Data center air conditioning system energy-saving control method and system based on AI and fault diagnosis

Abstract

The invention relates to the technical field of system energy-saving control processing, in particular to a data center air conditioning system energy-saving control method and system based on AI and fault diagnosis, and the method comprises model prediction and fault tracing. The method solves the problem that the real-time temperature control requirement of the data center is difficult to realize, so that early warning and fault traceability cannot be realized, and has the beneficial technical effects of realizing early warning of faults and realizing traceability of fault points in combination with a contribution graph method.

Description

Technical Field

[0001] The present invention relates to the technical field of system energy-saving control processing, and particularly relates to an energy-saving control method and system for a data center air-conditioning system based on AI and fault diagnosis. Background Art

[0002] A large number of servers are installed in the data center, generating extremely high heat. To keep each server in the data center working properly, it is necessary to control the temperature of the data center so that the temperature of the servers remains in the performance range for efficient operation. Currently, the control of the data center air-conditioning system is mainly based on manual operation or preset fixed operation programs. At the same time, the current fault diagnosis methods for the data center air-conditioning system all require professional technicians to rely on the alarm signals that appear after a fault occurs for fault location. To sum up, the existing water-cooled air-conditioning system in the data center mainly relies on manual adjustment or fixed program adjustment, making it difficult to meet the real-time temperature control requirements of the data center, resulting in the problems of inability to give early warnings and trace the source of faults. Summary of the Invention

[0003] In view of the problems existing in the prior art, the present invention provides an energy-saving control method and system for a data center air-conditioning system based on AI and fault diagnosis.

[0004] To achieve the above object, the technical solutions adopted by the present invention are as follows:

[0005] In a first aspect, an energy-saving control method for a data center air-conditioning system based on AI and fault diagnosis includes:

[0006] Model prediction: Based on the collected temperature, pressure, and flow parameters of the pipeline flowing working medium and the temperature and humidity parameters of the internal and external environments of the data center, through model prediction by artificial neural network modeling, a control strategy is generated to optimize the operation of the main equipment of the water-cooled air-conditioning system.

[0007] Fault tracing: Based on the collected fault data, through a data-driven fault diagnosis method using principal component analysis, the important operation indicators of the system are monitored in real time, and the empirical data of manual fault diagnosis is superimposed to achieve early warning of faults, and the contribution graph method is combined to trace the fault points.

[0008] Preferably, the model prediction by artificial neural network modeling includes: data preprocessing, prediction model establishment, and model-based prediction.

[0009] The data preprocessing includes: based on the historical data set, through normalization processing, a normalized data set is generated, where the historical data set includes a training sample subset, a valid sample subset, and a test sample subset, and the proportion of the training sample subset is 70%, the proportion of the valid sample subset is 15%, and the proportion of the test sample subset is 15%;

[0010] The establishment of the prediction model includes: based on the normalized data set, through training with a multi-layer artificial neural network, a data prediction model is established;

[0011] The model-based prediction includes generating an optimal model prediction result and outputting the prediction accuracy through real-time prediction of the data prediction model based on the current data set.

[0012] Preferably, the training of the multi-layer artificial neural network includes:

[0013] S1: Construct a multi-layer artificial neural network;

[0014] S2: Determine the number of hidden neurons in the model;

[0015] S3: Determine the corresponding optimization algorithm and loss function for each layer of the neural network according to the specific model.

[0016] Preferably, the construction of the multi-layer artificial neural network includes: constructing multiple neuron structures and constructing a multi-layer artificial neural network;

[0017] The construction of multiple neuron structures includes: based on the input data of the neurons in the previous layer output to the neurons in the next layer, the magnitude of the output value is controlled by the activation function on the neuron, where the output value is a non-linear value, and the value is obtained through the activation function, and it is judged whether to activate the neuron according to the limit value;

[0018] The construction of the multi-layer artificial neural network includes: based on the input signal received by the input layer, through processing by the hidden layer, an output signal is sent at the output layer, where the multi-layer artificial neural network is a network structure with three input layers and two output layers, and the network structure corresponds to the hot operating conditions and energy consumption levels of the data center respectively, and the hidden layer is a network layer composed of a large number of neurons arranged in parallel;

[0019] Preferably, the determination of the number of hidden neurons in the model includes:

[0020] Based on the determined desired boundary for separating classifications, each time a part of the line segments is connected according to the designer's design, and each time a hidden layer is added, where the desired boundary is a group of line segments, and the number of line segments is equal to the number of hidden layer neurons in the first hidden layer, to seek to determine the optimal number of hidden neurons;

[0021] Preferably, determining the corresponding optimization algorithms and loss functions for each layer of the neural network according to the specific model includes:

[0022] The loss function is:

[0023]

[0024] where y is the actual value of the data center operation, f(x) is the predicted value of the multi-layer artificial neural network model, which is a hyperparameter learned through training, δ is the neural threshold, and the loss function is a compromise between the mean square error and the mean absolute error loss function. By introducing an error balance parameter on the basis of smoothness, the robustness to abnormal fitting values is maintained;

[0025] The optimization algorithm includes: based on a small batch of data samples, updating the model through gradients to obtain the final predicted value; where the gradient calculation includes: based on the zero time step before the start of iteration and randomly initialized parameters, through the time step model, in the current time step, randomly and uniformly sample a small batch of samples composed of training data sample indices from the small batch. Each sample in the small batch is obtained by repeated sampling or non-repeated sampling. Repeated sampling allows repeated samples in the same small batch, while non-repeated sampling does not allow repeated samples in the same small batch.

[0026] The gradient update model is:

[0027]

[0028] where |B| represents the number of samples in the small batch, g t is the model gradient, is the predicted value of the network model of the small batch of samples, is the final predicted value of the gradient update model; i ∈ Bt is used for the small batch of samples composed of training data sample indices.

[0029] Preferably, the real-time prediction through the data prediction model includes:

[0030] Preprocessing the data: Based on a dataset containing time and acceleration data, input the prediction dataset into the data prediction model through low-pass filtering and normalization processing;

[0031]

[0032] where ax is the vector of filtered original signal values, min(ax) is the minimum value in the vector ax, max(ax) is the maximum value in the vector ax, ax normis the signal value within the range of 0 and 1 after normalization.

