Control method and system of evaporative cooling device

Abstract

The present invention provides a control method and system for an evaporative cooling device. In the training phase of a neural network, for a monitoring time series data set associated with a target observation element, the core data items of the target observation element are denoised to generate monitoring time series data set samples, and monitoring data of the target evaporative cooling device in an abnormal state is constructed. Then, based on the monitoring time series data set samples, a state monitoring network is debugged, so that the debugged state monitoring network can identify the situation where the core data items corresponding to the monitoring element drift, so as to increase the monitoring reliability of the state monitoring network. Furthermore, because the debugged state monitoring network can identify the core data items corresponding to the monitoring element drift, the debugged state monitoring network can also timely identify abnormal situations that have not occurred, and help identify the abnormal operating state of the target evaporative cooling device, so as to increase the robustness of the state monitoring network.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a control method and system for an evaporative cooling device. Background Art

[0002] In modern industry and many specific environmental applications, evaporative cooling devices play a vital role. For example, in the computer room of a data center, evaporative cooling devices are responsible for maintaining appropriate temperature and humidity to ensure the normal operation of equipment such as servers; in some industrial production processes, evaporative cooling devices are used to cool production equipment or adjust the temperature and humidity of the production environment. However, there are many problems in the operation of existing evaporative cooling devices. In terms of status monitoring, traditional monitoring methods can often only perform simple threshold judgments on a single or a few fixed parameters. For example, only monitoring whether the temperature exceeds a certain set maximum temperature value, or whether the humidity is lower than a certain lower limit value. This simple monitoring method cannot fully and deeply understand the operating status of the evaporative cooling device. Because the operating status of the evaporative cooling device is affected by multiple factors, such as the flow rate of the coolant, the flow rate of the air, the operating efficiency of each component of the equipment, etc., it is difficult to accurately judge whether the device as a whole is in normal operation by only focusing on individual parameters. Secondly, the existing monitoring system lacks the ability to effectively analyze the dynamic evolution of data. For data that changes over time, such as the fluctuation trend of temperature within a day, the change pattern of the coolant flow rate at different operating stages of the equipment, etc., traditional methods are difficult to capture the device operating status information hidden behind these time-series information. This results in that, in some cases, even if a potential problem has occurred in the device, the monitoring system cannot identify the abnormality of the time series evolution of the data and cannot issue an alarm or take corresponding measures in time. Furthermore, when faced with complex operating environments and various possible abnormal situations, the robustness of existing monitoring and control technologies is poor. Summary of the invention

[0003] The object of the present invention is to provide a control method and system for an evaporative cooling device. The present invention is implemented as follows:

[0004] In a first aspect, the present invention provides a control method for an evaporative cooling device, the method comprising: obtaining one or more monitoring time series data sets of a target evaporative cooling device, the monitoring time series data sets comprising a plurality of monitoring data clusters associated with a target observation element arranged in time series; obtaining a state monitoring result of the monitoring time series data set based on a state monitoring network, the state monitoring result of the monitoring time series data set being used to indicate whether a time series evolution of a characterization vector associated with a core data item of the target observation element in the monitoring time series data set meets a standard, the state monitoring network being debugged based on a monitoring time series data set sample, for u monitoring data clusters in the monitoring time series data set sample, the core data item of the target observation element being noised, and the time series evolution of a characterization vector associated with the core data item of the target observation element in the monitoring time series data set sample meets a standard, u≥1; obtaining a state monitoring result of the target evaporative cooling device based on the state monitoring results corresponding to the one or more monitoring time series data sets, the state monitoring result being used to indicate whether the target evaporative cooling device meets the standard; obtaining a control strategy of the target evaporative cooling device based on the state monitoring result of the target evaporative cooling device.

[0005] In a second aspect, the present invention provides a computer 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.

[0006] The beneficial effects of the present invention include: in the present invention, in the training link of the neural network, for the monitoring time series data set associated with the target observation element, based on adding noise to the core data items of the target observation element to generate the monitoring time series data set samples, the monitoring data of the target evaporative cooling device in abnormal state is constructed, and then the state monitoring network is debugged based on the monitoring time series data set samples, so that the debugged state monitoring network can identify the situation where the core data items corresponding to the monitoring element drift, so as to increase the monitoring reliability of the state monitoring network. Furthermore, because the debugged state monitoring network can identify the drift of the core data items corresponding to the monitoring element, the debugged state monitoring network can also timely identify the abnormal situation that has not occurred, and help identify the abnormal operation state of the target evaporative cooling device, so as to increase the robustness of the state monitoring network. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is a flow chart of a control method of an evaporative cooling device provided by an embodiment of the present invention.

[0008] Figure 2It is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The control method of the evaporative cooling device in the embodiment of the present invention is executed by a computer system, including but not limited to a server, a personal computer, a notebook computer, a tablet computer, a smart phone, etc. Figure 1 As shown, the method includes:

[0010] Step S110: Acquire one or more monitoring time series data sets of the target evaporative cooling device, wherein the monitoring time series data sets include a plurality of monitoring data clusters associated with the target observation elements and arranged in time series.

[0011] In an embodiment of the present application, a monitoring time series data set is a set of data arranged in chronological order, which includes multiple monitoring data clusters associated with target observation elements. The target observation element refers to a specific physical quantity or performance indicator of concern when monitoring the evaporative cooling device, such as temperature, humidity, cooling efficiency, etc. A monitoring data cluster refers to a set of related data collected for the target observation element at a specific moment or time period. For example, for the target observation element of temperature, the temperature value collected at a certain moment, the working status of the sensor related to the temperature, environmental factors (such as the air flow rate in the machine room and other factors that may affect the temperature) and other data constitute a monitoring data cluster.

[0012] The computer system obtains these monitoring time series data sets, for example, through a sensor network. Various sensors, such as temperature sensors, humidity sensors, etc., are deployed in the evaporative cooling device and its surrounding environment. These sensors collect data at certain time intervals. For example, the temperature sensor collects temperature data in the computer room every 5 minutes, and the humidity sensor collects humidity data in the computer room every 10 minutes. Then, the data collected by these sensors is transmitted to the computer system. The transmission method can be wired transmission, such as transmitting data to the computer system via an Ethernet cable; or wireless transmission, such as using wireless communication protocols such as ZigBee and Wi-Fi to send data to the computer system. After receiving these data, the computer system organizes and stores them to form a monitoring time series data set. During the sorting process, the computer system arranges each monitoring data cluster according to the time sequence of data collection.

[0013] For a complex evaporative cooling device, there may be multiple target observation elements, each of which has its corresponding monitoring data cluster, so the computer system may obtain multiple monitoring time series data sets. For example, in addition to temperature and humidity, you may also pay attention to target observation elements such as the energy consumption of the evaporative cooling device and the flow rate of the coolant. Each target observation element will form its own monitoring data cluster, which will then constitute multiple monitoring time series data sets. In addition, in order to ensure the accuracy and completeness of the data, some data preprocessing operations may also be performed in the process of obtaining the monitoring time series data set. For example, the validity of the collected data is checked. If it is found that a certain data is obviously beyond the normal range (such as the temperature value collected by the temperature sensor is several times higher than the normal room temperature, which may be an abnormal value caused by a sensor failure), the computer system may mark the data or try to re-collect it. At the same time, in order to facilitate subsequent data processing and analysis, the computer system may convert the received data into a unified format, such as converting analog signals collected by different sensors into digital signals and storing them in a unified encoding method.

[0014] Step S120: Acquire the status monitoring result of the monitoring time series data set based on the status monitoring network. The status monitoring result of the monitoring time series data set is used to indicate whether the time series evolution of the representation vector associated with the core data item of the target observation element in the monitoring time series data set meets the standard. The status monitoring network is debugged based on the monitoring time series data set sample. For u monitoring data clusters in the monitoring time series data set sample, the core data item of the target observation element is noised, and the time series evolution of the representation vector associated with the core data item of the target observation element in the monitoring time series data set sample meets the standard, u≥1.

[0015] In the embodiment of the present application, the state monitoring network is a specially constructed and debugged network structure, which can analyze and process the input monitoring time series data set to obtain relevant results about the state of the data set. This network is debugged based on the monitoring time series data set sample, and its purpose is to enable the network to accurately identify various state characteristics related to the target observation element in the data set. The target observation element is a key factor for measuring the operating state of the evaporative cooling device, such as cooling efficiency, inlet and outlet temperature difference, coolant flow rate, etc. The core data item is a key component of the target observation element. Taking cooling efficiency as an example, if cooling efficiency is the target observation element, some key parameters directly related to cooling efficiency, such as the heat transfer coefficient of the evaporator, air flow rate, etc. may be core data items. These core data items are of key significance in reflecting the operating state of the evaporative cooling device. A representation vector is a vector form that abstractly represents data features. For the core data item of the target observation element, it can be converted into a representation vector through a specific mathematical transformation or feature extraction method. For example, for the core data item of coolant flow, a representation vector is constructed based on its flow value at different time points, flow change rate and other factors. This characterization vector can reflect the comprehensive characteristics of the core data item, the coolant flow rate, to a certain extent.

[0016] The technical means by which a computer system uses a condition monitoring network to obtain the condition monitoring results of a monitoring time series data set involves multiple complex processes. One possible technical means is to use a deep learning algorithm to build a condition monitoring network. For example, a convolutional neural network (CNN) or a recurrent neural network (RNN) structure can be used. Assume that a convolutional neural network is used to build a condition monitoring network. The computer system first needs to preprocess the monitoring time series data set to meet the input requirements of the network. For each monitoring data cluster in the monitoring time series data set, standardization may be required to map the data to a specific interval, such as normalizing the data to the [0,1] interval. The formula is: , where x is the original data, x' is the normalized data, min(x) and max(x) are the minimum and maximum values ​​in the original data, respectively.

[0017] Then, the preprocessed monitoring time series data set is input into the convolutional neural network. The convolution layer in the convolutional neural network performs convolution operations by sliding the convolution kernel on the data, which can automatically extract local features in the data. For example, for a monitoring time series data set that reflects the operating status of an evaporative cooling device, the convolution layer can extract some local feature patterns related to the core data items, such as the fluctuation pattern of the coolant flow in a short period of time. The pooling layer downsamples the features extracted by the convolution layer to reduce the dimension of the data while retaining the main feature information. For example, the maximum pooling layer selects the maximum value in the local area as the result of pooling, which helps to reduce the amount of data, improve computational efficiency, and highlight important features.

[0018] After being processed by multiple convolutional layers and pooling layers, the fully connected layer integrates the features extracted by the previous layers to obtain a preliminary result of the state level of the representation vector associated with the core data item of the target observation element. This result can be regarded as a comprehensive description of the state of the monitoring time series dataset, which reflects the overall state characteristics of the core data items in the dataset.

[0019] Next, based on this preliminary result and the structure and parameters of the network, the computer system further analyzes whether the time series evolution of the characterization vector associated with the core data item of the target observation element in the monitoring time series data set meets the standard. The "standard" here is pre-set, which is based on a large amount of data statistics and analysis of the normally operating evaporative cooling device. For example, for the time series evolution of the coolant flow rate, if it is in normal operation, the coolant flow rate should fluctuate within a relatively stable range, and the fluctuation amplitude does not exceed a certain threshold. The setting of this threshold can be based on the statistical analysis of historical data, such as calculating the mean and standard deviation of the coolant flow fluctuation amplitude in the historical data, and then setting a reasonable threshold range according to actual needs, such as the mean ± 3 times the standard deviation.

[0020] When a computer system analyzes a monitoring time series data set through a condition monitoring network, it is also necessary to consider the dynamic nature of the data. Since the monitoring time series data set is arranged in chronological order, the representation vectors of the core data items may have different performances at different time points. For example, as the operating time of the evaporative cooling device increases, the temperature of the coolant may gradually increase, thereby affecting the changes in core data items such as its flow rate. The condition monitoring network needs to be able to capture this trend over time and compare it with the preset standard.

