An air compressor operation, maintenance and tuning method and system based on Informer model

Abstract

The present invention relates to an air compressor operation and maintenance and optimization method and system based on an Informer model, wherein the method comprises obtaining parameter information in an air compressor management system, constructing an original training data set, determining the output and input parameters of the Informer model, dividing the original training data set, training based on the training set, calculating a loss function, obtaining loss adjustment model parameters, and stopping training when the loss corresponding to the validation set is lower than a preset value or reaches a preset number of cycles; inputting the air compressor station parameters under the working conditions into the model to obtain a strategy table to achieve the lowest operation and maintenance and optimization cost of the air compressor system. Compared with the prior art, the present invention optimizes the air compressor's air replenishment time and air replenishment pressure end-to-end by processing a neural network algorithm for long time series data, thereby determining the optimal scheduling strategy and achieving the lowest operation and maintenance and optimization cost of the air compressor system, avoiding the problems of slow speed, gradient disappearance, gradient explosion, etc. in the prior art when predicting long time series problems.

Description

Technical Field

[0001] The present invention relates to the field of intelligent operation, maintenance and optimization of air compressor systems, and in particular to a method and system for operation, maintenance and optimization of air compressors. Background Art

[0002] As a power energy consumer, air compressors are widely used in the industrial field and are an important energy source for industrial products. Intelligent operation and maintenance of air compressors is also a key part of modern air compressor system management. However, in actual application scenarios, the rational operation and maintenance of air compressors is often very difficult due to the wide distribution of air compressors, uneven management personnel levels, and different usage methods. Irrational operation and maintenance will lead to increased power consumption of air compressor operation and maintenance, weakened air compressor functions, and ultimately waste costs in air compressor system management. Therefore, it is necessary and urgent to design methods and systems for intelligent operation and maintenance and optimization of air compressors. Using intelligent operation and maintenance and optimization strategies for air compressors, air compressor system management can be completed uniformly in a short time, significantly reducing management costs while optimizing management results.

[0003] However, in the existing technology, solving the intelligent operation and optimization of air compressors is still an NP-hard problem. If you want to solve it accurately, it is difficult to solve the problem in polynomial time, and it takes a lot of time and computing power. Therefore, this method is difficult to apply to actual scenarios. At present, the main algorithms used for the intelligent operation and maintenance optimization of air compressor stations are search algorithms represented by TPE algorithms, heuristic algorithms represented by ant colony algorithms, and meta-heuristic algorithms represented by genetic algorithms. However, these algorithms also have problems such as slow operation and optimization speed, difficulty in finding better solutions, and differences between algorithm theoretical results and practical applications. The system for intelligent operation and optimization of air compressors based on the traditional Transformer model has problems such as slow speed, gradient disappearance, and gradient explosion when predicting long time series problems. Summary of the Invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method and system for air compressor operation, maintenance and tuning, wherein the input is the relevant parameters of air compressor scheduling (including time, temperature, and user needs), and the air compressor's air replenishment time and air replenishment pressure are optimized end-to-end by processing a neural network algorithm (i.e., an Informer model) for long time series data, thereby determining the optimal scheduling strategy and achieving the lowest operation and maintenance and optimization cost of the air compressor system, avoiding the problems of slow speed, gradient disappearance, gradient explosion, etc. in the prior art when predicting long time series problems.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] The purpose of the present invention is to provide an air compressor operation, maintenance and tuning method based on the Informer model, comprising the following steps:

[0007] Obtain parameter information from the air compressor management system and construct the original training data set;

[0008] Selecting target parameters to be optimized from the original training data set as output parameters of the informer model, using the remaining parameters in the original training data set as input parameters of the informer model, and using the corresponding times of the input parameters of the informer model as timing parameters in the input parameters;

[0009] Dividing the original training data set into a training set, a test set, and a validation set;

[0010] Randomly extract a portion of the input parameters from the training set and input them into the informer model to obtain informer model output parameters, compare the informer model output parameters with the training set input parameters, adjust the model parameters according to the difference between the two, and repeat the input, output, and comparison process until all the data in the training set is extracted. After one cycle is completed, extract the input parameters from the validation set and the test set and input them into the informer model to obtain output parameters, compare the output parameters with the output parameters in the validation set and the test set, calculate the loss function, and obtain the loss-adjusted model parameters.

