Energy-saving control method and system for data center computer room

Abstract

The present invention relates to the field of data processing, and in particular provides a data center computer room energy-saving control method and system, the data center computer room energy-saving control method comprising the following steps: using an implicit representation model to perform implicit representation mining on a monitoring data stream template, obtaining a template data implicit representation array, obtaining a key data detection result based on the template data implicit representation array and a key data detection model, obtaining a linkage type detection result based on the key data detection result and a linkage type identification model, iterating the template data implicit representation array based on each result and model, and performing loop optimization until a preset judgment standard is met, obtaining a target implicit representation model, a target key data detection model, and a target linkage type identification model. The present invention has high detection accuracy for key data and linkage types in the computer room monitoring data stream, is convenient for formulating an accurate computer room energy-saving control strategy, and helps save computer room energy consumption.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to an energy-saving control method and system for a data center computer room. Background Art

[0002] With the rapid development of information technology, data centers, as core facilities for information processing and storage, have become increasingly energy-consuming. A large number of servers, storage devices, cooling systems and other equipment are integrated in the data center room. The operation of these devices consumes a lot of electricity, which not only increases operating costs but also affects the environment. Therefore, how to effectively control the energy consumption of the data center room has become a technical problem that needs to be solved urgently. Traditional energy-saving control methods for data center rooms rely on manual experience and simple rule judgments, such as setting a fixed cooling strategy based on the room temperature and adjusting power management based on server load. Although these methods can reduce energy consumption to a certain extent, they often fail to achieve the best energy-saving effect due to the lack of in-depth understanding of the complex relationship between data. In recent years, machine learning technology has received widespread attention in the field of energy-saving control of data center rooms. Machine learning models can automatically identify and predict key factors affecting energy consumption and their linkage relationships by learning patterns and rules in a large amount of historical data, thereby providing a scientific basis for the formulation of energy-saving control strategies. However, the application of machine learning technology to energy-saving control of data center rooms still faces many challenges. Most of the existing technical solutions have problems such as difficulty in capturing the complex features and structures in monitoring data streams, limited model adaptability, and lack of generalization capabilities for different types of monitoring data streams and energy-saving control scenarios. Summary of the invention

[0003] The object of the present invention is to provide a data center computer room energy-saving control method and system, and the embodiments of the present invention are implemented as follows:

[0004] In a first aspect, the present invention provides an energy-saving control method for a computer room in a data center, the method comprising: obtaining a computer room monitoring data stream, performing implicit representation mining on the computer room monitoring data stream using a pre-optimized target implicit representation model, and obtaining a data representation array corresponding to the computer room monitoring data stream; detecting key data of the computer room monitoring data stream based on the data representation array and a pre-optimized target key data detection model to obtain a key data detection result; detecting a data linkage type of the computer room monitoring data stream based on the key data detection result and a pre-optimized target linkage type recognition model to obtain a linkage type detection result; and formulating an energy-saving control strategy based on the key data detection result and the linkage type detection result.

[0005] Among them, the target implicit representation model, the target key data detection model and the target linkage type identification model are optimized through the following steps: obtaining a monitoring data flow template and a template tag corresponding to the monitoring data flow template; the template tag includes a key data tag corresponding to the monitoring data flow template and a linkage type tag corresponding to the monitoring data flow template; using an implicit representation model to perform implicit representation mining on the monitoring data flow template to obtain an implicit representation array of template data corresponding to the monitoring data flow template; based on the template data implicit representation array and the key data detection model, detecting the key data of the monitoring data flow template to obtain a key data detection result; based on the key data detection result and the linkage type identification model, detecting the data linkage type of the monitoring data flow template to obtain a linkage type detection result; determining a first optimization error based on the difference between the key data detection result and the key data tag, determining a second optimization error based on the difference between the linkage type detection result and the linkage type tag, determining a first directional derivative based on the first optimization error, and determining a second directional derivative based on the second optimization error. Determine a second directional derivative, sum the first directional derivative and the second directional derivative to obtain a combined directional derivative, optimize the implicit representation model, the key data detection model and the linkage type recognition model based on the combined directional derivative, and iterate the implicit representation model, the key data detection model and the linkage type recognition model; input the monitoring data flow template into the iterated implicit representation model, use the iterated implicit representation model to perform implicit representation mining on the monitoring data flow template to obtain an implicit representation array, and merge the linkage type detection result with the implicit representation array to iterate the template data implicit representation array; return to the process of detecting the key data of the monitoring data flow template based on the template data implicit representation array and the key data detection model for cyclic optimization until the preset judgment criteria are met, and obtain the optimized target implicit representation model, target key data detection model and target linkage type recognition model, the target implicit representation model, target key data detection model and target linkage type recognition model are used to collaboratively detect the data linkage type of the computer room monitoring data flow.

[0006] 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.

[0007] The beneficial effects of the present invention are as follows: the energy-saving control method and system for a data center computer room provided by the present invention, after acquiring the computer room monitoring data stream, uses a pre-optimized target implicit representation model to perform implicit representation mining on the computer room monitoring data stream, obtains a data representation array corresponding to the computer room monitoring data stream, detects key data of the computer room monitoring data stream based on the data representation array and a pre-optimized target key data detection model, obtains key data detection results, detects data linkage types of the computer room monitoring data stream based on the key data detection results and a pre-optimized target linkage type recognition model, obtains linkage type detection results, and performs energy-saving control strategies based on the key data detection results and the linkage type detection results. Formulate, wherein the target implicit representation model, the target key data detection model and the target linkage type recognition model are optimized through the monitoring data flow template. In the process, one round of optimization is based on the linkage type detection result output by the linkage type recognition model after the previous round of optimization, and it is iteratively input into the template data implicit representation array in the key data detection model obtained in the previous round of optimization. The involvement of the key data detection model and the linkage type recognition model is measured. The key data detection model and the target linkage type recognition model obtained have a wide range of adaptability, and have high detection accuracy for key data and linkage types in the computer room monitoring data flow, which is convenient for formulating accurate computer room energy-saving control strategies to help save computer room energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for describing the embodiments of the present invention are briefly introduced below.

[0009] Figure 1 The present invention provides a flowchart of a data center computer room energy saving control method.

[0010] Figure 2 It is a schematic diagram of a model optimization process provided by an embodiment of the present invention.

[0011] Figure 3 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The execution subject of the energy-saving control method for a data center computer room in the embodiment of the present invention is a computer system, including but not limited to a server, a personal computer, a laptop computer, a tablet computer, a smart phone, etc. The computer system can be operated alone to implement the present invention, or it can be connected to a network and implement the present invention through interactive operations with other computer systems in the network. The network in which the computer system is located includes but is not limited to the Internet, a wide area network, a metropolitan area network, a local area network, a VPN network, etc.

[0013] The embodiment of the present invention provides a data center computer room energy saving control method, such as Figure 1 As shown, the method includes:

[0014] Step S110: Acquire the computer room monitoring data stream, use the pre-optimized target implicit representation model to perform implicit representation mining on the computer room monitoring data stream, and obtain a data representation array corresponding to the computer room monitoring data stream.

[0015] For example, the monitoring data stream of the computer room includes various types of data, such as temperature, humidity, power consumption, server load, air conditioning system status, etc. These data are collected in real time by various sensors and devices deployed in the computer room. These data are continuously generated in the form of time series, forming a continuous monitoring data stream. The role of the target implicit representation model is to convert the complex monitoring data stream into a more compact and easy-to-process data representation array. Implicit representation models are built based on deep learning technologies, such as autoencoders or variational autoencoders (VAE).

[0016] In step S110, the computer system inputs the acquired computer room monitoring data stream into the pre-trained target implicit representation model. The encoder part of the model processes the input data stream and generates a corresponding implicit representation, namely a data representation array. This data representation array is a low-dimensional vector or matrix that captures the key features and structural information in the original data stream. The data representation array is the result of implicit representation mining. It contains the key information of the original monitoring data stream, but is presented in a more compact and efficient form. Each element in this array can be regarded as an implicit feature of the original data stream.

[0017] Step S120: Detect key data of the computer room monitoring data stream based on the data representation array and the pre-optimized target key data detection model to obtain key data detection results.

[0018] In step S110, the computer system has converted the original computer room monitoring data stream into a data representation array through the target implicit representation model. The target key data detection model identifies key data items from the data representation array. This model is built based on deep learning or traditional machine learning algorithms, such as support vector machines (SVM), random forests, or neural networks. The specific choice of the model depends on the characteristics of the data and the requirements of the task.

[0019] Taking a neural network as an example, the target key data detection model may include an input layer, multiple hidden layers and an output layer. The input layer receives a data representation array as input, the hidden layer extracts the deep features of the data through nonlinear transformation, and the output layer outputs the probability or score of each data item being identified as key data. The training process of the model, for example, involves using labeled data (i.e., it is known which data items are key data) to minimize prediction errors (such as cross entropy loss). In step S120, the computer system inputs the data representation array into the target key data detection model, and the model evaluates each data item and outputs the probability or score of them being identified as key data. This evaluation process may involve complex calculations, such as matrix multiplication, the application of nonlinear activation functions, etc.

[0020] For example, suppose the data representation array is a 100×5 matrix, where 100 rows represent observations at 100 time points and 5 columns represent 5 implicit features. The target key data detection model may be a neural network with two hidden layers, each containing 10 neurons. When the data representation array is input into the model, the input layer first passes the data to the first hidden layer, and each neuron calculates an output value through weighted summation and activation function (such as ReLU). This process is repeated on each neuron in the first hidden layer, generating a new matrix as the output of the first hidden layer. Then, this output matrix is ​​passed to the second hidden layer, and the same calculation process is repeated. Finally, each neuron in the output layer calculates the probability or score of each data item being identified as key data based on the output of the second hidden layer.

[0021] The key data detection result is the evaluation result of the target key data detection model for each data item, which is presented in the form of probability, score or label. In practical applications, key data may include abnormal temperature, humidity fluctuation, power consumption peak, etc. in the computer room, which have an important impact on the energy consumption and stability of the computer room. By identifying these key data, the computer system can further analyze the relationship between them and the energy consumption of the computer room and formulate corresponding energy-saving control strategies.

[0022] Step S130: Detect the data linkage type of the computer room monitoring data flow based on the key data detection result and the pre-optimized target linkage type recognition model to obtain the linkage type detection result.

[0023] In step S120, the computer system has identified key data items from the computer room monitoring data stream through the target key data detection model. The target linkage type recognition model is an optimized and trained machine learning model that identifies the linkage type between data items from the key data detection results. The linkage type refers to the relationship and influence pattern between different data items in the time series, such as causal relationship, correlation or synchronous change. This model may be built based on deep learning, time series analysis or other machine learning techniques, and the specific choice depends on the characteristics of the data and the requirements of the task. Taking deep learning as an example, the target linkage type recognition model may be a recurrent neural network (RNN) or its variants, such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU).

[0024] In step S130, the computer system determines multiple estimated key data items in the computer room monitoring data stream based on the key data detection results. These estimated key data items are data items with higher probabilities or scores identified in step S120. Then, the system combines these estimated key data items to form multiple estimated key data groups. Each data group contains two or more estimated key data items for analyzing the linkage relationship between them. Next, the system extracts data representation sequences corresponding to each estimated key data group from the data representation array. These sequences are data in the time period corresponding to the estimated key data items in the data representation array, and they contain linkage information between data items. Then, each data representation sequence is input into the target linkage type recognition model, the input data representation sequence is processed, and the time dependency and mutual relationship between different data items in the sequence are analyzed, so as to identify the linkage type between the data items.

