. Pressure parameters such as the air pressure inside the equipment are related to factors such as air flow resistance. Abnormal air pressure may indicate problems such as blockage or fan failure in the equipment.

[0025] The remote control system can use a variety of technical means to obtain these equipment monitoring parameter sequences. For some devices that have intelligent sensors and support data transmission protocols, the remote control system can receive these parameter data through network communication protocols (such as Modbus protocol, BACnet protocol, etc.). Taking the Modbus protocol as an example, the device acts as a slave station and the remote control system acts as a master station. The master station obtains the equipment monitoring parameters by sending a request frame to the slave station. The request frame contains information such as the device address, function code, data start address, and data length. The slave station returns the corresponding monitoring parameter data based on the content of the request frame.

[0026] In actual operation, in order to ensure that the acquired data is accurate and reliable, the remote control system may need to perform data verification. For example, a cyclic redundancy check (CRC) method is used. Assuming that the acquired device monitoring parameter sequence is represented in byte form as D=[d_1, d_2, …, d_n], a check code C is calculated through a specific CRC algorithm (such as CRC-16). When the device sends data, it sends data D and check code C at the same time. After receiving the data, the remote control system recalculates the check code C' according to the same CRC-16 algorithm. If C=C', it means that there is no error in the data transmission process and it can be reliably used for subsequent processing; if C≠C', it is necessary to re-request the device to send data or take error correction measures.

[0027] Once these equipment monitoring parameter sequences are successfully acquired by the remote control system, they become the basic data source for the subsequent analysis of the equipment operating status. For example, in the subsequent steps, these parameters will be loaded into the embedded branch network in the equipment status monitoring network for further processing. Through a series of complex implicit representation extraction and other operations, the equipment operating status is finally inferred and a suitable control strategy is generated. In short, the acquisition of the equipment monitoring parameter sequence in step S100 is an important prerequisite for the remote control system to effectively and centrally control the evaporative cooling equipment. Accurately and completely acquiring these parameters is crucial to the success of the entire control process.

[0028] Step S200: Load the equipment monitoring parameter sequence into x embedded branch networks, perform implicit representation extraction based on each of the x embedded branch networks one by one, and obtain embedded implicit representations output by the x embedded branch networks.

[0029] Among them, each embedded branch network is used to extract implicit representations based on the normalization component and the branch component one by one, and the implicit representation extraction results of the embedded branch network are processed based on the jump connections of the entry node and the exit node of the normalization component. The entry node is the input node, the exit node is the output node of the network, and the jump connection is the residual connection. When the embedded branch network is used for implicit representation extraction, it is assumed that an embedded branch network is processing the implicit representation extraction related to the monitoring parameters such as temperature, flow and pressure of the equipment. After the data passes through the normalization component, the traditional network structure can only pass the normalized result to the next layer (such as the branch component) for processing. However, the jump connection allows the original data before entering the normalization component (the data entering the node) and the data after being processed by the normalization component (the data leaving the node) to be considered together. For example, the original temperature data may contain some feature information that may be weakened during the normalization process. Through the jump connection, this information can be retained and participate in the subsequent implicit representation extraction together with the normalized information. In the back propagation process of the network, the stable transmission of the gradient is crucial for the training of the network. For deep embedded branch networks, as the number of network layers increases, traditional network structures are prone to gradient vanishing or gradient exploding problems. The jump connection provides an additional path for the transfer of gradients. When calculating the gradient, the gradient back-propagated from the subsequent network layer can be directly propagated to the entry node of the normalization component through the jump connection, rather than just through the regular path inside the normalization component. This helps to stabilize the transfer of gradients and avoid excessive attenuation or amplification of gradients in multi-layer networks. For example, assuming that in an embedded branch network with multiple hidden layers, without jump connections, the gradient may become very small after passing through several layers of normalization components and branch components, resulting in slow network training or failure to converge; with jump connections, the gradient can be more effectively transferred back, allowing the network to adjust parameters faster to optimize the results of implicit representation extraction. When processing the results of implicit representation extraction of embedded branch networks, jump connections enable the final implicit representation to integrate more information from more aspects. Taking the operating status monitoring of evaporative cooling equipment as an example, it is assumed that the implicit representation extraction aims to obtain feature representations related to the operating efficiency of the equipment. The raw data entering the node of the normalized component may contain the initial state information when the device is started, while the data leaving the node contains information that is more suitable for the current network layer calculation after normalization. By fusing these two parts of information through jump connections, the implicit representation obtained can more comprehensively and accurately reflect the characteristics related to the equipment's operating efficiency, thereby improving the accuracy of the entire embedded branch network in extracting the implicit representation related to the equipment's operating status.

[0030] In the embodiment of the present application, each embedded branch network includes a normalization component and a branch component connected in sequence. When processing the temperature data in the equipment monitoring parameters, assume that there is a set of monitoring values ​​of the cooling water inlet temperature, and its value range may fluctuate between 10°C and 30°C. The normalization component can use a method such as Min-MaxNormalization, and the formula is , where x is the original temperature monitoring value, x min is the minimum value in this set of temperature data, x max is the maximum value, x norm This is the normalized result. In this way, the temperature data is mapped to the range of 0-1, so that data of different scales can be processed at the same scale, which helps to improve the efficiency and accuracy of data processing by subsequent branch components.

[0031] The branch component is responsible for further feature extraction and conversion operations on the normalized data. Taking the multi-layer perceptron (MLP) as an example of a branch component, it is a common neural network structure. For the normalized equipment monitoring parameter sequence, the neurons in the MLP are weighted summed and activated by a function such as a sigmoid function. The function of is to transform the data nonlinearly. Assuming that the equipment monitoring parameter sequence contains multiple parameters such as inlet temperature, outlet temperature, cooling water flow, etc., these parameters are normalized as the input of MLP. MLP combines and transforms these input parameters through its internal neuron structure and weight parameters, thereby mining the hidden relationship between these parameters. This hidden relationship helps to infer the operating status of the equipment in the subsequent steps. When the remote control system loads the equipment monitoring parameter sequence into x embedded branch networks, an implicit representation extraction operation is performed for each embedded branch network. The purpose of this operation is to extract more abstract and representative feature information from the original equipment monitoring parameter sequence. For example, for an evaporative cooling device, there are complex relationships between the various parameters in its equipment monitoring parameter sequence. Implicit representation extraction can represent these relationships in a more compact and meaningful form. Assuming that the equipment monitoring parameter sequence contains temperature, flow and pressure data at multiple times, these data are just a series of discrete values ​​in their original form, but through implicit representation extraction, they may be converted into a vector representation that can reflect the overall operating trend of the equipment.

[0032] Taking a simple case with three embedding branch networks (x=3) as an example, the remote control system loads the equipment monitoring parameter sequence into the first embedding branch network. In this network, the equipment monitoring parameter sequence is first processed by the normalization component, such as the normalization method mentioned above to convert the data to a suitable range, and then enters the branch component for feature extraction. Assume that this branch component is a network structure based on the attention mechanism, which can assign different weights according to the importance of each parameter in the equipment monitoring parameter sequence to the equipment operating status. For example, at a certain moment, the outlet temperature of the equipment may be more critical to judging the cooling effect of the equipment, then the attention mechanism will assign a higher weight to this parameter, thereby highlighting the influence of this parameter in the implicit representation extraction process. After being processed by the first embedding branch network, a preliminary embedded implicit representation is obtained, which already contains the extraction of some key features in the equipment monitoring parameter sequence.

[0033] Next, the equipment monitoring parameter sequence (actually the result after processing by the first embedding branch network, which is a converted data form for the subsequent embedding branch network) is loaded into the second embedding branch network. Similarly, normalization is performed first. The normalization here may be adjusted according to the characteristics of the output data of the first embedding branch network, and then enters the branch component for further implicit representation extraction. This branch component may be a neural network with a different structure, such as a partial structure of a convolutional neural network (CNN). CNN is good at processing data with local correlation, and can better extract local features that may exist in the equipment monitoring parameter sequence (such as temperature change trends at adjacent moments). Through the processing of the second embedding branch network, a new embedded implicit representation is obtained. This representation further mines the feature information in the equipment monitoring parameter sequence based on the first embedding branch network.

[0034] Finally, the equipment monitoring parameter sequence (the result after being processed by the first two embedding branch networks) is loaded into the third embedding branch network. This embedding branch network also contains normalization components and branch components, and performs similar operations. For example, the branch component can be a recursive neural network (RNN) structure, which can process the time series information in the equipment monitoring parameter sequence. For evaporative cooling equipment, the changes of some parameters may have time dependencies. For example, as the equipment running time increases, the cooling efficiency of the equipment may gradually decrease. RNN can capture this characteristic relationship in time series. After being processed by the third embedding branch network, the final embedded implicit representation is obtained. This process of extracting implicit representations one by one through x embedding branch networks enables the remote control system to analyze and process the equipment monitoring parameter sequence from multiple angles and multiple levels. Each embedding branch network extracts and transforms unique features of the equipment monitoring parameter sequence based on the characteristics of its own normalization component and branch component. The embedded implicit representation output by the x embedding branch networks finally contains rich information related to the equipment operating status. This information will be used to further analyze the equipment operating status in the subsequent steps, thereby providing a basis for generating accurate equipment control strategies.

[0035] Step S300: Load the embedded implicit representation into y state mapping branch networks, perform implicit representation extraction based on each of the y state mapping branch networks one by one, and obtain the state mapping implicit representation output by the y state mapping branch networks.

[0036] Among them, each state mapping branch network is used to extract implicit representations based on normalization components and branch components one by one, and the implicit representation extraction results of the state mapping branch network are processed based on jump connections. When the state mapping branch network extracts implicit representations, it is assumed that the state mapping branch network is processing a variety of feature information related to the operating state of the equipment, such as the energy efficiency state of the equipment, the collaborative working state of the components, and other information. After the data passes through the normalization component, the traditional processing method can only use the normalized results for subsequent branch component processing. However, the jump connection allows the original embedded implicit representation before entering the normalization component (data entering the node) and the embedded implicit representation after processing by the normalization component (data leaving the node) to be considered simultaneously. For example, the original embedded implicit representation may contain the original feature relationship of the operating state of the equipment at a specific moment. The data processed by the normalization component is more suitable for branch component processing in terms of numerical range, but some original feature relationship information may be lost. Through jump connections, these different aspects of information can be integrated, making the final implicit representation extraction result more comprehensive. During the back propagation of the state-mapping branch network, the stable propagation of the gradient is critical for the training of the network. As the number of layers of the state-mapping branch network increases, if there is no skip connection, in the traditional network structure, the gradient may disappear or explode during the propagation process. The skip connection provides another path for the gradient propagation. When the gradient is back-propagated from the later network layer, the gradient can be directly propagated to the entry node of the normalization component through the skip connection, rather than just through the regular path inside the normalization component. This helps to maintain the stable propagation of the gradient in the network, allowing the network to be trained more efficiently. For example, in a complex structure with multiple continuous state-mapping branch networks, if there is no skip connection, the gradient may become extremely small or extremely large after passing through the normalization components and branch components of several state-mapping branch networks, making it difficult for the network to converge; with the skip connection, the gradient can be more stably returned, and the network can adjust the parameters faster to optimize the results of implicit representation extraction. When processing the implicit representation extraction results of the state-mapping branch network, the skip connection helps to improve the accuracy of the final state mapping implicit representation. Taking evaporative cooling equipment as an example, suppose we want to obtain an implicit representation of whether the equipment is in the best operating state through the state mapping branch network. The original embedded implicit representation of the normalized component entering the node may contain some basic feature information of the overall operating state of the equipment. The data leaving the node is more in line with the current network layer's requirements for data distribution after normalization. By fusing these two parts of information through jump connections, the final state mapping implicit representation can more accurately reflect whether the equipment is in the best operating state, thereby improving the accuracy of the entire state mapping branch network in mapping the equipment's operating state.