[0011] When the loss corresponding to the validation set is lower than the preset value, or the preset number of cycles is reached, the training process is terminated, and the trained model is obtained. The model is validated and tested on the test set to evaluate the model effect;

[0012] The air compressor station parameters under the working conditions are input into the trained model to obtain a strategy table. The air compressor operation, maintenance and optimization are performed based on the strategy table to achieve the lowest operation, maintenance and optimization costs of the air compressor system.

[0013] Furthermore, the output parameters of the Informer model include the air supply pressure and air supply time of the air compressor;

[0014] The input parameters of the Informer model include time, temperature, and user needs.

[0015] Furthermore, the input parameter data set composed of the input parameters of the Informer model is:

[0016]

[0017] Where: n represents the number of data contained in the data set, m represents the number of system users, t i(i=1,2,...,n) represents the time when the i-th data record is recorded, T i (i=1,2,...,n) represents the temperature when the i-th data record is recorded, x i,j (i=1,2,...,n;j=1,2,...,m) represents the demand of the jth user for the air compressor system in the i-th data.

[0018] Furthermore, the output parameter data set composed of the input parameters of the Informer model is:

[0019]

[0020] Where: m represents the number of air compressor stations in the system, n represents the number of data in the data set, They represent the first air replenishment time and the second air replenishment time of the jth air compressor station in the i-th data, They respectively represent the air supply pressure required by the j-th air compressor station in the i-th data during two air supply cycles.

[0021] Furthermore, the data sizes of the training set, validation set, and test set are s1, s2, and s3 respectively, and the proportions of the three are:

[0022] s1:s2:s3=7:1:2.

[0023] Furthermore, the process of calculating the loss function is:

[0024] Calculate sum variance SSE:

[0025]

[0026] Where n represents the sample size, y i Represents the label values ​​of all parameters in the optimization target dataset, The values ​​of all parameters in the strategy table representing the model output

[0027] Calculate the MSE loss:

[0028]

[0029] Calculate MAE loss:

[0030]

[0031] The second aspect of the present invention provides an air compressor operation, maintenance and optimization system based on the Informer model, including a front-end module, a perception module, an optimization module, an air compressor management module, and a general control center, wherein specifically:

[0032] Front-end module, used for inputting air compressor management requirements;

[0033] A sensing module is used to obtain environmental parameters from the environment where the air compressor station is located;

[0034] The optimization module uses the above method to obtain the strategy table;

[0035] The air compressor management module performs operation, maintenance, and optimization on the air compressor;

[0036] The general control center is respectively connected to the front-end module, the sensing module, the general control center, and the air compressor management module to realize the scheduling of each module.

[0037] Furthermore, the scheduling process of each module by the general control center is as follows:

[0038] Read the user demand data obtained by the front-end module and obtain the environmental data provided by the perception module;

[0039] Sending the user demand data and environment data to the optimization module;

[0040] Obtaining a strategy table calculated by the optimization module;

[0041] A policy table is sent to the air compressor management module, instructing the air compressor management module to manage the air compressor according to the policy table.

[0042] Furthermore, the sensing module includes a timer and a temperature sensor, and the timer and the temperature sensor are respectively connected to the main control center for communication.

[0043] Furthermore, the optimization module can deploy the Informer model, read the data provided by the central control center as the model input, optimize the air compressor operation and maintenance strategy table and output it to the central control center;

[0044] The Informer model in the optimization module is first trained to obtain model weights, and the model weights are loaded into the Informer model.