[0021] In order to enable the condition monitoring network to accurately perform condition monitoring, when building and debugging the network, the core data items of the target observation elements are denoised for the u monitoring data clusters (u≥1) in the monitoring time series data set samples. Noising is a means of simulating data abnormalities. By adding a certain amount of noise to the core data items, the network can learn how to accurately judge the state when there is interference in the data. For example, for the core data item of coolant flow, if the flow value is 10L / min in the normal monitoring data cluster, then noise can be added according to a certain probability distribution (such as normal distribution) to make the flow value fluctuate randomly within a certain range, such as setting the fluctuation range to ±1L / min. In this way, the monitoring time series data set samples after noise processing contain more possible situations, which helps to improve the generalization ability of the condition monitoring network and enable it to better adapt to various data changes that may occur in actual operation.

[0022] Step S130: based on the status monitoring results corresponding to one or more monitoring time series data sets, the status monitoring results of the target evaporative cooling device are obtained, and the status monitoring results are used to indicate whether the target evaporative cooling device meets the standards.

[0023] In an embodiment of the present application, step S130 obtains the state monitoring result of the target evaporative cooling device based on the state monitoring results corresponding to one or more monitoring time series data sets, and this result is used to indicate whether the target evaporative cooling device meets the standard. These monitoring time series data sets are data sets related to the target evaporative cooling device, and each data set contains multiple monitoring data clusters associated with the target observation elements arranged in time series. The target observation elements are factors that have important indicative significance for the operating status of the evaporative cooling device, for example, may include cooling efficiency, equipment power consumption, air humidity regulation effect, etc. The state monitoring result reflects whether the time series evolution of the representation vector associated with the core data item of the target observation element in each monitoring time series data set in the data set meets the standard.

[0024] When the computer system performs this step, it needs to comprehensively consider the state monitoring results corresponding to each monitoring time series data set. For example, suppose there are three monitoring time series data sets, which are monitoring data on the three target observation elements of cooling efficiency, equipment power consumption and air humidity adjustment effect. For the monitoring time series data set of cooling efficiency, its state monitoring results may show that over a period of time, the time series evolution of the characterization vector of the core data items related to cooling efficiency (such as the temperature difference of the evaporator, the coolant flow rate, etc.) basically meets the standard, but there are some small fluctuations. For the monitoring time series data set of equipment power consumption, the state monitoring results show that the equipment power consumption has a peak value beyond the normal range in certain time periods, which may indicate that there are some abnormal conditions in the equipment. For the monitoring time series data set of air humidity adjustment effect, the state monitoring results show that the time series evolution of the characterization vector of its core data items (such as outlet humidity, ambient humidity change rate, etc.) fully meets the standard.

[0025] In order to obtain the overall status monitoring results of the target evaporative cooling device, the computer system needs to integrate and analyze the status monitoring results of these different monitoring time series data sets. A possible technical means is to use the weighted average method. Assume that the weights corresponding to the three monitoring time series data sets of cooling efficiency, equipment power consumption and air humidity regulation effect are w1, w2 and w3 respectively (these weights can be determined according to the importance of each observation factor to the overall performance of the evaporative cooling device. For example, if the cooling efficiency is crucial to the performance of the device, a higher weight can be given, such as w1=0.5; equipment power consumption is relatively important, w2=0.3; air humidity regulation effect is relatively minor, w3=0.2).

[0026] Assume that the state monitoring result corresponding to the monitoring time series data set of cooling efficiency is r1 (r1 can be a quantitative value, for example, if it fully meets the standard, it is 1, if there is a small fluctuation, it is 0.8, etc.), the state monitoring result corresponding to the monitoring time series data set of equipment power consumption is r2 (for example, it is 0.6 when there is an abnormal peak), and the state monitoring result corresponding to the monitoring time series data set of air humidity adjustment effect is r3 (assuming it is 1). Then the state monitoring result R of the target evaporative cooling device can be calculated by the following formula: R=w1r1+w2r2+w3r3=0.5×0.8+0.3×0.6+0.2×1=0.4+0.18+0.2=0.78.

[0027] The result 0.78 here represents the overall status monitoring result of the target evaporative cooling device. If a standard value is set, for example, 0.8 or more represents that the device meets the standard, then according to this calculation result, it can be judged that the device does not fully meet the standard under the current situation.

[0028] Another technical means is to establish a decision tree model. The computer system uses the status monitoring results of each monitoring time series data set as the input features of the decision tree. For example, whether the status monitoring result of the cooling efficiency is higher than a certain threshold is used as a branch node of the decision tree. If it is higher than the threshold, the status monitoring result of the equipment power consumption is further analyzed to see whether it is within another threshold range, and so on. Through the layer-by-layer judgment of the decision tree, the result of whether the target evaporative cooling device meets the standard is finally obtained.

[0029] For another example, fuzzy logic methods can also be used. The computer system regards the state monitoring results corresponding to each monitoring time series data set as elements in a fuzzy set and performs reasoning based on pre-defined fuzzy rules. For example, if the state monitoring result of cooling efficiency is in the "better" fuzzy set, the state monitoring result of equipment power consumption is in the "slightly worse" fuzzy set, and the state monitoring result of air humidity regulation effect is in the "good" fuzzy set, then the fuzzy rules (such as comprehensive consideration of the membership of each fuzzy set and the relationship between them) are used to determine whether the target evaporative cooling device as a whole meets the standards.

[0030] In practical applications, the computer system takes into account the correlation between different monitoring time series data sets. For example, changes in cooling efficiency may affect the power consumption of the equipment, and abnormal power consumption of the equipment may in turn affect the cooling efficiency. Therefore, when integrating these state monitoring results, each result cannot be simply viewed in isolation, but the relationship between them needs to be considered. For example, if the state monitoring result of cooling efficiency is found to be abnormal, and the power consumption of the equipment is also abnormal, this may indicate that there is a systemic problem in the evaporative cooling device, and a deeper analysis of the cause is required, such as a failure of a component of the cooling system (such as a compressor or a radiator), which affects the cooling efficiency and causes abnormal power consumption of the equipment. In addition, the computer system also considers the timeliness of the data. The state monitoring results of the newer monitoring time series data set may be more representative of the current state of the target evaporative cooling device. For example, if the most recent cooling efficiency monitoring time series data set shows that the time series evolution of the representation vector of the core data item begins to deviate from the standard, and the previous results are normal, then when obtaining the state monitoring results of the target evaporative cooling device, more attention needs to be paid to the newer data. A time-weighted method can be used, for example, a higher weight is given to the most recent monitoring time series data set, and the weight gradually decreases over time.

[0031] Step S140: based on the status monitoring result of the target evaporative cooling device, obtaining a control strategy of the target evaporative cooling device.

[0032] In the embodiment of the present application, the state monitoring result of the target evaporative cooling device is obtained in the aforementioned step (such as step S130). This result reflects whether the evaporative cooling device as a whole meets the standard. It is obtained by comprehensively considering multiple monitoring time series data sets related to the target observation elements and their corresponding state monitoring results. The target observation elements cover factors such as cooling efficiency, equipment power consumption, and environmental temperature and humidity adjustment capabilities that are key to the operating status of the evaporative cooling device.

[0033] When the computer system obtains the control strategy, it first determines the current operating status category of the evaporative cooling device based on the existing status monitoring results. For example, if the status monitoring results show that the overall operation of the device meets the standards and each target observation element is within the ideal range, this is a normal operating state. Taking cooling efficiency as an example, if the relevant data of cooling efficiency shows that the temperature difference between the inlet and outlet of the evaporator is stable near the design value and the coolant flow is normal, this indicates that the cooling subsystem is operating normally. For equipment power consumption, if the power consumption value is stable within the rated power consumption range, it means that the energy consumption of the equipment is at a normal level. For the ability to adjust the ambient temperature and humidity, if the deviation between the actual adjusted temperature and humidity values ​​and the set values ​​is within an acceptable range, it indicates that the temperature and humidity adjustment function is normal.

[0034] In this normal operating state, the computer system can adopt a control strategy to maintain the status quo. For example, for the operating parameters of the device, keep the current fan speed, coolant pump flow setting, etc. unchanged. From the perspective of technical implementation, the computer system can achieve this by sending instructions to the relevant devices to maintain the current operating parameters. Assuming that the control instructions of the device are transmitted in a specific encoding format, for example, the control instruction format for the fan speed is [device ID: fan, instruction type: maintain speed, speed value: current speed value], the computer system sends such instructions to the corresponding device to ensure that the device continues to operate in the current state.

[0035] However, if the condition monitoring results show that the evaporative cooling device is in an abnormal state, the computer system needs to further analyze the specific situation of the abnormality to determine the appropriate control strategy. For example, if the cooling efficiency is found to be reduced, it may be due to the evaporator scaling that reduces the heat exchange efficiency. The computer system can determine the approximate degree of scaling based on the pre-established relationship model between the cooling efficiency and the degree of evaporator scaling. Assume that this relationship model is based on an empirical formula established based on a large amount of experimental data and theoretical analysis, such as the cooling efficiency loss rate E loss There is a relationship between E and the evaporator scaling thickness d: loss =kd 2+bd+c (where k, b, and c are constants determined experimentally). Through the known cooling efficiency loss rate, the computer system can infer the scaling thickness and determine the measures that need to be taken. In response to this situation, the control strategy that the computer system can formulate may be to increase the cleaning cycle or adjust the chemical composition of the coolant to enhance the descaling ability. If it is decided to increase the cleaning cycle, the computer system can send a notification to the maintenance management system, and the notification content includes information such as the specific location of the evaporator and the recommended cleaning time interval. For example, the notification format is [Evaporator location: No. 3, Area A, Machine Room, recommended cleaning cycle: adjusted from once a month to once every half month].

[0036] If the status monitoring results show that the power consumption of the equipment is abnormally high, the computer system needs to analyze the possible reasons. It may be that the efficiency of some components in the equipment (such as the motor) has decreased or the load of the cooling system is unbalanced. Assuming that the motor efficiency has decreased, the computer system can evaluate the deviation between the current motor efficiency and the normal efficiency based on the power-efficiency characteristic curve of the motor (this is a relationship curve provided by the motor manufacturer or obtained through experimental testing, usually with power as the horizontal axis and efficiency as the vertical axis). If the deviation exceeds a certain threshold, the control strategy may be to adjust the power supply voltage of the motor or replace the motor. For the adjustment of the power supply voltage, the computer system can calculate the appropriate voltage value based on the voltage-power relationship of the motor (assuming a simple linear relationship P=kV, where P is power, V is voltage, and k is a proportional constant), and then send instructions to the power supply system, such as [Device ID: Motor, Instruction Type: Adjust Voltage, Voltage Value: Calculated New Voltage Value].

[0037] In the case of abnormalities in the ability to adjust the ambient temperature and humidity, such as humidity adjustment exceeding the set range. If it is due to a humidifier or dehumidifier failure, the computer system can determine the type of failure by monitoring the operating parameters of the humidifier or dehumidifier (such as the spray volume of the humidifier, the condensed water volume of the dehumidifier, etc.). If it is determined that the humidity is low due to insufficient humidifier spray volume, the control strategy can be to check whether the humidifier nozzle is blocked. If so, send a command to the maintenance robot to clean the nozzle. The command format is such as [Maintenance Task: Nozzle Cleaning, Equipment Location: Humidifier in Area B of the Computer Room].

[0038] In addition to formulating control strategies for abnormal conditions of a single target observation element, the computer system also needs to consider comprehensive strategies when multiple abnormal conditions occur at the same time. For example, when the cooling efficiency decreases and the power consumption of the equipment is high, there may be comprehensive problems in the entire cooling system, such as the cooling pipe is blocked and the motor load is too large. At this time, the computer system needs to formulate a comprehensive solution, which may include cleaning the cooling pipe first, then re-evaluating the motor load, adjusting the motor operating parameters or maintaining the motor according to the evaluation results. In addition, when formulating control strategies, the computer system also considers the overall stability and safety of the system. For example, when adjusting equipment parameters, the equipment cannot be operated beyond its safe operating range. For some key equipment, there may be upper and lower limit constraints on parameters, such as the temperature of the coolant cannot be lower than its freezing point or higher than its boiling point. The computer system must follow these constraints when formulating control strategies. At the same time, in order to optimize the long-term operating performance of the evaporative cooling device, predictive control can be performed based on condition monitoring results and historical data.