[0037] Specifically, each state mapping branch network contains a normalization component and a branch component connected in sequence. The normalization component continues to play a role in normalizing the data at this stage. For example, assuming that there are parameter values ​​related to device performance in the embedded implicit representation obtained in step S200, these values ​​may have different magnitudes and distribution ranges. If the Z-score normalization method is used, for a certain parameter x, its normalization formula is , where μ is the mean of the parameter and σ is the standard deviation. Through such normalization operation, the data is transformed into a distribution with a mean of 0 and a standard deviation of 1, so that the data has better stability and comparability when entering the branch component, which helps to improve the accuracy of subsequent processing.

[0038] The branch component performs deep feature extraction and mapping on the normalized data. Take a kernel-based support vector machine (SVM) as an example of a branch component. For the normalized embedded implicit representation, the SVM constructs a hyperplane to classify or regress the data. In the context of evaporative cooling equipment, assuming that certain features in the embedded implicit representation are related to the normal operation, inefficient operation, or fault state of the equipment, the SVM can learn the boundaries between these states based on pre-labeled training data (for example, by labeling feature samples in different states through historical equipment operation data), thereby mapping the input embedded implicit representation to the corresponding equipment operation state space.

[0039] When the remote control system executes the operation of loading the embedded implicit representation into y state mapping branch networks, an implicit representation extraction operation is performed for each state mapping branch network. This operation aims to deeply explore the information in the embedded implicit representation to construct a state mapping implicit representation that more accurately reflects the operating status of the equipment. Taking an actual evaporative cooling device as an example, its operating status is affected by a combination of factors, such as the temperature and flow rate of the cooling medium, the pressure inside the equipment, and the external ambient temperature. The embedded implicit representation obtained in step S200 may have preliminarily extracted some relationships between these factors, but in step S300, the state mapping branch network further clarifies the mapping relationship between these relationships and the specific operating status of the equipment (such as whether the cooling efficiency meets the standard, whether the equipment is in energy-saving operation mode, etc.).

[0040] For example, in a network with four state mapping branches (y=4), the remote control system loads the embedded implicit representation into the first state mapping branch network. The normalization component in this network performs data normalization on the input embedded implicit representation, for example, normalizing the features related to the cooling efficiency of the equipment to meet the input requirements of the subsequent branch components. Then, the branch component (such as the kernel function-based SVM mentioned above) processes the normalized data. Assuming that the cooling efficiency of the equipment is related to factors such as the inlet water temperature and air flow, the SVM establishes a mapping relationship model between these factors and the cooling efficiency by learning historical data. When a new embedded implicit representation is input, the SVM calculates according to this model and preliminarily obtains a state mapping implicit representation related to the cooling efficiency. This representation can reflect the current cooling efficiency status of the equipment to a certain extent, such as a preliminary judgment of whether it is in a high efficiency, medium efficiency or low efficiency state.

[0041] Next, the result processed by the first state mapping branch network (including the preliminary state mapping implicit representation related to cooling efficiency) is loaded into the second state mapping branch network. In this network, the normalization component adjusts the data again to adapt to the characteristics of the branch component in this network. Assume that this branch component is a deep neural network (DNN) with multiple hidden layers. DNN further mines the deep features in the data by performing multi-layer nonlinear transformations on the input data. For evaporative cooling equipment, in addition to cooling efficiency, the energy consumption of the equipment is also an important operating status indicator. DNN can use the cooling efficiency-related information obtained in the first state mapping branch network and other energy consumption-related factors embedded in the implicit representation (such as equipment operating power, operating time of different components, etc.) to establish a more complex relationship model between cooling efficiency and energy consumption. After being processed by the second state mapping branch network, a state mapping implicit representation containing the comprehensive relationship between cooling efficiency and energy consumption is obtained. This representation reflects the operating status of the equipment more comprehensively than the result output by the first state mapping branch network.

[0042] Subsequently, the data (the result of processing by the first two state mapping branch networks) is loaded into the third state mapping branch network. The normalization component and branch components of this network continue to work together. For example, the branch component adopts an integrated learning model, such as Random Forest. Random Forest consists of multiple decision trees. For the input state mapping implicit representation containing cooling efficiency and energy consumption information, each decision tree can analyze the operating status of the equipment from a different perspective. For example, some decision trees may pay more attention to the changes in the operating status of the equipment in different seasons (external ambient temperature changes), while other decision trees may focus on analyzing the impact of the collaborative work of different components of the equipment on the overall operating status. Through the comprehensive decision of Random Forest, a more accurate and comprehensive implicit representation of the state mapping is obtained, which takes into account more factors and situations and describes the operating status of the equipment in a more detailed manner.

[0043] Finally, the data (the result of processing by the first three state mapping branch networks) is loaded into the fourth state mapping branch network. This network also contains normalization components and branch components. Assume that the branch component is a structure based on a Bayesian Network. The Bayesian Network can infer the various operating states of the equipment based on probabilistic relationships. Based on the cooling efficiency, energy consumption and other related factors of the equipment mined by the first three state mapping branch networks, the Bayesian Network can use prior probabilities (for example, the probability distribution of the equipment under different operating states obtained based on historical data) and conditional probabilities (such as the probability of the equipment being in a certain operating state under specific cooling efficiency and energy consumption conditions) to infer the final operating state of the equipment and obtain the final state mapping implicit representation. This state mapping implicit representation contains a very comprehensive and in-depth description of the equipment's operating state, including whether the equipment is operating normally, whether maintenance is required, whether there are potential faults, etc., which provides a solid foundation for accurately inferring the equipment's operating state and generating effective equipment control strategies in the subsequent steps.

[0044] Through the process of extracting implicit representations one by one through y state mapping branch networks, the remote control system can analyze and process the embedded implicit representations from multiple dimensions and multiple levels. Each state mapping branch network performs unique feature extraction and transformation on the embedded implicit representation based on the characteristics of its own normalized components and branch components. The state mapping implicit representation output by the y state mapping branch networks finally contains rich and in-depth information related to the equipment operation status. This information will be used in subsequent steps to accurately judge the equipment operation status, thereby providing a strong basis for generating reasonable equipment control strategies.

[0045] Step S400: implicitly representing the cooling device operating state of the inference device monitoring parameter sequence based on the state mapping.

[0046] In the embodiment of the present application, the implicit representation of the state mapping is the result obtained after the processing of the y state mapping branch networks in step S300, which contains rich information related to the operation status of the equipment. This information exists in a form of deep mining and abstraction, and is no longer a direct measurement value like the original equipment monitoring parameter sequence, but a more representative representation of the characteristics of the equipment operation status.

[0047] In order to infer the operating status of the equipment based on the implicit representation of the state mapping, the remote control system can adopt a variety of technical means. For example, the rule-based reasoning method. Assume that there are some known rules in the evaporative cooling equipment, which are based on the physical principles of the equipment, engineering experience and historical data. If a certain feature value in the implicit representation of the state mapping is compared with the standard range of the feature when the equipment is operating normally, and there is an out-of-range situation, it can be inferred that the equipment may be in an abnormal operating state.

[0048] Taking the temperature control in evaporative cooling equipment as an example, the temperature-related features of different locations inside the equipment may be included in the implicit representation of the state mapping. Under normal circumstances, the temperature of the cooled air should be within a specific range. If the temperature feature value is found to be higher than the upper limit of the normal range through the analysis of the implicit representation of the state mapping, the remote control system can infer that the cooling efficiency of the equipment may be reduced, the equipment may be in an inefficient operating state, or there may be hidden dangers of failure based on the relationship between temperature and equipment cooling effect.

[0049] In addition, the remote control system can also use machine learning models for reasoning. For example, the decision tree model can be used to classify and reason about the operating status of the equipment. The decision tree is based on a series of conditional judgments to divide different categories. When building a decision tree model, various features in the implicit representation of the state mapping are used as input, and the operating status of the equipment (such as normal operation, maintenance required, failure, etc.) is used as output for training.

[0050] Assume that there are features such as equipment flow, pressure, and temperature change rate in the implicit representation of the state mapping. The decision tree model may first make a judgment based on the flow characteristics. If the flow is lower than a certain threshold, it will further judge based on the pressure characteristics. If the pressure is also abnormal, it will finally combine the temperature change rate and other features to finally determine that the equipment is in a fault state. This decision tree-based reasoning method can systematically process the relationship between multiple features in the implicit representation of the state mapping, thereby accurately inferring the operating status of the equipment.

[0051] For another example, a neural network model (such as a multilayer perceptron) is used for reasoning. The multilayer perceptron processes the implicit representation of the input state mapping through multiple neuron layers. Assume that the input state mapping is implicitly represented as X=[x_1, x_2, …,x_n], where x_i represents different eigenvalues. The first layer of neurons in the multilayer perceptron performs a weighted sum operation on the input, such as , where w ij is the weight, b j is a bias term, and then a nonlinear transformation is performed through an activation function (such as the ReLU function y=max(0,x)). After multiple layers of such processing, a result representing the operating status of the device is output. For example, if the output value is close to 0, it means that the device is operating normally, and if it is close to 1, it means that the device is in a fault state, etc.

[0052] Step S500: Generate a corresponding device control strategy based on the operating state of the cooling device, so as to control the target evaporative cooling device based on the device control strategy.

[0053] After the operating state of the equipment is inferred, different operating states require different control strategies to maintain the normal operation of the equipment, improve efficiency, or perform fault repair. For example, if the equipment is inferred to be in a normal operating state, the control strategy may focus on optimizing the operating parameters of the equipment to improve energy efficiency. For evaporative cooling equipment, this may involve fine-tuning parameters such as the cooling medium flow rate and fan speed.

[0054] Suppose that through analysis, it is found that the equipment is in normal operation, but there is room for improvement in cooling efficiency. The remote control system can formulate a control strategy based on the operating principle of the equipment and relevant engineering knowledge. Taking the fan speed as an example, according to the principle of heat transfer, there is a certain relationship between air flow rate and cooling efficiency, which can generally be expressed by the formula Q=hAΔT, where Q is the amount of heat transferred, h is the heat transfer coefficient, A is the heat transfer area, and ΔT is the temperature difference. Increasing the fan speed can increase the air flow rate, which may increase the heat transfer coefficient h, and thus improve the cooling efficiency. In this case, the control strategy may be to increase the fan speed by a certain percentage, such as 10%. This decision is based on a comprehensive consideration of the current operating status of the equipment (cooling efficiency can be improved), the operating principle of the equipment (heat transfer relationship), and actual engineering experience (the relationship between fan speed and air flow rate).

[0055] If the equipment is in an inefficient operating state, the control strategy is more complex and targeted. It may be necessary to adjust multiple equipment parameters and even troubleshooting. For example, if it is found that the inefficient operation of the equipment is caused by insufficient cooling medium flow, the remote control system can formulate a control strategy to increase the flow of the cooling medium. This may involve adjusting the valve opening. Assume that there is a linear relationship between valve opening and flow rate Q=k×d, where Q is the flow rate, k is the proportionality factor, and d is the valve opening. If the current flow rate is Q_1 and the expected flow rate is Q_2, then the valve opening that needs to be adjusted Δd=(Q2-Q1) / k can be calculated, and then the remote control system sends a command to increase the valve opening by Δd.

[0056] When the equipment is in a faulty state, the control strategy must first ensure the safety of the equipment, and then try to repair the fault or switch the equipment to a backup device (if any). For example, if a key sensor failure is detected in the equipment, the control strategy of the remote control system may be to stop the equipment from running to avoid possible dangerous situations, and at the same time issue an alarm to notify maintenance personnel to replace or repair the sensor. This is based on the principle that equipment failure may lead to unpredictable operating risks, and ensuring the safety of equipment and personnel is the top priority.

[0057] To generate these equipment control strategies, remote control systems can use rule-based methods. These rules are pre-set based on equipment characteristics, operating principles, historical data, and engineering experience. For example, for different equipment operating state ranges (such as different temperature ranges, pressure ranges, etc.), there are corresponding control strategy rules.

