Monitoring method and device based on digital twin model, electronic equipment and medium

Abstract

The invention relates to the technical field of equipment detection, and provides a monitoring method and device based on a digital twin model, electronic equipment and a medium. The method comprises the following steps: acquiring a digital twin model and operation state information of target equipment; inputting the operation state information into the trained multilayer deep learning model to obtain an output result as a monitoring result of the digital twin model; the training process of the multi-layer deep learning model comprises the following steps: initializing each member individual in a hyper-parameter population; the optimization searching operation is cycled until the iteration ending condition is met, and the current optimal candidate model serves as a multi-layer deep learning model; the optimization searching operation comprises the following steps: generating a candidate model corresponding to each member individual; determining the fitness value of each candidate model, and taking the candidate model with the minimum fitness value as the current optimal candidate model; and updating each member individual. According to the method, nonlinear and complex-mode equipment faults can be analyzed foreseeably, and potential problems in operation can be found more accurately and deeply.

Description

Technical Field

[0001] The present application relates to the field of equipment monitoring technology, and in particular to a monitoring method, device, electronic equipment and medium based on a digital twin model. Background Art

[0002] With the rapid development of industrial automation, the performance requirements and stability monitoring of measurement and control equipment have become key indicators in research and production. These devices are usually in a state of continuous high-load operation, and their failure is likely to cause production line shutdown or even safety accidents.

[0003] The failure modes of measurement and control equipment are complex, and the current fault diagnosis of measurement and control equipment mostly relies on offline data analysis and processing, which cannot respond to and handle potential problems in operation in real time. Even if online monitoring systems are applied, these systems mainly use threshold judgment and simple logistic regression methods for fault alarms, which makes it difficult to predict nonlinear and complex equipment failures.

[0004] Therefore, how to provide a solution to the above technical problems is a problem that technicians in this field need to solve. Summary of the invention

[0005] In view of this, the embodiments of the present application provide a monitoring method, device, electronic device and medium based on a digital twin model to solve the problem that fault diagnosis in the prior art cannot cope with complex modes and potential faults.

[0006] A first aspect of an embodiment of the present application provides a monitoring method based on a digital twin model, comprising:

[0007] Obtain the digital twin model of the target device;

[0008] Obtain the operating status information of the digital twin model;

[0009] Input the operating status information into the trained multi-layer deep learning model to obtain the output result of the multi-layer deep learning model as the monitoring result of the digital twin model;

[0010] Among them, the training process of the multi-layer deep learning model includes:

[0011] Obtain a training sample data set and a test sample data set;

[0012] Initialize each member individual in the hyperparameter population, each member individual includes a set of candidate hyperparameters;

[0013] The optimization search operation is repeated until the iteration end condition is met, and the current optimal candidate model is used as the multi-layer deep learning model;

[0014] Among them, the optimization search operation includes:

[0015] Based on the training sample data set and each member individual, a candidate model corresponding to each member individual is generated;

[0016] Determine the fitness value of each candidate model based on the test sample data set, and take the candidate model with the smallest fitness value as the current optimal candidate model;

[0017] Update each member based on all current member individuals.

[0018] A second aspect of an embodiment of the present application provides a monitoring device based on a digital twin model, including:

[0019] An acquisition module is used to acquire a digital twin model of a target device and also to acquire operating status information of the digital twin model;

[0020] A monitoring module is used to input the operating status information into the trained multi-layer deep learning model to obtain the output result of the multi-layer deep learning model as the monitoring result of the digital twin model;

[0021] Among them, the training process of the multi-layer deep learning model includes:

[0022] Obtain a training sample data set and a test sample data set;

[0023] Initialize each member individual in the hyperparameter population, each member individual includes a set of candidate hyperparameters;

[0024] The optimization search operation is repeated until the iteration end condition is met, and the current optimal candidate model is used as the multi-layer deep learning model;

[0025] Among them, the optimization search operation includes:

[0026] Based on the training sample data set and each member individual, a candidate model corresponding to each member individual is generated;

[0027] Determine the fitness value of each candidate model based on the test sample data set, and take the candidate model with the smallest fitness value as the current optimal candidate model;

[0028] Update each member based on all current member individuals.