[0039] As an implementation manner, step S120, obtaining the status monitoring result of the monitoring time series data set based on the status monitoring network, may include:

[0040] Step S121: acquiring a first characterization vector set of a monitoring time series data set based on a state monitoring network, wherein the first characterization vector set is used to indicate evolution information of characterization vectors associated with core data items of target observation elements in the monitoring time series data set at a state level;

[0041] Step S122: obtaining a second set of characterization vectors based on the first set of characterization vectors based on the condition monitoring network, wherein the second set of characterization vectors is used to indicate evolution information of characterization vectors associated with core data items of target observation elements in the monitoring time series data set at the time series level;

[0042] Step S123: obtaining a state monitoring result of the monitoring time series data set based on the state monitoring network according to the second characterization vector set.

[0043] In step S121, the computer system obtains a first characterization vector set of the monitoring time series data set based on the condition monitoring network. The first characterization vector set is used to indicate the evolution information of the characterization vector associated with the core data item of the target observation element in the monitoring time series data set at the state level.

[0044] As mentioned above, the target observation elements are key indicators for monitoring the operating conditions of the evaporative cooling device, such as the cooling efficiency, air flow, humidity regulation capacity, etc. in the evaporative cooling device. The core data items are the key components of these target observation elements. Taking cooling efficiency as an example, they may include the temperature difference of the evaporator, the flow rate of the coolant, etc. These core data items can essentially reflect the state characteristics of the target observation elements. The evolution information at the state level reflects the changes of the core data items under different states. The computer system uses a state monitoring network to obtain a first set of characterization vectors. The state monitoring network can adopt a neural network structure in deep learning, such as a convolutional neural network (CNN) or a recurrent neural network (RNN).

[0045] Taking the convolutional neural network as an example, the computer system inputs the monitoring time series data set into the convolutional neural network. The convolution kernel in the convolution layer slides on the data to perform a convolution operation, which can automatically extract local features in the data. For the data part related to the core data items of the target observation elements in the monitoring time series data set, the convolution operation can extract the characteristic patterns of these data in different local areas. For example, for the coolant flow rate data, the convolution operation may extract the fluctuation pattern characteristics of the flow rate in a short period of time.

[0046] After being processed by multiple convolutional layers and pooling layers (pooling layers are used to reduce data dimensions and retain the main feature information), the fully connected layer integrates the features extracted by the previous layers to obtain the first set of representation vectors. Each representation vector in this set reflects the comprehensive characteristics of the core data items of the target observation element in a certain state. For example, in different operating stages of the evaporative cooling device (such as the startup stage, stable operation stage, load reduction stage, etc.), core data items such as coolant flow rate and evaporator temperature difference will have different performances. The first set of representation vectors can capture these evolutionary information at the state level.

[0047] In step S122, the computer system obtains a second set of characterization vectors based on the condition monitoring network according to the first set of characterization vectors, which is used to indicate the evolution information of the characterization vectors associated with the core data items of the target observation elements in the monitoring time series data set at the time series level.

[0048] The evolution information at the time series level focuses on the changing rules of the core data items over time. The computer system uses the state monitoring network to further process the first set of representation vectors to obtain the second set of representation vectors. If the state monitoring network adopts a recurrent neural network (RNN) structure, the RNN has the ability to process time series data, and it can process the vectors in the first set of representation vectors in chronological order. In the RNN, the hidden state of each time step is updated according to the current input (i.e., the vector in the first set of representation vectors) and the hidden state of the previous time step. This update mechanism enables the RNN to capture the dependencies of data in time series.

[0049] Let's take the network structure that combines the convolutional neural network (CNN) and the long short-term memory network (LSTM, a special RNN) as an example. First, CNN processes the monitoring time series data set to obtain the first set of representation vectors, and then inputs the first set of representation vectors into the LSTM network. The LSTM network processes time series information through its special gating mechanism (including input gate, forget gate, and output gate). The input gate determines which new information can enter the cell state, the forget gate determines which information can be forgotten from the cell state, and the output gate controls which information in the cell state can be used as output. In this way, LSTM can effectively process the time series information in the first set of representation vectors to obtain the second set of representation vectors.

[0050] In step S123, the computer system obtains the state monitoring result of the monitoring time series data set based on the state monitoring network according to the second characterization vector set. This state monitoring result is used to indicate whether the time series evolution of the characterization vector associated with the core data item of the target observation element in the monitoring time series data set meets the standard. The "standard" here is pre-set, which is based on a large amount of data statistics and analysis of the normally operating evaporative cooling device. For example, for the core data item of the coolant flow rate, if it is in normal operation, the flow rate should fluctuate within a relatively stable range, and the fluctuation amplitude should not exceed a certain threshold. The setting of this threshold can be based on the statistical analysis of historical data, such as calculating the mean and standard deviation of the coolant flow rate fluctuation amplitude in the historical data, and then setting a reasonable threshold range according to actual needs, such as the mean ± 3 times the standard deviation.

[0051] When the computer system uses the condition monitoring network to analyze the second characterization vector set, it compares the vectors in the set with the preset standard. If a condition monitoring network with a neural network structure is used, the last layer of the network can be a classification layer (for example, a fully connected layer using the Softmax function as the activation function), which maps the vectors in the second characterization vector set to different categories, which can represent different situations such as compliance with the standard and non-compliance with the standard.

[0052] Taking a simple binary classification as an example, if the vectors in the second representation vector set, after being processed by the state monitoring network, the output result is 0 or 1, where 0 indicates that the temporal evolution of the representation vectors associated with the core data items of the target observation elements in the monitored time series dataset conforms to the standard, and 1 indicates non - conformity. During the training (debugging) process of the network, a large number of labeled data (i.e., the second representation vector sets corresponding to the monitored time series datasets whose conformity to the standard is known) are used to adjust the parameters of the network, so that the network can accurately classify new input data (i.e., the second representation vector sets to be monitored). For example, assume that there are 1000 sets of labeled second representation vector set data, among which 800 sets indicate conformity to the standard and 200 sets indicate non - conformity. These data are input into the state monitoring network for training to adjust parameters such as weights and biases in the network. During the actual monitoring process, when a new second representation vector set is input, the network determines whether the state monitoring result of the corresponding monitored time series dataset conforms to the standard or not according to the trained model.

[0053] As an implementation manner, step S122, obtaining the second representation vector set based on the state monitoring network according to the first representation vector set, may include:

[0054] Step S1221: Perform dimensionality transformation on the first representation vector set to obtain the transformed first representation vector set;

[0055] Step S1222: Based on the state monitoring network, obtain the second representation vector set according to the transformed first representation vector set.

[0056] In the embodiment of the present application, in step S1221, dimensionality transformation is performed on the first representation vector set to obtain the transformed first representation vector set. The first representation vector set is obtained in step S121, and it is used to indicate the evolution information of the representation vectors associated with the core data items of the target observation elements in the monitored time series dataset at the state level. These representation vectors reflect the comprehensive characteristics of the core data items of the target observation elements in different states. Dimensionality transformation is a technical means for operating on the vector space. In practical applications, a computer system may perform dimensionality transformation for various reasons. For example, to make the data better adapt to the processing modules in the subsequent state monitoring network, or to reduce data redundancy and improve computational efficiency, etc.

[0057] Suppose the vectors in the first representation vector set are , and its dimension is n, that is . A common technical means for a computer system to perform dimensionality transformation is linear transformation. For example, through an m×n (m < n) transformation matrix A on to perform a linear transformation to obtain the transformed vector , the formula is: The element values ​​of the transformation matrix A here are determined according to the specific task requirements and data characteristics.

[0058] Taking the cooling efficiency in an evaporative cooling device as an example, it is assumed that the core data items of the cooling efficiency include the temperature difference of the evaporator, the flow rate of the coolant, etc. A vector in the first characterization vector set obtained in step S121 may represent the comprehensive characteristics of these core data items under a certain specific state. The dimension of the vector may be relatively high, containing multiple data dimensions related to these core data items. Through dimensionality transformation, the computer system can map this high-dimensional vector to a low-dimensional space. For example, the dimension of the original vector is 10, and after a linear transformation of a 5×10 transformation matrix, a vector with a dimension of 5 is obtained. This transformed vector still retains the key information related to the cooling efficiency, but is easier to process and analyze in a low-dimensional space.

[0059] Another technical means of dimensional transformation is nonlinear transformation. For example, kernel function method is used. Possible kernel functions such as Gaussian kernel function (where x and y are vectors and σ is a parameter). The kernel function is used to map the vectors in the original first characterization vector set to a new feature space, and the dimensionality transformation is realized in this new feature space. This nonlinear transformation is very useful in dealing with complex data distribution relationships. For example, when there is a complex nonlinear relationship between cooling efficiency and other related factors (such as ambient humidity, equipment load, etc.), nonlinear dimensionality transformation can better mine the intrinsic structure in the data, so that the transformed vector can more effectively reflect the evolution information of the target observation element at the state level.

[0060] After completing the dimensionality transformation of step S1221, the computer system enters step S1222, that is, obtaining a second set of characterization vectors based on the transformed first set of characterization vectors based on the condition monitoring network. This second set of characterization vectors is used to indicate the evolution information of the characterization vectors associated with the core data items of the target observation elements in the monitoring time series data set at the time series level.

[0061] The condition monitoring network can use a variety of structures to achieve this function, such as a neural network structure. If a network structure combining a convolutional neural network (CNN) and a recurrent neural network (RNN) is used, the computer system inputs the transformed first representation vector set into the network.

[0062] First, the convolutional neural network part can extract features from the transformed first representation vector set of the input. The convolution kernel in the convolution layer slides on the vector to perform convolution operations, which is similar to the convolution operation principle in image data, and can automatically extract local features in the vector. For example, for the transformed vector representing the core data items related to cooling efficiency, the convolution operation can extract the local pattern of cooling efficiency related features at different time points. The pooling layer downsamples the features extracted by the convolution layer to reduce the dimension of the data and retain the main feature information. Then, the feature vector processed by the convolutional neural network is input to the recurrent neural network part. Recurrent neural networks (taking the long short-term memory network LSTM as an example) have the ability to process time series data. In LSTM, the hidden state of each time step is updated according to the current input (i.e., the feature vector processed by the convolutional neural network) and the hidden state of the previous time step. The neurons in LSTM control the flow of information through special gating mechanisms (including input gates, forget gates, and output gates).

[0063] The input gate determines what new information can enter the cell state, the formula is: t =σ(W xi x t +W hi h t-1 +b i ) (where i t is the output of the input gate, σ is the Sigmoid function, W xi and W hi is the weight matrix, x t is the current input vector, h t-1 is the hidden state of the previous time step, b i is the bias term). The forget gate determines which information can be forgotten from the cell state, and the formula is: t =σ(W xf x t +W hf h t-1 +b f ) (where f_t is the output of the forget gate, W xf and W hf is the weight matrix, b f is a bias term). The output gate controls which information in the cell state can be used as output. The formula is: t =σ(W xo x t +W ho h t-1 +b o ) (where o t is the output of the output gate, W xo and W ho is the weight matrix, bo is the bias term).

[0064] In this way, LSTM can effectively process the time series information in the transformed first characterization vector set to obtain the second characterization vector set. The vectors in this second characterization vector set can reflect the evolution information of the core data items of the target observation elements at the time series level. For example, for the core data items related to cooling efficiency, the second characterization vector set can reflect the changing trend of cooling efficiency under different working conditions (such as different ambient temperatures, equipment loads, etc.) over time. For another example, if a simple recurrent neural network structure is used, the computer system inputs the transformed first characterization vector set into each time step of the RNN in sequence. Each neuron in the RNN is connected to the neuron of the previous time step. This connection structure enables the network to remember previous information. At each time step, the neuron calculates the current hidden state based on the current input vector and the hidden state of the previous time step. The formula is: h t =tanh(W xh x t +W hh h t-1 +b h ) (where h t is the hidden state of the current time step, tanh is the hyperbolic tangent function, W xh and W hh is the weight matrix, b h is the bias term). By processing the entire transformed first representation vector set at multiple time steps, RNN finally obtains the second representation vector set, which reflects the evolution information of the core data items of the target observation elements at the time series level.

[0065] As an embodiment, the state monitoring network includes a timing characterization component and a state characterization component, the timing characterization component is composed of a first pre-debugging component and a first fine-tuning component, the state characterization component is composed of a second pre-debugging component and a second fine-tuning component, the first fine-tuning component is used to obtain a first characterization vector set based on the first pre-debugging component, and the second fine-tuning component is used to obtain a second characterization vector set based on the second pre-debugging component.