[0033] Data real-time prediction: Based on the predicted data set after standardization processing, through weighted summation and non-linear transformation, the final predicted result after anti-normalization processing is output; the predicted result is a time series corresponding to the length of the input data;

[0034] Predicted result output: Based on the predicted result, through the analysis of the predicted result, the prediction accuracy based on physical interpretation is obtained.

[0035] Preferably, the data-driven fault diagnosis method based on principal component analysis includes:

[0036] Based on the given training samples, through the decomposition of the covariance matrix, eigenvalues and eigenvectors are obtained;

[0037] Based on the eigenvalues and eigenvectors, through the calculation of the online diagnosis samples, the number of principal components of the diagnosis samples is obtained;

[0038] Based on the number of principal components of the diagnosis samples, through the calculation of the index control limits, the index control values are obtained;

[0039] Judge the index control value. If the index control value is greater than the normal control value, it is determined that the system has a fault; otherwise, it is determined that the system is working normally.

[0040] Preferably, the data-driven fault diagnosis method based on principal component analysis further includes:

[0041] Fault method traceability:

[0042] Based on the index control value, through the contribution graph model, the fault variables are identified; wherein, the contribution graph model includes quantifying the contribution degree of each process variable to the fault, and for each process variable, adding up the contributions of the scores leading to the out-of-control state;

[0043] Based on the over-limit scores of each variable, through the total contribution degree calculation model, the fault variables are prioritized;

[0044] Based on the fault variables sorted by priority, through the state value analysis model, the reason for the out-of-control state is determined to achieve fault traceability;

[0045] The contribution graph model is:

[0046]

[0047] wherein, the σ m is the corresponding eigenvalue, the μ j is the mean value of variable j, the t test,m is the diagnosis sample; the p m,jFor the contribution of each process variable to the fault, the X j Each process variable; the cont m,j Is the contribution degree of the process variable to the fault;

[0048] The total contribution degree calculation model is:

[0049]

[0050] The CONT j Is the total contribution degree of the process variable to the fault.

[0051] Meanwhile, the present invention also provides an energy-saving control system for a data center air-conditioning system based on AI and fault diagnosis, including:

[0052] Based on the above energy-saving control method, the energy-saving control system includes:

[0053] Model prediction module: Based on the collected temperature, pressure, and flow parameters of the pipeline flowing working medium and the temperature and humidity parameters of the internal and external environments of the data center, through model prediction by artificial neural network modeling, generate a control strategy to achieve the operation optimization of the main equipment of the water-cooled air-conditioning system;

[0054] Fault traceability module: Based on the collected fault data, through a data-driven fault diagnosis method of principal component analysis, real-time monitor the important operation indicators of the system, superimpose the empirical data of artificial fault diagnosis, achieve early warning of faults, and combine the contribution graph method to achieve the traceability of the fault point.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] This energy-saving control method includes two main steps. Through model prediction by artificial neural network modeling, a control strategy is generated to optimize the operation of each main device in the water-cooled air-conditioning system. At the same time, important operation indicators of the system are monitored in real time, and empirical data of artificial fault diagnosis is superimposed to achieve early warning of faults, and the contribution graph method is combined to trace the fault points. Among them, the two steps of this method include: model prediction and fault tracing. Based on the above two steps, the present invention provides an energy-saving control method for a data center air-conditioning system based on AI and fault diagnosis. By using the AI energy-saving control algorithm, temperature, pressure, and flow sensors are used to collect the temperature, pressure, and flow parameters of the pipeline flowing working medium and the temperature and humidity parameters of the internal and external environments of the data center. Through the method of artificial neural network modeling, a control strategy is generated to optimize the operation of each main device in the water-cooled air-conditioning system. At the same time, for the collected data, a data-driven fault diagnosis method is used to monitor important operation indicators of the system in real time, and empirical data of artificial fault diagnosis is superimposed to achieve early warning of faults, and the contribution graph method is combined to trace the fault points. Compared with the traditional control methods of pure manual or fixed-program water-cooled hosts, the control method proposed by the present invention is more reliable and intelligent, avoiding the problems and errors easily occurring in traditional adjustment methods. Through verification in a pilot computer room, the method proposed by the present invention can completely replace manual adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is the overall flowchart of Embodiment 1 of the present invention;

[0058] Figure 2 is the specific flowchart of Embodiment 1 of the present invention;

[0059] Figure 3 is the system topology diagram of Embodiment 1 of the present invention;

[0060] Figure 4 is the system step-by-step diagram of Embodiment 1 of the present invention;

[0061] Figure 5 is the system flowchart of Embodiment 1 of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0062] It should be noted that the methods used in the present invention are all conventional methods without special regulations; the raw materials and devices used are all conventional commercially available products without special regulations, and their sources are not specifically limited.

[0063] The present invention will be further described below in conjunction with specific embodiments, but the protection scope of the present invention is not limited thereto.

[0064] Figure 1 is the overall flowchart according to the embodiment of the present invention, Figure 2This is a specific flowchart according to an embodiment of the present invention. As Figure 1 , 2 shown, the embodiment of the present invention includes a claim: an energy-saving control method for a data center air-conditioning system based on AI and fault diagnosis, including:

[0065] As Figures 1 to 5 shown, model prediction S101: Based on the collected temperature, pressure, and flow rate parameters of the pipeline flowing working medium and the temperature and humidity parameters of the internal and external environments of the data center, through model prediction by artificial neural network modeling, generate a control strategy to achieve the operation optimization S201 of the main equipment of the water-cooled air-conditioning system;

[0066] Fault traceability S102: Based on the collected fault data, through a data-driven fault diagnosis method based on principal component analysis, real-time monitor the important operation indicators of the system, superimpose the empirical data of artificial fault diagnosis, achieve early warning of faults, and combine the contribution graph method to achieve the traceability of the fault point S202.