[0025] Taking LSTM as an example, when the model processes each data representation sequence, it will split it into multiple time steps, each time step corresponding to a time point in the sequence. In each time step, the model reads the data representation of the current time point, and combines the hidden state and memory cell state of the previous time step to calculate the hidden state and output of the current time step. By iterating this process, the model can capture the time dependency in the sequence and the linkage relationship between data items. The linkage type detection result is the evaluation result of the target linkage type recognition model for each estimated key data group, which reveals the linkage type and mutual influence pattern between data items. This result is presented in the form of a label, category, or probability distribution, for example, indicating the possibility or certainty that each estimated key data group is identified as a specific linkage type.

[0026] Step S140: Formulate an energy-saving control strategy based on the key data detection results and the linkage type detection results.

[0027] In step S140, the computer system formulates an energy-saving control strategy based on the key data detection results and the linkage type detection results. This process involves in-depth interpretation and reasoning of the data analysis results to determine which measures can effectively reduce the energy consumption of the computer room.

[0028] The computer system analyzes the key data detection results to determine which data items are key drivers of energy consumption. For example, in the example above, the high load of the server is identified as a key data item, which may be the main reason for the increased energy consumption in the computer room. Next, the system will analyze the interaction and influence patterns between these key data items in combination with the linkage type detection results. For example, the system can find a positive correlation between server load and computer room temperature, which means that reducing server load may help reduce computer room temperature and thus reduce the energy consumption of the air conditioning system. Based on these analyses, the computer system can develop a series of energy-saving control strategies. These strategies may include adjusting the operating status of the air conditioning system, adjusting server load distribution, optimizing power management, and so on.

[0029] Please refer to Figure 2 The present invention provides an optimization process of a target implicit representation model, a target key data detection model, and a target linkage type recognition model, which is optimized by the following steps:

[0030] Step S11: Acquire a monitoring data flow template and a template tag corresponding to the monitoring data flow template; the template tag includes a key data tag corresponding to the monitoring data flow template and a linkage type tag corresponding to the monitoring data flow template.

[0031] Monitoring data flow templates are a part of data extracted from the actual computer room monitoring data. These data are representative and can reflect the typical characteristics and patterns of computer room monitoring data. These data flow templates contain monitoring data at multiple time points, and the data at each time point may include measurement values ​​of multiple dimensions, such as temperature, humidity, power consumption, server load, etc. For example, a monitoring data flow template may contain temperature, humidity, and power consumption data for a certain area in the computer room for 24 consecutive hours.

[0032] Template tags are labels or metadata corresponding to monitoring data flow templates, which provide information about which data items in the template are key data and how these data items are linked. Template tags include key data tags and linkage type tags. Key data tags indicate which data items in the monitoring data flow template have a significant impact on the energy consumption of the computer room, and are an important basis for the formulation of energy-saving control strategies. These tags are determined, for example, by the knowledge and experience of domain experts, or they can be automatically extracted through data analysis methods. The key data tag can be a binary vector, each element of the vector corresponds to a data item in the monitoring data flow template. If a data item is marked as key data, the element at the corresponding position is 1, otherwise it is 0.

[0033] The linkage type tag describes the linkage relationship between different data items in the monitoring data flow template. The linkage relationship refers to the mutual dependence and influence pattern between data items in the time series, such as causal relationship, correlation or synchronous change. These tags help machine learning models understand the intrinsic connection between data items, thereby improving the model's ability to identify data linkage types. The linkage type tag can be a matrix or graph structure that represents the linkage relationship between data items. For example, if there is a positive correlation between two data items, the element at the corresponding position in the linkage type tag matrix can be a positive number; if there is a negative correlation, the element at the corresponding position can be a negative number; if there is no significant linkage relationship, the element at the corresponding position can be 0.

[0034] In step S11, the computer system extracts a portion of data from the computer room monitoring data set as monitoring data flow templates, and generates corresponding template tags for these templates.

[0035] For example, suppose there is a flow template that contains monitoring data in two dimensions: room temperature and server load. Through analysis by domain experts, it is determined that both temperature and load are key data items, and there is a positive correlation between them.

[0036] Step S12: Use the implicit representation model to perform implicit representation mining on the monitoring data flow template to obtain an implicit representation array of template data corresponding to the monitoring data flow template.

[0037] An implicit representation model is a machine learning model that can map raw data to a low-dimensional latent space to generate an implicit representation of the data. This implicit representation can, for example, capture key features and patterns in the raw data while removing redundant information, making the data more compact and easier to process. Implicit representation models are widely used in a variety of tasks, such as dimensionality reduction, feature extraction, data compression, etc. In an embodiment of the present invention, an implicit representation model is used to convert a monitoring data stream template into an implicit representation array of template data.

[0038] The implicit representation mining process includes, for example, that before inputting the monitoring data flow template into the implicit representation model, the computer system may pre-process the data, such as data cleaning, denoising, normalization, etc., to ensure the quality and consistency of the data. The implicit representation model needs to be initialized before training, including setting the parameters and structure of the model. The selection of these parameters and structures depends on the characteristics of the data and the requirements of the task. After the model is initialized, the computer system needs to input the monitoring data flow template into the model for training. During the training process, the model learns how to map the original data to the latent space and generate an implicit representation. This process may involve multiple iterations and optimizations to find the optimal model parameters and structure. After the model training is completed, the computer system can use the trained implicit representation model to perform implicit representation mining on the monitoring data flow template. This process converts the original data flow template into a low-dimensional template data implicit representation array, which contains the key features and patterns of the data flow template.

[0039] For example, suppose there is a stream template containing two-dimensional monitoring data of room temperature and server load, and an implicit representation model has been trained. This model may be an autoencoder, which contains an encoder and a decoder. The encoder is responsible for mapping the input data to the latent space and generating an implicit representation; the decoder is responsible for reconstructing the original data from the implicit representation.

[0040] In step S12, the monitoring data stream template is input into the encoder of the autoencoder. The encoder processes the input data to generate a low-dimensional implicit representation array. This array may be a two-dimensional vector, in which each element represents an implicit feature in the original data stream template. For example, the first element may represent the average value or change trend of the temperature in the computer room, and the second element may represent the average value or change trend of the server load. Through this process, the computer system can convert a complex monitoring data stream template into a concise and informative template data implicit representation array, which provides strong support for subsequent key data detection and linkage type identification.

[0041] Taking the autoencoder as an example of an implicit representation model, the encoder part may contain multiple fully connected layers or convolutional layers to extract deep features of the data; the decoder part may contain a structure symmetrical to the encoder to reconstruct the original data from the implicit representation. The loss function, such as mean square error (MSE) or cross-entropy loss, is used to measure the difference between the reconstructed data and the original data. The optimization algorithm may choose stochastic gradient descent (SGD), Adam, etc. to update the parameters of the model to minimize the loss function.

[0042] Step S13: Detect key data of the monitoring data flow template based on the template data implicit representation array and the key data detection model to obtain key data detection results.

[0043] A key data detection model is a machine learning model that is used to identify data items from data that have a significant impact on a specific task or goal. In the context of energy-saving control of a data center computer room, the task of the key data detection model is to identify those data items that have a significant impact on the energy consumption of the computer room, such as the load of the server, the temperature of the computer room, etc.

[0044] The specific implementation of the key data detection model varies depending on the task requirements and data characteristics. For example, it may be a rule-based system that uses the knowledge of domain experts to define the criteria for key data; it may also be a statistical learning-based method that learns the characteristics of key data through training data; or it may be a deep learning model, such as a neural network, that identifies key data by automatically learning complex representations of data.

[0045] The template data implicit representation array is a low-dimensional vector or matrix that contains the key features and patterns of the monitoring data flow template, but in a more compact and easy-to-process form. For example, if the monitoring data flow template contains temperature and humidity data at multiple time points, then the template data implicit representation array may be a two-dimensional matrix, where each row represents the implicit representation of a time point and each column represents a specific implicit feature. In step S13, the computer system inputs the template data implicit representation array into the key data detection model, and the model evaluates each element (or vector) in the array to determine whether they represent key data. This process may involve complex calculations, such as matrix operations, nonlinear transformations, probabilistic reasoning, etc., depending on the specific implementation of the model.

[0046] Taking a neural network as an example, if the key data detection model is a multi-layer perceptron (MLP), the following steps are performed: Receive the template data implicit representation array as input. Extract the deep features of the data through a series of nonlinear transformations (such as ReLU activation function) and linear transformations (such as fully connected layers). Generate an output vector of the same length as the input array, where each element represents the probability or score of the corresponding input element being identified as key data.

[0047] The key data detection result is the output of the key data detection model after evaluating each element in the template data implicit representation array. This result is, for example, a vector or matrix of the same length as the input array, where each element represents the probability, score, or label of the corresponding input element being identified as key data. For example, if the output is a probability vector, then the value of each element will be between 0 and 1, indicating the probability of the corresponding input element being identified as key data.

[0048] Step S14: Detect the data linkage type of the monitoring data flow template based on the key data detection result and the linkage type identification model to obtain the linkage type detection result.

[0049] The linkage type recognition model is a machine learning model used to analyze the linkage relationship between data items. In an embodiment of the present invention, the task of the linkage type recognition model is to identify the correlation and mutual influence pattern between key data items, such as causal relationship, correlation or synchronous change. The linkage type recognition model is, for example, a method based on time series analysis, which identifies the linkage relationship by calculating the autocorrelation and cross-correlation functions between data items; it may also be a model based on graph neural network (GNN), which learns the linkage relationship by building a graph structure between data items; or it may be a model based on deep learning, such as recurrent neural network (RNN) or its variants (such as LSTM, GRU), which identifies the linkage relationship by capturing the time dependency and long-distance dependency between data items.

[0050] The key data detection result is the output obtained after processing the template data implicit representation array by the key data detection model in step S13. It contains an identifier indicating which data items are identified as key data, that is, those data items that have a significant impact on the energy consumption of the computer room. For example, the key data detection result may be a binary vector, where "1" indicates that the corresponding data item is key data and "0" indicates that it is not.

[0051] In step S14, the computer system determines multiple template key data in the monitoring data flow template based on the key data detection result. Then, it uses these template key data as input and calls the linkage type recognition model to detect the data linkage type.

[0052] The linkage type detection process includes, for example: extracting data fragments corresponding to the template key data from the template data implicit representation array. These data fragments contain the time series information of the key data items and are the basis for analyzing the linkage relationship. The extracted data fragments are input into the linkage type recognition model. If the model is based on time series analysis, the input may be a series of data points arranged in chronological order; if the model is based on a graph neural network, the input may be a graph structure representing the relationship between data items; if the model is based on deep learning, the input may be a preprocessed feature vector or matrix. The linkage type recognition model processes the input data to identify the linkage relationship between data items. This process may involve complex calculations, such as matrix operations, nonlinear transformations, probabilistic reasoning, etc., depending on the specific implementation of the model. After the model processing is completed, a result representing the data linkage type will be output. This result may be a label, category, or probability distribution used to describe the linkage relationship between data items.

[0053] The linkage type detection result is the final judgment of the linkage type recognition model on the data linkage relationship in the monitoring data flow template. This result may be presented in various forms, such as labels, categories, probability distributions, or graph structures, depending on the design of the model and the requirements of the task.

[0054] For example, if the linkage type recognition model is based on time series analysis and identifies a positive correlation between two key data items, it can output a label or category to represent this relationship. If the model is based on a graph neural network and builds a graph structure that represents the relationship between data items, it can output this graph structure as a result. If the model is based on deep learning and the output is a probability distribution, it can assign a probability value to each possible linkage relationship.