[0066] The time series characterization component and the state characterization component respectively undertake different functions and work together to realize a comprehensive analysis of the time series data set of the evaporative cooling device monitoring. The time series characterization component is composed of the first pre-debugging component and the first fine-tuning component. The first pre-debugging component plays a basic construction role in the entire time series characterization component. It is similar to building an initial framework for the entire time series characterization process, which determines some basic parameters, structures or rules. For example, when building a time series characterization component based on a neural network, the first pre-debugging component can determine the basic structure of the neural network, such as whether to use a recurrent neural network (RNN) and its variants (such as a long short-term memory network LSTM or a gated recurrent unit GRU), or other structures that can process time series data, such as a combination of a convolutional neural network (CNN) and RNN. The pre-debugging component is a pre-trained (i.e., pre-training) network component. Assuming that the LSTM structure is used to construct the first pre-debugging component in the time series characterization component, the computer system will initialize some basic parameters of the LSTM, such as the number of neurons, the number of layers, etc. For example, it is determined that the LSTM network has 3 layers, each with 100 neurons. This provides an initial network architecture for subsequent data processing. The first fine-tuning component makes more detailed adjustments based on the first pre-debugging component to obtain a first set of representation vectors based on the first pre-debugging component. Fine-tuning is a gradual optimization process that aims to make the network better suited to specific tasks and data.

[0067] Taking the temperature monitoring data in the evaporative cooling device as an example, the temperature data is collected in time series. The initial network structure constructed by the first pre-debugging component may not be accurate enough to process the temperature data. The first fine-tuning component will fine-tune the parameters in the network according to the time series characteristics of the temperature data, such as the temperature fluctuation frequency and fluctuation amplitude. Assume that in the LSTM network, the first fine-tuning component will adjust the connection weights between neurons. These connection weights determine how information is transmitted and processed in the network. For the input temperature time series data x=(x1,x2,...,x t ) (where t represents the time step), at each time step t of LSTM, the update formula of the neuron is as follows:

[0068] (input gate);

[0069] (Forget Gate);

[0070] ;

[0071] (output gate);

[0072] ;

[0073] Among them, W represents various weight matrices (such as W xi is the weight matrix of the input gate to the hidden layer, etc.), b represents the bias term, σ is the Sigmoid function, and tanh is the hyperbolic tangent function. The first fine-tuning component adjusts the values ​​of these weight matrices W and the bias term b according to the characteristics of the temperature data, so that the network output results can accurately reflect the characteristics of the temperature data at the time series level, and then obtain the first characterization vector set. This first characterization vector set can represent the comprehensive characteristics of the temperature data at different time steps, such as the temperature change trend, periodicity and other information.

[0074] The state characterization component is composed of the second pre-debugging component and the second fine-tuning component. Its principle is similar to that of the timing characterization component, but its function focuses on the characterization of the state level.

[0075] The second pre-debugging component also builds the initial framework for the state representation component. For example, when performing state representation on the target observation factor of the cooling efficiency of the evaporative cooling device, if a neural network structure is used, the second pre-debugging component may determine that the type of network is a multi-layer perceptron (MLP), and determine the basic structural parameters such as the number of layers of the network and the number of neurons in each layer. Assume that a three-layer MLP structure is determined, the first layer has 50 neurons, the second layer has 30 neurons, and the third layer determines the number of neurons according to the dimension of the state representation vector to be output.

[0076] The second fine-tuning component obtains the second characterization vector set based on the second pre-debugging component. Taking the data related to cooling efficiency as an example, the cooling efficiency may be affected by multiple factors, such as the temperature difference of the evaporator, the flow rate of the coolant, etc. These factors constitute the state data related to the cooling efficiency. The second fine-tuning component will fine-tune the network according to the characteristics of these state data.

[0077] As an implementation method, the debugging process of the condition monitoring network includes the following steps:

[0078] Step S10: Acquire a monitoring time series data set, where the monitoring time series data set includes a plurality of monitoring data clusters associated with target observation elements arranged in time series;

[0079] Step S20: for u monitoring data clusters in the monitoring time series data set, add noise to the core data items of the target observation elements in the monitoring data clusters to obtain monitoring time series data set samples, and the time series evolution of the representation vector associated with the core data items of the target observation elements in the monitoring time series data set samples meets the standard, u≥1;

[0080] Step S30: Debugging the state monitoring network according to the monitoring time series data set sample to obtain a debugged state monitoring network, and the debugged state monitoring network is used to monitor whether the target evaporative cooling device meets the standard.

[0081] Step S10 obtains a monitoring time series data set, which contains multiple monitoring data clusters associated with target observation elements arranged in time series. The target observation elements are key factors for evaluating the operating status of the evaporative cooling device, such as cooling efficiency, equipment power consumption, air humidity regulation capacity, etc. Taking cooling efficiency as an example, the computer system may obtain relevant data by installing temperature sensors and flow sensors at key parts of the evaporative cooling device. The temperature sensor measures the temperature of the evaporator inlet and outlet at regular time intervals (such as every 5 minutes), and the flow sensor measures the flow of the coolant at the same time. These data collected at different time points constitute a monitoring data cluster related to cooling efficiency. Multiple such monitoring data clusters are arranged in chronological order to form a monitoring time series data set for cooling efficiency.

[0082] For monitoring of equipment power consumption, the computer system can obtain data from the power monitoring device. The power monitoring device records the power consumption of each component of the evaporative cooling device (such as the compressor, fan, pump, etc.) in real time. Over a period of time (such as a day), these power consumption data at different times form a monitoring data cluster related to equipment power consumption, which becomes a monitoring time series data set of equipment power consumption after being arranged in chronological order.

[0083] The monitoring data of air humidity regulation capacity may be obtained through humidity sensors. Humidity sensors collect air humidity data at different locations in the operating environment of the evaporative cooling device, and these data are associated with the collection time. Within a time period, these humidity data at different times and other related data that may affect humidity regulation (such as ambient temperature, air flow rate, etc.) constitute a monitoring data cluster about air humidity regulation capacity. Multiple such monitoring data clusters are arranged in chronological order to form a corresponding monitoring time series data set.

[0084] The computer system can obtain these monitoring time series data sets by directly reading the data from the sensor device, and then organizing and storing them to ensure that the data is arranged in the correct time sequence and that different monitoring data clusters have clear correlations, providing a complete data foundation for the subsequent debugging process.

[0085] In step S20, the computer system performs a noise addition operation on some monitoring data clusters (u) in the acquired monitoring time series data set. The core data items of the target observation element are the data parts that have a key impact on the target observation element. Taking cooling efficiency as an example, if the evaporator inlet and outlet temperature difference and the coolant flow rate are the key factors affecting the cooling efficiency, then they are the core data items.

[0086] Noising is a means of simulating data anomalies. The computer system creates more challenging data set samples by adding noise to core data items. For example, for the core data item of the evaporator inlet and outlet temperature difference, if the original temperature difference data sequence is T=(t1, t2,…, t n ) (where n is the number of data points), the computer system can add noise according to certain rules. A simple way to add noise is to add random noise according to the normal distribution. Let the noise sequence be N=(n1,n2,…,n n ), where n i It follows a normal distribution with a mean of 0 and a standard deviation of σ (i.e., n i ~N(0,σ)), then the temperature difference data sequence after noise addition is T'=(t1+n1,t2+n2,…,t n +n n ). For the core data item of coolant flow rate, assume that the original flow rate data sequence is V=(v1,v2,…,v n ), the computer system can also add noise according to a certain distribution (such as uniform distribution). Let the interval of uniform distribution be [a,b], and the generated noise sequence be U=(u1,u2,…,u n ), where u i Randomly select values ​​in the interval [a,b], and the velocity data sequence after noise addition is V'=(v1+u1,v2+u2,…,v n +u n ).

[0087] After the noise addition operation, although the data has changed, it is necessary to ensure that the time series evolution of the characterization vector associated with the core data item of the target observation element in the monitoring time series data set sample meets the standard. The standard here is based on a large amount of data statistics and analysis of the normally operating evaporative cooling device. For example, for the core data items related to cooling efficiency, although the data fluctuates after noise addition, the fluctuation trend of the overall cooling efficiency within a certain range must still meet the expected pattern during normal operation. This requirement of meeting the standard is to enable the debugged condition monitoring network to accurately judge the status of the device in the face of small data anomalies that may actually occur.

[0088] Through this noise addition operation, the computer system obtains a monitoring time series data set sample, which contains normal data and abnormal data with noise. It can more comprehensively cover various situations that may occur in actual operation and provide a richer data basis for the debugging of the condition monitoring network.

[0089] In step S30, the computer system uses the monitoring time series data set samples to debug the condition monitoring network. Assume that the condition monitoring network adopts a neural network structure (such as a combination of a convolutional neural network and a recurrent neural network). First, the computer system inputs the monitoring time series data set samples into the condition monitoring network. The initial parameters of the network are initialized randomly or according to some predefined rules. During the forward propagation process of the network, the data is processed by each layer in turn.

[0090] Taking the convolutional neural network (CNN) part as an example, for each monitoring data cluster in the input monitoring time series data set sample, the convolution kernel in the convolution layer slides on the data to perform a convolution operation. Assuming that the input data is a two-dimensional matrix (the time series data can be converted into a two-dimensional form in a certain way, such as taking the data at each time point as a row of the matrix), the convolution kernel is K, and the size is m×n (m and n are the dimensions of the convolution kernel), then the convolution operation can be expressed as: ; Where C(i,j) is the element in the convolution result matrix, and I(i,j) is the element in the input matrix. Through the convolution operation, the network automatically extracts local features in the data.

[0091] The pooling layer downsamples the features extracted by the convolution layer. For example, the maximum pooling operation is used to select the maximum value in the local area as the result of pooling, which helps to reduce the dimension of the data while retaining the main feature information. The feature data processed by the CNN part then enters the recurrent neural network (RNN) part (taking the long short-term memory network LSTM as an example). In LSTM, the hidden state of each time step is updated according to the current input and the hidden state of the previous time step.

[0092] After completing the forward propagation, the computer system calculates the loss function based on the network output and the known targets in the monitoring time series data set samples (i.e., whether they meet the standards or not). For example, if the cross entropy loss function is used, for a binary classification problem (meeting the standards or not), assuming that the network output is y (the probability of being predicted to meet the standards), and the true label is y' (0 means not meeting the standards, 1 means meeting the standards), then the cross entropy loss function is: L=-[y'log(y)+(1-y')log(1-y)]; then, the computer system uses the back propagation algorithm to update the parameters of the network to minimize the loss function. The back propagation algorithm updates the parameters according to a certain learning rate based on the gradient of the loss function for each parameter in the network. For example, for the weight parameter W, the update formula is: ; where α is the learning rate, is the gradient of the loss function with respect to the weight W.

[0093] Through multiple iterations of this process, the parameters of the network are continuously adjusted until the network's performance on the samples of the monitoring time series data set reaches a satisfactory effect, such as the value of the loss function converges to a smaller value or reaches a preset number of iterations. Finally, the computer system obtains a debugged state monitoring network that can accurately monitor whether the target evaporative cooling device meets the standards, thereby providing reliable protection for the operating status monitoring of the evaporative cooling device.

[0094] As an implementation manner, step S20, adding noise to the core data items of the target observation elements in the monitoring data cluster to obtain a monitoring time series data set sample, may include:

[0095] Step S21: adding noise to each data item in the monitoring data cluster to obtain a noisy monitoring data cluster, wherein there is an error between the data items in the noisy monitoring data cluster and the data items in the monitoring data cluster;

[0096] Step S22: Integrate the noise-added monitoring data cluster and the monitoring data cluster to obtain an abnormal monitoring data cluster, wherein there is an error between the core data items of the target observation element in the abnormal monitoring data cluster and the core data items of the target observation element in the monitoring data cluster, and there is no error between the common data items of the target observation element in the abnormal monitoring data cluster and the common data items of the target observation element in the monitoring data cluster;

[0097] Step S23: Change the u monitoring data clusters into the abnormal monitoring data clusters corresponding to the u monitoring data clusters, and obtain the monitoring time series data set samples.