[0067] The model prediction by artificial neural network modeling includes: data preprocessing, prediction model establishment, and model prediction based on the model;

[0068] The data preprocessing includes: Based on the historical data set, through normalization processing, generate a normalized data set, where the historical data set includes a training sample subset, a valid sample subset, and a test sample subset, where the proportion of the training sample subset is 70%, the proportion of the valid sample subset is 15%, and the proportion of the test sample subset is 15%;

[0069] The prediction model establishment includes: Based on the normalized data set, through multi-layer artificial neural network training, establish a data prediction model;

[0070] The model prediction based on the model includes generating an optimal model prediction result based on the current data set through real-time prediction of the data prediction model and outputting the prediction accuracy.

[0071] The multi-layer artificial neural network training includes:

[0072] S1: Construct a multi-layer artificial neural network;

[0073] S2: Determine the number of hidden neurons in the model;

[0074] S3: According to the specific model, determine the corresponding optimization algorithm and loss function for each layer of the neural network.

[0075] As Figures 1 to 5 shown, the construction of the multi-layer artificial neural network includes: constructing multiple neuron structures and constructing a multi-layer artificial neural network;

[0076] The construction of multiple neuron structures includes: based on the input data output from the neurons of the previous layer network to the neurons of the next layer, controlling the magnitude of the output value through the activation function on the neuron, where the output value is a non-linear value, and the value is obtained through the activation function, and it is judged whether to activate the neuron according to the limit value;

[0077] The construction of a multi-layer artificial neural network includes: based on the input signal received by the input layer, processed through the hidden layer, and the output signal is sent at the output layer, where the multi-layer artificial neural network is a network structure with three input layers and two output layers, and the network structure corresponds to the hot operation conditions and energy consumption levels of the data center respectively, and the hidden layer is a network layer composed of a large number of neurons arranged in parallel;

[0078] The determination of the number of hidden neurons in the model includes:

[0079] Based on the determined expected boundaries for separating and classifying, each time connect some line segments according to the designer's design, and each time add a hidden layer, where the expected boundaries are a set of line segments, and the number of line segments is equal to the number of hidden layer neurons in the first hidden layer, in order to seek to achieve the determined optimal number of hidden neurons;

[0080] The determination of the corresponding optimization algorithms and loss functions for each layer of the neural network according to the specific model includes:

[0081] The loss function is:

[0082]

[0083] Among them, y is the actual value of the operation of the data center, f(x) is the predicted value of the multi-layer artificial neural network model, which is a hyperparameter learned through training, δ is the neural threshold, where the loss function is a compromise between the mean square error and the mean absolute error loss function, and by introducing an error balance parameter on the basis of smoothness, the robustness to abnormal fitting values is maintained;

[0084] The optimization algorithm includes: based on small batches of data samples, updating the model through gradients to obtain the final predicted value; among them, the gradient calculation includes: based on the zero time step before the start of iteration and randomly initialized parameters, through the time step model, in the current time step, a small batch sample composed of indexes of training data samples is randomly and uniformly sampled by mini-batch stochastic gradient descent, where the small batch sample obtains each sample in a small batch through repeated sampling or non-repeated sampling, where repeated sampling allows repeated samples to appear in the same small batch, while non-repeated sampling does not allow repeated samples to appear in the same small batch.

[0085] The gradient update model is:

[0086]

[0087] Among them, |B| represents the number of samples in a small batch, and g t is the model gradient, and is the predicted value of the network model for the small batch of samples, and is the final predicted value of the gradient update model; i ∈ Bt is the small batch of samples composed of training data sample indices.

[0088] The real-time prediction through the data prediction model includes:

[0089] Preprocessing data: Based on a dataset containing time and acceleration data, through low-pass filtering and normalization processing, input the prediction dataset into the data prediction model;

[0090]

[0091] Among them, ax is the vector of the filtered original signal values, min(ax) is the minimum value in the vector ax, max(ax) is the maximum value in the vector ax, and ax norm is the signal value within the range of 0 and 1 after normalization.

[0092] Data real-time prediction: Based on the standardized prediction dataset, through weighted summation and non-linear transformation, output the final prediction result after anti-normalization processing; the prediction result is a time series corresponding to the length of the input data;

[0093] Prediction result output: Based on the prediction result, through prediction result analysis, obtain the prediction accuracy based on physical interpretation.

[0094] As Figures 1 to 5 shown, the data-driven fault diagnosis method through principal component analysis includes:

[0095] Based on the given training samples, through the decomposition of the covariance matrix, obtain the eigenvalues and eigenvectors;

[0096] Based on the eigenvalues and eigenvectors, through online diagnosis sample calculation, obtain the number of principal components of the diagnosis samples;

[0097] Based on the number of principal components of the diagnosis samples, through index control limit calculation, obtain the index control value;

[0098] Judge the index control value. If the index control value is greater than the normal control value, it is determined that the system has a fault; otherwise, it is determined that the system is working normally.

[0099] As Figures 1 to 5 shown, the data-driven fault diagnosis method through principal component analysis also includes:

[0100] Fault tracing method:

[0101] Based on the index control value, through the contribution graph model, identify the fault variables; wherein, the contribution graph model includes quantifying the contribution degree of each process variable to the fault, and for each process variable, adding up the contributions of the scores that lead to the out-of-control state;

[0102] Based on the over-limit scores of each variable, through the total contribution degree calculation model, prioritize the fault variables;

[0103] Based on the prioritized fault variables, through the state value analysis model, determine the cause of the out-of-control state and achieve fault tracing;

[0104] The contribution graph model is:

[0105]

[0106] Wherein, the σ m is the corresponding eigenvalue, the μ j is the mean value of variable j, the t test,m is the diagnostic sample; the p m,j is the contribution of each process variable to the fault, the X j is each process variable; the cont m,j is the contribution degree of the process variable to the fault;

[0107] The total contribution degree calculation model is:

[0108]

[0109] The CONT j is the total contribution degree of the process variable to the fault.

[0110] Meanwhile, the present invention also provides an energy-saving control system for a data center air conditioning system based on AI and fault diagnosis, including:

[0111] Based on the energy-saving control method described in claims 1-9, the energy-saving control system includes:

[0112] Model prediction module: Based on the collected temperature, pressure, and flow parameters of the pipeline flowing working medium and the temperature and humidity parameters of the internal and external environments of the data center, through model prediction using artificial neural network modeling, generate a control strategy to achieve the operation optimization of each main equipment of the water-cooled air conditioning system S201;

[0113] Fault tracing module: Based on the collected fault data, the important operating indicators of the system are monitored in real time through the data-driven fault diagnosis method based on principal component analysis. The empirical data of manual fault diagnosis is superimposed to achieve early warning of faults, and the contribution graph method is combined to achieve fault point tracing S202.