[0055] For example, suppose there is a flow template containing two-dimensional monitoring data of the room temperature and server load, and it has been determined through step S13 that both temperature and load are key data. Now, the linkage type recognition model is used to detect the linkage relationship between these two key data items. If the linkage type recognition model is a model based on time series analysis, it can calculate the autocorrelation and cross-correlation functions between temperature and load, and identify that there is a positive correlation between them. This means that when the temperature rises, the load may also increase; and vice versa. The model can output a label or category to represent this positive correlation.

[0056] If the linkage type recognition model is a model based on a graph neural network, it can build a graph structure representing the relationship between temperature and load, and learn the linkage relationship between them. This graph structure can show the edges between the temperature and load nodes, as well as the weights or types of the edges to represent the strength and nature of the linkage between them. If the linkage type recognition model is a model based on deep learning, such as an LSTM network, it can learn the linkage relationship between temperature and load time series by capturing the temporal dependency and long-distance dependency between them. The model can output a probability distribution to represent the probability of a positive correlation between temperature and load.

[0057] Step S15: Determine a first optimization error based on the difference between the key data detection result and the key data mark, determine a second optimization error based on the difference between the linkage type detection result and the linkage type mark, determine a first directional derivative based on the first optimization error, determine a second directional derivative based on the second optimization error, sum the first directional derivative and the second directional derivative to obtain a merged directional derivative, optimize the implicit representation model, the key data detection model and the linkage type recognition model based on the merged directional derivative, and iterate the implicit representation model, the key data detection model and the linkage type recognition model.

[0058] In step S15, the first optimization error is determined based on the difference between the key data detection result and the key data mark. The key data mark is pre-defined, indicating which data items in the monitoring data flow template are identified as key data. The key data detection result is the output obtained by the model after processing the monitoring data flow template, indicating which data items the model considers to be key data. The first optimization error measures the difference between the key data predicted by the model and the actual key data.

[0059] For example, suppose the monitoring data stream template contains 10 data items, and the key data label indicates that the 3rd, 5th, and 7th data items are key data. The key data detection result of the model may also indicate these 3 data items, but may also include the 2nd and 8th data items, or omit the 5th data item. The first optimization error will be calculated based on this difference, for example, using a loss function (such as cross entropy loss, mean square error, etc.) to measure the difference between the predicted result and the actual label.

[0060] The second optimization error is determined based on the difference between the linkage type detection result and the linkage type tag. The linkage type tag indicates the linkage relationship between different data items in the monitoring data flow template (such as positive correlation, negative correlation, no correlation, etc.). The linkage type detection result is the prediction of the linkage relationship after the model processes the monitoring data flow template. The second optimization error measures the difference between the linkage relationship predicted by the model and the actual linkage relationship.

[0061] Continuing with the example above, suppose the linkage type flag indicates that there is a positive correlation between the 3rd data item (assumed to be temperature) and the 5th data item (assumed to be load). The linkage type detection result of the model may also indicate this positive correlation, but may also incorrectly indicate that there is a correlation between the 2nd data item and the 7th data item. The second optimization error will be calculated based on this difference, also measured using an appropriate loss function.

[0062] After calculating the optimization error, the computer system needs to determine the optimization direction, that is, how to adjust the model parameters to reduce the error. This is achieved by calculating the directional derivative.

[0063] The first directional derivative is determined based on the first optimization error and indicates how to adjust the model parameters to reduce the key data detection error. This may involve, for example, taking partial derivatives of the model parameters to understand the effect of each parameter on the error. For example, in a neural network, this may involve taking partial derivatives of the weights and bias terms.

[0064] The second directional derivative is determined based on the second optimization error, which indicates how to adjust the model parameters to reduce the linkage type detection error. Similar to the first directional derivative, this is also achieved by taking partial derivatives of the model parameters.

[0065] After determining the first directional derivative and the second directional derivative, the computer system adds them together to obtain a combined directional derivative. This combined directional derivative combines the needs of reducing key data detection errors and linkage type detection errors, and provides comprehensive guidance for adjusting model parameters.

[0066] Next, the computer system iteratively updates the implicit representation model, the key data detection model, and the linkage type recognition model using the merged directional derivatives. This involves, for example, fine-tuning the model parameters along the merged directional derivatives to gradually reduce the optimization error. The iterative process may involve multiple updates until the model performance reaches a preset standard or is no longer significantly improved.

[0067] For example, suppose there is a simple neural network model that includes an input layer, a hidden layer, and an output layer. In the key data detection task, the output layer may contain neurons equal to the number of data items in the monitoring data flow template, and each neuron corresponds to the criticality prediction of a data item. In the linkage type detection task, the output layer may contain neurons representing different linkage types.

[0068] At a certain stage of model training, the key data detection results and linkage type detection results are obtained, and the first optimization error and the second optimization error are calculated. Assume that the first optimization error indicates that the model has a large error in predicting whether the fifth data item is key data, and the second optimization error indicates that the model also has an error in predicting the linkage relationship between the third data item and the fifth data item. Next, the first directional derivative and the second directional derivative are calculated, which indicate how to adjust the model parameters to reduce these two errors respectively. Then, the two directional derivatives are added to obtain the combined directional derivative, and it is used to update the weights and bias terms of the model. This process can be repeated many times until the model's performance in both key data detection and linkage type detection is improved. In practical applications, model optimization may use optimization algorithms such as stochastic gradient descent (SGD) and Adam to update model parameters. In addition, in order to prevent overfitting and improve the generalization ability of the model, regularization, dropout and other techniques may also be used.

[0069] Step S16: input the monitoring data flow template into the iterated implicit representation model, use the iterated implicit representation model to perform implicit representation mining on the monitoring data flow template to obtain an implicit representation array, and merge the linkage type detection result with the implicit representation array to iterate the template data implicit representation array.

[0070] In step S15, the computer system has optimized the implicit representation model, the key data detection model and the linkage type recognition model to obtain iterative models. The performance of these iterative models in identifying key data and linkage types has been improved because they have adjusted parameters based on previous optimization errors and reduced prediction errors.

[0071] In step S16, the computer system inputs the monitoring data flow template into the iterated implicit representation model and performs implicit representation mining again. This process is similar to step S12, but uses an optimized implicit representation model. By processing the monitoring data flow template again, the model can generate a more accurate implicit representation array that better captures the key features and patterns in the monitoring data flow template.

[0072] After the iterated implicit representation model processes the monitoring data flow template, a new implicit representation array is generated. This array is similar to the template data implicit representation array generated in step S12, but because the model has been optimized, this new array may contain richer information and more accurately reflect the key features and patterns in the monitoring data flow template.

[0073] In step S16, the computer system also merges the linkage type detection result with the implicit representation array. The linkage type detection result represents the linkage relationship between different data items in the monitoring data flow template, which is essential for understanding the interaction and influence between data items. By merging these linkage relationship information into the implicit representation array, more comprehensive data support can be provided for subsequent model analysis and decision-making.

[0074] The merging process may involve attaching the linkage type detection result to the corresponding position of the implicit representation array in some form (such as a label, probability distribution, or graph structure). For example, if the linkage type detection result is a matrix representing the positive or negative correlation between data items, then this matrix can be multiplied or added element by element with the implicit representation array to embed the linkage relationship information into the implicit representation array.

[0075] By merging the linkage type detection results and the implicit representation array, the computer system obtains a new template data implicit representation array that contains richer information. This array not only contains the key features and patterns in the monitoring data flow template, but also contains the linkage relationship information between data items. This iterative template data implicit representation array will be used in subsequent model training and analysis to improve the model's understanding of data representation and linkage relationships.

[0076] For example, suppose there is a flow template containing two-dimensional monitoring data of room temperature and server load, and the implicit representation model has been optimized through step S15. In step S16, this monitoring data flow template is input into the iterated implicit representation model to generate a new implicit representation array. This array may be a two-dimensional matrix, in which each row represents the implicit representation of a time point and each column represents a specific implicit feature.

[0077] At the same time, the linkage type detection result is also obtained, which shows that there is a positive correlation between temperature and load. During the merging process, this positive correlation information can be attached to the implicit representation array in some form. For example, a matrix of the same size as the implicit representation array can be created, in which all elements are set to 1 (indicating positive correlation), and then this matrix can be multiplied element by element with the implicit representation array. In this way, each element in the new implicit representation array will contain the original implicit eigenvalue and the linkage relationship information.

[0078] When merging the linkage type detection result and the implicit representation array, a suitable method can be selected according to the characteristics of the data and the requirements of the task. For example, if the linkage type detection result is a probability distribution, the probability information can be embedded into the implicit representation array using a weighted average method. In addition, during the iteration process, the computer system monitors the performance of the model and adjusts the number of iterations and parameters according to performance changes. For example, methods such as cross-validation can be used to evaluate the generalization ability of the model, and decide whether to continue iteration based on the evaluation results.

[0079] Step S17: Return to the process of detecting the key data of the monitoring data flow template based on the template data implicit representation array and the key data detection model for loop optimization until the preset judgment criteria are met, and obtain the optimized target implicit representation model, target key data detection model and target linkage type identification model. The target implicit representation model, target key data detection model and target linkage type identification model are used to collaboratively detect the data linkage type of the computer room monitoring data flow.

[0080] In step S17, the computer system returns to the process of detecting the key data of the monitoring data flow template based on the template data implicit representation array and the key data detection model, and starts a loop optimization process. This process will continue until the performance of the model meets the preset judgment criteria. The core idea of ​​the loop optimization is to enable the model to more accurately identify the key data and linkage types when processing new monitoring data flow templates by continuously iterating and adjusting the model parameters. Each iteration includes: inputting the monitoring data flow template into the current implicit representation model to generate the template data implicit representation array. Based on the generated template data implicit representation array, the computer system uses the key data detection model to identify the key data in the monitoring data flow template. Then, the computer system uses the linkage type recognition model to detect the linkage type between the identified key data. According to the key data detection results and the linkage type detection results, the computer system calculates the difference between the actual label (i.e., the key data label and the linkage type label), i.e., the optimization error. Based on the calculated optimization error, the computer system adjusts the parameters of the implicit representation model, the key data detection model, and the linkage type recognition model to reduce the optimization error. After each iteration, the computer system will check whether the model's performance has reached the preset judgment standard. This standard may include the accuracy of key data detection, the accuracy of linkage type recognition, etc.

[0081] The preset judgment criteria are a series of indicators used to evaluate the performance of the model. These criteria are set according to the requirements of the specific task, and may include: accuracy, recall, F1 score, loss function value, etc. When the performance of the model reaches or exceeds these preset criteria, the loop optimization process will stop, and the model obtained at this time is the optimized target implicit representation model, target key data detection model and target linkage type recognition model.

[0082] As an implementation scheme, before obtaining the monitoring data flow template and the template tag corresponding to the monitoring data flow template in step S11, the method further includes:

[0083] Step SA: split the monitoring data flow template into data items to obtain a corresponding data item sequence;

[0084] Step SB: Compare the data item sequence with the key data comparison table, and determine a plurality of first key data tags corresponding to the monitoring data flow template based on the comparison result; the key data comparison table is obtained based on an empirical key database established in advance;

[0085] Step SC: Determine the data linkage types corresponding to the plurality of first key data tags based on the empirical key database, and obtain the first linkage type tag corresponding to the monitoring data flow template.

[0086] In step SA, the computer system first splits the monitoring data flow template into data items. The monitoring data flow template is a part of the data extracted from the actual computer room monitoring data. It contains monitoring data at multiple time points. The data at each time point may include measurement values ​​of multiple dimensions, such as temperature, humidity, power consumption, server load, etc. The process of data item splitting is to split these multi-dimensional data into a series of separate data items in chronological order. Each data item represents the measurement value of a monitoring indicator at a specific time point.