[0098] In step S21, the computer system performs a noise adding operation on each data item in the monitoring data cluster. The monitoring data cluster contains multiple data items associated with the target observation element, which reflect the information related to the operating status of the evaporative cooling device at a certain moment or in a certain time period. For example, for the target observation element of the cooling efficiency of the evaporative cooling device, its monitoring data cluster may contain data items such as the inlet and outlet temperatures of the evaporator, the flow rate of the coolant, and the ambient temperature. Noise adding is an operation of artificially introducing interference, the purpose of which is to make the monitoring data closer to the actual complex situations that may occur, thereby improving the robustness of the state monitoring network. The computer system can add noise to the data items in a variety of ways.

[0099] One feasible approach is to add noise based on probability distribution. For example, for a data item representing the evaporator inlet temperature, assuming that its original data value is T original , the computer system can follow the normal distribution N(μ,σ 2) to add noise. Here μ is the mean and σ is the standard deviation. Assume μ=0 (meaning that the center value of the original data is not changed on average) and σ=1 (the standard deviation can be adjusted according to the actual situation to control the intensity of the noise). Generate a random number n that conforms to the normal distribution. Then the data item T after adding noise is noisy =T original +n.

[0100] For the data item of coolant flow, assume that its original value is F original , the computer system can also use uniform distribution to add noise. Suppose the interval of uniform distribution is [a,b], for example [-0.5,0.5]. The computer system randomly generates a number u in this interval, then the coolant flow data item after adding noise is F noisy =F original +u. Taking the ambient temperature data item as an example, if the original ambient temperature value is E_{original}, the computer system can also use other distributions to add noise, such as exponential distribution. Let the parameter of the exponential distribution be λ, and generate a random number e that conforms to the exponential distribution. Then the ambient temperature data item after adding noise is E noisy =E original +e.

[0101] By performing such a noise addition operation on each data item in the monitoring data cluster, the computer system obtains a noisy monitoring data cluster. An error is generated between each data item after noise addition and the corresponding data item in the original monitoring data cluster. This error simulates the data fluctuations that may be caused by sensor errors, environmental interference and other factors in actual operation.

[0102] In step S22, after obtaining the noise-added monitoring data cluster, the computer system performs an integration operation to obtain an abnormal monitoring data cluster. The purpose of this step is to make the core data items produce errors while keeping some data items unchanged (ordinary data items), thereby constructing a special abnormal data structure for debugging the state monitoring network.

[0103] Taking the target observation factor of the equipment power consumption of the evaporative cooling device as an example, the core data items include the power values ​​of the main power-consuming equipment (such as compressors, fans, etc.), which have a greater impact on the equipment power consumption; while ordinary data items include the power values ​​of some auxiliary equipment or some fixed parameters related to the power consumption calculation (such as line resistance, etc.), which have relatively little impact on the overall power consumption.

[0104] During the integration process, the computer system keeps the original values ​​of common data items unchanged, that is, directly obtains the corresponding values ​​from the monitoring data cluster. For core data items, the noisy core data items in the noisy monitoring data cluster replace the core data items in the original monitoring data cluster.

[0105] For example, for the monitoring data cluster of equipment power consumption, assume that the original monitoring data cluster is [P1, P2, r1, r2, ...], where P1 and P2 are the power of the compressor and fan (core data items), and r1 and r2 are the power of some auxiliary equipment or related fixed parameters (ordinary data items). The noise-added monitoring data cluster is [P1', P2', r1, r2, ...], where P1' and P2' are the power of the compressor and fan after noise addition.

[0106] The integrated abnormal monitoring data cluster is [P1', P2', r1, r2, ...]. In this abnormal monitoring data cluster, the core data items (compressor and fan power) of the target observation element (equipment power consumption) have errors with the core data items in the original monitoring data cluster (due to the noise addition operation), while the common data items (auxiliary equipment power or related fixed parameters) have no errors with the common data items in the original monitoring data cluster.

[0107] The abnormal monitoring data cluster constructed by this integration method can more specifically simulate the situation in which the core data items are abnormal while the common data items are normal in practice, which helps the condition monitoring network learn to distinguish the abnormal changes of the core data items, thereby improving the accuracy of the condition monitoring of the evaporative cooling device.

[0108] In step S23, based on the previous steps, the computer system processes u monitoring data clusters (u≥1) one by one, and changes each monitoring data cluster into a corresponding abnormal monitoring data cluster. These monitoring data clusters are arranged in time series, and they constitute a monitoring time series data set.

[0109] For example, suppose the initial monitoring time series data set contains 10 monitoring data clusters, namely Cluster1, Cluster2,..., Cluster 10 The computer system performs noise addition and integration operations on each monitoring data cluster according to the method of steps S21 and S22, and changes Cluster1 to the corresponding abnormal monitoring data cluster AbnormalCluster1, changes Cluster2 to AbnormalCluster2, and so on, until Cluster 10 Changed to AbnormalCluster 10 .

[0110] After such changes, these 10 abnormal monitoring data clusters constitute a new monitoring time series data set sample. The data in this sample contains both normal information (ordinary data item part) in the original monitoring data cluster and abnormal information (error part of the core data item) introduced by noise addition and integration. In the actual operation of the evaporative cooling device, various factors may cause abnormal changes in the core data items, and this sample can simulate these possible situations. By using this sample to debug the condition monitoring network, the network can learn how to identify abnormal evolution of core data items, thereby improving its accuracy and reliability in actual monitoring. For example, when a core data item (such as the key temperature value in the cooling efficiency) of the actual evaporative cooling device fluctuates abnormally, the debugged condition monitoring network can detect this abnormality more accurately and issue an alarm or take corresponding control measures in time.

[0111] As an implementation manner, step S22, integrating the noise-added monitoring data cluster and the monitoring data cluster to obtain the abnormal monitoring data cluster, may include:

[0112] Step S221: obtaining multiple core data item subsets corresponding to the target observation element in the monitoring data cluster, each core data item subset including multiple core data items;

[0113] Step S222: for any one of the plurality of core data item subsets, generating a shielding data set corresponding to the monitoring data cluster according to the core data item subset, wherein the shielding data set is used to indicate the position of the core data item subset in the monitoring data cluster;

[0114] Step S223: Under the constraints of the shielded data sets corresponding to the plurality of core data item subsets, the noise-added monitoring data cluster and the monitoring data cluster are integrated to obtain an abnormal monitoring data cluster.

[0115] In step S221, in the monitoring data cluster of the evaporative cooling device, the target observation element is a key factor for measuring the operating status of the device, such as cooling efficiency, equipment power consumption, humidity adjustment capability, etc. For each target observation element, there are some data items that have a more critical impact on it, and these data items are divided into multiple core data item subsets.

[0116] Taking the cooling efficiency as an example, the computer system identifies the corresponding core data items in the monitoring data cluster. Assume that the monitoring data cluster contains data items such as the inlet and outlet temperatures of the evaporator, the flow rate of the coolant, the temperature of the coolant, and the air flow rate. The computer system divides the inlet and outlet temperatures of the evaporator into a subset of core data items, because the difference between the inlet and outlet temperatures directly reflects the heat exchange and has a key impact on the cooling efficiency; the flow rate and temperature of the coolant can be divided into another subset of core data items, because the flow rate and temperature of the coolant work together to affect the heat carrying capacity, thereby affecting the cooling efficiency; the air flow rate can also be used as a core data item subset alone, because it affects the efficiency of heat exchange.

[0117] The multiple core data items contained in each core data item subset are inherently related and together have an important impact on the target observation elements. This division method helps the computer system process data more specifically in subsequent steps and perform corresponding operations based on the characteristics of different subsets.

[0118] In step S222, for each core data item subset, the computer system needs to generate a shielding data set. The function of this shielding data set is to mark the position information of the core data item subset in the entire monitoring data cluster.

[0119] Still taking cooling efficiency as an example, assume that the core data item subset of the evaporator inlet and outlet temperature previously divided is {T in ,T out}(T in Indicates the inlet temperature, T out represents the outlet temperature), this subset may be located at a specific position in the entire monitoring data cluster. Assume that the monitoring data cluster is a sequentially arranged data structure, such as (T ambient Indicates the ambient temperature, F coolant Indicates the coolant flow rate, T coolant Indicates the coolant temperature, V air represents air flow rate), then the masked data set for the core data item subset of the evaporator inlet and outlet temperatures may be a vector with the same length as the monitoring data cluster, for example, [0,1,1,0,0,0], where 1 indicates that the corresponding position is a data item in the core data item subset, and 0 indicates that it is not.

[0120] A technical means for a computer system to generate a masked data set can be based on index comparison. First, determine the starting index and ending index of the core data item subset in the monitoring data cluster. For the above-mentioned evaporator inlet and outlet temperature core data item subset, the starting index is 1 and the ending index is 2. Then, create an all-zero vector with the same length as the monitoring data cluster, and set the elements from the starting index to the ending index position to 1, thereby obtaining a masked data set.

[0121] In step S223, the computer system uses the previously generated shielded data set to integrate the noisy monitoring data cluster and the monitoring data cluster to obtain an abnormal monitoring data cluster.

[0122] Let’s continue with the cooling efficiency example. Assume that there are two core data item subsets, namely, the evaporator inlet and outlet temperature subsets {T in ,T out} and the coolant flow and temperature subset {F coolant , T coolant}, and their corresponding masked data sets are [0,1,1,0,0,0] and [0,0,0,1,1,0] respectively.

[0123] For the noisy monitoring data cluster (The data items with ' represent the noised data) and the original monitoring data cluster , the computer system performs the following integration operations.

[0124] For the shielded data set [0,1,1,0,0,0] corresponding to the subset of the evaporator inlet and outlet temperatures, the computer system replaces the data in the original monitoring data cluster with the data at the corresponding position in the noisy monitoring data cluster (i.e., indexes 1 and 2), and obtains an intermediate result .

[0125] For the shielded data set [0,0,0,1,1,0] corresponding to the coolant flow and temperature subsets, the above intermediate results are operated again, and the data at the corresponding positions in the noise monitoring data cluster (i.e., indexes 4 and 5) are replaced with the data in the intermediate results, and finally the abnormal monitoring data cluster is obtained. .

[0126] In this abnormal monitoring data cluster, the core data items (evaporator inlet and outlet temperatures, coolant flow rate and temperature) of the target observation element (cooling efficiency) have errors with the core data items in the original monitoring data cluster (due to noise addition and replacement operations), while other common data items (ambient temperature and air flow rate) have no errors with the common data items in the original monitoring data cluster.

[0127] This integration operation mode under the restriction of shielded data sets enables the computer system to accurately control which data items are replaced, so as to construct abnormal monitoring data clusters in a targeted manner. The abnormal monitoring data clusters constructed in this way can more accurately simulate the abnormalities of core data items in actual situations, which helps to improve the accuracy and reliability of the condition monitoring network for evaporative cooling device condition monitoring. For example, in actual operation, if the evaporator inlet and outlet temperatures or the coolant flow and temperature change abnormally, this abnormal monitoring data cluster can enable the condition monitoring network to better learn how to identify such abnormalities, and then judge the status of the evaporative cooling device in a timely and accurate manner.

[0128] As an implementation mode, step S222, generating a shielded data set corresponding to the monitoring data cluster according to the core data item subset, may include:

[0129] Step S2221: for any data item in the monitoring data cluster, obtaining a first spatial interval between the data item and the location of the core data item subset;

[0130] Step S2222: determining a masking value of the data item based on the first spatial interval, wherein the masking value of the data item is inversely associated with the first spatial interval;

[0131] Step S2223: compose a masked data set according to the masked value of each data item in the monitoring data cluster.

[0132] In step S2221, the monitoring data cluster of the evaporative cooling device includes multiple data items related to the target observation element. For each data item, the computer system needs to determine the spatial relationship between it and the location of the specific core data item subset, which is described here by the first spatial interval.

[0133] Assuming that the cooling efficiency in the evaporative cooling device is taken as the target observation factor, the monitoring data cluster contains the data items: ambient temperature T env , evaporator inlet temperature T in , evaporator outlet temperature T out , Coolant flow F coolant , Coolant temperature T c , air velocity V air Etc. Divide the evaporator inlet temperature and outlet temperature into a core data item subset {T in ,T out}. With ambient temperature T env For example, its position in the monitoring data cluster is the first data item, and the core data item subset {T in ,T out}Starting from the second data item. The computer system can calculate the first spatial interval by simply subtracting the index. In this example, the first spatial interval between the ambient temperature and the core data item subset is 1 because the difference between it and the starting position of the core data item subset (index 2) is 1.