[0045] Compared with the prior art, the method and system of the present invention have the following advantages:

[0046] First, the Informer long-time series training, prediction method and system for intelligent operation and maintenance and optimization of air compressors described in the present invention solve the intelligent operation and maintenance and optimization problems in air compressor systems, and use an end-to-end strategy. Therefore, compared with traditional methods (such as search algorithms, heuristic algorithms, meta-heuristic algorithms, etc.), the present invention is more suitable for application in the practice of specific air compressor system operation and maintenance optimization. In traditional methods, it is often necessary to manually extract features from the collected data, and then reorganize the data form and send it into the model for prediction; in addition, the output of the traditional model often needs to continue to be processed in a certain way before it can be applied to the operation and maintenance strategy of the air compressor station. In the strategy described in the present invention, it is only necessary to send the original collected data and labels into the Informer long-time series model, and the model can learn the relationship between the data by itself, and can directly output the strategy table of the air compressor station.

[0047] Second, the Informer model for intelligent operation and maintenance and optimization of air compressors described in the present invention has a strong ability to process time series data, which traditional optimization models do not have. The model used in the present invention takes time as a special variable and samples it according to the time frequency. The time frequency can be set according to specific needs. In addition, the model directly predicts the data for a period of time in the future based on the current existing data (i.e., the strategy table). Compared with the traditional model, the Informer model takes into account the relationship between the input data in the time dimension and can learn the deep connection of the input data at a deep level; compared with the original Transformer model, the Informer model avoids the cumulative error caused by step-by-step reasoning, so that the problems of gradient disappearance and gradient explosion will not occur. Overall, the present invention improves the accuracy and effectiveness of the intelligent operation and maintenance and optimization results of air compressors.

[0048] Third, the Informer model for intelligent operation, maintenance and optimization of air compressors described in the present invention is different from the traditional Transformer model. Both the Informer model and the Transformer model belong to the encoder-decoder model architecture, that is, they are composed of two parts: Encoder and Decoder. However, the Decoder part of the Transformer model must use the output of the unit at the previous moment as the input of the unit at the next moment, that is, perform dynamic decoding, which greatly slows down the prediction efficiency. The present invention uses a generative Decoder to obtain long sequence outputs, and only one forward step is required to output the entire decoded sequence. Therefore, the present invention can quickly calculate the strategy table of the air compressor station from the input, greatly reducing the time cost of intelligent operation, maintenance and optimization of the air compressor system.

[0049] Fourth, the present invention designs a relevant architecture for the Informer long-term training and prediction system for intelligent operation and maintenance and optimization of air compressors described in the present invention. This architecture includes multiple modules, each of which is used to complete independent and specific functions. The air compressor management module can manage multiple air compressors, and the number of air compressors is not limited, and can be flexibly deployed in actual scenarios. The perception module only needs a timer and a temperature sensor to work. Similarly, if there are more sensors, they can also be flexibly deployed in the perception module. The front-end module provides users with a visual air compressor system management page, which enhances the user-friendliness and interactivity of the entire system. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a deployment diagram of an actual air compressor system in an embodiment of the present invention.

[0051] Figure 2 Flowchart of the informer model training process in an embodiment of the present invention.

[0052] Figure 3 Flowchart of the informer model prediction process in an embodiment of the present invention.

[0053] Figure 4 This is the basic architecture of the informer model network in the embodiment of the present invention. DETAILED DESCRIPTION

[0054] The overall concept of the present invention is further elaborated in detail from several aspects below:

[0055] The Informer long-time series training method for intelligent operation, maintenance and optimization of air compressors in this invention includes the following steps:

[0056] The first step is to obtain relevant parameters of the air compressor management system through on-site research and recording in an actual large-scale air compressor management system, which will serve as a data set for model training.

[0057] Step 2: Select the parameters related to the air compressor station operation in the data set used in the first step as labels, that is, the parameters that the informer model needs to output; select other parameters in the data set as input parameters of the informer model; in particular, select time as the timing parameter in the input parameters of the model;

[0058] The third step is to divide the data set into a training set, a validation set, and a test set in a certain proportion;

[0059] Step 4: Randomly select a portion of the input parameters from the training set of the dataset and input them into the Informer model to obtain the output parameters. The output parameters are compared with the labels (i.e., true values) in the training set, and the model parameters are adjusted according to the difference between the two. The above process is repeated until all the data in the training set are obtained.