[0087] In step SB, the computer system compares the data item sequence with the key data comparison table. The key data comparison table is a table based on a pre-established empirical key database, which lists the monitoring indicators considered to be key data and their corresponding thresholds or ranges. The comparison process is to match each data item in the data item sequence with the entries in the key data comparison table to determine which data items meet the key data conditions.

[0088] In step SB, the computer system compares the split temperature data item sequence and humidity data item sequence with the key data comparison table. In this example, since all values ​​in the temperature data item sequence do not exceed 25°C, no temperature data item is marked as key data. However, one value (52%) in the humidity data item sequence exceeds the 60% threshold, so this humidity data item is marked as key data.

[0089] The comparison process involves, for example, simple conditional judgment, that is, checking whether the value of the data item meets the conditions listed in the key data comparison table. If the value of the data item meets the conditions, it is marked as key data; otherwise, it is considered non-key data. This process does not require complex calculations, but it does require accurate maintenance of the key data comparison table to ensure that it reflects the latest experience and knowledge.

[0090] In step SC, the computer system determines the data linkage type corresponding to the first key data tag determined in step SB based on the empirical key database. The data linkage type describes the interaction and influence mode between two or more key data items, such as causal relationship, correlation relationship or synchronous change. The empirical key database contains not only the conditions of the key data, but also the possible linkage type information between these key data.

[0091] In step SC, the computer system determines the data linkage type corresponding to the first key data mark determined in step SB (in this example, the data item of humidity>60%) based on this information. Therefore, this humidity data item is marked as the linkage type of "causing increased energy consumption of the air conditioning system".

[0092] The process of determining the data linkage type may involve querying and matching the empirical key database. The computer system needs to find the linkage type information related to the first key data tag and assign it as the first linkage type tag to the monitoring data flow template. This process can perform structured processing on the empirical key database to quickly and accurately query and match the linkage type information.

[0093] Steps SA, SB, and SC are key steps in preparing and preprocessing monitoring data flow templates and their corresponding tags during model training. Through these steps, the computer system can ensure that subsequent model training is based on accurate and meaningful data. Step SA splits the data into a sequence of separate data items, step SB compares these data items with the key data comparison table to determine the key data tags, and step SC determines the data linkage type corresponding to these key data tags based on the empirical key database. Together, these steps provide the necessary input data and information for model training, thereby improving the accuracy and reliability of the model.

[0094] As an implementation solution, before obtaining the monitoring data flow template, the method further includes:

[0095] Step Sa: determining candidate data items in the data item sequence, the candidate data items being data items other than the plurality of first key data markers;

[0096] Step Sb: obtaining the independence coefficient of the candidate data items;

[0097] Step Sc: If the obtained independence coefficient is higher than the set value, the candidate data item is determined as the second key data tag corresponding to the monitoring data flow template;

[0098] Step Sd: Obtain the data linkage type corresponding to the second key data mark, and obtain the second linkage type mark corresponding to the monitoring data flow template.

[0099] In step Sa, the computer system selects candidate data items from the data item sequence obtained in the previous step SA. Candidate data items refer to those data items that have not been marked as key data (i.e., the first key data mark). These data items may also have a significant impact on the energy consumption of the computer room, but because they do not meet the conditions in the initial key data comparison table, they are not directly marked as key data.

[0100] For example, suppose there is a data item sequence containing monitoring data in four dimensions: room temperature, humidity, power consumption, and server load. In step SB, temperature and humidity have been determined as key data according to the key data comparison table (assuming that temperature and humidity are marked as key data when they exceed a certain threshold). However, although power consumption and server load have not yet reached the threshold of key data, they may still have an impact on the energy consumption of the room. Therefore, in step Sa, the computer system will determine power consumption and server load as candidate data items.

[0101] In step Sb, the computer system needs to calculate the independence coefficient of each candidate data item. The independence coefficient is a parameter that measures the independence between data items, which reflects the strength of the association between data items. The lower the value of the independence coefficient, the stronger the association between the data items, that is, the stronger the linkage between them.

[0102] In order to calculate the independence coefficient, the computer system can use a variety of methods, one of which is to calculate the ratio of co-occurrence probabilities. Co-occurrence probability refers to the ratio of the probability of two data items appearing at the same time to the product of their respective probabilities of appearing independently. If the co-occurrence probability of two data items is significantly higher than the product of their respective probabilities of appearing independently, it means that there is a strong correlation between them. Assuming there are two data items X and Y, their independence coefficient can be calculated by the following formula: ; where P(X, Y) is the probability that data items X and Y appear at the same time, and P(X) and P(Y) are the probabilities that data items X and Y appear independently. For example, continuing the above example, suppose there are two candidate data items, power consumption (E) and server load (L), and the key data items temperature (T) and humidity (H) determined previously. The independence coefficients between power consumption and server load, as well as their independence coefficients with temperature and humidity can be calculated. Suppose that through statistical analysis, the following probability values ​​are obtained: P(E, L)=0.3; P(E)=0.5, P(L)=0.4; P(E, T)=0.2, P(E) ×P(T)=0.15; P(E, H)=0.15, P(E) ×P(H)=0.1; P(L, T)=0.25, P(L) ×P(T)=0.2; P(L, H)=0.1, P(L) ×P(H)=0.08.

[0103] According to the calculation formula of the independence coefficient, we can get the independence coefficient between power consumption and server load: ; Independence coefficient between power consumption and temperature: ; Independence coefficient between electricity consumption and humidity: ;Independence coefficient between server load and temperature: ;Independence coefficient between server load and humidity: .

[0104] In step Sc, the computer system determines whether the candidate data item should be marked as key data according to the preset independence coefficient threshold. If the independence coefficient of the candidate data item is higher than the set value, it means that it has a strong correlation with other data items, and therefore it may be regarded as a data item that has a significant impact on the energy consumption of the computer room. At this time, the computer system will mark the candidate data item as the second key data mark.

[0105] For example, continuing with the above example, assume that the set independence coefficient threshold is 1.2. According to the independence coefficient calculated in step Sb, it can be seen that the independence coefficient between power consumption and server load is 1.5, which is higher than the threshold value of 1.2, so they can both be regarded as key data items. Although the independence coefficient between power consumption and temperature and humidity is also higher than 1, considering that temperature and humidity have been marked as the first key data, power consumption is no longer repeatedly marked as key data. Similarly, the independence coefficient of server load with temperature and humidity is also slightly higher than the threshold, but because they are more closely related to power consumption and power consumption has been marked as key data, server load is no longer repeatedly marked.

[0106] In step Sd, the computer system needs to determine the data linkage type corresponding to the second key data marker. This involves, for example, analyzing the interaction and influence pattern between the second key data marker and other data items (including the first key data marker and other alternative data items). This process can rely on empirical knowledge, statistical analysis, or machine learning models to identify the linkage relationship between data items. For example, continuing the example above, power consumption and server load have been determined to be the second key data markers. Next, it is necessary to analyze the linkage relationship between them and other data items.

[0107] For power consumption, it may be found that there is a positive correlation between it and the server load, that is, when the server load increases, the power consumption will also increase accordingly. This may be because the server needs to consume electricity to run, so the higher the load, the greater the power consumption. For server load, it may be found that there is a positive correlation between it and the temperature of the computer room, that is, when the temperature of the computer room rises, the server can increase the speed of the cooling fan or turn on more cooling equipment to maintain a suitable operating temperature, which leads to an increase in load.

[0108] Based on these analyses, corresponding linkage type tags can be assigned to power consumption and server load. For example: power consumption: positively correlated with server load; server load: positively correlated with computer room temperature.

[0109] As an implementation scheme, the method further includes:

[0110] Step SI: obtain a new monitoring data flow template, use an implicit representation model to perform implicit representation mining on the new monitoring data flow template, and obtain a data representation array corresponding to the new monitoring data flow template;

[0111] Step SII: detecting key data of the new monitoring data flow template based on the data representation array corresponding to the new monitoring data flow template and the key data detection model, and obtaining a key data detection result corresponding to the new monitoring data flow template;

[0112] Step SIII: Detect the data linkage type of the new monitoring data flow template based on the key data detection result corresponding to the new monitoring data flow template and the linkage type identification model, and obtain the linkage type detection result corresponding to the new monitoring data flow template;

[0113] Step SIV: if the new monitoring data flow template is determined to be a difficult-to-classify template based on one or more of the key data detection results corresponding to the new monitoring data flow template and the linkage type detection results corresponding to the new monitoring data flow template, the new monitoring data flow template is uploaded to the manual labeling platform;

[0114] Step SV: When obtaining the key data tags and linkage type tags corresponding to the new monitoring data flow template returned by the manual labeling platform, continue to optimize the target implicit representation model, target key data detection model and target linkage type recognition model based on the new monitoring data flow template.

[0115] In step S1, the computer system first obtains a new monitoring data stream template. This new template may come from the real-time data collection of the computer room monitoring system, or it may be specially collected to test the adaptability of the model to new data. The new template contains multi-dimensional monitoring data similar to the existing template, but may contain new data patterns or abnormal situations.

[0116] After obtaining the new template, the computer system uses the previously trained implicit representation model to mine the implicit representation of the new template. The role of the implicit representation model is to convert the original high-dimensional monitoring data stream into a low-dimensional data representation array, which retains the key information and structure of the original data but removes redundancy and noise. Through this step, the new template is converted into a compact and easy-to-process data representation array.

[0117] For example, suppose a new monitoring data stream template contains the temperature, humidity, power consumption, and server load data of the computer room over a period of time. Through the processing of the implicit representation model, these data are converted into a two-dimensional matrix, where each row represents the data representation of a time point, and each column represents an implicit feature (such as average temperature, humidity change trend, power consumption peak, etc.).

[0118] In step SII, the computer system uses the key data detection model to analyze the data representation array of the new monitoring data flow template to identify the key data items therein. The key data detection model has been previously trained to identify data items that have a significant impact on the energy consumption of the computer room. It determines which data items are key data based on the implicit features in the data representation array.

[0119] For example, continuing the example above, the critical data detection model may analyze that at a certain time point in the new template, the server load reached an abnormally high value, and the power consumption also peaked. Based on these analyses, the model marks the server load and power consumption as critical data items.

[0120] In step SIII, the computer system further uses the linkage type recognition model to analyze the linkage relationship between key data items in the new monitoring data flow template. The linkage type recognition model can identify the interaction and influence pattern between different data items, such as causal relationship, correlation relationship or synchronous change.

[0121] For example, continuing with the example above, the linkage type recognition model may analyze that in the new template, there is a positive correlation between the increase in server load and the peak value of power consumption. That is, when the server load rises, the power consumption also increases. This linkage relationship is crucial for understanding the changes in energy consumption in the computer room and formulating energy-saving strategies.

[0122] In step SIV, the computer system will evaluate the key data detection results and linkage type detection results of the new monitoring data flow template. If the model is found to have difficulties in processing this new template, that is, the recognition results of key data or linkage type are significantly different from expectations, or the model's recognition results for the new template are unstable, then this template is considered a "difficult to classify template". For difficult-to-classify templates, the computer system uploads them to the manual annotation platform. The manual annotation platform is a team of domain experts who will carefully analyze the new template and give accurate key data tags and linkage type tags.

[0123] In step SV, after the computer system obtains the key data tags and linkage type tags of the new monitoring data flow template from the manual annotation platform, it will use these tags to further optimize the target implicit representation model, the target key data detection model and the target linkage type recognition model. The optimization process, for example, involves the following: adding the new monitoring data flow template and its tags to the training data set to increase the training sample size of the model. Fine-tune the model using the new training data set to adjust the parameters and structure of the model so that it can better adapt to the new data pattern. After fine-tuning, the performance of the model is evaluated using the validation data set to ensure that the accuracy of the model in identifying key data and linkage types is improved. If the performance of the model still does not meet expectations, repeat the above optimization process until the performance of the model meets the requirements. Through this process, the accuracy and generalization ability of the model are continuously improved, so that it can better adapt to the changes and complexity of the computer room monitoring data. At the same time, this also provides more reliable and accurate data support for the formulation of energy-saving control strategies for computer rooms.