[0058] In addition, intelligent algorithms can also be used to generate control strategies. For example, a reinforcement learning algorithm is used to take the operating state of the equipment as the environmental state and the control operation of the equipment (such as adjusting the fan speed, valve opening, etc.) as the action. Through continuous trial and error and learning, the equipment can achieve the best operating effect in long-term operation. In reinforcement learning, a reward function R(s,a) is defined, where s represents the state and a represents the action. If an action causes the equipment to change from an inefficient operating state to an efficient operating state, then this action will receive a higher reward value. Through continuous exploration and optimization, a control strategy that maximizes the reward function is found.

[0059] In one implementation, the equipment status monitoring network is debugged by the following steps:

[0060] Step S10: Obtain equipment monitoring parameter sequence samples;

[0061] Step S20: Load the equipment monitoring parameter sequence samples into x embedded branch networks, perform implicit representation extraction based on each of the x embedded branch networks one by one, and obtain embedded implicit representations output by the x embedded branch networks; wherein each embedded branch network is used to perform implicit representation extraction based on the normalization component and the branch component one by one, and process the implicit representation extraction results of the embedded branch network based on the jump connections of the entry node and the exit node of the normalization component;

[0062] Step S30: Load the embedded implicit representation into y state mapping branch networks, perform implicit representation extraction based on each of the y state mapping branch networks, and obtain the state mapping implicit representation output by the y state mapping branch networks; wherein each state mapping branch network is used to perform implicit representation extraction based on the normalization component and the branch component one by one, and process the implicit representation extraction result of the state mapping branch network based on the jump connection;

[0063] Step S40: implicitly representing the sample cooling device operating state of the inference device monitoring parameter sequence sample based on the state mapping;

[0064] Step S50: Optimizing network parameters of the equipment status monitoring network based on the difference between the prior cooling equipment operating state of the equipment monitoring parameter sequence sample and the sample cooling equipment operating state.

[0065] In step S10, the remote control system obtains a sequence of samples of equipment monitoring parameters. These samples are data sets that can reflect various characteristics of the evaporative cooling equipment under different operating conditions. For example, for a typical evaporative cooling equipment, the monitoring parameters may include water inlet temperature, water outlet temperature, air flow, humidity, equipment internal pressure, etc. These parameters will have different value ranges or change trends when the equipment is operating normally, operating inefficiently, or in a fault state.

[0066] In the embodiment of the present application, the remote control system can obtain these sample data by communicating with the sensor network of the device. The sensor network can transmit data to the remote control system based on various communication protocols, such as Modbus, BACnet, etc. Assuming that the sensor of a device collects data at a certain interval (for example, 5 minutes), a large number of device monitoring parameter sequence samples can be accumulated within a period of time (such as one day, one week or one month).

[0067] In step S20, the remote control system loads the acquired equipment monitoring parameter sequence samples into x embedded branch networks. Each embedded branch network contains a normalization component and a branch component, and this structure helps to process the data effectively. For example, for the parameter of inlet water temperature, if its value range is between 10℃-30℃, and the numerical range of other parameters such as air flow is quite different from the numerical range of temperature, the normalization component can use the minimum-maximum normalization method, which can map parameter values ​​of different magnitudes to a relatively uniform interval, which is convenient for subsequent branch components to process. Branch components can be in various forms, such as multi-layer perceptron (MLP). Take a simple MLP as an example, assuming that it has an input layer, a hidden layer and an output layer. The input layer receives the normalized equipment monitoring parameter sequence samples, and the neurons in the hidden layer perform nonlinear transformation on the input data through weighted summation and activation function (such as ReLU function y=max(0,x)), and the output layer outputs the processed implicit representation.

[0068] For each embedding branch network, implicit representation extraction is performed one by one in this way. Assume that the first embedding branch network may focus more on feature extraction of temperature-related parameters, while the second embedding branch network may focus on feature extraction of flow-related parameters. Through the processing of x embedding branch networks, the embedded implicit representations of the outputs of x embedding branch networks are finally obtained.

[0069] In step S30, the remote control system then loads the embedded implicit representation obtained in step S20 into y state mapping branch networks. Each state mapping branch network also has a structure of a normalization component and a branch component. In the state mapping branch network, the normalization component continues to normalize the input embedded implicit representation. For example, if the distribution of a certain feature value in the embedded implicit representation is relatively scattered, normalization can make it more in line with the processing requirements of subsequent branch components.

[0070] The branch component can adopt a structure such as a support vector machine (SVM). The basic principle of SVM is to classify or regress data by constructing a hyperplane. Assume that in the state mapping of evaporative cooling equipment, it is necessary to distinguish between the normal operating state and the fault state of the equipment. SVM can learn an optimal hyperplane based on pre-labeled training data (i.e., sample data of which embedded implicit representations correspond to normal operation and which are in fault state), map the input embedded implicit representation to the corresponding equipment operation state space, and thus obtain a preliminary state mapping implicit representation.

[0071] Such an implicit representation extraction operation is performed on each state mapping branch network. Since different state mapping branch networks may focus on different aspects of device state characteristics, for example, one may focus on the energy efficiency status of the device, and another may focus on the coordination of the operation of the device components, etc., through the processing of y state mapping branch networks, the state mapping implicit representation output by y state mapping branch networks is finally obtained.

[0072] In step S40, the remote control system uses the state mapping implicit representation obtained in step S30 to infer the sample cooling device operating state of the device monitoring parameter sequence sample.

[0073] One way is to perform reasoning by constructing logical rules. For example, if a feature value in the implicit representation of the state mapping is compared with the threshold of the feature under the pre-set normal operating state, if it exceeds the threshold range, it may indicate that the device is in an abnormal operating state. Assume that there is a feature related to the cooling efficiency of the device in the implicit representation of the state mapping. Under normal circumstances, the feature value should be between 0.8 and 1. If the reasoning finds that this value is less than 0.8, then it can be inferred that the cooling efficiency of the device may be low and the device is in a non-ideal operating state.

[0074] Another way is to use machine learning models for reasoning. For example, use a decision tree model. Decision trees classify data based on a series of conditional judgments. Multiple features in the implicit representation of the state mapping are used as inputs to the decision tree, and the operating status of the equipment (such as normal, inefficient, faulty, etc.) is used as output for training. During reasoning, the operating status of the equipment is gradually determined according to the node judgment rules of the decision tree.

[0075] In step S50, the remote control system optimizes the network parameters of the equipment status monitoring network according to the difference between the a priori cooling equipment operating state of the equipment monitoring parameter sequence sample (i.e., the known actual equipment operating state, which can be determined by manual inspection, more accurate equipment detection means or historical data) and the sample cooling equipment operating state inferred in step S40.

[0076] If it is found that the inferred operating state is significantly different from the prior operating state, it means that the current parameter settings of the equipment condition monitoring network may not be accurate. For example, in the branch components of the neural network structure, the network parameters include the connection weights and bias terms between neurons. The back propagation algorithm can be used to adjust these parameters.

[0077] Assume that in a simple neural network structure, for a certain layer of neuron output , where x i is the input value, w ij is the connection weight, b jis the bias term. The back propagation algorithm calculates the gradient from the output layer to the input layer layer by layer based on the error (the difference between the prior state and the inference state), and then updates the weight w according to the gradient descent method ij and the bias term b j , the formula is and , where α is the learning rate and E is the error function.

[0078] By continuously using the equipment monitoring parameter sequence samples to perform the above steps (S10-S50) in a loop, the remote control system can gradually optimize the parameters of the equipment status monitoring network, so that the network can more accurately monitor the operating status of the evaporative cooling equipment, thereby providing a reliable basis for the subsequent generation of accurate control strategies based on the equipment operating status. This training process is an iterative process that requires a large amount of sample data and a reasonable network structure and parameter settings to ensure the effectiveness and accuracy of the equipment status monitoring network.

[0079] In practical applications, in order to improve the training effect, some data enhancement techniques can also be used, such as adding random noise to sample data, data scaling and other operations to increase the diversity of sample data. At the same time, methods such as cross-validation can also be used to evaluate the performance of the equipment status monitoring network to avoid problems such as overfitting. For example, using k-fold cross-validation, the sample data is divided into k subsets, and k-1 subsets are used as training sets each time, and the remaining subset is used as a validation set. Repeat k times, and take the average validation result to evaluate the network performance. According to the evaluation results, the network structure or training parameters are further adjusted to improve the generalization ability and accuracy of the equipment status monitoring network.

[0080] In addition, in the process of optimizing network parameters, the mutual influence between different parameters also needs to be considered. For example, in the embedded branch network and the state mapping branch network, there may be synergy between the parameters of the normalization component and the parameters of the branch component. If the parameters of the normalization component are not set reasonably, it may affect the data processing effect of the branch component, and then affect the accuracy of the entire network's monitoring of the device operation status. Therefore, when optimizing network parameters, it is necessary to comprehensively consider the relationship between the various components and adopt an overall optimization strategy instead of optimizing a single parameter.

[0081] Moreover, as the equipment is continuously operated and used, the characteristics of the equipment may change. For example, the wear of equipment components may cause changes in the operating parameters of the equipment. This requires the equipment status monitoring network to have a certain degree of adaptive ability. The remote control system can regularly re-acquire equipment monitoring parameter sequence samples, re-train and optimize the network to adapt to changes in the equipment. For example, for long-term evaporative cooling equipment, the equipment status monitoring network is retrained every quarter or every six months to ensure that the network can accurately monitor the current operating status of the equipment.

[0082] As an implementation mode, the equipment status monitoring network includes t embedded networks connected in sequence, each embedded network includes a plurality of embedded branch networks connected in sequence in x embedded branch networks, t>1. Based on this, step S20, loading the equipment monitoring parameter sequence samples into the x embedded branch networks, performing implicit representation extraction one by one based on each embedded branch network in the x embedded branch networks, and obtaining the embedded implicit representation output by the x embedded branch networks, may include:

[0083] Step S21: Load the equipment monitoring parameter sequence samples into the first embedding network, perform implicit representation extraction based on each embedding branch network in the first embedding network one by one, and obtain the output implicit representation of the first embedding network;

[0084] Step S22: Load the output implicit representation of the e-th embedding network into the f-th embedding network, perform implicit representation extraction based on each embedding branch network in the f-th embedding network one by one, and obtain the output implicit representation of the f-th embedding network, f=e+1, and e≥1, f≤t.

[0085] Further, the equipment state monitoring network includes t state mapping networks connected in sequence, and each state mapping network includes two or more state mapping branch networks connected in sequence in y state mapping branch networks. Based on this, step S30, loading the embedded implicit representation into the y state mapping branch networks, performing implicit representation extraction one by one based on each state mapping branch network in the y state mapping branch networks, and obtaining the state mapping implicit representation output by the y state mapping branch networks, may include:

[0086] Step S31: Load the output implicit representation of the t-th embedding network into the first state mapping network, perform implicit representation extraction based on each state mapping branch network in the first state mapping network one by one, and obtain the output implicit representation of the first state mapping network;

[0087] Step S32: Load the output implicit representation of the t-th embedding network and the output implicit representation of the g-th state mapping network into the h-th state mapping network, perform implicit representation extraction one by one based on each state mapping branch network in the h-th state mapping network, and obtain the output implicit representation of the h-th state mapping network, h=g+1, and g≥1, h≤t.

[0088] In step S21, the remote control system loads the equipment monitoring parameter sequence sample to the first embedded network. This equipment monitoring parameter sequence sample contains parameter information related to various operating states of the evaporative cooling equipment, such as temperature, flow, pressure and other parameter values. For example, for an industrial evaporative cooling equipment, the equipment monitoring parameter sequence sample may contain data such as the cooling water inlet temperature value between 15-30°C, the air flow rate between 5-15 cubic meters per second, and the internal pressure of the equipment between 0.5-1.5 atmospheres.