[0134] Then take the coolant temperature T c For example, its position in the monitoring data cluster is assumed to be the fifth data item, then it is related to the core data item subset {T in ,T out The first spatial interval of} is 3 (5-2=3).

[0135] In step S2222, after obtaining the first spatial interval between each data item and the core data item subset, the computer system needs to determine the masking value of the data item based on this interval. Since the masking value is inversely related to the first spatial interval, the closer the data item is to the core data item subset, the higher its masking value; the farther the distance is, the lower its masking value.

[0136] For example, suppose a simple linear relationship is set to determine the masking value. Let the maximum masking value be 1, the minimum masking value be 0, the total length of the monitoring data cluster be n, and the first spatial interval between the data item and the core data item subset be d. The formula can be used: Masking value = .

[0137] Let’s continue with the cooling efficiency example above. Assume that the monitoring data cluster has a total of 6 data items (n=6). env , which is related to the core data item subset {T in ,T out The first space interval of} is 1. According to the formula, its shielding value is .

[0138] For coolant temperature , whose first spatial interval with the core data item subset is 3, then the masked value is .

[0139] This masking value setting method reflects the degree of relevance of the data item to the core data item subset. Data items that are closer to the core data item subset may receive more attention in subsequent integration operations, while data items that are farther away are relatively less affected.

[0140] In step S2223, a masked data set is formed based on the masked values ​​of each data item in the monitoring data cluster. After the computer system determines the masked value of each data item in the monitoring data cluster, it combines these masked values ​​to form a masked data set.

[0141] Still taking the cooling efficiency monitoring data cluster as an example, the ambient temperature T has been calculated. env The shielding value is 0.8, and the evaporator inlet temperature T in and outlet temperature T out Belonging to the core data item subset, their mask value can be set to 1 (because they are data items in the core data item subset), coolant flow F coolant The masking value of is assumed to be 0.2 (calculated based on its spatial separation from the core data item subset), and the coolant temperature T c The shielding value is 0.4, and the air velocity V air The masking value of is assumed to be 0.1.

[0142] Then, the masked data set composed of these masked values ​​is [0.8, 1, 1, 0.2, 0.4, 0.1]. This masked data set reflects the degree of association or importance of each data item in the monitoring data cluster relative to a specific subset of core data items.

[0143] In actual operation, the masked data set will be used as a weight or identifier for subsequent operations when integrating the noisy monitoring data cluster and the monitoring data cluster. For example, during the integration process, data items with higher masking values ​​may retain more data in the original monitoring data cluster during integration, while data items with lower masking values ​​may be more easily replaced or affected by data in the noisy monitoring data cluster.

[0144] As an implementation mode, step S223, under the restriction of the shielded data sets corresponding to the plurality of core data item subsets, integrating the noise-added monitoring data cluster and the monitoring data cluster to obtain the abnormal monitoring data cluster, may include:

[0145] Step S2231: for any one of the shielded data sets corresponding to the plurality of core data item subsets, updating the noise-added monitoring data cluster according to the shielded data set to obtain a first integrated monitoring data cluster, and updating the monitoring data cluster according to the inverse shielded data set corresponding to the shielded data set to obtain a second integrated monitoring data cluster, wherein the sum of the shielded values ​​of the data items at the same position in the shielded data set and the inverse shielded data set is equal to 1;

[0146] Step S2232: obtaining an integrated monitoring data cluster corresponding to the shielded data set according to the first integrated monitoring data cluster and the second integrated monitoring data cluster;

[0147] Step S2233: Integrate the integrated monitoring data clusters corresponding to the multiple shielded data sets to obtain abnormal monitoring data clusters.

[0148] In step S2231, during the monitoring data processing of the evaporative cooling device, the computer system operates on the shielded data set corresponding to each core data item subset. Assuming that the cooling efficiency of the evaporative cooling device is the target observation element, the monitoring data cluster includes, for example, the evaporator inlet temperature T in , evaporator outlet temperature T out , Coolant flow F coolant , Coolant temperature T c 、Ambient temperature T env and air velocity V air The evaporator inlet temperature and outlet temperature are divided into a core data item subset. Assuming that the masked data set generated for this core data item subset is [0.2, 1, 1, 0.1, 0.3, 0.1], then its inverse masked data set is [0.8, 0, 0, 0.9, 0.7, 0.9] (because the masked values ​​of the data items at the same position add up to 1).

[0149] For a noisy monitoring data cluster, for example, [T env ',T in ',T out ',F coolant ',T c ',V air '] (with ' indicating the data after adding noise), the computer system updates the noisy monitoring data cluster according to the shielding data set to obtain the first integrated monitoring data cluster. The specific technical means is to multiply each data item in the noisy monitoring data cluster with the shielding value at the corresponding position in the shielding data set, and then add them to obtain a new data item.

[0150] The evaporator inlet temperature T in 'For example, in the masked data set [0.2, 1, 1, 0.1, 0.3, 0.1], the corresponding masked value is 1, so the updated value is T in '×1=T in '. For ambient temperature T env ', its corresponding shielding value is 0.2, and the updated value is T env '×0.2. In this way, each data item in the noise-added monitoring data cluster is operated to finally obtain the first integrated monitoring data cluster.

[0151] At the same time, the computer system updates the monitoring data cluster according to the inverse masking data set to obtain a second integrated monitoring data cluster. env ,T in ,T out ,F coolant ,T c ,V air ], with the evaporator inlet temperature T inFor example, in the inverse masked data set [0.8, 0, 0, 0.9, 0.7, 0.9], the corresponding masked value is 0, so the updated value is T in ×0=0. For ambient temperature T env , the corresponding shielding value is 0.8, and the updated value is T env × 0.8. Similarly, this operation is performed on each data item in the monitoring data cluster to obtain a second integrated monitoring data cluster.

[0152] In step S2232, after obtaining the first integrated monitoring data cluster and the second integrated monitoring data cluster, the computer system needs to combine them into a new integrated monitoring data cluster, which corresponds to the specific shielded data set.

[0153] For example, the first integrated monitoring data cluster is , the second integrated monitoring data cluster is .

[0154] The computer system adds the data items at corresponding positions in the two integrated monitoring data clusters to obtain the integrated monitoring data cluster corresponding to the shielded data set. That is, the new integrated monitoring data cluster is .

[0155] This combination method integrates the adjustment results of noisy monitoring data clusters and monitoring data clusters based on shielded data sets and inverse shielded data sets, so that the integrated monitoring data clusters not only take into account the impact of noisy data, but also retain the information of the original monitoring data. Moreover, this impact and retention are based on the degree of association between the data items and the core data item subsets (reflected by the shielding value).

[0156] In step S2233, in the monitoring data processing of the evaporative cooling device, there may be multiple core data item subsets, each of which has a corresponding shielded data set, a first integrated monitoring data cluster, a second integrated monitoring data cluster, and an integrated monitoring data cluster. The computer system needs to integrate these integrated monitoring data clusters corresponding to different shielded data sets together to obtain an abnormal monitoring data cluster.

[0157] Assume that there are two subsets of core data items. The integrated monitoring data cluster corresponding to the first subset of core data items is [A1, B1, C1, D1, E1, F1], and the integrated monitoring data cluster corresponding to the second subset of core data items is [A2, B2, C2, D2, E2, F2].

[0158] The computer system integrates these two integrated monitoring data clusters. A simple integration method is to directly add the data items at corresponding positions (if the properties of the data items allow this), and the obtained abnormal monitoring data cluster is [A1+A2, B1+B2, C1+C2, D1+D, E1+E, F1+F2].

[0159] The process of integrating the integrated monitoring data clusters corresponding to multiple shielded data sets is to construct an abnormal monitoring data cluster that fully reflects the impact of various core data item subsets. This abnormal monitoring data cluster is more comprehensive and accurate in simulating abnormal conditions of evaporative cooling devices because it comprehensively considers the characteristics of different core data item subsets and their interrelationships.

[0160] Through steps S2231-S2233, the computer system can construct an abnormal monitoring data cluster through a series of updating, combining and integrating operations based on the degree of association between the data item and the core data item subset. This helps to more accurately simulate various abnormal situations that may occur in the evaporative cooling device during the debugging process of the condition monitoring network, thereby improving the accuracy and reliability of the condition monitoring network for the condition monitoring of the evaporative cooling device. For example, in actual operation, if a certain core data item subset (such as evaporator temperature data closely related to cooling efficiency) undergoes abnormal changes, the condition monitoring network trained with the abnormal monitoring data cluster constructed in this way can better identify such anomalies, thereby providing guarantees for the effective monitoring and control of the evaporative cooling device.

[0161] As an implementation manner, in step S2231, updating the noisy monitoring data cluster according to the shielded data set to obtain the first integrated monitoring data cluster, and updating the monitoring data cluster according to the inverse shielded data set corresponding to the shielded data set to obtain the second integrated monitoring data cluster may include:

[0162] Step S22311: for any data item in the noisy monitoring data cluster, multiply the data item by the corresponding masking value of the data item in the masking data set to obtain a first integrated monitoring data cluster;

[0163] Step S22312: For any data item in the monitoring data cluster, multiply the data item by the masking value corresponding to the data item in the inverse masking data set to obtain a second integrated monitoring data cluster.

[0164] In step S22311, the computer system needs to operate the noise-added monitoring data cluster to obtain the first integrated monitoring data cluster. First, the concept of the noise-added monitoring data cluster is clarified, which is a data set obtained by performing noise-added processing on the data items in the original monitoring data cluster.

[0165] Assume that a certain operating state monitoring in an evaporative cooling device is taken as an example, the target observation factor is the cooling efficiency, and the original monitoring data cluster contains the evaporator inlet temperature T in , evaporator outlet temperature T out , Coolant flow F coolant , Coolant temperature T c 、Ambient temperature Tenv and air velocity V air After the noise processing, the noise monitoring data cluster becomes [T' env ,T' in ,T' out ,F' coolant ,T' c ,V' air ] (' here means the data after adding noise). At the same time, for a specific subset of core data items, there is a corresponding masked data set. For example, assuming that the evaporator inlet temperature and outlet temperature are divided into a core data item subset, the corresponding masked data set is [0.2, 1, 1, 0.1, 0.3, 0.1].

[0166] The computer system operates on each data item in the noise monitoring data cluster. in For example, its corresponding masking value in the masked data set is 1, then according to the operation requirements, the calculated value is T' in ×1=T' in . Let’s look at the ambient temperature T’ env , the corresponding shielding value in the shielded data set is 0.2, so the calculated result is T' env ×0.2.

[0167] In this way, the computer system multiplies each data item in the noisy monitoring data cluster by its corresponding masked value in the masked data set. coolant , assuming that the corresponding shielding value in the shielded data set is 0.1, the calculated result is F' coolant ×0.1.

[0168] After performing the above calculations on all data items in the noise-added monitoring data cluster, these calculation results are combined to obtain the first integrated monitoring data cluster. For example, if the results of the data items after calculation are , then the first integrated monitoring data cluster is .

[0169] This operation weights the noisy data according to the shielding value by multiplying the data items in the noisy monitoring data cluster with the shielding value. The shielding value reflects the degree of association between the data item and the core data item subset. The data items with high association (large shielding value) retain more original noisy values ​​in the first integrated monitoring data cluster, while the data items with low association (small shielding value) are weakened to a certain extent. This helps to better reflect the impact of different data items on the core data item subset in the subsequent construction of the abnormal monitoring data cluster, and thus more accurately simulate various possible abnormal situations.

[0170] In step S22312, the computer system performs similar operations on the monitoring data cluster, but uses an inverse masked data set. The inverse masked data set corresponds to the masked data set, and the masked values ​​of the data items at the same position add up to 1.

[0171] Still taking the cooling efficiency monitoring of the above evaporative cooling device as an example, the original monitoring data cluster is [T env ,T in ,T in ,F coolant ,T c ,V air ], and the inverse masked data set corresponding to the previously mentioned masked data set [0.2, 1, 1, 0.1, 0.3, 0.1] is [0.8, 0, 0, 0.9, 0.7, 0.9].