[0114] Embodiment 1:

[0115] Combine the following Figure 1 , 2 The working principle of the energy-saving control method shown in the embodiment is explained.

[0116] like Figure 1 As shown, the energy-saving control method includes two main steps, among which the first step is model prediction; the second step is fault tracing; specifically, it also includes two specific steps, and corresponds to the above two steps, among which the first step corresponds to the first specific step, namely: based on the collected temperature, pressure, flow parameters of the pipeline flowing medium and the temperature and humidity parameters of the environment inside and outside the data center, the model prediction of artificial neural network modeling is used to generate a control strategy to achieve operation optimization of each major equipment of the water-cooled air-conditioning system; the second step corresponds to the second specific step, namely: based on the collected fault data, through the data-driven fault diagnosis method based on principal component analysis, the important operating indicators of the system are monitored in real time, the empirical data of artificial fault diagnosis is superimposed, the early warning of the fault is achieved, and the contribution diagram method is combined to achieve the tracing of the fault point; based on the above two steps and two specific steps, the present invention proposes a first step: the temperature of the pipeline flowing medium is collected through temperature, pressure and flow sensors , pressure, flow parameters and temperature and humidity parameters of the environment inside and outside the data center; the second step: input the parameters obtained in the first step into the AI energy-saving algorithm model built into the energy-saving control cabinet based on the system, generate a control strategy according to the "temperature, pressure and flow" results, and output the "valve opening, flow control, etc." control results; the intelligent energy-saving system provided by the present invention is advanced, safe, economical and easy to use, etc., which can effectively improve the energy-saving level of the data center, reduce the system operating load and extend the service life of the system. At the same time, by adopting the ANN model to model the intelligent energy-saving system, compared with the traditional method, the method of the present invention has considerable improvements in convergence speed, accuracy, network generalization ability, algorithm stability and timeliness of calculation. The fault detection method based on PCA and contribution graph realizes early warning of faults and fault tracing, reduces the cost of manual detection and troubleshooting, and avoids the economic losses caused by faults to a certain extent, thereby improving the safety of the air-conditioning system.

[0117] Combine the following Figure 1 , 2 The working principle of the energy-saving control method shown in the embodiment is explained.

[0118] As Figure 1 , 2 shown, the model prediction by artificial neural network modeling includes three aspects: (1) data preprocessing, (2) prediction model establishment, and (3) model-based prediction. Among them, the data preprocessing: based on the historical data set, through normalization processing, a normalized data set is generated. Among them, the historical data set includes a training sample subset, a valid sample subset, and a test sample subset. Among them, the proportion of the training sample subset is 70%, the proportion of the valid sample subset is 15%, and the proportion of the test sample subset is 15%. The prediction model establishment: based on the normalized data set, through multi-layer artificial neural network training, a data prediction model is established. The model-based prediction: based on the current data set, through real-time prediction of the data prediction model, an optimal model prediction result is generated and the prediction accuracy is output.

[0119] In the above steps, the operating parameters of the key links and positions of the data center air conditioning system model, such as load, temperature, humidity, pressure, flow rate, etc., can be used as input layer variables to construct an artificial neural network; the output variable is the thermal operating condition (thermal environment level) and energy consumption level of the data center. Specifically, a machine model for predicting the thermal environment level and energy consumption level of the data center based on operating parameters can be constructed in advance. In this solution, the machine model for prediction can be trained based on training data, such as historical operating parameters and historical thermal environment levels and energy consumption levels, with an artificial neural network (Artificial Neural Networks, abbreviated as ANN) as the framework. Here, the artificial neural network is also simply referred to as the neural network (NN) or the connection model (Connection Model). It is an algorithmic mathematical model that mimics the behavioral characteristics of animal neural networks and performs distributed parallel information processing. This network relies on the complexity of the system and adjusts the relationship between a large number of internal nodes to achieve the purpose of establishing a model to process information. The typical structure of an artificial neural network is an input layer, some hidden layers, and an output layer. The layers of the network are connected to each other in sequence from the input layer to the output layer and the connections are weighted. When a pair of learning patterns is provided to the network, the activation values of the neurons are propagated from the input layer through the intermediate layer to the output layer, and each neuron in the output layer obtains the input response of the network. In the direction of reducing the error between the expected output and the actual output value, the connection weights of each layer are corrected layer by layer from the output layer through the intermediate layer, and finally return to the input layer. As this error backpropagation correction continues, the correct rate of the network's response to the input pattern also continuously increases. When the network reaches the set number of iterations or the output error reaches the allowable range, the network stops running;

[0120] Step 1: Data preprocessing. Determine the input and output variables of the neural network, and based on the idea of cross-validation and the characteristics of the new cooperative co-evolution model, randomly divide the dataset (i.e., historical operating parameters, historical thermal environment levels, and energy consumption levels) into 3 subsets: a training sample subset, a validation sample subset, and a test sample subset, which are used for network learning, effectiveness verification, and testing respectively. Compared with the data processing methods in traditional neural networks, this method can better improve the generalization ability of the neural network. 70% is used as the training set; 15% is used as the validation sample set; 15% is used as the test set, which are used for training, testing, and verification respectively. Subsequently, the ANN model data can be trained, tested, and verified according to the characteristics of the dataset containing the original system. Preferably, feature scaling processing, i.e., normalization processing, is performed before the data is input into the artificial neural network to improve the efficiency of the ANN model.

[0121] The working principle of the energy-saving control method shown in the following combined with Figure 1 、 2 the embodiments will be described.