[0089] The first embedding network contains x embedding branch networks, and for each of the embedding branch networks, implicit representation extraction is performed. In this process, as mentioned above, each embedding branch network contains a normalization component and a branch component. The normalized parameters enter the branch component for processing. The branch component can be in various forms, such as using part of the structure of a convolutional neural network (CNN) as a branch component. The convolution kernel in the CNN performs a convolution operation on the normalized parameter data to extract local features in the data. For example, for a 3x3 convolution kernel, it slides on the data matrix (assuming that the equipment monitoring parameter sequence samples are constructed in the form of a data matrix), calculates the weighted sum of the data in the area covered by the convolution kernel, and thus extracts the local feature relationships in the data. These feature relationships are implicit in the original equipment monitoring parameter sequence samples, and are extracted in this way and gradually constructed into implicit representations. After such processing of each embedding branch network in the first embedding network, the output implicit representation of the first embedding network is finally obtained. This output implicitly indicates that preliminary feature extraction and information conversion have been performed on the original equipment monitoring parameter sequence samples. It is no longer in the form of original parameter values, but a more abstract representation that better reflects the characteristics of the equipment operating status.

[0090] In step S22, when the remote control system executes this step, it is performing a cascade process of multiple rounds of embedding networks. Here, e and f represent the sequence numbers of the embedding networks, where f=e+1, indicating that the output implicit representation of the previous embedding network is used as the input of the next embedding network. For example, when e=1, the output implicit representation of the first embedding network (which has been processed in step S21) is loaded into the second embedding network (i.e., f=2).

[0091] For each embedded branch network in the f-th embedded network, the implicit representation extraction operation is also performed. This process is similar to the operation in the first embedded network in step S21, but since the input is the output implicit representation of the previous embedded network, its data features have undergone a certain degree of transformation. Assume that the output implicit representation of the e-th embedded network is in the form of a matrix containing multiple feature vectors (for example, these feature vectors are obtained by combining and processing different parameters in the equipment monitoring parameter sequence samples by the previous embedded branch network). When this output implicit representation enters the embedded branch network of the f-th embedded network, the normalization component still plays the role of adjusting the data range. Taking a simple linear normalization as an example, if the range of an element value y in the input feature vector is [-1, 1], but the branch component expects the input to be in the range of [0, 1], the formula can be used. Normalize. The branch component continues to process the normalized data. For example, a recursive neural network (RNN) is used as a branch component. RNN can process the time dependency in sequence data. For equipment operation status monitoring, some operation status features of the equipment may be related in time. Assuming that the cooling efficiency of the equipment is interrelated at different time points, RNN can process the input implicit representation of the current moment according to the implicit representation features of the previous moment, so as to further explore the temporal feature relationship in the data. After being processed by each embedded branch network in the fth embedded network, the output implicit representation of the fth embedded network is obtained. This output implicit representation further extracts and transforms the features related to the equipment operation status on the basis of the previous embedded network. With the continuous cascading of the embedded network (e continues to increase, f increases accordingly), the feature representation of the equipment operation status becomes more and more comprehensive and in-depth.

[0092] In step S31, the remote control system loads the output implicit representation of the tth embedded network into the first state mapping network at this step. The output implicit representation of the tth embedded network has been processed by multiple previous embedded networks and embedded branch networks, and contains rich characteristics related to the equipment operation status. For example, this implicit representation may include the complex relationship characteristics between parameters such as temperature, flow, pressure, etc. of the equipment under different operating periods, as well as the characteristics of the changing trends of these parameters over time. The first state mapping network contains y state mapping branch networks, and for each state mapping branch network, implicit representation extraction must be performed. In each state mapping branch network, the normalization component first processes the input data. Assuming a variant of maximum-minimum normalization, for a certain eigenvalue z in the implicit representation of the input, if the minimum value of the vector in which it is located is z min , the maximum value is z max , and the value range of z is [z min ,z max], the normalized formula is ( is a small positive number to prevent the denominator from being zero). For example, if z=0.5, z min =0,z max =1, ,but .

[0093] Then, the branch component processes the normalized data. Taking the support vector machine (SVM) as a branch component as an example, the SVM constructs a hyperplane to classify or regress the data. In the context of device state mapping, assuming that the features in the input implicit representation are related to the normal operation, inefficient operation, and faulty operation of the device, the SVM learns an optimal hyperplane based on pre-labeled training data (for example, which implicit representations correspond to normal operation and which are faulty operation in historical data) to map the input implicit representation to the corresponding device operation state space. For example, if some feature values ​​in the input implicit representation are combined on one side of the hyperplane, it may indicate that the device is in a normal operation state, while on the other side it may indicate that the device is in an abnormal operation state. After such processing by each state mapping branch network in the first state mapping network, the output implicit representation of the first state mapping network is obtained, and this output implicit representation begins to map the implicit representation obtained by the embedding network to the actual operation state of the device.

[0094] In step S32, when the remote control system executes this step, it is cascading multiple state mapping networks. Here, g and h represent the serial numbers of the state mapping networks, and h is 1 greater than g, indicating that the output implicit representation of the t-th embedded network and the output implicit representation of the g-th state mapping network are used together as the input of the h-th state mapping network. For example, when g=1, the output implicit representation of the t-th embedded network and the output implicit representation of the first state mapping network are loaded into the second state mapping network (i.e., h=2).

[0095] For each state mapping branch network in the hth state mapping network, an implicit representation extraction operation is performed. Since the input contains implicit representations from the embedding network and the output of the previous state mapping network, the data is more complex and contains more information. In each state mapping branch network, the normalization component adjusts the input data. For example, using the batch normalization technique, a feature in a batch of input data (here, a batch of input data composed of implicit representations from two sources) is processed according to the batch normalization formula, which is , where x is the original input eigenvalue, μ B is the mean value in this batch of data, σ B is the standard deviation, is a very small number. is the normalized result.

[0096] The branch component processes the normalized data. Taking the multi-layer perceptron (MLP) as an example, the MLP performs nonlinear transformation on the input data through multiple neuron layers. Assuming that the implicit representation of the input is a high-dimensional vector, the input layer of the MLP receives this vector, and then performs weighted summation and activation function (such as ReLU function y=max(0,x)) operations through the neurons of the hidden layer, and finally the output layer outputs the processed implicit representation. For example, if the input vector has n elements [x_1,x_2,…,x_n], for the jth neuron in the first hidden layer, its output , where w ij is the connection weight, b j is the bias term. After being processed by each state mapping branch network in the hth state mapping network, the output implicit representation of the hth state mapping network is obtained. This output implicit representation further integrates more information on the basis of the previous state mapping network, and the mapping of the device operation status is more accurate and comprehensive. With the continuous cascading of the state mapping network (g continues to increase, h increases accordingly), the final output implicit representation can more accurately reflect the operation status of the device.

[0097] As an implementation mode, each embedding network includes a first embedding branch network and a second embedding branch network connected in sequence, the branch component of the first embedding branch network is a parallel weight focusing component, and the branch component of the second embedding branch network is a multi-layer perceptron component. Based on this, step S22, loading the output implicit representation of the e-th embedding network into the f-th embedding network, performing implicit representation extraction one by one based on each embedding branch network in the f-th embedding network, and obtaining the output implicit representation of the f-th embedding network, may include:

[0098] Step S221: Loading the output implicit representation of the e-th embedding network into the first embedding branch network of the f-th embedding network;

[0099] Step S222: extracting implicit representations based on the normalization components and the parallel weight focusing components of the first embedding branch network one by one;

[0100] Step S223: summing the output implicit representation of the e-th embedding network and the output implicit representation of the normalized component of the first embedding branch network based on the skip connection and the output implicit representation of the parallel weighted focusing component of the first embedding branch network to obtain the output implicit representation of the first embedding branch network;

[0101] Step S224: loading the output implicit representation of the first embedding branch network into the second embedding branch network of the f-th embedding network;

[0102] Step S225: extracting implicit representations based on the normalization components and multi-layer perceptron components of the second embedding branch network one by one;

[0103] Step S226: The output implicit representation of the first embedding branch network and the output implicit representation of the normalized component of the second embedding branch network are summed based on the jump connection and the output implicit representation of the multilayer perceptron component of the second embedding branch network to obtain the output implicit representation of the f-th embedding network.

[0104] In step S221, the output implicit representation obtained after a series of processing of the e-th embedded network is passed to the first embedded branch network in the f-th embedded network. The output implicit representation of the e-th embedded network is a result of feature abstraction and conversion of the equipment monitoring parameter sequence samples after processing by the previous network. For example, suppose that the e-th embedded network processes the monitoring parameter sequence samples related to the temperature and flow of the evaporative cooling equipment. After being processed by e embedded networks, the output implicit representation may be a vector containing the complex relationship characteristics between temperature and flow. This vector is no longer the original temperature and flow values, but exists in a form that can reflect the coordinated characteristics of temperature and flow under the operating state of the equipment.

[0105] When this output implicit representation is loaded into the first embedding branch network of the f-th embedding network, it becomes the input data for further processing by this branch network. This operation is similar to the transfer of data between network layers, laying the foundation for deeper feature extraction and transformation in the subsequent specific branch network in the f-th embedding network.

[0106] In step S222, implicit representation extraction is performed based on the normalization component and the parallel weight focusing component of the first embedding branch network one by one. In the first embedding branch network, the normalization component first processes the input implicit representation. The normalization component plays an important role in the entire network structure, and it can adjust the input data to a range that is more suitable for subsequent component processing.

[0107] For example, suppose that the value range of a certain eigenvalue in the implicit representation output by the e-th embedding network fluctuates greatly. If the Z-score normalization method is used, for the eigenvalue x, the normalization formula is , where μ is the mean of the eigenvalue in the current input data and σ is the standard deviation. Through this normalization operation, the distribution of the eigenvalue can be converted to a standard normal distribution with a mean of 0 and a standard deviation of 1 (approximately). This helps to improve the stability and efficiency of subsequent parallel weight focusing components in processing data.

[0108] The parallel weighted focusing component is a structure that can perform weighted processing on the input data to focus on the importance of different features, such as the Multi-Head Self-Attention Network. When processing implicit representations, it can dynamically assign weights based on the relationship between different elements in the input implicit representation.

[0109] Taking evaporative cooling equipment as an example, assume that the implicit representation of the input contains information such as temperature characteristics and flow characteristics of different parts of the equipment. The parallel weighted focusing component can assign different weights to different features based on the correlation between these features under the operating state of the equipment. For example, if under a certain operating state, the temperature of a key part of the equipment is more important for judging the overall operating state, then the parallel weighted focusing component will assign a higher weight to this temperature feature. It calculates the correlation between the elements in the implicit representation of the input to determine the weight of each element, and then performs weighted summation and other operations on the input to achieve further extraction of the implicit representation. This method can effectively mine the hidden feature relationships in the implicit representation of the input, making it possible to more accurately represent the operating state of the equipment in the future.

[0110] In step S223, the jump connection allows the output implicit representation of the e-th embedded network and the output implicit representation of the normalized component of the first embedded branch network to be directly summed with the output implicit representation of the parallel weight focusing component. This jump connection mechanism breaks the limitation that information can only be transmitted layer by layer in the traditional network structure. In the equipment status monitoring network, taking the monitoring of evaporative cooling equipment as an example, it is assumed that the traditional network structure may cause information loss or gradient disappearance due to too many layers. The jump connection enables information to be directly transferred from the shallower layer (such as the output implicit representation of the e-th embedded network) and the output of the normalized component to the deeper layer (the output of the parallel weight focusing component), thereby ensuring the integrity of the information and the stability of the network.

[0111] The summation operation fuses the implicit representations from different sources. The output implicit representation of the e-th embedding network contains the feature information related to the device operating status obtained by the previous network processing, the output implicit representation of the normalization component of the first embedding branch network is the intermediate result after normalization processing, and the output implicit representation of the parallel weighted focusing component is the result after weighted focusing processing.

[0112] For example, suppose the output of the e-th embedding network is implicitly represented as a vector A=[a_1,a_2,…,a_n], the output of the normalized component of the first embedding branch network is implicitly represented as a vector B=[b_1,b_2,…,b_n], and the output of the parallel weight focusing component is implicitly represented as a vector C=[c_1,c_2,…,c_n]. The output of the first embedding branch network obtained by the summation operation is implicitly represented as a vector D=[a_1 + b_1+ c_1,a_2 + b_2+ c_2,…,a_n + b_n+ c_n]. The result of this summation integrates the information obtained from different stages of processing and can more comprehensively reflect the characteristics related to the operating status of the device.