[0124] As an implementation scheme, the implicit representation model includes an array projection module and an implicit representation extraction module; the monitoring data flow template is composed of multiple data items arranged in time sequence; step S12, using the implicit representation model to perform implicit representation mining on the monitoring data flow template, and obtain the template data implicit representation array corresponding to the monitoring data flow template, including:

[0125] Step S121: using an array projection module to perform array space projection on each data item to obtain a data item array representation corresponding to each data item;

[0126] Step S122: inputting the array representation of each data item into the implicit representation extraction module to obtain the data item implicit representation corresponding to each data item;

[0127] Step S123: obtaining an array of implicit representations of template data corresponding to the monitoring data flow template based on the implicit representations of the data items corresponding to the respective data items.

[0128] In step S121, the computer system first uses the array projection module (a mapping network) in the implicit representation model to process each data item in the monitoring data flow template. The function of the array projection module is to map the original high-dimensional data item into a low-dimensional array space, thereby generating an array representation of the data item. This process can be regarded as a form of data dimensionality reduction, which aims to retain the key information of the original data while removing redundancy.

[0129] For example, suppose that the monitoring data stream template contains three data items: timestamp (T), temperature (T°C), and humidity (RH%). The values ​​of each data item at different time points constitute a time series. The array projection module processes these time series and maps them from the original high-dimensional space (for example, the value of each data item at each time point can be regarded as a dimension) to a low-dimensional array space. For example, the time series of each data item can be converted into a fixed-length vector, each element of which is a low-dimensional representation of the original time series.

[0130] The specific implementation of the array projection module may rely on a variety of techniques, such as principal component analysis (PCA), linear discriminant analysis (LDA) or autoencoder. Taking the autoencoder as an example, it consists of two parts: encoder and decoder. In step S121, the encoder part is mainly focused on. The encoder receives the original data item as input, and through a series of nonlinear transformations (such as fully connected layers, convolutional layers, etc.) and activation functions (such as ReLU, Sigmoid, etc.), it finally outputs a low-dimensional vector or matrix as the array representation of the data item.

[0131] In step S122, the computer system inputs the array representation of each data item obtained in step S121 into the implicit representation extraction module. The function of the implicit representation extraction module is to extract deeper implicit features from these array representations, which can better reflect the internal connections and potential laws between data items.

[0132] For example, continuing the above example, assume that array representations of timestamp, temperature, and humidity have been obtained. In step S122, these array representations will be input into the implicit representation extraction module. This module may be a multi-layer perceptron (MLP), a recurrent neural network (RNN), or its variants (such as LSTM, GRU), etc. Regardless of the structure adopted, the implicit representation extraction module will further process and analyze these array representations to extract deeper implicit features. For example, it can learn features such as the trend of temperature changes over time and the periodic fluctuations of humidity.

[0133] The specific implementation of the implicit representation extraction module depends on the characteristics of the data and the requirements of the task. For time series data, recurrent neural networks (RNNs) and their variants may be a suitable choice because they can capture temporal dependencies and long-distance dependencies in the data. For non-time series data or data that needs to capture spatial structure, convolutional neural networks (CNNs) or graph neural networks (GNNs) may be more appropriate. Regardless of the structure used, the implicit representation extraction module processes the input data item array representation through a series of nonlinear transformations and activation functions, and finally outputs a vector or matrix containing implicit features as the implicit representation of the data item.

[0134] In step S123, the computer system combines or concatenates the implicit representation of each data item obtained in step S122 to form a complete template data implicit representation array. This array contains the implicit features of all data items in the monitoring data flow template and is the basis for subsequent key data detection and linkage type identification.

[0135] For example, continuing the above example, assume that the implicit representations of timestamp, temperature and humidity have been obtained. In step S123, these implicit representations will be combined into a complete template data implicit representation array. This array may be a two-dimensional matrix, in which each row corresponds to an implicit representation of a data item, and each column corresponds to a feature in the implicit representation. For example, if the implicit representations of timestamp, temperature and humidity are vectors of length 10 respectively, then the combined template data implicit representation array is a 3x10 matrix.

[0136] When combining implicit representations, the computer system can choose the appropriate combination method based on the characteristics of the data and the requirements of the task. A simple method is to directly concatenate or stack the implicit representations of each data item. However, in some cases, the correlation or dependency between data items can be considered, and a more complex combination method (such as weighted summation, attention mechanism, etc.) can be used to generate an array of implicit representations of template data. Regardless of the method used, the ultimate goal is to generate an array containing the implicit features of all data items for subsequent key data detection and linkage type recognition.

[0137] As an implementation scheme, step S122, inputting the array representation of each data item into the implicit representation extraction module to obtain the data item implicit representation corresponding to each data item, including:

[0138] Step S1221: Based on the time series distribution of the data items corresponding to the array representations of each data item, the array representations of each data item are mined for implicit representation based on the positive direction to obtain the positive implicit representation arrays corresponding to each data item;

[0139] Step S1222: Based on the time series distribution of the data items corresponding to the array representations of each data item, perform implicit representation mining on the array representations of each data item in the reverse direction to obtain the inverse implicit representation arrays corresponding to each data item;

[0140] Step S1223: Merge the positive implicit representation array and the negative implicit representation array corresponding to each data item respectively, to obtain the data item implicit representation corresponding to each data item respectively.

[0141] In step S1221, the computer system focuses on the time series distribution characteristics of the data items. Time series distribution refers to the changing rules and trends of data items in the time series. For each data item array representation, the system performs implicit representation mining in the positive direction of its time series (i.e., from front to back). The process of implicit representation mining can be regarded as a process of nonlinear transformation and feature extraction of the data item array representation. In the implementation of the computer system, this involves, for example, the application of one or more neural network layers (such as fully connected layers, convolutional layers, recurrent neural network layers, etc.). These network layers generate implicit representations that can capture the key features and time series dependencies of data items by learning the complex patterns and structures in the array representations of data items.

[0142] For example, suppose there is an array representation of data items that records the changes in the temperature of the computer room within a week. This array representation may be a vector of length 7, with each element representing the average temperature of the day. In step S1221, the computer system will perform implicit representation mining on this vector in the positive direction of the time series (from Monday to Sunday). By applying the neural network layer, the system may learn features such as the trend of temperature changes over time (such as gradual increase or decrease), periodic fluctuations (such as higher temperatures on weekends), and encode these features into a positive implicit representation array.

[0143] In the mining of positive implicit representations, commonly used neural network structures include recurrent neural networks (RNNs) and their variants (such as LSTM, GRU), which can process time series data well and capture the temporal dependencies therein. In addition, convolutional neural networks (CNNs) can also be used for feature extraction of time series data, especially when processing local features. Regardless of the structure used, the goal of implicit representation mining is to generate a positive implicit representation array that can accurately reflect the key features and temporal dependencies of the data items.

[0144] In step S1222, the computer system uses a method similar to step S1221, but this time it performs implicit representation mining in the reverse direction of the time series (i.e., from back to front). The purpose of reverse implicit representation mining is to capture the temporal dependencies and potential features of data items from another perspective.

[0145] Different from forward implicit representation mining, reverse mining may reveal the mutual influence and feedback mechanism of data items at different time points. For example, in the temperature monitoring of the computer room, the high temperature on the weekend can affect the cooling strategy in the following days, thereby affecting the temperature change trend. Through reverse mining, the computer system can learn this reverse temporal dependency and encode it into an array of reverse implicit representations.

[0146] For example, continuing the above example, suppose there is still an array representation that records the temperature changes in the computer room within a week. In step S1222, the computer system will perform implicit representation mining on this vector in the reverse direction of the time series (from Sunday to Monday). By applying the same neural network structure, the system may learn features such as the reverse trend of temperature changes (such as from high to low) and the impact of high temperatures on the following days, and encode these features into an inverse implicit representation array.

[0147] Reverse implicit representation mining can also use neural network structures such as RNN, LSTM, GRU or CNN. Compared with forward mining, reverse mining pays more attention to the reverse dependencies and feedback mechanisms between data items. By combining the results of forward and reverse implicit representation mining, the computer system can more comprehensively capture the temporal dependencies and potential characteristics of data items. In step S1223, the computer system merges the positive implicit representation array and the reverse implicit representation array obtained in steps S1221 and S1222. The merging process may involve a variety of methods such as simple splicing, weighted summation, attention mechanism, etc. The purpose of the merging is to combine the information obtained from the forward and reverse implicit representation mining to generate a more comprehensive and accurate implicit representation of the data item.

[0148] For example, continuing the above example, assume that the positive implicit representation array and the negative implicit representation array of the room temperature are obtained in steps S1221 and S1222, respectively. In step S1223, the computer system merges the two arrays. The merging method may depend on the shape and size of the arrays. For example, if both arrays are one-dimensional and have the same length, they can be directly concatenated or weighted summed; if the arrays are of different shapes, dimension transformation or feature selection can be performed first.

[0149] The method of merging the positive implicit representation array and the negative implicit representation array depends on the specific application scenario and data characteristics. A common method is concatenation, which simply connects the two arrays together to form a longer array. Another method is weighted sum, which adds the elements of the two arrays according to certain weights. In addition, the attention mechanism can also be used to merge the two implicit representation arrays, and the merging process is dynamically adjusted by calculating the attention weight of each element. Regardless of the method used, the goal of the merging is to generate an implicit representation of the data item that can fully reflect the key features and temporal dependencies of the data item.

[0150] Step S122 extracts a more comprehensive and accurate implicit representation from the data item array representation through a forward and reverse implicit representation mining method based on temporal distribution. The forward implicit representation mining captures the forward temporal dependencies and characteristics of the data items, while the reverse implicit representation mining reveals the reverse dependencies and potential characteristics of the data items. By merging the implicit representations of these two directions, the computer system can generate an implicit representation of the data item that contains rich information, which not only improves the model's ability to capture the temporal characteristics of the data items, but also enhances the model's adaptability to data complexity and diversity.

[0151] As an implementation scheme, the key data detection model includes a reconstruction module and a decision module; step S13, based on the template data implicit representation array and the key data detection model, detects the key data of the monitoring data flow template to obtain the key data detection result, including:

[0152] Step S131: in the reconstruction module, based on the time series distribution of the data items corresponding to the implicit representations of the data items, the implicit representations of the data items are reconstructed to obtain the data item reconstruction arrays corresponding to the data items;

[0153] Step S132: Perform key data decision on each data item reconstruction array through the decision module to obtain key data detection results.

[0154] In step S131, the computer system first uses the reconstruction module (a decoding network) in the key data detection model to process the template data implicit representation array. The role of the reconstruction module is to convert the implicit representation array back to the original data space to reconstruct the temporal distribution of each data item. This process can be regarded as a process of decoding or inverse transformation of the implicit representation array.

[0155] For example, suppose there is an implicit representation array of template data, which contains implicit representations of three data items: room temperature, humidity, and power consumption. In step S131, the reconstruction module processes these implicit representations respectively to reconstruct the time series distribution of each data item. For example, for the temperature data item, the reconstruction module outputs a reconstruction array, which contains the reconstructed values ​​of the room temperature at different time points, and these reconstructed values ​​should be as close as possible to the true value of the original temperature data item. The specific implementation of the reconstruction module relies on a variety of technologies, such as fully connected layers, convolutional layers, recurrent neural network layers, etc. These network layers can convert the implicit representation back to the original data space by learning the mapping relationship between the implicit representation array and the original data items. During the reconstruction process, the reconstruction module can adopt some optimization strategies to improve the accuracy of the reconstruction, such as using a loss function to measure the difference between the reconstructed value and the true value, and updating the network parameters through the back propagation algorithm to minimize the loss function.