[0060] Step 5: After completing one cycle of step 4, take the input parameters of the validation set and the test set and input them into the Informer model to obtain the output parameters. The output parameters are compared with the output parameters of the validation set and the test set to calculate the loss.

[0061] Step 6: If the loss on the validation set in step 5 is low enough, the training process can be terminated early; otherwise, steps 4 to 6 are repeated until the pre-set number of training rounds is reached;

[0062] Step 7. After the training is completed, save the weight data obtained from the training.

[0063] Specifically, in the first step, the relevant parameters of the air compressor management system that need to be recorded are time, temperature, and user needs, and the data set can be divided into an input parameter data set and an optimization target data set.

[0064] Specifically, as follows, Data represents the input parameter data set, n represents the number of data contained in the data set, m represents the number of system users, and t i (i=1,2,...,n) represents the time when the i-th data record is recorded, T i (i=1,2,...,n) represents the temperature when the i-th data record is recorded, x i,j (i=1,2,...,n;j=1,2,...,m) represents the demand of the jth user for the air compressor system in the i-th data.

[0065]

[0066] Specifically, as follows, Target represents the optimization target data set, that is, the strategy table, m represents the number of air compressor stations in the system, and n represents the number of data in the data set. They represent the first air replenishment time and the second air replenishment time of the j-th air compressor station in the i-th data respectively. They respectively represent the air supply pressure required by the j-th air compressor station in the i-th data during two air supply cycles.

[0067]

[0068] Specifically, in the practice of air compressor system operation and maintenance, some better scheduling strategies can be obtained. These better scheduling strategies and their corresponding parameters can be recorded in the Target and Data data sets for model training.

[0069] Specifically, in the second step, the input parameters are the parameters in the above Data dataset, the output parameters are the parameters in the above Target dataset, and the timing parameters are the t i (i=1,2,...,n) parameters.

[0070] Specifically, in the third step, assuming that the sizes of the training set, validation set, and test set are s1, s2, and s3 respectively, the proportions of the three are generally:

[0071] s1:s2:s3=7:1∶.2

[0072] Adjustments can also be made on this basis, but the training set must be guaranteed to account for the largest proportion.

[0073] Specifically, in the fourth step, the model parameters are adjusted according to the loss calculated based on the relative comparison between the labels in the training set and the model output.

[0074] Specifically, the MSE loss function is used to calculate the difference between the label and the output. MSE (Mean-Squared Loss) refers to the mean square error loss. To calculate MSE, we must first calculate SSE, that is, the sum of the squares. The calculation formula is as follows, where n represents the sample size and y represents the sum of the squares. i Represents the label values ​​of all parameters in the optimization target data set (strategy table), The values ​​of all parameters in the strategy table representing the model output. Specifically, SSE is calculated for each output parameter:

[0075]

[0076] MSE can be calculated from SSE:

[0077]

[0078] Specifically, in the fifth step, in addition to the MSE loss, the calculated loss also needs to calculate the MAE loss. MAE (Mean Absolute Error) is the mean absolute error. The MAE loss calculation method is as follows, which can be compared with the performance of the MAE loss evaluation model in the fifth step:

[0079]

[0080] Specifically, in the sixth step, the minimum loss on the validation set is recorded. Thereafter, if the loss calculated on the validation set is greater than the minimum value multiple times (which can be set as needed), the model can be considered to have been trained to the optimal value and training can be stopped.

[0081] On the other hand, the informer long time series prediction method for intelligent operation, maintenance and optimization of air compressors in the present invention includes the following steps:

[0082] The first step is to obtain relevant parameters in the entire air compressor management system. The current time, temperature, and the needs of all users of the system are used as input parameters, and the air filling time and air filling pressure of all air compressor stations in the air compressor management system (i.e., the strategy table) are used as output values.

[0083] Step 2: Load the model weight parameters obtained from the last training saved in step 7 of the informer model training process into the new model, and then set the model to perform prediction instead of training;

[0084] The third step is to input the input parameters in the first step into the model, calculate and obtain the output of the model as the strategy table, that is, the optimization result, to complete the prediction process.