[0113] In step S224, the remote control system passes the output implicit representation of the first embedding branch network obtained in step S223 as input to the second embedding branch network of the fth embedding network. This output implicit representation has been processed by the normalization component, parallel weight focusing component, and skip connection summation in the first embedding branch network, and contains rich feature information related to the device operation status.

[0114] For example, this output implicit representation may have more deeply mined and integrated the relationship between key parameters such as temperature and flow of the equipment. When it is loaded into the second embedding branch network, it provides basic data for the second embedding branch network to further extract and transform the characteristics of the equipment operation status.

[0115] In step S225, in the second embedding branch network, the normalization component also processes the input implicit representation (i.e., the output implicit representation of the first embedding branch network). Since the input data has been processed by the first embedding branch network, its data features may be different from the original data.

[0116] The multi-layer perceptron component performs nonlinear transformations on the input data through multiple neuron layers. For the normalized implicit representation, the input layer of the multi-layer perceptron receives this data, and then performs weighted summation and activation function (such as ReLU function y=max(0,x)) operations through the neurons in the hidden layer.

[0117] Taking evaporative cooling equipment as an example, assuming that the implicit representation of the input contains multiple feature relationships under the operating status of the equipment, the multi-layer perceptron component can learn more complex nonlinear relationships between these feature relationships through its internal neuron weights and activation functions. For example, if a feature in the implicit representation of the input is related to the cooling efficiency of the equipment, and another feature is related to the energy consumption of the equipment, the multi-layer perceptron component can mine the nonlinear relationship between the two features, thereby further improving the representation of the operating status of the equipment.

[0118] Step S226 is similar to step S223, and the skip connection plays a role here again. It allows the output implicit representation of the first embedding branch network and the output implicit representation of the normalization component of the second embedding branch network to be directly summed with the output implicit representation of the multilayer perceptron component.

[0119] This skip connection mechanism also ensures the integrity of information and the stability of the network in the second embedded branch network. For example, when processing equipment operating status monitoring, there may be some complex situations, such as the large changes in the operating status of the equipment under different working conditions. The skip connection can ensure that the information obtained at different processing stages can be effectively integrated to avoid information loss or over-processing.

[0120] The output implicit representation of the first embedding branch network contains the feature information related to the equipment operation status previously processed in the first embedding branch network, the output implicit representation of the normalization component of the second embedding branch network is the intermediate result after the normalization processing in the second embedding branch network, and the output implicit representation of the multi-layer perceptron component is the result after the nonlinear transformation of the multi-layer perceptron. Assume that the output of the first embedding branch network is implicitly represented as a vector E=[e_1,e_2,…,e_n], the output of the normalization component of the second embedding branch network is implicitly represented as a vector F=[f_1,f_2,…,f_n], and the output of the multi-layer perceptron component is implicitly represented as a vector G=[g_1,g_2,…,g_n]. The output of the fth embedding network obtained by the summation operation is implicitly represented as a vector H=[e_1 + f_1+ g_1,e_2 + f_2+ g_2,…,e_n + f_n+ g_n]. This output implicit representation integrates the information processed by the two embedded branch networks in the f-th embedded network, and represents the characteristics of the equipment operation status more comprehensively and deeply, providing a richer information basis for subsequent network processing or the final judgment of the equipment operation status.

[0121] Through the operations of steps S221-S226, the remote control system effectively further processes and extracts features of the output implicit representation of the e-th embedding network in the f-th embedding network, and gradually constructs an implicit representation that can more accurately reflect the operating status of the device.

[0122] As an implementation manner, step S223, summing the output implicit representation of the e-th embedding network and the output implicit representation of the normalized component of the first embedding branch network based on the skip connection and the output implicit representation of the parallel weighted focusing component of the first embedding branch network to obtain the output implicit representation of the first embedding branch network, may include:

[0123] Step S2231: determining a first multiplication result of an output implicit representation of the e-th embedding network and a first important coefficient; and determining a second multiplication result of an output implicit representation of a normalized component of the first embedding branch network and a second important coefficient;

[0124] Step S2232: summing the first multiplication result and the second multiplication result based on the output implicit representation of the parallel weight focusing component of the skip connection and the first embedding branch network.

[0125] Step S226, summing the output implicit representation of the first embedding branch network and the output implicit representation of the normalized component of the second embedding branch network based on the skip connection and the output implicit representation of the multilayer perceptron component of the second embedding branch network to obtain the output implicit representation of the f-th embedding network, may include:

[0126] Step S2261: determining a third multiplication result of the output implicit representation of the first embedding branch network and the third important coefficient; and determining a fourth multiplication result of the output implicit representation of the normalized component of the second embedding branch network and the fourth important coefficient;

[0127] Step S2262: summing the third multiplication result and the fourth multiplication result based on the skip connection and the output implicit representation of the multilayer perceptron component of the second embedded branch network.

[0128] In step S2231, in the equipment status monitoring network, the remote control system performs weighted processing on the implicit representations from different sources, which is the significance of determining the multiplication result. For the output implicit representation of the e-th embedded network, it is a feature representation related to the equipment operating status obtained after a series of previous network processing. For example, assuming that the output implicit representation of the e-th embedded network is a vector A=[a_1, a_2, …, a_n], and the first important coefficient is α, then the first multiplication result is the vector The first important coefficient α here is an adjustable parameter, and its value determines the importance of the implicit representation of the output of the e-th embedding network in the final summation result.

[0129] Similarly, for the implicit representation of the output of the normalized component of the first embedding branch network, assume that it is a vector B=[b_1, b_2, …, b_n], the second important coefficient is β, and the second multiplication result is the vector The second most important coefficient β also plays a similar role, which is used to adjust the weight of the normalized component output implicitly represented in the final result. These important coefficients can be adjusted according to the training of the network. For example, in the initialization stage of the network, both α and β can be set to smaller values, such as α=0.1, β=0.1. As the network training progresses, the values ​​of these coefficients are optimized according to algorithms such as error back propagation.

[0130] In the embodiment of the present application, this weighting operation is a common method for adjusting the importance of data in a neural network. In practical applications, the initial important coefficient values ​​can be determined according to the specific characteristics of the device and the monitoring requirements, and optimization algorithms such as stochastic gradient descent (SGD) or its variants Adagrad, Adadelta, etc. can be used to adjust these coefficients. For example, when updating the coefficients, the Adagrad algorithm takes into account the historical gradient information of each parameter, and its update formula is , where θ represents the parameter (here it can be α or β), η is the learning rate, G t is the sum of squared gradients of the parameters up to time t, is a very small number to prevent the denominator from being 0, g t is the gradient at the current time t.

[0131] In step S2232, the remote control system performs this step to fuse the implicit representations of different sources after weighted processing with the output implicit representation of the parallel weighted focusing component. Assume that the output implicit representation of the parallel weighted focusing component of the first embedding branch network is a vector C=[c_1, c_2, …, c_n]. The first multiplication result αA and the second multiplication result βB obtained in the previous step are summed with C based on the jump connection.

[0132] Taking the evaporative cooling equipment as an example, assume that the output implicit representation A of the e-th embedding network contains the features related to the temperature trend of different parts of the equipment, the output implicit representation B of the normalization component contains the features related to the internal pressure of the equipment after normalization, and the output implicit representation C of the parallel weighted focusing component contains the comprehensive features after re-weighting different features according to the operating status of the equipment. Through this summation operation, a new implicit representation that integrates temperature trends, pressure-related information, and the processing results of the parallel weighted focusing component can be obtained.

[0133] The specific summation operation is to add the corresponding elements, and the resulting vector is This summation operation is performed under the mechanism of jump connection, which allows the results of different stages to be directly fused, avoiding the limitations of information transmission between layers in traditional network structures. This helps to retain more original information and improve the accuracy of the network's representation of the device's operating status characteristics. During the network training process, the result of this summation operation will be used as input for subsequent operations, and based on the difference between the final device operating status inference result and the actual status, the back propagation algorithm is used to adjust various parameters in the network, including the first important coefficient α and the second important coefficient β.

[0134] In step S2261, the remote control system also performs a weighted operation on the implicit representations from different sources. For the output implicit representation of the first embedding branch network, assuming that it is a vector D=[d_1, d_2, …, d_n], and the third important coefficient is γ, then the third multiplication result is the vector The third important coefficient γ here is used to adjust the importance of the output implicit representation of the first embedding branch network in the final summation result.

[0135] For the implicit representation of the output of the normalized component of the second embedding branch network, assume that it is a vector E=[e_1, e_2, …, e_n], the fourth important coefficient is δ, and the fourth multiplication result is a vector The fourth important coefficient δ is used to adjust the weight of the normalized component output of the second embedding branch network implicitly represented in the final result. The determination and adjustment of these important coefficients are similar to those in the previous steps. They can be initialized according to the operating characteristics of the equipment, the characteristics of the monitoring data, and the goals of network training, and adjusted using an optimization algorithm during the training process.

[0136] For example, when monitoring evaporative cooling equipment, if the output of the first embedded branch network implicitly indicates that it focuses more on the short-term operating state characteristics of the equipment, and the output of the normalized component of the second embedded branch network implicitly indicates that it focuses more on the long-term operating state characteristics of the equipment, then the values ​​of γ and δ can be adjusted according to the needs of judging the operating state of the equipment. If you pay more attention to the impact of the short-term operating state on the final result, you can set γ to be relatively large; if the long-term operating state characteristics are also important, you need to set the value of δ reasonably.

[0137] In step S2262, the remote control system fuses the weighted implicit representations of different sources with the output implicit representation of the multilayer perceptron component of the second embedding branch network. Assume that the output implicit representation of the multilayer perceptron component of the second embedding branch network is a vector F=[f_1, f_2, …, f_n].

[0138] Taking evaporative cooling equipment as an example, assume that the output implicit representation D of the first embedded branch network contains the characteristics of the relationship between the temperature, flow and other parameters of the equipment at the current moment, the output implicit representation E of the normalization component of the second embedded branch network contains the normalized characteristics related to the energy consumption of the equipment, and the output implicit representation F of the multi-layer perceptron component contains the comprehensive characteristics after nonlinear transformation of different characteristics according to the operating status of the equipment. Through this summation operation, a new implicit representation that integrates multiple characteristics related to the operating status of the equipment can be obtained.

[0139] The specific summation operation is to add the corresponding elements, and the resulting vector is . This skip connection-based summation operation is similar to the principle in the previous step. It can effectively fuse information from different sources, retain more original information and improve the accuracy of the network's representation of the equipment's operating status characteristics. During the network training process, this summation result will be used as the output implicit representation of the f-th embedded network, and according to the difference with the actual operating status of the equipment, various parameters in the network, including the third important coefficient γ and the fourth important coefficient δ, are adjusted through algorithms such as back propagation, thereby optimizing the performance of the entire equipment status monitoring network, enabling it to more accurately monitor and analyze the operating status of the evaporative cooling equipment.

[0140] As an implementation method, the method may further include:

[0141] Step S227: Based on the difference between the a priori cooling device operating state of the device monitoring parameter sequence sample and the sample cooling device operating state, one or more of the first important coefficient, the second important coefficient, the third important coefficient and the fourth important coefficient are adjusted.

[0142] Step S227 is to adjust one or more of the first important coefficient, the second important coefficient, the third important coefficient and the fourth important coefficient based on the difference between the prior cooling device operating state of the equipment monitoring parameter sequence sample and the sample cooling device operating state.