[0172] The computer system operates on each data item in the monitoring data cluster. in For example, its corresponding masking value in the inverse masking data set is 0, so the calculation result is T in ×0=0. For ambient temperature T env , the corresponding shielding value in the inverse shielding data set is 0.8, so the calculated result is T env ×0.8. In the same way, all data items in the monitoring data cluster are calculated. For example, for the coolant flow F coolant , assuming that the corresponding masking value in the inverse masking data set is 0.9, the calculated result is F coolant ×0.9. Combining these calculation results, we get the second integrated monitoring data cluster. For example, if the results of each data item after calculation are , then the second integrated monitoring data cluster is .

[0173] The purpose of this operation is to perform weighted processing on the original monitoring data cluster according to the inverse masking value. The inverse masking value reflects another weight relationship between the data item and the core data item subset. Through this weighted processing, when constructing the second integrated monitoring data cluster, the data items with a low degree of association with the core data item subset (large inverse masking value) retain more values ​​of the original monitoring data in the result, while the data items with a high degree of association (small inverse masking value) are weakened. This cooperates with the operation in step S22311, and provides data information with different focuses for the subsequent synthesis of the two integrated monitoring data clusters, which helps to construct the abnormal monitoring data cluster more comprehensively and accurately, thereby improving the accuracy and reliability of the state monitoring network for the state monitoring of the evaporative cooling device.

[0174] As an implementation manner, step S21, adding noise to each data item in the monitoring data cluster to obtain the noisy monitoring data cluster, may include:

[0175] Step S211: obtaining a linear transformation array, where array elements in the linear transformation array are noise adding variables, and the noise adding variables are used to transform the spatial distribution of data items;

[0176] Step S212: performing linear transformation on the monitoring data cluster based on the linear transformation array to obtain a noisy monitoring data cluster.

[0177] In step S211, in the monitoring data processing of the evaporative cooling device, the computer system first obtains a linear transformation array, which plays a key role in the process of adding noise to the data items in the monitoring data cluster. The linear transformation array is an array composed of noise adding variables. The noise adding variable is a special variable that is designed to change the spatial distribution of the data items. The spatial distribution of the data items here can be understood as the arrangement and distribution of the data in the data space to which it belongs.

[0178] Taking the cooling efficiency monitoring in the evaporative cooling device as an example, it is assumed that the monitoring data cluster contains data items such as evaporator inlet temperature, outlet temperature, coolant flow rate, etc. For the evaporator inlet temperature data item, its possible value range under normal operation is [20°C, 30°C], and this range constitutes a distribution area of ​​it in the data space.

[0179] The noise-added variables in the linear transformation array obtained by the computer system can transform the spatial distribution of such data items. For example, the linear transformation array can be a one-dimensional array, such as [a1, a2, ..., a n ], where each a i (i=1,2,...,n) are all noise variables. Assume there is a simple two-dimensional data space, where the coordinates of a data item x are (x1,x2). Its coordinate position can be changed by linear transformation. The formula can be expressed as: ; Here It can be regarded as a transformation matrix composed of appropriate noise variables selected from the linear transformation array. (x'1, x'2) is the coordinate after linear transformation, that is, the position of the noised data item in the transformed data space.

[0180] In practical applications, for different data items of the evaporative cooling device, the linear transformation array can be determined according to the characteristics of the data and the expected noise effect. For example, for the temperature data item, since the temperature changes relatively smoothly, the noise variable in the linear transformation array may choose a smaller value to avoid excessive changes. For data items such as coolant flow, which may have a large fluctuation range, the value of the noise variable can be relatively large to simulate a wider range of actual conditions.

[0181] There are many ways for a computer system to obtain a linear transformation array. One way is to determine it based on statistical analysis of historical data. If historical data shows that the fluctuation range of a data item is small during normal operation, then when constructing the linear transformation array, the corresponding noise variable will choose a smaller value. Another way is to determine it based on experience or predefined rules. For example, based on the understanding of the operating mechanism of the evaporative cooling device, for some key data items, a specific noise variable range is set to ensure that the result after noise addition is still within a reasonable physical range.

[0182] In step S212, after obtaining the linear transformation array, the computer system performs a linear transformation on the monitoring data cluster based on the array, thereby obtaining a noisy monitoring data cluster.

[0183] Continuing with the cooling efficiency monitoring data in the evaporative cooling device as an example, assuming that the monitoring data cluster is [T in , F coolant , T out ], respectively representing the evaporator inlet temperature, coolant flow rate and evaporator outlet temperature. Assuming that the linear transformation array is [a1, a2, a3] (here, in order to simplify the example, it is assumed to correspond to the data items one by one), the computer system performs a linear transformation on each data item.

[0184] For the evaporator inlet temperature T in , assuming its original value is t in , the value after linear transformation T in ' can be calculated as follows (assuming a simple linear relationship): For example, if (Indicates a 10% increase in fluctuation based on the original value), then T_{in}'=1.1×25=27.5°C. For the coolant flow rate F coolant , let its original value be f coolant , the value after linear transformation For example, if (meaning a 10% reduction in fluctuations based on the original value), then F coolant '=0.9×10=9L / min. For the evaporator outlet temperature T out , let its original value be t out , the value after linear transformation By performing such a linear transformation operation on each data item in the monitoring data cluster, the computer system obtains a noisy monitoring data cluster, such as .

[0185] This linear transformation operation based on the linear transformation array enables the computer system to add noise to the data items in the monitoring data cluster in a predetermined manner. It is not just a simple addition of random noise to the original data, but a targeted transformation of the spatial distribution of the data items based on the characteristics of the data items and the predetermined noise adding variables. The noise added monitoring data cluster obtained in this way can better simulate various data changes that may occur in actual operation, and provide more representative data for the subsequent construction of abnormal monitoring data clusters and the debugging of status monitoring networks.

[0186] In addition, when performing linear transformations, the computer system also needs to consider the boundary conditions of the data. For example, for temperature data items, their values ​​have physical limits and cannot be lower than absolute zero or higher than the maximum temperature that the device can withstand. Therefore, when calculating the value after adding noise, if it exceeds this range, the computer system needs to make adjustments, such as setting the value that exceeds the upper limit to the upper limit value, and setting the value that is lower than the lower limit to the lower limit value.

[0187] As an implementation manner, step S30, debugging the state monitoring network according to the monitoring time series data set sample to obtain the debugged state monitoring network, may include:

[0188] Step S31: obtaining a first characterization vector set of the monitoring time series data set samples based on the state monitoring network, where the first characterization vector set is used to indicate the evolution information of the corresponding characterization information of the target observation element in the monitoring time series data set samples at the state level;

[0189] Step S32: obtaining a second characterization vector set based on the state monitoring network according to the first characterization vector set, wherein the second characterization vector set is used to indicate the evolution information of the characterization information corresponding to the target observation element in the monitoring time series data set sample at the time series level;

[0190] Step S33: obtaining a state monitoring result of the monitoring time series data set sample based on the state monitoring network according to the second characterization vector set, wherein the state monitoring result of the monitoring time series data set sample is used to indicate whether the time series evolution of the characterization vector associated with the core data item of the target observation element in the monitoring time series data set sample meets the standard;

[0191] Step S34: Debugging the state monitoring network based on the state monitoring result to obtain the debugged state monitoring network.

[0192] In step S31, the state monitoring network is a key part for analyzing the data in the monitoring time series data set sample. The computer system first obtains a first characterization vector set based on the state monitoring network.

[0193] The monitoring time series data set sample is constructed through the previous steps (such as noise addition, integration, etc.), and it contains various data information reflecting the operating status of the evaporative cooling device. The target observation elements are factors that have key characterization significance for the operating status of the evaporative cooling device, such as cooling efficiency, equipment power consumption, humidity regulation capacity, etc. Taking cooling efficiency as an example, the monitoring time series data set sample may contain data related to cooling efficiency at different time points, such as the changes in evaporator inlet and outlet temperatures, coolant flow rate, and other data over time. The computer system inputs these data related to cooling efficiency into the state monitoring network.

[0194] The condition monitoring network can adopt a neural network structure, such as a combination of a convolutional neural network (CNN) and a recurrent neural network (RNN). First, the CNN part extracts features from the input data. Assuming that part of the data in the input monitoring time series data set sample is in the form of a two-dimensional matrix (for example, the evaporator inlet and outlet temperature data at different time points are arranged in a matrix form), the convolution layer of the CNN performs a convolution operation by sliding the convolution kernel on this matrix. For example, for a 3×3 convolution kernel K and an input matrix I, the calculation of the convolution operation at a certain position (i, j) can be expressed as: ;

[0195] Among them, C(i,j) is the element in the convolution result matrix. Through the convolution operation, CNN can automatically extract local features in the data, such as the fluctuation pattern of the inlet and outlet temperatures of the evaporator in a short period of time. The pooling layer downsamples the features extracted by the convolution layer to reduce the dimension of the data. For example, by using the maximum pooling operation, the maximum value in the local area is selected as the result after pooling.

[0196] The feature data processed by CNN then enters the RNN part. RNN has the ability to process time series data, and it can process the features extracted by CNN in chronological order. In RNN, the hidden state of each time step will be updated according to the current input (i.e., the features processed by CNN) and the hidden state of the previous time step. Through the RNN processing of the entire monitoring time series data set samples, the computer system obtains a series of hidden state vectors, which constitute the first characterization vector set. Each vector in this set reflects the comprehensive characteristics of the target observation element (such as cooling efficiency) at a certain point in time at the state level, such as a comprehensive state representation of the cooling efficiency under the combined action of factors such as the evaporator inlet and outlet temperature and the coolant flow rate at a certain moment. The evolution information of these vectors reflects the changes in the cooling efficiency at the state level over time, such as whether it is stable, whether there are periodic state changes, etc.

[0197] In step S32, after obtaining the first characterization vector set, the computer system continues to process it based on the state monitoring network to obtain a second characterization vector set. The second characterization vector set focuses on the evolution information of the target observation element at the time series level.

[0198] Still taking the cooling efficiency of the evaporative cooling device as an example, the vectors in the first characterization vector set have reflected the characteristics of the cooling efficiency at the state level to a certain extent. In order to further explore its evolution law in time series, the state monitoring network adopts a specific structure and algorithm for processing.

[0199] If the state monitoring network uses a long short-term memory network (LSTM) to process the first set of representation vectors, LSTM is a special RNN that processes temporal information through a special gating mechanism. For each vector in the first set of representation vectors, the input gate, forget gate, and output gate in the LSTM work together.

[0200] By using LSTM to process the vectors in the first characterization vector set one by one, the computer system obtains the second characterization vector set. The vectors in this set can better reflect the evolution information of the cooling efficiency at the time series level. For example, it can capture the time series characteristics such as the rising or falling trend of the cooling efficiency over time and the rate of change in different time periods. This kind of evolution information at the time series level is very critical for judging whether the operating status of the evaporative cooling device is normal. For example, if the cooling efficiency continues to decrease over a period of time, it may indicate that there are some problems with the device, such as coolant leakage or evaporator scaling.

[0201] In step S33, after obtaining the second characterization vector set, the computer system further obtains the state monitoring results of the monitoring time series data set samples based on the state monitoring network. This state monitoring result mainly focuses on whether the time series evolution of the characterization vector associated with the core data item of the target observation element meets the standard. Taking the equipment power consumption of the evaporative cooling device as the target observation element, the core data items may include the power values ​​of the main power-consuming devices (such as compressors, fans, etc.). In the second characterization vector set, the characterization vectors associated with these core data items contain the evolution information of the equipment power consumption at the time series level. The state monitoring network can determine whether the time series evolution of these characterization vectors meets the standard by setting a classifier. For example, a support vector machine (SVM) is used as a classifier. The basic idea of ​​​​SVM is to find an optimal hyperplane in the feature space to separate data of different categories.

[0202] Assume that the temporal evolution of the representation vector that meets the standard is marked as the positive class, and the one that does not meet the standard is marked as the negative class. For each vector in the second representation vector set, SVM determines its category by calculating the distance between the vector and the hyperplane. Let the vector be , the hyperplane equation is (in is the normal vector of the hyperplane, b is the offset), then the distance from the vector to the hyperplane is: ;

[0203] If d is greater than a certain threshold (this threshold can be determined by training data), the vector is classified as a positive class, that is, it is considered that the time series evolution of the representation vector associated with the core data item of the target observation element in the sample of the monitoring time series data set meets the standard; otherwise, it is classified as a negative class, indicating that it does not meet the standard. For example, for the monitoring of equipment power consumption, if it is operating normally, the power of the compressor and fan should remain relatively stable in time series or fluctuate within a certain range. Through the classification of the second set of representation vectors by SVM, if the representation vector within a certain period of time is classified as a negative class, it may mean that the power consumption of the equipment has abnormally fluctuated, such as a sudden increase in the power consumption of the compressor or an abnormal decrease in the power consumption of the fan, which may indicate that the equipment is faulty or the operating environment has changed.