[0085] Specifically, in the first step, unlike the training process, labels are not required in the policy table. That is, the policy table is not an input but an output of the model. There is no need to compare the model output with the labels to calculate the loss. Simply input the input parameters into the model and calculate the corresponding policy table.

[0086] On the other hand, the informer long time series training and prediction system for intelligent operation, maintenance and optimization of air compressors in the present invention has the following specific steps:

[0087] The first step is to collect real-world data from air compressor intelligent operation and maintenance practices as a dataset. This dataset is then used in the training process using the informer long-series training method for air compressor intelligent operation and maintenance and optimization. The model weights obtained from the training are then saved.

[0088] In the second step, the model weight parameters obtained in the first step are loaded according to the Informer long-series prediction method for intelligent air compressor operation and optimization to obtain a pre-trained model. Then, air compressor station parameters are collected during the practice of intelligent air compressor operation and maintenance, and these parameters are input into the model to calculate the strategy table.

[0089] The third step is to perform air compressor system operation and maintenance according to the strategy table calculated in the second step.

[0090] Specifically, the system can be divided into five modules, namely, front-end module, air compressor management module, perception module, optimization module, and general control center.

[0091] More specifically, the front-end module is the module that interacts with users. Specifically, it's a web-based application. Users can enter their requirements for the air compressor management system on a webpage, and the front-end module reads these requirements and sends them to the central control center.

[0092] More specifically, the air compressor management module is a module that directly controls the air compressors. The air compressor management module obtains relevant signals for controlling the air compressors from the central control center, and then directly performs operation and maintenance control on different air compressors based on the signals.

[0093] More specifically, the sensing module directly senses relevant environmental parameters within the air compressor station's environment. At a minimum, the sensing module should include a timer to determine precise time and a temperature sensor to record accurate temperature. The sensing module can transmit the recorded data to the central control center.

[0094] More specifically, the optimization module is responsible for developing the air compressor system's operation and maintenance strategy. This module deploys the Informer model, reading data from the central control center as model input, optimizing the air compressor operation and maintenance strategy table, and then outputting it to the central control center. It's worth noting that the Informer model in this module must first undergo the aforementioned training process to obtain model weights, which are then loaded into the Informer model.

[0095] More specifically, the central control center must plan the system's operational logic according to a predetermined sequence. Specifically, the first step involves reading user requirements from the front-end module and acquiring environmental data from the perception module. The second step involves sending the data from the first step to the optimization module. The third step involves receiving the strategy table calculated by the optimization module. The fourth step involves sending the strategy table to the air compressor management module, which then manages the compressors accordingly.

[0096] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. Any features such as structure / module name, control mode, algorithm, process or composition ratio not clearly described in this technical solution are regarded as common technical features disclosed in the prior art.

[0097] Example 1

[0098] like Figure 1 As shown, the entire system can be divided into air compressor station management module, front-end module, perception module, optimization module and general control center.

[0099] Specifically, the front-end module is responsible for processing user input. In this embodiment, this module is a web-based application. Users can enter their requirements for the air compressor management system on a webpage. The front-end module can process input from up to 15 users simultaneously and transmit these user inputs to the central control center.

[0100] Specifically, the air compressor management module is a module that directly controls the air compressor. In this embodiment, the air compressor management module directly controls three air compressor stations. Figure 1 The three air compressor stations are air compressor station 1, air compressor station 2, and air compressor station 3. The air compressor management module obtains the relevant signals for controlling the air compressor from the main control center and controls the three air compressor stations respectively.

[0101] Specifically, the perception module directly senses relevant environmental parameters from the air compressor station's environment. In this embodiment, the module includes a timer and a temperature sensor. The timer must provide time data accurate to the hour at a minimum. In this embodiment, the timestamp format is "xxxx / xx / xx xx:xx." The temperature sensor measures the temperature of the air compressor system's environment at the time the record is generated, accurate to at least the nearest integer. The perception module can transmit the recorded data to the central control center.