[0122] As Figure 1 、 2 shown, the training of the multi-layer artificial neural network includes three steps, namely the first step: constructing a multi-layer artificial neural network; the second step: determining the number of hidden neurons in the model; the third step: determining the corresponding optimization algorithm and loss function for each layer of the neural network according to the specific model; the construction of the multi-layer artificial neural network includes: constructing multiple neuron structures and constructing a multi-layer artificial neural network; the determination of the number of hidden neurons in the model includes: based on the determined expected boundary of the separated classification, each time connect some line segments according to the designer's design, and each time add a hidden layer. The determination of the corresponding optimization algorithm and loss function for each layer of the neural network according to the specific model includes:

[0123] The loss function is:

[0124]

[0125] The optimization algorithm includes: based on a mini-batch of data samples, update the model through gradients to obtain the final predicted value; among them, the gradient calculation includes: based on the zero time step before the start of iteration and randomly initialized parameters, through the time step model, in the current time step, randomly and uniformly sample a mini-batch of samples composed of training data sample indices from the mini-batch, where the mini-batch samples are obtained by repeated sampling or non-repeated sampling to get each sample in a mini-batch. Repeated sampling allows repeated samples in the same mini-batch, while non-repeated sampling does not allow repeated samples in the same mini-batch.

[0126] The gradient update model is:

[0127]

[0128] Step 2: Establish an artificial intelligence model based on ANN: The first specific step: ANN consists of multiple neuron structures. Each layer of neurons has inputs and outputs, and each layer is composed of multiple neurons. The output of the neurons in the upper layer of the network is the input of the neurons in the lower layer. The input data controls the magnitude of the output value through the activation function on the neurons. This output value is a non-linear value, and the value is obtained through the activation function, and it is judged whether to activate the neuron according to the limit value. Generally, a multi-layer artificial neural network ANN consists of an input layer, an output layer, and a hidden layer. Input layer (InputLayer): Receives the input signal as the input of the input layer. Output layer (Output Layer): After the signal is transmitted, inner-producted, and activated by the neurons in the neural network, an output signal is formed for output. Hidden layer (Hidden Layer): The hidden layer is also called the hidden layer. It is located between the input layer and the output layer and is a network layer composed of a large number of neurons arranged in parallel. Usually, an artificial neural network can have multiple hidden layers. Through sensitivity analysis, an ANN model with three input layers and two output layers is established here, corresponding to the hot operating conditions (hot environment level) and energy consumption level of the data center respectively; The second specific step: Determine the number of hidden neurons in the ANN model. In the classification problem, the general step for determining the hidden neurons is to first draw the expected boundary for separating the classification on the data, represent the expected boundary as a set of line segments, and the number of line segments is equal to the number of hidden layer neurons in the first hidden layer. Finally, some of these line segments are connected (which line segments to choose each time depends on the designer), and a new hidden layer is added. That is to say, every time some line segments are connected, a new hidden layer is added, and the number of connections each time is equal to the number of neurons in the newly added hidden layer. This system selects the relatively simple trial-and-error method, which pursues the goal by continuously experimenting and eliminating errors. The trial-and-error method is a purely empirical learning method. The main body applying the trial-and-error method changes the parameters of the ANN model intermittently or continuously, and tests the responses made by the model output to seek to determine the optimal number of hidden neurons. Since the number of input and output variables in this system is small and the model is simple, the optimal number of hidden neurons in each layer can be obtained through the simple trial-and-error method; The third specific step: Determine the corresponding optimization algorithm and loss function for the neural network according to the specific model. The loss function selects the Huber loss function, which is a compromise between the common mean square error and mean absolute error loss functions. It can achieve robustness to abnormal fitting values by introducing an error balance parameter while ensuring smoothness. The calculation formula is as follows:

[0129]

[0130] where y is the actual value of the data center operation, f(x) is the predicted value of the ANN model, and δ is a hyperparameter learned through training; the optimization algorithm uses the mini-batch stochastic gradient descent optimization algorithm, which is an improvement of the gradient descent algorithm that combines the advantages of batch gradient descent and stochastic gradient descent. By calculating the gradient using a small batch of data each time, a balance is achieved between computational efficiency and stability. Assume the final predicted value is f(x). The time step before the start of iteration is set to 0. The model parameters at this time step are obtained by random initialization. In each subsequent time step t > 0, the mini-batch stochastic gradient descent randomly and uniformly samples a mini-batch B consisting of indices of training data samples t . We can obtain each sample in a mini-batch by resampling or non-resampling. The former allows duplicate samples in the same mini-batch, while the latter does not and is more common. For either of these two methods, the model gradient g t is updated according to the following formula:

[0131]

[0132] where |B| represents the batch size, i.e., the number of samples in the mini-batch

[0133] The working principle of the energy-saving control method shown in the following combined with Figure 1 and 2 the embodiments will be described

[0134] As shown in Figure 1 and 2 , the real-time prediction by the data prediction model includes: preprocessing the data: based on a dataset containing time and acceleration data, through low-pass filtering and normalization processing, inputting the prediction dataset into the data prediction model Real-time data prediction: based on the normalized prediction dataset, through weighted summation and non-linear transformation, outputting the final prediction result after inverse normalization processing; the prediction result is a time series corresponding to the length of the input data; prediction result output: based on the prediction result, through prediction result analysis, obtaining the prediction accuracy based on physical interpretation

[0135] Step 3: Model Prediction: Predict the above model. Consider using commercial software for model prediction, such as Matlab, etc., for solving. Among them, the "model" refers to the artificial neural network model based on ANN established in Step 2. Among them, for the "model prediction", choose to use the M language to call the ANN model to predict new data. First, it is necessary to fully preprocess the original input data. Read the dataset containing time and acceleration data from the Excel file. In order to remove the interference of high-frequency noise on the prediction, use a 4th-order Butterworth low-pass filter to filter the signal, and set the filter parameters as the sampling frequency of 20 Hz and the cut-off frequency of 2 Hz. After filtering, in order to ensure that the neural network input data is within an appropriate numerical range, the acceleration signal is linearly normalized and scaled between 0 and 1. The normalized data can be regarded as the input vector of the neural network for subsequent prediction analysis. The normalization formula is as follows:

[0136]