[0143] The equipment monitoring parameter sequence samples are a data set used to train the equipment status monitoring network, which contains various parameters that can reflect the operating status of the evaporative cooling equipment, such as temperature, flow, pressure, etc. The prior cooling equipment operating status is the actual operating status of the equipment that is known and determined by other reliable methods (such as manual precise detection, high-precision instrument measurement, or accurate annotation in historical data, etc.). The sample cooling equipment operating status is the operating status obtained by processing and inferring the equipment monitoring parameter sequence samples based on the current network parameters (including various important coefficients, etc.) through the equipment status monitoring network.

[0144] When there is a difference between these two operating states, it means that the performance of the current equipment status monitoring network needs to be optimized, and the important coefficients play a key role in weight adjustment in the implicit representation extraction and fusion process of the network, so they need to be adjusted.

[0145] For example, for evaporative cooling equipment, it is assumed that the water inlet temperature, air flow and other parameter values ​​in the equipment monitoring parameter sequence sample indicate that the equipment should be in an efficient operating state (this is the prior cooling equipment operating state), but the sample cooling equipment operating state obtained by the equipment status monitoring network is an inefficient operating state. This difference indicates that the network has a deviation in processing these parameter information.

[0146] In the previous steps (such as S223 and S226 related sub-steps), the first important coefficient, the second important coefficient, the third important coefficient and the fourth important coefficient respectively perform weighted processing on the implicit representations from different sources. Assume that in S223, the first important coefficient is used to adjust the weight of the output implicit representation of the e-th embedding network in the summation result, and the second important coefficient is used to adjust the weight of the output implicit representation of the normalized component of the first embedding branch network; in S226, the third important coefficient and the fourth important coefficient also perform similar weighting on different implicit representations.

[0147] If the remote control system finds that the inferred operating state is different from the prior operating state, these important coefficients need to be adjusted. One feasible technical means is to use a gradient-based optimization algorithm, such as the stochastic gradient descent (SGD) algorithm. Taking the adjustment of the first important coefficient as an example, assuming that the loss function L represents the difference between the prior cooling device operating state and the sample cooling device operating state, according to the SGD algorithm, the update formula of α is , where α new is the updated α value, α old is the current α value, η is the learning rate, and represents the step size of each update.

[0148] If it is found that the sample cooling device operating state does not make sufficient use of the implicit representation of the output of the e-th embedding network (perhaps because the current α value is small), resulting in a difference from the prior operating state, then The value of is calculated according to the back propagation of the network, and the value of α is increased by the above formula, thereby increasing the influence of the output implicit representation of the e-th embedding network in the final result.

[0149] Similarly, other important coefficients can be adjusted in a similar way.

[0150] This adjustment of important coefficients based on the difference between the prior state and the sample state is a key step in optimizing the performance of the equipment condition monitoring network. By continuously adjusting these coefficients, the equipment condition monitoring network can more accurately process the equipment monitoring parameter sequence samples, thereby improving the accuracy of monitoring and reasoning of the operating state of the evaporative cooling equipment, and laying a solid foundation for the subsequent generation of effective equipment control strategies.

[0151] In actual operation, it may not be necessary to adjust only one important coefficient at a time, but rather to adjust multiple important coefficients at the same time according to the specific differences and characteristics of the network structure. For example, if it is found that the overall deviation of the sample cooling device operating state is large, there may be a problem with the fusion of multiple implicit representations. At this time, it may be necessary to adjust other coefficients at the same time, and it may require multiple iterations of the adjustment process. After each adjustment, the difference between the prior cooling device operating state and the sample cooling device operating state is re-evaluated until the difference is reduced to an acceptable range, indicating that the equipment status monitoring network can monitor and infer the equipment operating state more accurately.

[0152] As an implementation method, the method may further include:

[0153] Step S60: performing a normalization operation on the state mapping implicit representations output by the y state mapping branch networks to obtain a normalized state mapping implicit representation.

[0154] Step S60 is to perform a normalization operation on the state mapping implicit representation output by the y state mapping branch networks to obtain a normalized state mapping implicit representation.

[0155] The implicit representation of the state mapping is the result obtained after processing by y state mapping branch networks. It contains rich information related to the operating status of the equipment, but this information may differ in terms of numerical range, data distribution, etc. The normalization operation helps to adjust this data to a form that is more suitable for subsequent processing or analysis.

[0156] For example, in the monitoring of evaporative cooling equipment, the implicit representation of the state mapping may contain information of multiple dimensions, such as features related to the cooling efficiency of the equipment, features of pressure changes inside the equipment, and features related to energy consumption. These features may have different magnitudes and data distributions. Assume that the eigenvalues ​​related to cooling efficiency range from [0.1, 0.9], while the eigenvalues ​​related to energy consumption range from [100, 500] (this is just a simple example, the actual situation may be more complicated). This difference may affect the subsequent judgment of the equipment operation status or other analysis operations based on these implicit representations of the state mapping.

[0157] The remote control system can use a variety of normalization techniques to implement step S60. A common method is Min-Max Normalization. For a feature x in the implicit representation of the state mapping, let its minimum value be x min , the maximum value is x max , then the normalized eigenvalue x norm The formula can be Calculated.

[0158] Taking the evaporative cooling device mentioned above as an example, if there is a feature value x=300 in the implicit representation of the state mapping that represents the operating status indicator of a certain component of the device, after analysis, it is found that the minimum value x of this feature in all samples is min =100, maximum value x max =500, then according to the minimum-maximum normalization formula, x norm =0.5. Through this normalization operation, the eigenvalues ​​that originally had a large value range and were different in magnitude from other features were converted to the interval [0, 1], making different features more comparable in value.

[0159] Another possible normalization method is Z-score normalization (also called standardization). For a feature x, let its mean be μ and standard deviation be σ, then the normalized feature value x norm The formula can be For example, suppose that a feature related to the stability of equipment operation in the implicit representation of the state mapping has a mean μ=50 and a standard deviation σ=10. If the value of the feature at a certain moment is x=60, then This normalization method converts the eigenvalues ​​into a distribution with a mean of 0 and a standard deviation of 1, which helps to process data with different distribution characteristics.

[0160] The normalized implicit representation of the state mapping, i.e. the normalized implicit representation of the state mapping, has many advantages in the subsequent equipment operation status analysis, equipment control strategy generation and other links. For example, when diagnosing equipment faults based on machine learning algorithms, normalized data can improve the convergence speed and accuracy of the algorithm. If a neural network is used for fault diagnosis, the normalized input data (i.e. the normalized implicit representation of the state mapping) can make the training process of the neural network more stable and avoid problems such as gradient vanishing or gradient exploding due to large differences in data magnitude.

[0161] In addition, when conducting a comprehensive multi-indicator evaluation of the equipment's operating status, the normalized state mapping implicit representation can ensure that the weight distribution of each indicator in the comprehensive evaluation is more reasonable. For example, when considering multiple indicators such as cooling efficiency, energy consumption, and operating stability to judge the overall operating status of the equipment, since these indicators are in the same numerical range after normalization, a certain indicator will not occupy an unreasonable dominant position in the comprehensive evaluation due to its excessive value, thereby being able to more accurately reflect the actual operating status of the equipment.

[0162] As an implementation method, each state mapping network includes a first state mapping branch network, a second state mapping branch network, and a third state mapping branch network connected in sequence, the branch component of the first state mapping branch network is a first parallel weight focusing component, the branch component of the second state mapping branch network is a second parallel weight focusing component, and the branch component of the third embedded branch network is a multi-layer perceptron component.

[0163] Based on this, step S32, loading the output implicit representation of the t-th embedding network and the output implicit representation of the g-th state mapping network into the h-th state mapping network, performing implicit representation extraction one by one based on each state mapping branch network in the h-th state mapping network, and obtaining the output implicit representation of the h-th state mapping network, may include:

[0164] Step S321: Loading the output implicit representation of the g-th state mapping network into the first state mapping branch network of the h-th state mapping network;

[0165] Step S322: extracting implicit representations based on the normalization components and the first parallel weight focusing components of the first state mapping branch network one by one;

[0166] Step S323: summing the output implicit representation of the g-th state mapping network and the output implicit representation of the normalized component of the first state mapping branch network based on the skip connection and the output implicit representation of the first parallel weighted focusing component of the first state mapping branch network to obtain the output implicit representation of the first state mapping branch network;

[0167] Step S324: loading the output implicit representation of the t-th embedding network and the output implicit representation of the first state mapping branch network into the second state mapping branch network of the h-th state mapping network;

[0168] Step S325: extracting implicit representations based on the normalization components of the second state mapping branch network and the second parallel weight focusing components one by one;

[0169] Step S326: summing the output implicit representation of the first state mapping branch network and the output implicit representation of the normalized component of the second state mapping branch network based on the skip connection and the output implicit representation of the second parallel weighted focusing component of the second state mapping branch network to obtain the output implicit representation of the second state mapping branch network;

[0170] Step S327: Loading the output implicit representation of the second state mapping branch network into the third state mapping branch network of the hth state mapping network;

[0171] Step S328: extracting implicit representations based on normalized components and multi-layer perceptron components of the third state mapping branch network one by one;

[0172] Step S329: The output implicit representation of the second state mapping branch network and the output implicit representation of the normalized component of the third state mapping branch network are summed based on the jump connection and the output implicit representation of the multilayer perceptron component of the third state mapping branch network to obtain the output implicit representation of the hth state mapping network.

[0173] In step S321, the output implicit representation obtained after processing of the g-th state mapping network is passed to the first state mapping branch network in the h-th state mapping network. The output implicit representation of the g-th state mapping network is the result of complex processing of the equipment monitoring parameter sequence by the previous network structure (including multiple state mapping branch networks and possible embedded networks, etc.), which contains abstract feature information related to the equipment operation status.

[0174] For example, for evaporative cooling equipment, it is assumed that the output implicit representation of the g-th state mapping network may contain the characteristics of the relationship between the cooling efficiency and energy consumption of the equipment under different working conditions. This relationship is obtained after processing the implicit representation of a large number of equipment monitoring parameters (such as temperature, flow, pressure and other parameters). When this output implicit representation is loaded into the first state mapping branch network of the h-th state mapping network, it becomes the input data for further processing of this branch network, laying the foundation for subsequent mining of deeper equipment operation status characteristics.

[0175] In the first state mapping branch network, the normalization component first processes the implicit representation of the input. The purpose of the normalization component is to adjust the input data to a range that is more suitable for subsequent component processing and improve the stability and consistency of the data.

[0176] The first parallel weighted focusing component is a structure that can weight the input data to focus on the importance of different features. It can dynamically assign weights based on the relationship between different elements in the input implicit representation.

[0177] Taking evaporative cooling equipment as an example, assume that the implicit representation of the input contains information such as the temperature characteristics of different parts of the equipment, flow characteristics, and the relationship between them and cooling efficiency. The first parallel weighted focusing component can assign different weights to different features based on the correlation between these features under the operating state of the equipment. For example, if under a certain operating state, the temperature of a key part of the equipment is more important for the judgment of the overall operating state, then the component will assign a higher weight to this temperature feature. It calculates the correlation between the elements in the implicit representation of the input to determine the weight of each element, and then performs weighted summation and other operations on the input to achieve further extraction of the implicit representation. This method can effectively mine the hidden feature relationships in the implicit representation of the input, providing the possibility of more accurate representation of the operating state of the equipment.

[0178] In step S323, the skip connection allows the output implicit representation of the g-th state mapping network and the output implicit representation of the normalized component of the first state mapping branch network to be directly summed with the output implicit representation of the first parallel weight focusing component. This skip connection mechanism breaks the limitation that information can only be transmitted layer by layer in the traditional network structure.

[0179] In the equipment status monitoring network, taking evaporative cooling equipment as an example, it is assumed that the traditional network structure may cause information loss or gradient disappearance due to too many layers. However, the jump connection enables information to be directly transferred from the shallower layer (such as the output implicit representation of the g-th state mapping network and the output of the normalization component) to the deeper layer (the output of the first parallel weight focusing component), thereby ensuring the integrity of the information and the stability of the network.