[0102] Specifically, the optimization module formulates the operation and maintenance strategy of the air compressor system. In this embodiment, the optimization module deploys a pre-trained Informer model (i.e., the pre-selected and trained model weights need to be loaded). In this embodiment, the time, temperature, and needs of users 1-15 provided by the general control center, a total of 17 parameters, are used as model inputs. The model calculates the strategy table of the air compressor station 1-3 and sends it to the general control module. In this embodiment, the strategy table includes the first air replenishment time of the air compressor station 1-3, the second air replenishment time of the air compressor station 1-3, the first air replenishment pressure of the air compressor station 1-3, and the second air replenishment pressure of the air compressor station 1-3, a total of 12 results.

[0103] Specifically, the central control center plans the system's operational logic according to a predetermined sequence. In this embodiment, the system first reads the 12 user requirements provided by the front-end module and obtains environmental data (time, temperature) from the perception module. This data is then sent to the optimization module, which then calculates the strategy table. Finally, the air compressor management module manages air compressor stations 1-3 based on the strategy table.

[0104] like Figure 2 As shown, the training process of this embodiment is described in detail below.

[0105] S101 is the first step. In this embodiment, on-site investigation and recording are required to obtain the initial training set. The training set is divided into two parts: Data (input parameter data set) and Target (optimization target data set). Data includes 17 dimensions of data, including time, temperature, and the needs of users 1-15. Target includes 12 dimensions of data, including the first air replenishment time and pressure, and the second air replenishment time and pressure of air compressor stations 1-3.

[0106] Specifically, Data records data every hour, with 24 data items recorded per day, for a total of 365 days of data, resulting in a total of 8,760 data items. Target performs maintenance only once a day, so each of the 24 data items in Data for one day corresponds to only one maintenance strategy in the Target dataset. In this example, these two datasets are concatenated as the model training set. To ensure dimensionality consistency, each data item in Target is repeated 24 times, corresponding to each hour of data in Data.

[0107] S102 is the second step. In this embodiment, the designated labels are the first air replenishment time and pressure, the second air replenishment time and pressure of air compressor stations 1-3, the designated inputs are data of 17 dimensions, including time, temperature, and the needs of users 1-15, and the designated time series variable is time.

[0108] S103 is the third step. In this embodiment, the original training set is further divided into a training set, a test set, and a validation set according to a ratio of 7:2:1.

[0109] S104 is the fourth step. In this example, 32 (batch_size) data points are taken from the training set per batch. These are fed into the model for prediction. The predictions are compared with the labels to calculate the loss value. Based on this loss value, a backpropagation strategy is used to update the model parameters. This process repeats until all data in the training set are taken. It is worth noting that if the last batch of data contains less than 32 data points, it is discarded.

[0110] S105, S106, and S107 are the fifth, sixth, and seventh steps. In this embodiment, training is performed on the validation set, and the loss obtained on the validation set each time is compared with the minimum loss. If the former is larger than the latter for more than three consecutive times, training can be terminated early and the weight data can be saved.

[0111] Comparative Example 1

[0112] The following compares a multi-layer perceptron (MLP) prediction algorithm with the algorithm proposed in this paper. Their overall MAE and MSE losses are compared (the comparison is based on the losses calculated on the test set after training). It is worth noting that both use the same dataset, and the informer is configured according to the deployment method described in the previous embodiment. The MLP is configured as a multi-layer perceptron with three hidden layers, and the batch size is 8 (batch_size).

[0113] Table 1 Algorithm comparison results

[0114] algorithm MAE MSE MLP 63.67819941485126 5947.98090862966 informer 0.717556044 1.087911338

[0115] The comparison results above clearly show that using an informer-based long-sequence prediction model can achieve significantly better prediction accuracy than traditional methods. In fact, MLPs have a very low probability of finding correlations between data.

[0116] Specifically, the algorithm comparison results given above compare the losses calculated by comparing the overall prediction results with the labeled data. Shown below are the losses (referring to the final loss on the test set) achieved by the informer when predicting the output values ​​of the 12 dimensions in the strategy table of this embodiment.