[0137] where ax represents the original signal value after filtering (a vector), min(ax) represents the minimum value in the vector ax, max(ax) represents the maximum value in the vector ax, and ax norm represents the normalized signal value (ranging between 0 and 1); Subsequently, create a neural network model (ANN). This neural network model contains a hidden layer with 10 neurons, and uses the hyperbolic tangent (tansig) activation function to capture the non-linear features in the input signal. During the prediction process, the neural network first standardizes the input data, then sends it to the hidden layer, and outputs to the next layer after weighted summation and non-linear transformation. The output layer calculates the final prediction result using a linear combination method and performs anti-normalization on it to restore it to the scale of the original physical quantity. The entire neural network structure is fixed, and the input data no longer affects the network parameters. Therefore, only one forward propagation needs to be run to complete the prediction of new data; After completing the above model definition and data preparation, package the normalized new input data into the cell format as required by the neural network, call the myNeuralNetworkFunction(X) function, input the data for prediction, and the returned result Y is the model output (predicted value). The prediction result is a time series corresponding to the length of the input data, which is used for subsequent physical interpretation and fault analysis application scenarios; Step 4: Result Output. Through model prediction, the optimal model prediction result will be obtained. At the same time, the corresponding prediction output will also be obtained to achieve model prediction. In addition, the loss function calculates the accuracy of this ANN model: Through the above steps, the prediction of the dataset is completed, and finally, by calculating the accuracy of this prediction, the accuracy of this prediction system is represented.

[0138] The working principle of the energy-saving control method shown in the following embodiments will be described below. Figure 1 and 2 The working principle of the energy-saving control method shown in the embodiments will be described.

[0139] As Figure 1 and 2 shown, the data-driven fault diagnosis method based on principal component analysis includes 4 steps. Among them, the first step: based on the given training samples, through the decomposition of the covariance matrix, eigenvalues and eigenvectors are obtained; the second step: based on the eigenvalues and eigenvectors, through the calculation of the online diagnosis samples, the number of principal components of the diagnosis samples is obtained; the third step: based on the number of principal components of the diagnosis samples, through the calculation of the index control limit, the index control value is obtained; the fourth step: judge the index control value. If the index control value is greater than the normal control value, it is determined that the system has a fault, otherwise it is determined that the system is working normally;

[0140] While the present invention models the neural network, according to the key original data to be monitored such as the temperature, pressure, and flow rate of the data center system, a PCA fault diagnosis method for online real-time fault prediction is established, and the fault point tracing is realized by combining the contribution graph calculation method. The present invention is an energy-saving control method for a data center air-conditioning system based on AI model reduction and fault diagnosis. Through the trained model, according to the operating parameters of the data center air-conditioning system, the thermal environment level and energy consumption level of the data center are predicted and adjusted, and through the real-time monitoring of key indicators, the early warning and tracing of fault points are realized;

[0141] PCA fault diagnosis: Given training samples x1, x2,....., x N ∈R M , where N is the number of samples and M is the number of variables. The covariance matrix C of the data is calculated as follows:

[0142]

[0143] The eigenvalue decomposition of the covariance matrix is calculated as follows:

[0144]

[0145] where λ is the eigenvalue and v represents the eigenvector; the principal components of the training sample set are calculated as follows:

[0146]

[0147] where t k,m represents the m-th principal component of the k-th training sample x k , v m is the m-th eigenvector, and p is the number of principal components. Similarly, the principal components of the online diagnosis sample x test are calculated as follows:

[0148]

[0149] where t test,m represents the m-th principal component of the online diagnostic sample x test and v m is the m-th eigenvector, and p is the number of principal components determined through the above training process; the T 2 index and the T 2 index control limits are calculated as follows:

[0150]

[0151] where Λ is the eigenvalue diagonal matrix. α is the confidence level determined artificially; if, then it is judged as a fault, otherwise, it is normal.

[0152] The working principle of the energy-saving control method shown in the following combined with Figure 1 and 2 the embodiments will be described.

[0153] As Figure 1 and 2 shown, the data-driven fault diagnosis method based on principal component analysis further includes: fault method traceability, where the fault method traceability includes a total of three steps. The first step: Based on the index control value, through the contribution graph model, identify the fault variables; where the contribution graph model includes quantifying the contribution degree of each process variable to the fault, and for each process variable, adding up the contributions of the scores that lead to the out-of-control state. The second step: Based on the over-limit scores of each variable, through the total contribution degree calculation model, rank the fault variables in terms of priority. The third step: Based on the fault variables ranked in terms of priority, through the state value analysis model, determine the cause of the out-of-control state to achieve fault traceability;

[0154] The contribution graph model is:

[0155]

[0156] The total contribution degree calculation model is:

[0157]

[0158] Step six: Fault traceability by the contribution graph method: After diagnosing a fault, conduct a traceability study on the fault point, and use the contribution graph method to identify the fault variables. This method is based on quantifying the contribution degree of each process variable to the fault, and for each process variable, adding up the contributions of the scores that lead to the out-of-control state. Calculate the over-limit score of each variable:

[0159]

[0160] where σ m is the corresponding eigenvalue, and μ j is the mean value of variable j

[0161] Calculate the total contribution degree

[0162]

[0163] Based on the total contribution value CONT j Rank the fault variables by priority to facilitate engineers to immediately focus on those CONT j variables with high values, and use process knowledge to determine the cause of the out-of-control state.

[0164] The working principle of the energy-saving control method shown in the following combined Figures 1 - 5 with the embodiments will be described.

[0165] The present invention also provides an energy-saving control system for a data center air-conditioning system based on AI and fault diagnosis. The energy-saving control system includes a model prediction module and a fault tracing module. This solution is implemented through an energy-saving control cabinet, which includes a control unit, a communication module, several protection units, and a set of system software. The control cabinet connects the control of the chilled water pump, the control of the terminal valve, and the start-stop control of the original air conditioner through on-site communication cables. Through communication with each control cabinet, comprehensive data of the air-conditioning system can be collected to achieve centralized monitoring, control, and management of the operation of the air-conditioning system. In this solution, a control strategy can be generated by the control unit, and the water pump can be controlled to work at the adjusted speed through a communication model, etc. At the same time, early warning of fault points and tracing and investigation of fault points can be carried out through a data-driven fault detection method, so as to comprehensively achieve the purpose of safe, efficient, and energy-saving. The artificial intelligence modeling method based on ANN in this system does not require complex parameter design and iterative calculation of equations, can significantly shorten the calculation time, and improve the modeling efficiency. Under the same computing resources, the average calculation time of the ANN model is 0.064s, the average fitting time of the response surface analysis method is 0.6s, and the average calculation time of the traditional modeling model is 46086s. The modeling method of ANN greatly reduces the time cost of modeling and improves the timeliness of calculation. For the collected data, a PCA-based data-driven fault diagnosis method is used to achieve early warning and diagnosis of faults 10-15 minutes in advance, which greatly saves the cost required for manual detection and troubleshooting, avoids economic losses caused by faults, and improves system safety. The fault tracing method of the contribution diagram only relies on data and realizes the specific tracing of fault points by capturing the characteristic changes in the fault data, without manual checking one by one.