[0180] The summation operation fuses the implicit representations from different sources. The output implicit representation of the g-th state mapping network contains the feature information related to the device operation status obtained by the previous network processing, the output implicit representation of the normalization component of the first state mapping branch network is the intermediate result after normalization processing, and the output implicit representation of the first parallel weight focusing component is the result after weighted focusing processing. For example, assume that the output implicit representation of the g-th state mapping network is a vector A=[a_1,a_2,…,a_n], the output implicit representation of the normalization component of the first state mapping branch network is a vector B=[b_1,b_2,…,b_n], and the output implicit representation of the first parallel weight focusing component is a vector C=[c_1,c_2,…,c_n]. The output implicit representation of the first state mapping branch network obtained by the summation operation is a vector D=[a_1 + b_1+ c_1,a_2 + b_2+ c_2,…,a_n + b_n+ c_n]. The summed result integrates the information obtained from different stages and can more comprehensively reflect the characteristics related to the equipment operation status.

[0181] In step S324, the remote control system passes the output implicit representation of the t-th embedded network and the output implicit representation of the first state mapping branch network obtained in step S323 as input to the second state mapping branch network of the h-th state mapping network. The output implicit representation of the t-th embedded network is the feature abstraction result of the equipment monitoring parameter sequence sample obtained after processing by the previous embedded network, while the output implicit representation of the first state mapping branch network is the result after processing by the first branch network in this state mapping network. For example, the output implicit representation of the t-th embedded network may contain more basic relationship features between the original monitoring parameters of the equipment, such as the original relationship features between temperature and flow. When loaded into the second state mapping branch network together with the output implicit representation of the first state mapping branch network (which already contains a deeper feature mining result of the equipment operation status), it provides a richer data source for this network to further mine the equipment operation status features.

[0182] In step S325, in the second state mapping branch network, the normalization component processes the input implicit representation (including the output implicit representation of the tth embedding network and the output implicit representation of the first state mapping branch network). Since this input data is a combination of multiple sources, its data features may be more complex and diverse.

[0183] The second parallel weighted focusing component also performs weighted processing on the input data to mine feature relationships. Similar to the first parallel weighted focusing component, it dynamically assigns weights based on the element relationships in the input implicit representation.

[0184] Taking evaporative cooling equipment as an example, assume that the implicit representation of the input now contains more information about the equipment's operating stability, energy consumption, and the coordinated relationship between the temperatures of different parts. The second parallel weighted focusing component can reallocate weights to these features based on the current operating status requirements of the equipment. For example, if the equipment is currently more concerned with the relationship between energy consumption and operating stability, then this component will assign higher weights to features related to energy consumption and operating stability, and further extract implicit representations through operations such as weighted summation, thereby more accurately mining features related to the equipment's operating status.

[0185] In step S326, the skip connection allows the output implicit representation of the first state mapping branch network and the output implicit representation of the normalization component of the second state mapping branch network to be directly summed with the output implicit representation of the second parallel weight focusing component.

[0186] This skip connection mechanism also ensures the integrity of information and the stability of the network in the second state mapping branch network. For example, when processing equipment operating status monitoring, there may be some complex situations, such as the large changes in the operating status of the equipment under different working conditions. The skip connection can ensure that the information obtained at different processing stages can be effectively integrated to avoid information loss or over-processing.

[0187] The output implicit representation of the first state mapping branch network includes characteristic information related to the equipment operating state previously processed in the first state mapping branch network. The output implicit representation of the normalization component of the second state mapping branch network is the intermediate result after normalization processing in the second state mapping branch network. The output implicit representation of the second parallel weight focusing component is the result after weighted processing by the second parallel weight focusing component.

[0188] Assume that the output of the first state mapping branch network is implicitly represented as a vector E=[e_1,e_2,…,e_n], the output of the normalized component of the second state mapping branch network is implicitly represented as a vector F=[f_1,f_2,…,f_n], and the output of the second parallel weight focusing component is implicitly represented as a vector G=[g_1,g_2,…,g_n]. The output of the second state mapping branch network obtained by the summation operation is implicitly represented as a vector H=[e_1 + f_1+ g_1,e_2 + f_2+ g_2,…,e_n + f_n+ g_n]. This output implicit representation integrates the information obtained at different processing stages in the second state mapping branch network, and has a more comprehensive and in-depth representation of the characteristics of the device operation status, providing a richer information basis for subsequent network processing or the final device operation status judgment.

[0189] In step S327, the remote control system passes the output implicit representation of the second state mapping branch network obtained in step S326 to the third state mapping branch network of the hth state mapping network. This output implicit representation has been processed by the previous two state mapping branch networks and contains rich feature information related to the device operation status.

[0190] For example, this output implicit representation may have deeply mined and integrated the cooling efficiency, energy consumption, operation stability, and complex relationships between temperature and flow in different parts of the equipment. When it is loaded into the third state mapping branch network, it provides basic data for the third state mapping branch network to further extract and transform the characteristics of the equipment operation status.

[0191] In step S328, in the third state mapping branch network, the normalization component processes the implicit representation of the input (i.e., the output implicit representation of the second state mapping branch network). Since this input data has been through the previous complex processing, its data features may need to be further adjusted to meet the processing requirements of the multi-layer perceptron component. Assuming that a variant of maximum-minimum normalization is adopted, for a certain feature value x in the input implicit representation, if the minimum value of the vector in which it is located is x min , the maximum value is x max , and the value range of x is [x min ,x max ], the normalized formula is ( is a small positive number to prevent the denominator from being zero). For example, if x=0.5, x min =0,x max =1, ,but This normalization operation adjusts the input data to a range suitable for processing by the multi-layer perceptron component.

[0192] The multi-layer perceptron component is a classic neural network structure that performs nonlinear transformations on input data through multiple neuron layers. For the normalized implicit representation, the input layer of the multi-layer perceptron receives this data, and then performs weighted summation and activation function (such as ReLU function y=max(0,x)) operations through the neurons in the hidden layer.

[0193] Taking evaporative cooling equipment as an example, assuming that the implicit representation of the input contains multiple feature relationships under the operating status of the equipment, the multi-layer perceptron component can learn more complex nonlinear relationships between these feature relationships through its internal neuron weights and activation functions. For example, if a feature in the implicit representation of the input is related to the cooling efficiency of the equipment, and another feature is related to the energy consumption of the equipment, the multi-layer perceptron component can mine the nonlinear relationship between the two features, thereby further improving the representation of the operating status of the equipment.

[0194] In step S329, the jump connection allows the output implicit representation of the second state mapping branch network and the output implicit representation of the normalization component of the third state mapping branch network to be directly summed with the output implicit representation of the multi-layer perceptron component. This jump connection mechanism also ensures the integrity of information and the stability of the network in the third state mapping branch network. For example, in the process of monitoring the operating status of the equipment, there may be some special cases, such as the change of the operating status of the equipment when it is suddenly disturbed by external interference. The jump connection can ensure that the information obtained at different processing stages can be effectively integrated to avoid information loss or over-processing. The output implicit representation of the second state mapping branch network contains the feature information related to the operating status of the equipment previously processed in the second state mapping branch network. The output implicit representation of the normalization component of the third state mapping branch network is the intermediate result after the normalization processing in the third state mapping branch network. The output implicit representation of the multi-layer perceptron component is the result after the nonlinear transformation of the multi-layer perceptron.

[0195] Assume that the output of the second state mapping branch network is implicitly represented as a vector I=[i_1,i_2,…,i_n], the output of the normalized component of the third state mapping branch network is implicitly represented as a vector J=[j_1,j_2,…,j_n], and the output of the multilayer perceptron component is implicitly represented as a vector K=[k_1,k_2,…,k_n]. The output of the h-th state mapping network obtained by the summation operation is implicitly represented as a vector L=[i_1 + j_1+ k_1,i_2 + j_2+ k_2,…,i_n + j_n+ k_n]. This output implicitly represents the information processed by the three state mapping branch networks in the h-th state mapping network, and the characteristic representation of the equipment operation status is more comprehensive and in-depth, providing a richer basis for subsequent equipment operation status judgment, control strategy generation, etc.

[0196] Through the operations of these sub-steps S321-S329, the remote control system effectively further processes and extracts features of the output implicit representation of the t-th embedding network and the output implicit representation of the g-th state mapping network in the h-th state mapping network, and gradually constructs an implicit representation that can more accurately reflect the operating status of the equipment.

[0197] As an implementation method, the method may further include:

[0198] Step SA: if the debugging round of the device state monitoring network is less than or equal to the first set round, limiting the number of network parameter variables optimized in each debugging round of the device state monitoring network to be less than or equal to a preset number;

[0199] Step SB: If the debugging round of the equipment status monitoring network is greater than or equal to the second set round, the preset number is no longer limited, and the second set round is greater than the first set round.

[0200] During the training process of the equipment status monitoring network, the debugging round refers to the number of times the network parameters are optimized and adjusted. Each round of debugging will adjust the network parameters based on the equipment monitoring parameter sequence samples and the difference between the sample cooling equipment operating status output by the current network and the prior cooling equipment operating status, so that the network can monitor the equipment operating status more accurately.

[0201] For example, in the early debugging rounds, the network parameters may include various weights, biases, and other parameters embedded in the branch network and the state mapping branch network. The initial values ​​of these parameters may be randomly set, and as the debugging rounds increase, these parameters are gradually adjusted to reduce the error between the network output and the actual device operation state.

[0202] When the debugging round is less than or equal to the first set round, the network is in the early stage of training. At this time, it is an effective strategy to limit the number of network parameters optimized in each round of debugging equipment status monitoring network to be less than or equal to the preset number. This is because in the early stage of network training, the relationship between network parameters has not been fully explored. If too many parameters are adjusted at the same time, the network may become unstable and difficult to converge to a better state.

[0203] Taking a simple linear regression model as an example (although the equipment condition monitoring network is more complex, the principle is similar), suppose there is a temperature prediction model for evaporative cooling equipment , where θ1 and θ2 are weight parameters, b is a bias term, x1 and x2 are relevant monitoring parameters of the equipment (such as water inlet temperature and air flow), and y is the predicted temperature of a certain part of the equipment. In the early stage of training, if θ1, θ2, and b are adjusted significantly at the same time, the model may fluctuate significantly in the parameter space and fail to effectively learn the patterns in the data.

[0204] The determination of the preset number needs to be comprehensively considered based on the structural complexity of the equipment status monitoring network, the amount of data, and other factors. For networks with relatively simple structures and small amounts of data, the preset number can be relatively small; while for complex networks and large amounts of data, the preset number can be appropriately increased.

[0205] The remote control system can use some technical means to limit the number of network parameter optimizations. For example, an update flag can be set for each parameter. At the beginning of each round of debugging, a counter is initialized. When a parameter is ready to be updated, first check whether the value of the counter is less than the preset number. If so, the update flag of the parameter is set to updateable and the counter value is increased; if the counter value has reached the preset number, the flag of the parameter is not updated, thereby achieving the purpose of limiting the number of network parameter optimizations.

[0206] In the equipment status monitoring network, assume that the first set round is 10 rounds. If the current is the 5th round of debugging, and the preset number is 5 parameters. During this round of debugging, the remote control system will select up to 5 parameters for optimization from all optimizable network parameters (such as certain parameters of the normalized component embedded in the branch network, weights in the branch component, etc.). For example, a scaling factor of a normalized component embedded in a branch network, weights in two branch components, and bias items in two state mapping branch networks may be selected as the parameters to be optimized in this round. In this way, in the early stage of network training, the network parameters are adjusted step by step and stably to avoid excessive adjustment that leads to degradation of network performance.

[0207] When the debugging round of the equipment status monitoring network is greater than or equal to the second set round, the network has been trained for a certain number of rounds. At this time, the relationship between the network parameters has been mined to a certain extent, and the network is gradually approaching a convergence state.

[0208] For example, after the previous debugging rounds, the equipment status monitoring network has learned the rules in the equipment monitoring parameter sequence samples. In the embedding branch network, the ability of each embedding branch network to extract the relevant features of the equipment operation status has been improved to a certain extent; in the state mapping branch network, the implicit representation can also be accurately mapped to the equipment operation status space.