[0156] In step S132, the computer system processes the reconstructed array of each data item obtained in step S131 using the decision module in the key data detection model. The role of the decision module is to evaluate and analyze the reconstructed array to determine which data items are key data. This process may involve a series of complex calculations and decision rules.

[0157] For example, continuing the above example, assume that the reconstructed arrays of the three data items of temperature, humidity and power consumption have been obtained. In step S132, the decision module will evaluate these reconstructed arrays. It can analyze the characteristics of each data item at different time points, such as the change trend, abnormal values, fluctuation range, etc., and determine which data items are key data according to the preset key data definition or threshold. For example, if a value in the temperature reconstruction array exceeds the preset high temperature threshold, the decision module can mark the temperature data item at that time point as key data.

[0158] The specific implementation of the decision module may vary depending on the task requirements and data characteristics. A common implementation method is to use a classifier (such as logistic regression, support vector machine, neural network, etc.) to classify the reconstructed array. The classifier can classify and predict the new reconstructed array by learning the mapping relationship between the reconstructed array and the key data labels. In the decision-making process, the classifier can consider multiple feature dimensions (such as reconstruction value, change trend, outliers, etc.) and use weighted sum, decision tree, neural network and other algorithms to make comprehensive decisions.

[0159] In addition to classifiers, the decision module can also use other methods to make key data decisions. For example, it can calculate the mean, variance, kurtosis and other statistics of the reconstructed array based on statistical analysis, and judge whether the data item is key data based on the comparison of these statistics with the preset threshold. In addition, the decision module can also combine domain knowledge and expert experience to formulate more complex decision rules to improve the accuracy and reliability of key data detection.

[0160] Step S13 is the core step of key data detection, which uses the reconstruction module and decision module in the key data detection model to process the template data implicit representation array to obtain the key data detection result. The reconstruction module is responsible for converting the implicit representation array back to the original data space and reconstructing the time series distribution of each data item; while the decision module is responsible for evaluating and analyzing the reconstructed array to determine which data items are key data. Through the collaborative work of these two modules, the computer system can accurately identify the key data items in the monitoring data flow template, which not only improves the accuracy and reliability of key data detection, but also enhances the adaptability of the model to data complexity and diversity.

[0161] As an implementation scheme, step S14 detects the data linkage type of the monitoring data flow template based on the key data detection result and the linkage type identification model to obtain the linkage type detection result, including:

[0162] Step S141: determining multiple template key data corresponding to the monitoring data flow template based on the key data detection result;

[0163] Step S142: determining a corresponding template key data group based on a plurality of template key data;

[0164] Step S143: determining the template array representation fragment corresponding to the template key data group in the template data implicit representation array;

[0165] Step S144: input the template array representation fragment into the linkage type recognition model to perform linkage type detection, and obtain the linkage type detection result corresponding to the template key data group.

[0166] In step S141, the computer system first uses the key data detection result to determine multiple template key data in the monitoring data flow template. The key data detection result is obtained by the key data detection model in step S13, which identifies which data items in the monitoring data flow template are identified as key data. These data items have a significant impact on the energy consumption of the computer room, for example, and are the focus of subsequent analysis.

[0167] For example, suppose there is a monitoring data flow template that contains time series data of four data items: room temperature, humidity, power consumption, and server load. In step S13, the key data detection model may identify temperature and power consumption as key data items because they exceed a preset threshold or have a direct impact on the energy consumption of the room. Therefore, in step S141, the computer system determines temperature and power consumption as template key data.

[0168] In step S142, the computer system further determines the combination relationship between the multiple template key data determined in step S141 to form template key data groups. These data groups represent the potential linkage relationship between key data items and are the basis for subsequent linkage type detection.

[0169] For example, continuing the above example, assume that temperature and power consumption have been determined as template key data. In step S142, the computer system may combine these key data items to form multiple template key data groups. For example, it may form a data group containing temperature and power consumption to explore the linkage relationship between them. In addition, if there are other key data items (such as humidity, server load, etc.), they can also be added to different data groups for analysis.

[0170] In step S143, the computer system finds the template array representation fragments corresponding to the template key data groups determined in step S142 in the template data implicit representation array. The template data implicit representation array is obtained by the implicit representation model in step S12, and it contains the low-dimensional representation information of the monitoring data flow template. By locating the data fragments corresponding to the template key data groups, the computer system can provide the necessary data input for the subsequent linkage type detection.

[0171] For example, continuing the above example, assume that a template key data group containing two key data items, temperature and power consumption, has been obtained. In step S143, the computer system will find a data segment corresponding to this data group in the template data implicit representation array. This data segment may be a sub-matrix of a two-dimensional matrix, which contains implicit representation information related to temperature and power consumption. This sub-matrix will be used as input data for subsequent linkage type detection.

[0172] In step S144, the computer system inputs the template array representation fragment obtained in step S143 into the linkage type recognition model for linkage type detection. The linkage type recognition model is a trained machine learning model that can analyze the correlation and mutual influence between data items and output the linkage type between them. This output result will serve as the linkage type detection result corresponding to the template key data group, providing an important basis for the subsequent energy-saving control strategy formulation.

[0173] For example, continuing the above example, assume that a template array representation fragment related to temperature and power consumption has been obtained. In step S144, the computer system inputs this fragment into the linkage type recognition model for detection. The model may identify the linkage relationship between temperature and power consumption by analyzing the implicit representation information in the fragment. For example, it may find that when the temperature rises, the power consumption will also increase accordingly, indicating that there is a positive correlation between temperature and power consumption. Based on this relationship, the linkage type recognition model will output a linkage type detection result indicating a positive correlation.

[0174] The specific implementation of the linkage type recognition model may rely on a variety of machine learning technologies, such as neural networks, support vector machines, decision trees, etc. These models, for example, have complex internal structures and highly nonlinear processing capabilities, and can capture the complex correlations and mutual influences between data items. During the training process, the model will learn linkage type information from a large amount of labeled data and continuously optimize its own parameters to improve detection accuracy. During the detection process, the model will calculate the probability distribution or category label of the output result based on the characteristics of the input data to determine the linkage type between data items.

[0175] Step S14 reveals the factors affecting the energy consumption of the computer room by analyzing the key data items in the monitoring data flow template and the linkage relationship between them. Through the detailed execution process of steps S141 to S144, the computer system can accurately determine the template key data, construct the template key data group, locate the template array representation fragment and perform linkage type detection. These steps together constitute the core process of linkage type detection. By continuously optimizing the linkage type recognition model and improving the detection accuracy, the energy consumption behavior of the computer room can be better understood and controlled, and more efficient and sustainable energy management can be achieved.

[0176] As an implementation scheme, the linkage type recognition model includes an implicit module, a sliding window module and a downsampling module; step S144, inputting the template array representation fragment into the linkage type recognition model to perform linkage type detection, and obtaining the linkage type detection result corresponding to the template key data group, including:

[0177] Step S1441: input the template array representation fragment into the implicit module, and use the implicit module to process the template array representation fragment to obtain a corresponding first implicit array representation;

[0178] Step S1442: input the first implicit array representation into the sliding window module, and use the sliding window module to perform a sliding window operation on the first implicit array representation to obtain a corresponding second implicit array representation;

[0179] Step S1443: inputting the second implicit array representation into a downsampling module, and downsampling the second implicit array representation using the downsampling module;

[0180] Step S1444: performing data linkage type identification based on the downsampling result to obtain a linkage type detection result.

[0181] In step S1441, the computer system first inputs the template array representation fragment into the implicit module of the linkage type recognition model. The implicit module is a multi-layer neural network structure, such as composed of multiple fully connected layers or convolutional layers, which is used to extract deep features of the input data. These features are called the first implicit array representation, which are the basis for subsequent processing steps.

[0182] The processing of the implicit module is a process of nonlinear transformation of the input data. Through the layer-by-layer processing of the multi-layer neural network, the implicit module can capture the complex patterns and structures in the input data and encode these patterns into higher-level feature representations. These feature representations not only retain the key information of the input data, but also remove redundancy and noise, making subsequent processing more efficient and accurate. For example, assume that the template array representation fragment is a two-dimensional matrix, which contains implicit representation information related to the template key data group. In step S1441, this matrix is ​​input into the implicit module. The implicit module converts this matrix into a new two-dimensional matrix, namely the first implicit array representation, through layer-by-layer processing of the multi-layer neural network. Each element of this new matrix represents a deep feature of the input data, which will be used to identify the linkage type in subsequent steps.

[0183] The specific structure of the hidden module depends on the characteristics of the input data and the requirements of the task. For time series data or image data, convolutional layers may be an important part of the hidden module because they can capture local features and spatial structures. For unstructured data or situations where global features are required, fully connected layers may be more appropriate. Regardless of the structure used, the goal of the hidden module is to transform the input data into a higher-level feature representation for subsequent processing.

[0184] In step S1442, the computer system inputs the first implicit array representation obtained in step S1441 into a sliding window module. The sliding window module is a technology for processing sequence data, which can slide a fixed-size window on the sequence and operate on the data in the window. In the linkage type recognition model, the sliding window module is used to capture local features and temporal dependencies in the first implicit array representation.

[0185] The sliding window operation, for example, involves sliding a fixed-size window over the sequence and performing convolution, pooling or other types of operations on the data within the window. In step S1442, the sliding window module performs such operations on the first implicit array representation to generate a second implicit array representation. This new representation contains the local features and timing dependency information in the first implicit array representation, which is the basis for subsequent downsampling and linkage type identification.

[0186] For example, continuing the above example, assume that the first implicit array representation is a two-dimensional matrix. In step S1442, the sliding window module slides a fixed-size window (e.g., a 3×3 window) on this matrix and performs a convolution operation on the elements in the window. The convolution operation calculates the weighted sum of the elements in the window and the convolution kernel and generates a new element as the output. By sliding the window and performing such operations on the entire matrix, the sliding window module generates a new two-dimensional matrix as the second implicit array representation. This new matrix contains the local features and temporal dependency information in the first implicit array representation. The specific implementation of the sliding window module depends on the characteristics of the input data and the requirements of the task. For time series data or image data, convolution operations may be an important part of the sliding window module because they can capture local features and spatial structures. For other types of data, other types of operations (such as pooling, attention mechanisms, etc.) can be used to implement the sliding window function. Regardless of the operation used, the goal of the sliding window module is to capture local features and temporal dependency information in the input data.

[0187] In step S1443, the computer system inputs the second implicit array representation obtained in step S1442 into a downsampling module. The downsampling module is a technology for reducing data dimensions, which can reduce the complexity of the data while retaining important information of the data. In the linkage type recognition model, the downsampling module is used to reduce the dimension and complexity of the second implicit array representation so as to perform more efficient linkage type recognition later.

[0188] The downsampling operation, for example, involves downsampling or subsampling the input data to reduce the resolution or quantity of the data. In step S1443, the downsampling module performs such an operation on the second implicit array representation to generate a new, lower-dimensional array representation. This new representation retains important information about the input data, but has lower complexity and higher computational efficiency.

[0189] For example, continuing the above example, assume that the second implicit array representation is a two-dimensional matrix. In step S1443, the downsampling module performs a downsampling operation on this matrix. For example, it may halve the number of rows and columns of the matrix (or adopt a downsampling rate of other proportions) to generate a new, lower-dimensional matrix as output. This new matrix retains important information of the input data (such as local features and timing dependencies), but has lower complexity and higher computational efficiency.