[0166] Embodiment 2:

[0167] The method of the present invention will be introduced in detail with specific examples as follows:

[0168] As Figures 1 to 5 shown, the current water-cooled air-conditioning system in the data center is mainly controlled manually or by using a preset fixed operation program. Once the load in the computer room or the internal components of the air-conditioning system change, it is difficult to meet the temperature control requirements of the data center in real time, and it is easy to cause slow adjustment or energy waste, affecting the performance and safe operation of the server. The present invention intends to use an AI algorithm to model the water-cooled air-conditioning system in the data center by using an artificial neural network, and generate a control strategy through learning and training of relevant data to achieve automatic control of the system. At the same time, for the collected data, a fault early warning and fault tracing module using the PCA and contribution graph methods is established to realize real-time monitoring of the entire water-cooled air-conditioning system in the data center, further improving the safety and reliability of the system operation.

[0169] In the control strategy of this solution, the idea of cascade control is mainly adopted.

[0170] The difference between cascade control and simple control is that its structure contains two closed loops, namely the secondary loop and the primary loop, which can also be called the inner loop and the outer loop. Its control characteristics are: (1) The secondary loop acts quickly to overcome the secondary disturbance; (2) Since the inner loop improves the dynamic characteristics of the object, the gain of the primary regulator is increased to improve the working efficiency of the system.

[0171] The implementation of cascade control requires a set of intelligent control cabinets, control units in the cabinets, communication modules, several protection units, and a set of system software. The control cabinet is used for on-site communication cables to connect the control of the chilled water pump, the control of the terminal valve, and the start-stop control of the original air conditioner. Through communication with each control cabinet, comprehensive data of the air-conditioning system can be collected to achieve centralized monitoring, control, and management of the operation of the air-conditioning system.

[0172] First step, according to the original operation data, while comparing the real-time state of the acquisition system, collect parameters such as temperature, pressure, and flow in the acquisition system, and calculate the thermal operation conditions and energy consumption level of the data center based on the data. Independently learn the peak and valley power consumption conditions of the current unit, user usage habits, etc., and conduct multiple repeated trainings.

[0173] Secondly, after obtaining the thermal operation conditions and energy consumption level of the data center, adjust the control strategy according to the thermal operation conditions and energy consumption level of the data center, and conduct independent learning and verification according to the actual operating conditions.

[0174] Third step: According to the learning process of the target objects (hybrid terminal control system and chilled water system), after the training is completed, the specific internal control logic scheme is as follows.

[0175] For the hybrid terminal control system, the control strategy for the set value of the total water supply pressure is:

[0176] (1) Find the most unfavorable loop;

[0177] (2) If the most unfavorable loop is under continuous control, measure its pressure; if it is on-off control, measure the return water temperature of the device;

[0178] (3) If the pressure at the end of the most unfavorable loop under continuous control is lower than the design value, or the supply-return water temperature difference at the end of the most unfavorable loop under on-off control is higher than the design value, then increase the total water supply pressure setting value;

[0179] (4) If the pressure at the end of the most unfavorable loop under continuous control is not lower than the design value and the supply-return water temperature difference at the end of the most unfavorable loop under on-off control is not higher than the design value, and there is a high pressure value at the end of continuous control or a small temperature difference at the end of on-off control, then decrease the total water supply pressure setting value.

[0180] (5) Otherwise, the total water supply pressure setting value remains unchanged.

[0181] For the chilled water pump, the control principle is:

[0182] According to the cascade control design, the tasks of the chilled water pump system control are mainly the following three parts:

[0183] (I) Ensure sufficient chilled water flow in the chiller to ensure the safe operation of the chiller

[0184] (II) Meet the water supply pressure setting value

[0185] (III) Minimize the energy consumption of the chilled water pump

[0186] Step 4: Collect the important performance indicators under normal operation of the system, establish a PCA principal component space model based on the normal data, and calculate the T2 and SPE control limits of the normal data.

[0187] Step 5: Online monitor all the important performance indicator data at each time point (i.e., calculate the T2 and SPE values corresponding to each time point), and compare them with the control limits of the normal data. If it exceeds the limit, it indicates that a fault has occurred or is about to occur.

[0188] Step 6: Introduce the method of the contribution plot to trace the fault point, calculate the contribution degree of each indicator to the fault point, conduct fault tracing, and find the most likely factor causing the fault.

[0189] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than a limitation on the protection scope of the present invention. Any simple modification or equivalent replacement made by those of ordinary skill in the art to the technical solution of the present invention shall not depart from the essence and scope of the technical solution of the present invention.

Claims

1. An energy-saving control method for a data center air conditioning system based on AI and fault diagnosis, characterized in that, Including: Model prediction: Based on the temperature, pressure, and flow rate parameters of the pipeline flowing working medium collected, as well as the temperature and humidity parameters of the internal and external environments of the data center, through model prediction by artificial neural network modeling, generate control strategies to achieve the operation optimization of each main equipment of the water-cooled air-conditioning system; Fault traceability: Based on the collected fault data, through a data-driven fault diagnosis method of principal component analysis, real-time monitor the important operation indicators of the system, superimpose the empirical data of artificial fault diagnosis, achieve early warning of faults, and combine the contribution graph method to achieve the traceability of the fault point.