[0117] Table 2 Algorithm performance

[0118]

[0119]

[0120] As can be seen from Table 2, the performance of the informer model on the dataset is not bad, and it can be said that it has good accuracy and prediction stability.

[0121] like Figure 3 , which is the prediction process of this embodiment. S201: Obtain air compressor station environmental parameters, including temperature and time, from the air compressor system; S202: Obtain requests input by users 1-15 from the front end; S203: Load the weights of the model used in this embodiment obtained through the aforementioned training process into the model; S204: Use the model loaded with pre-trained weights to read the model inputs obtained in steps S201 and S202 and calculate a strategy table; S205: Perform operation and maintenance on air compressor stations 1-3 according to the strategy table obtained in step S204.

[0122] like Figure 4 , which is the specific network architecture of the informer model used in this embodiment, is analyzed in detail below.

[0123] Specifically, in this embodiment, the model is divided into two parts: the encoder and the decoder. The encoder processes the model to obtain input, and the decoder outputs the calculation results. The following will analyze the two separately.

[0124] In this example, the encoder first reads the input. As mentioned above, the input is a 17-dimensional vector, including time, temperature, and the requirements of users 1-15, which can be expressed as follows:

[0125] X token =(t, T, x1, x2, ..., x 15 )

[0126] Before reading the input, the encoder needs to normalize it. The normalization method is as follows. token [n] represents the nth item of input, Y n Represents the value after normalization of the input nth item, mean(X token )、std(X token ) are the mean and standard deviation of the input respectively.

[0127]

[0128] The Encoder then reads the input. Multiply the input by three hyperparameter matrices W Q 、W K 、W V , and obtain the three matrices Q, K, and V for multi-head attention processing.

[0129] Q=W Q X token

[0130] K=W K X token

[0131] V=W V X token

[0132] Randomly sample F elements from K (F is a hyperparameter). Then calculate the following formula:

[0133]

[0134] where K F represents a set of F K samples. d refers to the dimension of the sample data. L K is the number of K samples. Calculate the M value and select the u q with the largest M i (u is a hyperparameter). Based on the u q with the largest M value i Combine into new Finally, for this u q i , calculated according to the self-attention mechanism:

[0135]

[0136] For other q values ​​with smaller M i , then calculate:

[0137]

[0138] Then, these A’s are organized in their original order. Self-attention extraction and pooling are performed. Self-attention extraction requires a one-dimensional convolution and an exponential linear unit (ELU) activation function. The entire process is as follows:

[0139] X j+1 =MaxPool(ELU(Conv([A j ])))

[0140] X j+1 sent to the subsequent calculation process.

[0141] After multiple rounds of multi-head sparse self-attention processing and self-attention extraction (this process is called attention distillation), the output results are feature mapped and sent to the multi-head attention processing part of the Decoder.

[0142] The decoder input is a one-dimensional vector of length d+t. As mentioned above, d is the dimension of the encoder input. t is the length of the future period to be predicted. The input data is then fed into a masked multi-head self-attention mechanism. This mechanism adds a mask to prevent the prediction of data from earlier moments from being seen when predicting data from later moments. The output of the masked multi-head self-attention mechanism is used as Q, and the encoder outputs are used as K and V, continuing the multi-head attention process. The final output is then passed through a fully connected layer, completing all processing and outputting the resulting air compressor station strategy table.

[0143] Specifically, after obtaining the air compressor station strategy table, the air compressor station operation, maintenance and optimization can be performed according to the strategy table.

[0144] The above description of the embodiments is intended to facilitate understanding and use of the invention by those skilled in the art. It will be apparent that those skilled in the art can readily make various modifications to these embodiments and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above-described embodiments. Improvements and modifications made by those skilled in the art based on the disclosure of the present invention, without departing from the scope of the present invention, should be within the scope of protection of the present invention.