[0190] The specific implementation of the downsampling module depends on the characteristics of the input data and the requirements of the task. For image data or time series data, common downsampling methods include maximum pooling, average pooling, etc. These methods can retain the important information of the input data while reducing the complexity of the data. For other types of data, other types of downsampling methods (such as random downsampling, clustering-based downsampling, etc.) can be used to implement the downsampling function. Regardless of the method used, the goal of the downsampling module is to reduce the complexity of the data while retaining the important information of the input data.

[0191] In step S1444, the computer system performs data linkage type recognition based on the downsampling result obtained in step S1443. Linkage type recognition is the ultimate goal of the linkage type recognition model, which aims to determine the linkage type between the input data (i.e., the template key data group). In step S1444, the computer system may use a classifier, a regressor, or other types of machine learning models to implement the linkage type recognition function.

[0192] The specific implementation of linkage type recognition depends on the characteristics of the input data and the requirements of the task. For classification tasks, the computer system may use classifiers such as support vector machines, random forests, neural networks, etc. to determine which linkage type category the input data belongs to. For regression tasks or other types of tasks, other types of machine learning models can be used to implement linkage type recognition. Regardless of the model used, the goal of linkage type recognition is to determine the linkage type between the input data based on their characteristics.

[0193] For example, continuing the above example, assume that the downsampling module outputs a two-dimensional matrix with a lower dimension as input for linkage type identification. In step S1444, the computer system may use a neural network classifier to process the matrix and output a label or probability distribution representing the linkage type as a result. For example, the classifier may output a label representing a linkage type such as positive correlation, negative correlation, or no correlation; or output a distribution vector representing the probability of each linkage type. This output result is the linkage type detection result, which reveals the linkage type between the input data (i.e., the template key data group).

[0194] The specific implementation of linkage type identification may involve a variety of machine learning technologies and algorithms. For classification tasks, common algorithms include support vector machines, logistic regression, decision trees, random forests, neural networks, etc. These algorithms have their own advantages and disadvantages and are suitable for different types of data and tasks. For regression tasks or other types of tasks, other types of machine learning models (such as regression models, clustering models, etc.) can be used to implement linkage type identification. Regardless of the model or algorithm used, the goal of linkage type identification is to determine the linkage type between the input data based on the characteristics of the input data, and provide an important basis for the subsequent formulation of energy-saving control strategies.

[0195] Step S144 uses the linkage type recognition model to process the template array representation fragment to identify the linkage type between the key data groups of the template. Through the collaborative work of the implicit module, the sliding window module and the downsampling module, the computer system can convert the input data into a higher-level feature representation and capture the local features, timing dependencies and important information therein. Finally, in step S1444, the computer system performs linkage type recognition based on the downsampling result and outputs the linkage type detection result. This result reveals the linkage type relationship between the input data. By continuously optimizing the linkage type recognition model and improving the detection accuracy, we can better understand the energy consumption behavior of the computer room and formulate corresponding energy-saving measures.

[0196] As an implementation scheme, step S130 detects the data linkage type of the computer room monitoring data flow based on the key data detection result and the pre-optimized target linkage type recognition model, and obtains the linkage type detection result including:

[0197] Step S1301: determining a plurality of estimated key data corresponding to the computer room monitoring data stream based on the key data detection result;

[0198] Step S1302: determining target device data from a plurality of estimated key data, and forming an estimated key data group with the target device data and each of the other estimated key data;

[0199] Step S1303: determining the array representation sequences corresponding to each estimated key data group in the data representation array;

[0200] Step S1304: input each array representation sequence into the target linkage type recognition model to perform linkage type detection, and obtain the linkage type detection results corresponding to each estimated key data group.

[0201] In step S1301, the computer system uses the key data detection result to determine multiple estimated key data in the computer room monitoring data stream. The key data detection result is obtained by the key data detection model in the previous step, which identifies which data items in the monitoring data stream are identified as key data. These key data, for example, have a significant impact on the energy consumption of the computer room and are the focus of subsequent analysis.

[0202] Assume that there is a computer room monitoring data stream, which contains monitoring data of multiple devices (such as servers, air conditioners, UPS, etc.), such as temperature, humidity, power consumption, load, etc. In the key data detection step, the model may identify the load of a server, the temperature of the computer room, and the power consumption of the air conditioner as key data items. Therefore, in step S131, the computer system determines these data items as estimated key data.

[0203] In step S1302, the computer system selects one or more of the multiple estimated key data determined in step S1301 as target device data. The target device data is, for example, a data item of special concern because it may have a significant impact on the energy consumption of the computer room or have a complex linkage relationship with other data items. Then, the computer system combines each target device data with each of the remaining estimated key data to form multiple estimated key data groups.

[0204] Continuing with the above example, assume that the server load is selected as the target device data. In step S132, the computer system combines the server load with the estimated key data such as the temperature of the computer room and the power consumption of the air conditioner to form multiple estimated key data groups. For example, one data group may include the server load and the temperature of the computer room, and another data group may include the server load and the power consumption of the air conditioner.

[0205] In step S1303, the computer system finds the array representation sequence corresponding to the estimated key data groups formed in step S1302 in the data representation array. The data representation array is obtained through the implicit representation model in the previous step, and it contains the low-dimensional representation information of the computer room monitoring data stream. By locating the data fragments corresponding to the estimated key data groups, the computer system can provide the necessary data input for the subsequent linkage type detection.

[0206] Continuing with the above example, assume that three estimated key data, namely, server load, room temperature, and air conditioner power consumption, have been obtained, and multiple estimated key data groups have been formed. In step S1303, the computer system will find data segments corresponding to these data groups in the data representation array. For example, for a data group containing server load and room temperature, the computer system will locate the time series data segment corresponding to these two data items in the data representation array; for a data group containing server load and air conditioner power consumption, the computer system will locate the time series data segment corresponding to these two data items.

[0207] In step S1304, the computer system inputs each array representation sequence obtained in step S1303 into a pre-optimized target linkage type recognition model for linkage type detection. The target linkage type recognition model is a trained machine learning model that can analyze the correlation and mutual influence between data items and output the linkage type between them. For each input array representation sequence, the model will output a result representing the linkage type.

[0208] Continuing with the above example, it is assumed that array representation sequences corresponding to estimated key data groups such as server load and computer room temperature, server load and air conditioning power consumption have been obtained. In step S134, the computer system inputs these sequences into the target linkage type recognition model for detection. The model may determine the linkage type between data items by analyzing the time dependency, correlation or other features in the sequence. For example, for the data group of server load and computer room temperature, the model may identify that there is a positive correlation between them, that is, when the server load increases, the computer room temperature will also rise accordingly; for the data group of server load and air conditioning power consumption, the model may identify that there is a causal relationship between them, that is, an increase in server load will lead to an increase in air conditioning power consumption.

[0209] The specific implementation of the target linkage type recognition model may rely on a variety of machine learning technologies, such as neural networks, support vector machines, decision trees, etc. These models, for example, have complex internal structures and highly nonlinear processing capabilities, and can capture the complex correlations and mutual influences between data items. During the training process, the model will learn linkage type information from a large amount of labeled data and continuously optimize its own parameters to improve detection accuracy. During the detection process, the model will calculate the probability distribution or category label of the output result based on the characteristics of the input data to determine the linkage type between data items.

[0210] For time series data or situations where time dependencies need to be captured, recurrent neural networks (RNNs) and their variants (such as LSTMs and GRUs) may be important components of target linkage type recognition models. These models are able to handle time dependencies in sequence data and capture long-term dependencies between data items through hidden states. In addition, convolutional neural networks (CNNs) can also be used to process data with spatial structures or situations where local features need to be captured. Regardless of the model or technology used, the goal of the target linkage type recognition model is to accurately identify the linkage type between data items and provide an important basis for the subsequent formulation of energy-saving control strategies. By continuously optimizing the model structure and improving detection accuracy, we can better understand the correlation and mutual influence between different devices or data items in the computer room, so as to formulate more effective energy-saving control strategies.

[0211] As an implementation scheme, step S140, formulating an energy-saving control strategy based on the key data detection results and the linkage type detection results, includes:

[0212] Step S1401: determining the energy consumption key data and temperature key data in the computer room monitoring data stream based on the key data detection result;

[0213] Step S1402: determining the linkage relationship between the key energy consumption data and the key temperature data based on the linkage type detection result;

[0214] Step S1403: determining energy consumption influencing factors according to the values ​​and linkage relationships of key energy consumption data;

[0215] Step S1404: Formulate an energy-saving control strategy based on the factors affecting energy consumption, wherein the energy-saving control strategy includes at least one of adjusting the operating state of the air-conditioning system, adjusting the server load distribution, and optimizing power management.

[0216] In step S1401, the computer system first reviews the key data detection results obtained by analyzing the computer room monitoring data stream through the key data detection model in the previous step. The key data detection results identify the data items in the computer room monitoring data stream that have a significant impact on energy consumption. These data items may involve multiple aspects, such as power consumption, server load, ambient temperature, etc.

[0217] In this step, the goal of the computer system is to further filter out key energy consumption data and key temperature data. Key energy consumption data refers to data items that directly affect the total energy consumption of the computer room, such as power consumption, UPS (uninterruptible power supply) energy consumption, etc. Key temperature data refers to data items that are closely related to the temperature of the computer room and may have an indirect impact on energy consumption, such as ambient temperature, server internal temperature, etc. Assume that the key data detection result identifies the following data items as key data: power consumption of server A, overall temperature of the computer room, power consumption of UPS, and CPU load of server B. In step S1401, the computer system will further classify these data items into key energy consumption data and key temperature data. Among them, the power consumption of server A and the power consumption of UPS are obviously key energy consumption data because they directly reflect the energy consumption of the computer room. The overall temperature of the computer room and the ambient temperature may be classified as key temperature data because, although they do not directly consume energy, they will indirectly affect energy consumption by affecting the cooling requirements of the computer room.

[0218] In step S1402, the computer system uses the linkage type detection result to analyze the linkage relationship between the key energy consumption data and the key temperature data. The linkage type detection result is obtained by analyzing the linkage type identification model in the previous step, which reveals the interaction and influence mode between different data items.

[0219] In this step, the goal of the computer system is to identify the specific linkage between key energy consumption data and key temperature data, such as positive correlation, negative correlation, or no significant correlation, etc. These linkages are crucial for understanding the dynamic changes in energy consumption in the computer room and formulating effective energy-saving strategies.

[0220] Continuing with the above example, assume that the linkage type detection results reveal the following linkage relationship: There is a positive correlation between the overall temperature of the computer room and the power consumption of server A, that is, when the temperature of the computer room increases, the power consumption of server A will also increase, which may be because the server needs more cooling to maintain its normal operating temperature. On the other hand, there may not be a significant linkage relationship between the power consumption of the UPS and the overall temperature of the computer room, because the energy consumption of the UPS mainly depends on its own load and efficiency, rather than the temperature of the computer room.

[0221] In step S1403, the computer system further analyzes the factors affecting energy consumption in combination with the values ​​of the key energy consumption data and the linkage relationship determined in step S1402. These factors may include the physical environment in the computer room (such as temperature and humidity), the equipment operation status (such as server load and cooling system efficiency), and external conditions (such as climate conditions and electricity price fluctuations).

[0222] By comprehensively analyzing these factors, the computer system can have a deeper understanding of the changing patterns of energy consumption in the computer room and the potential energy saving space.