[0209] The reason why the number of presets is no longer limited at this stage is that the network is relatively stable and needs to be adjusted more finely to further improve the performance of the network. At this point, the network can withstand more simultaneous adjustments of parameters without serious instability.

[0210] Take the training curve of a neural network as an example. In the early stage of training, the error decreases quickly but fluctuates greatly. As the number of debugging rounds increases, the error gradually decreases steadily. When the second set round is reached, the network is already in a stage where the error decreases slowly but is relatively stable. In order to further reduce the error, more parameters need to be adjusted to tap the potential performance of the network.

[0211] Assume that the second set round is 50 rounds. After the debugging round reaches 50 rounds, the remote control system is no longer limited to the preset number. For example, in the 55th round of debugging, the remote control system can adjust multiple network parameters at the same time based on the performance of the current network and the difference between the equipment monitoring parameter sequence samples and the operating status of the sample cooling equipment. It may be necessary to comprehensively optimize the normalized component parameters embedded in multiple branch networks, the weights of branch components, and various parameters in the state mapping branch network to further improve the network's monitoring accuracy of the equipment's operating status. This may involve adjusting a large number of weights, biases, normalization parameters, etc., so that the network can better adapt to the monitoring needs of the equipment's operating status through more comprehensive optimization.

[0212] Through the strategy of limiting and adjusting the optimization of network parameters in stages, such as steps SA-SB, the remote control system can more effectively train the equipment status monitoring network, ensure the stability of the network in the early stage of network training, and fully tap the performance potential of the network in the later stage, thereby improving the accuracy of the equipment status monitoring network in monitoring the operating status of the evaporative cooling equipment.

[0213] The present application embodiment provides a remote control system, such as Figure 2 As shown, the remote control system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the remote control system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the remote control system 100 does not constitute a limitation on the embodiments of the present application.

[0214] An embodiment of the present application provides a remote control system. The remote control system in the embodiment of the present application includes: one or more processors; a memory; one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above-described method is implemented.

Claims

1. A centralized control method for evaporative cooling equipment, characterized in that: The method is applied to a remote control system, wherein the remote control system is deployed with a device state monitoring network, wherein the device state monitoring network includes x embedded branch networks connected in sequence and y state mapping branch networks connected in sequence, wherein each of the embedded branch networks and each of the state mapping branch networks includes a normalized component and a branch component connected in sequence, x>1, y>2, and the method includes: Obtaining a device monitoring parameter sequence of a target evaporative cooling device; The device monitoring parameter sequence is loaded into the x embedded branch networks, and implicit representation extraction is performed one by one based on each of the x embedded branch networks to obtain embedded implicit representations output by the x embedded branch networks; each of the embedded branch networks is used to perform implicit representation extraction based on the normalization component and the branch component one by one, and the implicit representation extraction results of the embedded branch network are processed based on the jump connections of the entry node and the exit node of the normalization component; The embedded implicit representation is loaded into the y state mapping branch networks, and implicit representation extraction is performed one by one based on each of the y state mapping branch networks to obtain the state mapping implicit representation output by the y state mapping branch networks; each of the state mapping branch networks is used to perform implicit representation extraction based on the normalization component and the branch component one by one, and the implicit representation extraction result of the state mapping branch network is processed based on the jump connection; Inferring the cooling device operating state of the device monitoring parameter sequence based on the state mapping implicit representation; A corresponding device control strategy is generated based on the operating state of the cooling device, so as to control the target evaporative cooling device based on the device control strategy.

2. The method according to claim 1, characterized in that The equipment status monitoring network is debugged by the following steps: Obtain equipment monitoring parameter sequence samples; The equipment monitoring parameter sequence samples are loaded into the x embedded branch networks, and implicit representation extraction is performed one by one based on each of the x embedded branch networks to obtain embedded implicit representations output by the x embedded branch networks; wherein each of the embedded branch networks is used to perform implicit representation extraction based on the normalization component and the branch component one by one, and based on the jump connection of the entry node and the exit node of the normalization component, the implicit representation extraction result of the embedded branch network is processed; The embedded implicit representation is loaded into the y state mapping branch networks, and implicit representation extraction is performed one by one based on each of the y state mapping branch networks to obtain the state mapping implicit representation output by the y state mapping branch networks; wherein each of the state mapping branch networks is used to perform implicit representation extraction based on the normalization component and the branch component one by one, and the implicit representation extraction result of the state mapping branch network is processed based on the jump connection; Inferring the sample cooling device operating state of the device monitoring parameter sequence sample based on the state mapping implicit representation; Based on the difference between the a priori cooling device operating state of the device monitoring parameter sequence sample and the sample cooling device operating state, the network parameters of the device state monitoring network are optimized.

3. The method according to claim 2, characterized in that The equipment status monitoring network includes t embedded networks connected in sequence, each of the embedded networks includes a plurality of the embedded branch networks connected in sequence in the x embedded branch networks, t>1; the equipment monitoring parameter sequence samples are loaded into the x embedded branch networks, implicit representation extraction is performed one by one based on each of the embedded branch networks in the x embedded branch networks, and embedded implicit representations output by the x embedded branch networks are obtained, including: Loading the equipment monitoring parameter sequence samples into a first embedding network, performing implicit representation extraction based on each of the embedding branch networks in the first embedding network one by one, and obtaining an output implicit representation of the first embedding network; Loading the output implicit representation of the e-th embedding network into the f-th embedding network, performing implicit representation extraction based on each of the embedding branch networks in the f-th embedding network one by one, and obtaining the output implicit representation of the f-th embedding network, f=e+1, and e≥1, f≤t; The equipment state monitoring network includes t state mapping networks connected in sequence, each of the state mapping networks includes two or more of the state mapping branch networks connected in sequence in the y state mapping branch networks; the embedded implicit representation is loaded into the y state mapping branch networks, and implicit representation extraction is performed one by one based on each of the y state mapping branch networks to obtain the state mapping implicit representation output by the y state mapping branch networks, including: Loading the output implicit representation of the t-th embedding network into the first state mapping network, performing implicit representation extraction based on each of the state mapping branch networks in the first state mapping network one by one, and obtaining the output implicit representation of the first state mapping network; The output implicit representation of the t-th embedding network and the output implicit representation of the g-th state mapping network are loaded into the h-th state mapping network, and implicit representation extraction is performed one by one based on each of the state mapping branch networks in the h-th state mapping network to obtain the output implicit representation of the h-th state mapping network, h=g+1, and g≥1, h≤t.

4. The method according to claim 3, characterized in that Each of the embedded networks comprises a first embedded branch network and a second embedded branch network connected in sequence, the branch components of the first embedded branch network are parallel weighted focusing components, and the branch components of the second embedded branch network are multi-layer perceptron components; The step of loading the output implicit representation of the e-th embedding network into the f-th embedding network, and performing implicit representation extraction one by one based on each of the embedding branch networks in the f-th embedding network to obtain the output implicit representation of the f-th embedding network comprises: Loading the output implicit representation of the e-th embedding network into the first embedding branch network of the f-th embedding network; Extracting implicit representations based on the normalization component and the parallel weight focusing component of the first embedding branch network one by one; Summing the output implicit representation of the e-th embedding network and the output implicit representation of the normalization component of the first embedding branch network based on the skip connection and the output implicit representation of the parallel weight focusing component of the first embedding branch network to obtain the output implicit representation of the first embedding branch network; Loading the output implicit representation of the first embedding branch network to the second embedding branch network of the f-th embedding network; Extracting implicit representations based on the normalization component and the multi-layer perceptron component of the second embedding branch network one by one; The output implicit representation of the first embedding branch network and the output implicit representation of the normalization component of the second embedding branch network are summed based on the jump connection and the output implicit representation of the multilayer perceptron component of the second embedding branch network to obtain the output implicit representation of the f-th embedding network.

5. The method according to claim 4, characterized in that The step of summing the output implicit representation of the e-th embedding network and the output implicit representation of the normalization component of the first embedding branch network based on the skip connection and the output implicit representation of the parallel weight focusing component of the first embedding branch network to obtain the output implicit representation of the first embedding branch network comprises: Determine a first multiplication result of an output implicit representation of the e-th embedding network and a first important coefficient; and determine a second multiplication result of an output implicit representation of the normalized component of the first embedding branch network and a second important coefficient; summing the first multiplication result and the second multiplication result based on the skip connection and the output implicit representation of the parallel weight focusing component of the first embedding branch network; The step of summing the output implicit representation of the first embedding branch network and the output implicit representation of the normalized component of the second embedding branch network based on the skip connection and the output implicit representation of the multilayer perceptron component of the second embedding branch network to obtain the output implicit representation of the f-th embedding network comprises: Determine a third multiplication result of the output implicit representation of the first embedding branch network and a third important coefficient; and determine a fourth multiplication result of the output implicit representation of the normalization component of the second embedding branch network and a fourth important coefficient; The third multiplication result and the fourth multiplication result are summed based on the skip connection and the output implicit representation of the multilayer perceptron component of the second embedding branch network.

6. The method according to claim 5, characterized in that The method further comprises: Based on the difference between the a priori cooling device operating state of the device monitoring parameter sequence sample and the sample cooling device operating state, one or more of the first importance coefficient, the second importance coefficient, the third importance coefficient and the fourth importance coefficient are adjusted.

7. The method according to any one of claims 2 to 6, characterized in that: The method further comprises: A normalization operation is performed on the state mapping implicit representations output by the y state mapping branch networks to obtain a normalized state mapping implicit representation.

8. The method according to any one of claims 3 to 6, characterized in that: Each of the state mapping networks comprises a first state mapping branch network, a second state mapping branch network and a third state mapping branch network connected in sequence, the branch component of the first state mapping branch network is a first parallel weight focusing component, the branch component of the second state mapping branch network is a second parallel weight focusing component, and the branch component of the third embedding branch network is a multi-layer perceptron component; The step of loading the output implicit representation of the t-th embedding network and the output implicit representation of the g-th state mapping network into the h-th state mapping network, and performing implicit representation extraction one by one based on each of the state mapping branch networks in the h-th state mapping network to obtain the output implicit representation of the h-th state mapping network comprises: Loading the output implicit representation of the g-th state mapping network into the first state mapping branch network of the h-th state mapping network; Extracting implicit representations based on the normalization component and the first parallel weight focusing component of the first state mapping branch network one by one; The output implicit representation of the g-th state mapping network and the output implicit representation of the normalization component of the first state mapping branch network are summed based on the skip connection and the output implicit representation of the first parallel weight focusing component of the first state mapping branch network to obtain the output implicit representation of the first state mapping branch network; Loading the output implicit representation of the t-th embedding network and the output implicit representation of the first state mapping branch network into the second state mapping branch network of the h-th state mapping network; Extracting implicit representations based on the normalization component and the second parallel weight focusing component of the second state mapping branch network one by one; The output implicit representation of the first state mapping branch network and the output implicit representation of the normalization component of the second state mapping branch network are summed based on the skip connection and the output implicit representation of the second parallel weight focusing component of the second state mapping branch network to obtain the output implicit representation of the second state mapping branch network; Loading the output implicit representation of the second state mapping branch network to the third state mapping branch network of the h-th state mapping network; Extracting implicit representations based on the normalization component and the multi-layer perceptron component of the third state mapping branch network one by one; The output implicit representation of the second state mapping branch network and the output implicit representation of the normalization component of the third state mapping branch network are summed based on the jump connection and the output implicit representation of the multilayer perceptron component of the third state mapping branch network to obtain the output implicit representation of the h-th state mapping network.

9. The method according to any one of claims 2 to 6, characterized in that: The method further comprises: If the debugging round of the device state monitoring network is less than or equal to the first set round, limiting the number of network parameter variables optimized by the device state monitoring network in each debugging round to be less than or equal to a preset number; If the debugging round of the device status monitoring network is greater than or equal to a second set round, the preset number is no longer limited, and the second set round is greater than the first set round.

10. A remote control system, characterized in that: include: one or more processors; Memory; One or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 9 is implemented.

Images