[0204] As an implementation mode, the state monitoring network includes a timing characterization component and a state characterization component, the timing characterization component is composed of a first pre-debugging component and a first fine-tuning component, the state characterization component is composed of a second pre-debugging component and a second fine-tuning component, the first fine-tuning component is used to obtain a first characterization vector set based on the first pre-debugging component, and the second fine-tuning component is used to obtain a second characterization vector set based on the second pre-debugging component;

[0205] Step S34, debugging the state monitoring network based on the state monitoring result to obtain the debugged state monitoring network, including:

[0206] Step S341: determining a debugging error value of the condition monitoring network based on the condition monitoring result;

[0207] Step S342: maintaining the component parameters of the first pre-debugging component and the component parameters of the second pre-debugging component unchanged, updating the component parameters of the first fine-tuning component and the component parameters of the second fine-tuning component based on the debugging error value, and obtaining a debugged state monitoring network.

[0208] There is a certain difference between the status monitoring result and the actual correct result (this correct result is obtained based on the statistics and analysis of a large amount of data of the normally operating evaporative cooling device, or determined by manual marking, etc.), and this difference is the debugging error value. For example, suppose that in the monitoring of equipment power consumption, the status monitoring results obtained through the previous steps show that 80% of the data are correctly classified (that is, judging whether the timing evolution meets the standard correctly), while the actual correct proportion should be 90%. Then the debugging error value can be expressed by some metric, such as the mean square error (MSE). Suppose the prediction result is (Classification results obtained through the condition monitoring network, 0 or 1 means non-compliance or compliance with the standard), the actual result is , the sample size is n, then the mean square error formula is: ; In this example, if the MSE value obtained according to actual calculation is 0.1 (this is just an example value), this value reflects the current debugging error value of the condition monitoring network, which indicates the degree of deviation between the network's predicted results and the actual results.

[0209] In the condition monitoring network, as mentioned above, it includes a timing characterization component (composed of a first pre-debugging component and a first fine-tuning component) and a condition characterization component (composed of a second pre-debugging component and a second fine-tuning component).

[0210] The component parameters of the first pre-tuning component and the second pre-tuning component remain unchanged in this step. This is to avoid over-adjusting the determined basic network structure and initialization settings and maintain the stability of the network. Taking the first fine-tuning component as an example, if it is a fully connected layer in a neural network, its component parameters include the weight matrix W and the bias vector . The computer system updates these parameters according to the debugging error value. A feasible updating method is to use the gradient descent algorithm. After multiple such update iterations, the state monitoring network is continuously optimized, and finally a debugged state monitoring network is obtained. This debugged network can more accurately analyze the monitoring time series data set of the evaporative cooling device, and more accurately determine whether the time series evolution of the representation vector associated with the core data item of the target observation element meets the standard, thereby improving the accuracy and reliability of the state monitoring of the evaporative cooling device.

[0211] The embodiment of the present invention also provides a computer system, such as Figure 2 As shown, the computer 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 computer 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 computer system 100 does not constitute a limitation on the embodiments of the present invention.

[0212] An embodiment of the present invention provides a computer system. The computer system in the embodiment of the present invention includes: one or more processors; a memory; and 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 method is implemented.

Claims

1. A control method for an evaporative cooling device, characterized in that: The method comprises: Acquire one or more monitoring time series data sets of a target evaporative cooling device, wherein the monitoring time series data sets include a plurality of monitoring data clusters associated with the target observation elements and arranged in time series; The state monitoring result of the monitoring time series data set is obtained based on the state monitoring network, and the state monitoring result of the monitoring time series data set is used to indicate whether the time series evolution of the representation vector associated with the core data item of the target observation element in the monitoring time series data set meets the standard. The state monitoring network is debugged based on the monitoring time series data set sample. For u monitoring data clusters in the monitoring time series data set sample, the core data item of the target observation element is noised, and the time series evolution of the representation vector associated with the core data item of the target observation element in the monitoring time series data set sample meets the standard, u≥1; Based on the status monitoring results respectively corresponding to the one or more monitoring time series data sets, obtaining the status monitoring result of the target evaporative cooling device, wherein the status monitoring result is used to indicate whether the target evaporative cooling device meets the standard; Based on the status monitoring result of the target evaporative cooling device, obtaining a control strategy of the target evaporative cooling device; Wherein, the obtaining of the state monitoring result of the monitoring time series data set based on the state monitoring network includes: Acquire a first characterization vector set of the monitoring time series data set based on the condition monitoring network, wherein the first characterization vector set is used to indicate evolution information of the characterization vector associated with the core data item of the target observation element in the monitoring time series data set at the state level; Based on the condition monitoring network, according to the first characterization vector set, a second characterization vector set is obtained, wherein the second characterization vector set is used to indicate evolution information of the characterization vector associated with the core data item of the target observation element in the monitoring time series data set at the time series level; Based on the condition monitoring network and according to the second characterization vector set, a condition monitoring result of the monitoring time series data set is obtained.

2. The method according to claim 1, characterized in that The obtaining a second characterization vector set based on the condition monitoring network according to the first characterization vector set includes: Performing a dimension transformation on the first characterization vector set to obtain a transformed first characterization vector set; The second characterization vector set is obtained based on the state monitoring network according to the transformed first characterization vector set.

3. The method according to claim 1 or 2, characterized in that The state monitoring network includes a timing characterization component and a state characterization component, the timing characterization component is composed of a first pre-debugging component and a first fine-tuning component, the state characterization component is composed of a second pre-debugging component and a second fine-tuning component, the first fine-tuning component is used to obtain the first characterization vector set based on the first pre-debugging component, and the second fine-tuning component is used to obtain the second characterization vector set based on the second pre-debugging component.

4. The method according to claim 1, characterized in that The debugging process of the condition monitoring network includes: Acquire a monitoring time series data set, wherein the monitoring time series data set includes a plurality of monitoring data clusters associated with target observation elements arranged in time series; For u monitoring data clusters in the monitoring time series data set, the core data items of the target observation elements in the monitoring data clusters are denoised to obtain monitoring time series data set samples, and the time series evolution of the representation vector associated with the core data items of the target observation elements in the monitoring time series data set samples meets the standard, u≥1; Debugging a state monitoring network according to the monitoring time series data set sample to obtain a debugged state monitoring network, wherein the debugged state monitoring network is used to monitor whether a target evaporative cooling device meets the standard; The step of adding noise to the core data items of the target observation elements in the monitoring data cluster to obtain a monitoring time series data set sample includes: Noising each data item in the monitoring data cluster to obtain a noisy monitoring data cluster, wherein the data items in the noisy monitoring data cluster have errors with the data items in the monitoring data cluster; Integrate the noise-added monitoring data cluster and the monitoring data cluster to obtain an abnormal monitoring data cluster, wherein there is an error between the core data items of the target observation element in the abnormal monitoring data cluster and the core data items of the target observation element in the monitoring data cluster, and there is no error between the common data items of the target observation element in the abnormal monitoring data cluster and the common data items of the target observation element in the monitoring data cluster; Changing the u monitoring data clusters into abnormal monitoring data clusters corresponding to the u monitoring data clusters respectively, to obtain the monitoring time series data set samples; The step of integrating the noise-added monitoring data cluster and the monitoring data cluster to obtain an abnormal monitoring data cluster includes: Acquire multiple core data item subsets corresponding to the target observation element in the monitoring data cluster, each of the core data item subsets including multiple core data items; For any one of the plurality of core data item subsets, generating a shielded data set corresponding to the monitoring data cluster according to the core data item subset, wherein the shielded data set is used to indicate a position of the core data item subset in the monitoring data cluster; Under the restriction of the shielded data sets respectively corresponding to the plurality of core data item subsets, the noise-added monitoring data cluster and the monitoring data cluster are integrated to obtain the abnormal monitoring data cluster.

5. The method according to claim 4, characterized in that The step of generating a shielded data set corresponding to the monitoring data cluster based on the core data item subset includes: For any data item in the monitoring data cluster, obtaining a first spatial interval between the data item and a location of the core data item subset; determining, based on the first spatial interval, a masking value for the data item, the masking value for the data item being inversely related to the first spatial interval; According to the masking value of each data item in the monitoring data cluster, the masking data set is formed; The step of integrating the noise-added monitoring data cluster and the monitoring data cluster under the restriction of the shielded data sets respectively corresponding to the plurality of core data item subsets to obtain the abnormal monitoring data cluster comprises: For any one of the shielded data sets corresponding to the plurality of core data item subsets, the noise-added monitoring data cluster is updated according to the shielded data set to obtain a first integrated monitoring data cluster, and the monitoring data cluster is updated according to an inverse shielded data set corresponding to the shielded data set to obtain a second integrated monitoring data cluster, wherein the sum of the shielded values ​​of the data items at the same position in the shielded data set and the inverse shielded data set is equal to 1; Obtaining an integrated monitoring data cluster corresponding to the shielded data set according to the first integrated monitoring data cluster and the second integrated monitoring data cluster; The integrated monitoring data clusters corresponding to the plurality of shielded data sets are integrated to obtain the abnormal monitoring data cluster.

6. The method according to claim 5, characterized in that The updating of the noise-added monitoring data cluster according to the shielding data set to obtain a first integrated monitoring data cluster, and the updating of the monitoring data cluster according to the inverse shielding data set corresponding to the shielding data set to obtain a second integrated monitoring data cluster, include: For any one data item in the noisy monitoring data cluster, multiply the data item by a masking value corresponding to the data item in the masking data set to obtain the first integrated monitoring data cluster; For any data item in the monitoring data cluster, the data item is multiplied by the masking value corresponding to the data item in the inverse masking data set to obtain the second integrated monitoring data cluster.

7. The method according to claim 6, characterized in that The step of adding noise to each data item in the monitoring data cluster to obtain the noisy monitoring data cluster includes: Acquire a linear transformation array, wherein array elements in the linear transformation array are noise-added variables, and the noise-added variables are used to transform the spatial distribution of the data items; The monitoring data cluster is linearly transformed based on the linear transformation array to obtain the noisy monitoring data cluster.

8. The method according to any one of claims 4 to 7, characterized in that: The debugging of the state monitoring network according to the monitoring time series data set sample to obtain the debugged state monitoring network includes: Acquire a first characterization vector set of the monitoring time series data set sample based on the state monitoring network, wherein the first characterization vector set is used to indicate evolution information of the characterization information corresponding to the target observation element in the monitoring time series data set sample at the state level; Based on the state monitoring network and the first characterization vector set, a second characterization vector set is obtained, wherein the second characterization vector set is used to indicate the evolution information of the characterization information corresponding to the target observation element in the monitoring time series data set sample at the time series level; Based on the condition monitoring network and according to the second characterization vector set, a condition monitoring result of the monitoring time series data set sample is obtained, wherein the condition monitoring result of the monitoring time series data set sample is used to indicate whether the time series evolution of the characterization vector associated with the core data item of the target observation element in the monitoring time series data set sample meets the standard; The state monitoring network is debugged based on the state monitoring result to obtain the debugged state monitoring network.

9. The method according to claim 8, characterized in that The state monitoring network includes a timing characterization component and a state characterization component, the timing characterization component is composed of a first pre-debugging component and a first fine-tuning component, the state characterization component is composed of a second pre-debugging component and a second fine-tuning component, the first fine-tuning component is used to obtain the first characterization vector set based on the first pre-debugging component, and the second fine-tuning component is used to obtain the second characterization vector set based on the second pre-debugging component; The debugging of the state monitoring network based on the state monitoring result to obtain the debugged state monitoring network includes: Based on the condition monitoring result, determining a debugging error value of the condition monitoring network; Maintaining component parameters of the first pre-debugging component and component parameters of the second pre-debugging component unchanged, updating component parameters of the first fine-tuning component and component parameters of the second fine-tuning component based on the debugging error value, and obtaining the debugged state monitoring network.

10. A computer system, characterized in that: include: one or more processors; Memory; one or more computer programs; 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