[0223] Continuing with the above example, assume that the current overall temperature of the computer room is 25°C and the power consumption of server A is 1000W. Based on the linkage relationship determined in step S142, the computer system can infer that if the temperature of the computer room rises further, the power consumption of server A may continue to increase because more cooling is required to keep the server temperature stable. At the same time, if the load of server A is also at a high level, its power consumption may rise further. Therefore, the temperature of the computer room, the load of server A, and the linkage relationship between them are all key factors affecting energy consumption.

[0224] In step S1404, the computer system formulates specific energy-saving control strategies according to the energy consumption influencing factors determined in step S1403. These strategies aim to reduce energy consumption and improve energy efficiency by optimizing the operating parameters and equipment configuration of the computer room.

[0225] Energy-saving control strategies may include a variety of measures, such as adjusting the operating status of the air-conditioning system to maintain the temperature of the computer room within an appropriate range, adjusting the server load distribution to balance the energy consumption of each server, and optimizing the power management strategy to reduce unnecessary energy consumption.

[0226] Continuing with the above example, suppose the computer system determines through analysis that the high temperature in the computer room is one of the main factors leading to increased energy consumption. In order to reduce energy consumption, the computer system can formulate the following energy-saving control strategy:

[0227] 1. Adjust the operating status of the air conditioning system: Increase the cooling capacity of the air conditioning system or adjust the air supply strategy to reduce the temperature of the computer room. For example, the set temperature of the air conditioning system can be lowered to 23°C to reduce the additional cooling demand caused by the high temperature of the computer room.

[0228] 2. Adjust server load distribution: Use load balancing technology to migrate high-load tasks from servers with higher energy consumption to servers with lower energy consumption. For example, if server A has a high load and consumes more energy, some tasks can be migrated to server B with lower load and less energy consumption.

[0229] 3. Optimize power management: Use dynamic power management technology (such as DVFS) to adjust the operating frequency and voltage of the server in real time according to the server load to reduce energy consumption. For example, when the server load is low, its operating frequency and voltage are automatically reduced to reduce unnecessary energy consumption.

[0230] When formulating energy-saving control strategies, computer systems can use a variety of technologies and algorithms to optimize the operating parameters and equipment configuration of the computer room. For example, in adjusting the operating status of the air-conditioning system, machine learning algorithms can be used to predict the changing trend of the room temperature and adjust the set parameters of the air-conditioning system in advance; in adjusting the server load distribution, load balancing algorithms can be used to optimize the task allocation strategy to reduce overall energy consumption; in optimizing power management, dynamic power management technology can be used to adjust the working status of the server according to the real-time load of the server to reduce energy consumption.

[0231] In addition, the computer system can consider the implementation cost and benefit evaluation of energy-saving strategies. For example, when adjusting the operating status of the air-conditioning system or adding cooling equipment, the equipment purchase cost and operating cost need to be considered; when adjusting the server load distribution or optimizing power management, the impact on server performance and stability needs to be considered. By comprehensively considering these factors and formulating reasonable energy-saving control strategies, the computer system can achieve energy-saving goals while maintaining stable operation and efficient performance of the computer room.

[0232] An embodiment of the present invention provides a computer system, such as Figure 3 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.

[0233] 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 data center computer room energy saving control method, characterized in that: The method comprises: Acquire a computer room monitoring data stream, use a pre-optimized target implicit representation model to perform implicit representation mining on the computer room monitoring data stream, and obtain a data representation array corresponding to the computer room monitoring data stream; Detect key data of the computer room monitoring data stream based on the data representation array and the pre-optimized target key data detection model to obtain key data detection results; Detect the data linkage type of the computer room monitoring data stream based on the key data detection result and the pre-optimized target linkage type recognition model to obtain a linkage type detection result; Formulate an energy-saving control strategy based on the key data detection results and the linkage type detection results; The target implicit representation model, the target key data detection model and the target linkage type recognition model are optimized by the following steps: Obtain a monitoring data flow template and a template tag corresponding to the monitoring data flow template; the template tag includes a key data tag corresponding to the monitoring data flow template and a linkage type tag corresponding to the monitoring data flow template; Using an implicit representation model to perform implicit representation mining on the monitoring data flow template to obtain an implicit representation array of template data corresponding to the monitoring data flow template; Detecting the key data of the monitoring data flow template based on the template data implicit representation array and the key data detection model to obtain a key data detection result; Detect the data linkage type of the monitoring data flow template based on the key data detection result and the linkage type identification model to obtain a linkage type detection result; Determine a first optimization error based on the difference between the key data detection result and the key data mark, determine a second optimization error based on the difference between the linkage type detection result and the linkage type mark, determine a first directional derivative based on the first optimization error, determine a second directional derivative based on the second optimization error, sum the first directional derivative and the second directional derivative to obtain a merged directional derivative, optimize the implicit representation model, the key data detection model and the linkage type recognition model based on the merged directional derivative, and iterate the implicit representation model, the key data detection model and the linkage type recognition model; Inputting the monitoring data flow template into the iterated implicit representation model, using the iterated implicit representation model to perform implicit representation mining on the monitoring data flow template to obtain an implicit representation array, and merging the linkage type detection result with the implicit representation array to iterate the template data implicit representation array; Return to the process of detecting the key data of the monitoring data flow template based on the template data implicit representation array and the key data detection model to perform loop optimization until the preset judgment criteria are met, and obtain the optimized target implicit representation model, target key data detection model and target linkage type identification model, the target implicit representation model, target key data detection model and target linkage type identification model, which are used to collaboratively detect the data linkage type of the computer room monitoring data flow; The data linkage type of the monitoring data flow template is detected based on the key data detection result and the linkage type identification model, and the linkage type detection result is obtained, including: Determine multiple template key data corresponding to the monitoring data flow template based on the key data detection result; Based on the plurality of template key data, determining a corresponding template key data group; Determining, in the template data implicit representation array, a template array representation fragment corresponding to the template key data group; The template array representation fragment is input into the linkage type recognition model to perform linkage type detection, and a linkage type detection result corresponding to the template key data group is obtained.

2. The method according to claim 1, characterized in that Before obtaining the monitoring data flow template and the template tag corresponding to the monitoring data flow template, the method further includes: Splitting the monitoring data flow template into data items to obtain a corresponding data item sequence; Comparing the data item sequence with a key data comparison table, and determining a plurality of first key data tags corresponding to the monitoring data flow template based on the comparison result; the key data comparison table is obtained based on an empirical key database established in advance; The data linkage types corresponding to the multiple first key data tags are determined based on the empirical key database, and the first linkage type tag corresponding to the monitoring data flow template is obtained.

3. The method according to claim 2, characterized in that Before obtaining the monitoring data flow template, the method further includes: Determining candidate data items in the data item sequence, the candidate data items being data items other than the plurality of first key data markers; Obtaining the independence coefficient of the candidate data item; If the obtained independence coefficient is higher than a set value, the candidate data item is determined as a second key data tag corresponding to the monitoring data flow template; Obtain the data linkage type corresponding to the second key data mark, and obtain the second linkage type mark corresponding to the monitoring data flow template.

4. The method according to claim 1, characterized in that The method further comprises: Acquire a new monitoring data flow template, and use an implicit representation model to perform implicit representation mining on the new monitoring data flow template to obtain a data representation array corresponding to the new monitoring data flow template; Detect key data of the new monitoring data flow template based on the data representation array and key data detection model corresponding to the new monitoring data flow template, and obtain key data detection results corresponding to the new monitoring data flow template; Detect the data linkage type of the new monitoring data flow template based on the key data detection result corresponding to the new monitoring data flow template and the linkage type identification model, and obtain the linkage type detection result corresponding to the new monitoring data flow template; If the new monitoring data flow template is determined to be a difficult-to-classify template based on one or more of the key data detection results corresponding to the new monitoring data flow template and the linkage type detection results corresponding to the new monitoring data flow template, the new monitoring data flow template is uploaded to the manual labeling platform; When the key data mark and linkage type mark corresponding to the new monitoring data flow template returned by the manual annotation platform are obtained, the target implicit representation model, the target key data detection model and the target linkage type recognition model are continuously optimized based on the new monitoring data flow template; The implicit representation model includes an array projection module and an implicit representation extraction module; the monitoring data flow template is composed of multiple data items arranged in time sequence; the implicit representation model is used to perform implicit representation mining on the monitoring data flow template to obtain the template data implicit representation array corresponding to the monitoring data flow template, including: Using an array projection module to perform array space projection on each of the data items to obtain data item array representations corresponding to each of the data items; Inputting the array representation of each of the data items into the implicit representation extraction module to obtain the data item implicit representation corresponding to each of the data items; Based on the implicit representations of the data items respectively corresponding to the data items, an array of implicit representations of template data corresponding to the monitoring data flow template is obtained.

5. The method according to claim 4, characterized in that The step of inputting the array representation of each data item into the implicit representation extraction module to obtain the data item implicit representation corresponding to each data item includes: Based on the time series distribution of the data items corresponding to the array representations of each data item, the array representations of each data item are mined based on the positive direction to obtain the positive implicit representation arrays corresponding to each data item; Based on the time series distribution of the data items corresponding to the array representations of each data item, the array representations of each data item are mined based on the implicit representation in the reverse direction to obtain the inverse implicit representation arrays corresponding to each data item; The positive implicit representation array and the negative implicit representation array corresponding to each of the data items are respectively merged to obtain the data item implicit representation corresponding to each of the data items.

6. The method according to claim 4, characterized in that The key data detection model includes a reconstruction module and a decision module; the key data of the monitoring data flow template is detected based on the template data implicit representation array and the key data detection model, and the key data detection result obtained includes: In the reconstruction module, based on the time series distribution of the data items corresponding to the implicit representations of the data items, the implicit representations of the data items are reconstructed to obtain data item reconstruction arrays corresponding to the data items; The decision module is used to perform key data decision on each of the data item reconstruction arrays to obtain key data detection results.

7. The method according to claim 1, characterized in that The linkage type recognition model includes an implicit module, a sliding window module and a downsampling module; the template array representation fragment is input into the linkage type recognition model to perform linkage type detection, and the linkage type detection result corresponding to the template key data group is obtained, which includes: Inputting the template array representation fragment into the implicit module, and processing the template array representation fragment using the implicit module to obtain a corresponding first implicit array representation; Inputting the first implicit array representation into the sliding window module, and using the sliding window module to perform a sliding window operation on the first implicit array representation to obtain a corresponding second implicit array representation; Inputting the second implicit array representation into the downsampling module, and downsampling the second implicit array representation using the downsampling module; The data linkage type is identified based on the downsampling result to obtain the linkage type detection result.

8. The method according to claim 1, characterized in that The data linkage type of the computer room monitoring data stream is detected based on the key data detection result and the pre-optimized target linkage type recognition model, and the linkage type detection result is obtained, including: Determine a plurality of estimated key data corresponding to the computer room monitoring data stream based on the key data detection result; Determine target device data from the plurality of estimated key data, and form an estimated key data group with the target device data and each of the remaining estimated key data; Determine in the data representation array the array representation sequences corresponding to each of the estimated key data groups; Inputting each of the array representation sequences into the target linkage type recognition model to perform linkage type detection, and obtaining linkage type detection results corresponding to each of the estimated key data groups; The energy-saving control strategy formulation based on the key data detection result and the linkage type detection result includes: Determine the energy consumption key data and temperature key data in the computer room monitoring data stream based on the key data detection result; Determine the linkage relationship between the energy consumption key data and the temperature key data based on the linkage type detection result; Determining energy consumption influencing factors according to the values ​​of the key energy consumption data and the linkage relationship; Based on the energy consumption influencing factors, an energy-saving control strategy is formulated, wherein the energy-saving control strategy includes at least one of adjusting the operating state of the air-conditioning system, adjusting the server load distribution, and optimizing power management.

9. 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 8 is implemented.

Images