2. The energy-saving control method according to claim 1, wherein, The model prediction by artificial neural network modeling includes: data preprocessing, prediction model establishment, and model-based prediction; The data preprocessing includes: Based on the historical data set, through normalization processing, generate a normalized data set, where the historical data set includes a training sample subset, a valid sample subset, and a test sample subset, and the proportion of the training sample subset is 70%, the proportion of the valid sample subset is 15%, and the proportion of the test sample subset is 15%; The prediction model establishment includes: Based on the normalized data set, through training of a multi-layer artificial neural network, establish a data prediction model; The model-based prediction includes: Based on the current data set, through real-time prediction of the data prediction model, generate the optimal model prediction result and output the prediction accuracy.

3. The energy-saving control method according to claim 2, wherein The training of the multi-layer artificial neural network includes: S1: Construct a multi-layer artificial neural network; S2: Determine the number of hidden neurons in the model; S3: Determine the corresponding optimization algorithm and loss function for each layer of the neural network according to the specific model.

4. The energy-saving control method according to claim 3, wherein The construction of the multi-layer artificial neural network includes: constructing multiple neuron structures and constructing a multi-layer artificial neural network; The construction of multiple neuron structures includes: Based on the input data of the neurons in the previous layer network output to the neurons in the next layer, control the magnitude of the output value through the activation function on the neuron, where the output value is a non-linear value, obtain the value through the activation function, and judge whether to activate the neuron according to the limit value; The construction of the multi-layer artificial neural network includes: Based on the input signal received by the input layer, through processing by the hidden layer, send an output signal at the output layer, where the multi-layer artificial neural network is a network structure with three input layers and two output layers, the network structure corresponds to the thermal operating conditions and energy consumption levels of the data center respectively, and the hidden layer is a network layer composed of a large number of neurons arranged in parallel.

5. The energy-saving control method according to claim 3, wherein The determination of the number of hidden neurons in the model includes: Based on the determined expected boundary for separating classifications, each time connect some line segments according to the designer's design, and each time add a hidden layer, where the expected boundary is a group of line segments, and the number of line segments is equal to the number of hidden layer neurons in the first hidden layer, to seek to achieve the determination of the optimal number of hidden neurons.

6. The energy-saving control method according to claim 3, wherein The determination of the corresponding optimization algorithm and loss function for each layer of the neural network according to the specific model includes: The loss function is: Wherein, y is the actual value of the data center operation, f(x) is the predicted value of the multi-layer artificial neural network model, which is a hyperparameter learned through training, and δ is the neural threshold. The loss function is a compromise between the mean square error and the mean absolute error loss function. By introducing an error balance parameter on the basis of smoothness, the robustness to abnormal fitting values is maintained; The optimization algorithm includes: based on a small batch of data samples, updating the model through gradients to obtain the final predicted value. The gradient calculation includes: based on the zero time step before the start of iteration and randomly initialized parameters, through the time step model, in the current time step, randomly and uniformly sample a small batch of samples composed of training data sample indices by mini-batch stochastic gradient descent. Each sample in a small batch is obtained by repeated sampling or non-repeated sampling. Repeated sampling allows repeated samples in the same small batch, while non-repeated sampling does not allow repeated samples in the same small batch; The gradient update model is: where |B| represents the number of samples in a mini-batch, and g t is the model gradient, and is the predicted value of the network model for the mini-batch samples, and is the final predicted value of the gradient update model; i ∈ Bt is a mini-batch sample composed of training data sample indices.

7. The energy-saving control method according to claim 2, wherein The real-time prediction through the data prediction model includes: Preprocessing data: Based on a data set containing time and acceleration data, through low-pass filtering and normalization processing, input the prediction data set into the data prediction model; wherein, ax is a vector of the filtered original signal values, min(ax) is the minimum value in the vector ax, max(ax) is the maximum value in the vector ax, and ax norm is a signal value with a range between 0 and 1 after normalization; Data real-time prediction: Based on the standardized prediction data set, through weighted summation and non-linear transformation, output the final prediction result after inverse normalization processing. The prediction result is a time series corresponding to the length of the input data; Prediction result output: Based on the prediction result, through prediction result analysis, obtain the prediction accuracy based on physical interpretation.

8. The energy-saving control method according to claim 1, characterized in that The data-driven fault diagnosis method through principal component analysis includes: Based on a given training sample, obtain eigenvalues and eigenvectors through the decomposition of the covariance matrix; Based on the eigenvalues and eigenvectors, obtain the number of principal components of the diagnostic sample through online diagnostic sample calculation; Based on the number of principal components of the diagnostic sample, obtain the index control value through index control limit calculation; Judge the index control value. If the index control value is greater than the normal control value, it is determined that the system has a fault, otherwise it is determined that the system is working normally.

9. The energy-saving control method according to claim 8, wherein The data-driven fault diagnosis method through principal component analysis further includes: Fault source tracing: Based on the index control value, identify the fault variables through the contribution graph model. The contribution graph model includes quantifying the contribution degree of each process variable to the fault, and for each process variable, adding up the scores of the contributions that lead to the out-of-control state; Based on the over-limit scores of each variable, prioritize the fault variables through the total contribution degree calculation model; Based on the prioritized fault variables, determine the cause of the out-of-control state through the state value analysis model to achieve fault source tracing; The contribution graph model is: Among them, the σ m is the corresponding eigenvalue, the μ j is the mean value of variable j, the t test,m p is the diagnostic sample; the p m,j is the contribution of each process variable to the fault, the X j is each process variable; the cont m,j is the contribution degree of the process variable to the fault; The total contribution degree calculation model is: The CONT j is the total contribution of the process variable to the fault.

10. An energy-saving control system for a data center air conditioning system based on AI and fault diagnosis, characterized in that, Including: Based on the energy-saving control method described in claims 1-9, the energy-saving control system includes: Model prediction module: Based on the temperature, pressure, and flow rate parameters of the pipeline flowing working medium collected, as well as the temperature and humidity parameters of the internal and external environments of the data center, through model prediction using artificial neural network modeling, generate control strategies to achieve operation optimization of the main equipment of the water-cooled air conditioning system; Fault tracing module: Based on the collected fault data, through a data-driven fault diagnosis method using principal component analysis, real-time monitor the important operation indicators of the system, superimpose the empirical data of artificial fault diagnosis, achieve early warning of faults, and combine the contribution graph method to trace the fault points.

Images