Claims

1. An air compressor operation, maintenance and tuning method based on the Informer model, characterized in that: The following steps are involved: Obtain parameter information from the air compressor management system and construct the original training data set; Selecting target parameters to be optimized from the original training data set as output parameters of the informer model, using the remaining parameters in the original training data set as input parameters of the informer model, and using the corresponding times of the input parameters of the informer model as timing parameters in the input parameters; Dividing the original training data set into a training set, a test set, and a validation set; Randomly extract a portion of the input parameters from the training set and input them into the informer model to obtain informer model output parameters, compare the informer model output parameters with the training set input parameters, adjust the model parameters according to the difference between the two, and repeat the input, output, and comparison process until all the data in the training set is extracted. After one cycle is completed, extract the input parameters from the validation set and the test set and input them into the informer model to obtain output parameters, compare the output parameters with the output parameters in the validation set and the test set, calculate the loss function, and obtain the loss-adjusted model parameters. When the loss corresponding to the validation set is lower than the preset value, or the preset number of cycles is reached, the training process is terminated, and the trained model is obtained. The model is validated and tested on the test set to evaluate the model effect; Input the air compressor station parameters under working conditions into the trained model to obtain a strategy table. Based on the strategy table, the air compressor operation and maintenance and optimization are carried out to achieve the lowest operation and maintenance and optimization costs of the air compressor system; The output parameters of the Informer model include the air supply pressure and air supply time of the air compressor; The input parameters of the Informer model include time, temperature, and user needs; The input parameter dataset composed of the input parameters of the Informer model is: Where: n represents the number of data contained in the data set, m represents the number of system users, Indicates the time when the i-th data record is recorded, Indicates the temperature when the i-th data record is recorded, Indicates the demand of the jth user for the air compressor system in the i-th data; The output parameter dataset composed of the input parameters of the Informer model is: Where: m represents the number of air compressor stations in the system, n represents the number of data in the data set, They represent the first air replenishment time and the second air replenishment time of the jth air compressor station in the i-th data, They respectively represent the air supply pressure required by the j-th air compressor station in the i-th data during two air supply cycles.

2. The air compressor operation, maintenance and optimization method based on the Informer model according to claim 1 is characterized in that: The data sizes of the training set, validation set, and test set are 、 、 , and the proportions of the three are: 。 3. The air compressor operation, maintenance and optimization method based on the Informer model according to claim 1 is characterized in that: The process of calculating the loss function is: Calculate sum variance SSE: Where n represents the sample size, Represents the label values ​​of all parameters in the optimization target dataset, The values ​​of all parameters in the strategy table representing the model output Calculate the MSE loss: Calculate MAE loss: 。 4. An air compressor operation, maintenance and tuning system based on the Informer model, characterized in that: include: Front-end module, used for inputting air compressor management requirements; A sensing module is used to obtain environmental parameters from the environment where the air compressor station is located; An optimization module, which obtains a strategy table using the method described in any one of claims 1 to 3; The air compressor management module performs operation, maintenance, and optimization on the air compressor; The general control center is respectively connected to the front-end module, the sensing module, the general control center, and the air compressor management module to realize the scheduling of each module.

5. The air compressor operation, maintenance and optimization system based on the Informer model according to claim 4 is characterized in that: The scheduling process of each module by the general control center is as follows: Read the user demand data obtained by the front-end module and obtain the environmental data provided by the perception module; Sending the user demand data and environment data to the optimization module; Obtaining a strategy table calculated by the optimization module; A policy table is sent to the air compressor management module, instructing the air compressor management module to manage the air compressor according to the policy table.

6. The air compressor operation, maintenance and optimization system based on the Informer model according to claim 4 is characterized in that: The sensing module includes a timer and a temperature sensor, and the timer and the temperature sensor are respectively connected to the main control center for communication.

7. The air compressor operation, maintenance and optimization system based on the Informer model according to claim 5 is characterized in that: The optimization module can deploy the Informer model, read the data provided by the central control center as the model input, optimize the air compressor operation and maintenance strategy table and output it to the central control center; The Informer model in the optimization module is first trained to obtain model weights, and the model weights are loaded into the Informer model.

Images