[0029] According to a third aspect of an embodiment of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0030] According to a fourth aspect of an embodiment of the present application, a readable storage medium is provided, which stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0031] Compared with the prior art, the beneficial effects of the embodiments of the present application include at least the following: the embodiments of the present application utilize a multi-layer deep learning model to analyze the operating status information in the digital twin model to obtain the monitoring results of the digital twin model. The training of the multi-layer deep learning model is implemented by a cyclic iteration of the optimization search operation based on the hyperparameter population. Compared with the training scheme based on empirical values, the multi-layer deep learning model that has been trained in the present application has higher recognition accuracy and better prediction effect. The multi-layer deep learning model is combined with the digital simulation characteristics of the digital twin model to predictably analyze nonlinear and complex mode equipment failures and more accurately and deeply discover potential problems in operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 It is a flow chart of a monitoring method based on a digital twin model provided in an embodiment of the present application;

[0034] Figure 2 It is a flowchart of a training process of a multi-layer deep learning model provided in an embodiment of the present application;

[0035] Figure 3 It is a flowchart of a training process of another multi-layer deep learning model provided in an embodiment of the present application;

[0036] Figure 4 It is a structural schematic diagram of a monitoring device based on a digital twin model provided in an embodiment of the present application;

[0037] Figure 5 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0038] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0039] To ensure the efficiency and safety of industrial production, the measurement and control equipment involved often needs to be diagnosed and maintained based on the results of the fault diagnosis. Timely diagnosis of real-time faults or prediction of potential faults can avoid the expansion of faults, reduce the scope of the fault's impact, and improve the reliability of equipment operation.

[0040] At present, fault analysis of measurement and control equipment includes online and offline. Offline fault analysis is based on offline data analysis and processing, and cannot respond to and handle potential problems in operation in real time. Online fault analysis mainly uses threshold judgment and simple logistic regression methods. The diagnostic logic is simple and superficial, and it is difficult to handle nonlinear and complex equipment failures, let alone predict potential failures.

[0041] With the emergence of digital twin technology, physical devices can be simulated and operated in a digital form. However, the operating logic of the digital twin model can only reflect the information processing effect of the physical device. It does not have the ability to deeply analyze data and predict faults. It also has the above two problems of fault analysis.

[0042] Based on the digital twin model, the embodiment of the present application analyzes the operating status information of the digital twin model through a multi-layer deep learning model, wherein the multi-layer deep learning model uses a hyperparameter population optimization operation loop to complete training, so that the trained multi-layer deep learning model has higher recognition accuracy for complex faults and the ability to more sensitively predict potential faults. At the same time, the digital twin model has the digital simulation characteristics of simulating the actions of physical equipment, and can simulate the operating conditions of different states on the timeline in a virtual environment. The use of the operating status information of the digital twin model further improves the accuracy of fault prediction by the multi-layer deep learning model, so that the present embodiment can achieve the effect of more sensitively predicting potential faults and in-depth analysis of complex faults.

[0043] A monitoring method, device, electronic device and medium based on a digital twin model according to an embodiment of the present application will be described in detail below with reference to the accompanying drawings.

[0044] Figure 1 It is a flow chart of a monitoring method based on a digital twin model provided in an embodiment of the present application. Figure 1 The monitoring method can be executed by a terminal processor device or a server. Figure 1 As shown, the monitoring method includes:

[0045] S101: Obtain a digital twin model of a target device;

[0046] S102: Obtaining operation status information of the digital twin model;

[0047] S103: Inputting the operation status information into the trained multi-layer deep learning model to obtain the output result of the multi-layer deep learning model as the monitoring result of the digital twin model;

[0048] Among them, the training process of the multi-layer deep learning model includes:

[0049] Obtain a training sample data set and a test sample data set;

[0050] Initialize each member individual in the hyperparameter population, each member individual includes a set of candidate hyperparameters;

[0051] The optimization search operation is repeated until the iteration end condition is met, and the current optimal candidate model is used as the multi-layer deep learning model;

[0052] Among them, the optimization search operation includes:

[0053] Based on the training sample data set and each member individual, a candidate model corresponding to each member individual is generated;

[0054] Determine the fitness value of each candidate model based on the test sample data set, and take the candidate model with the smallest fitness value as the current optimal candidate model;

[0055] Update each member based on all current member individuals.

[0056] Among them, the monitoring method of this embodiment is implemented based on the digital twin model. The digital twin model of the target device can simulate the action of the physical device, and the operating status information of the digital twin model is input into the multi-layer deep learning model. The monitoring results obtained can be used as the status analysis results of the target device. Furthermore, the multi-layer deep learning model in this embodiment is obtained by a training process based on multiple iterations of optimization of a hyperparameter population. Therefore, the hyperparameters of the multi-layer deep learning model finally determined in this embodiment are relatively ideal, and the accuracy of the multi-layer deep learning model that has completed training is relatively high. When analyzing the input operating status information, it can pay attention to potential features and in-depth and complex feature relationships, thereby improving the accuracy, efficiency and predictive ability of fault diagnosis.

[0057] It can be understood that in this embodiment, a digital twin model at least includes an information input interface, a logic processing module, and an information output interface. If the monitored object includes multiple related digital twin models, each digital twin model also includes an information input interface or an information output interface connected to other digital twin models. The logic processing module is provided with an internal operating logic, which is a digital expression of the working principle, parameter setting, and internal information flow of the equipment. The operation process of the digital twin model generally includes: the information input interface receives external stimulus information, the internal operating logic responds to the stimulus information transmitted by the information input interface according to the internal operating logic, generates response information, and the information output interface outputs the corresponding response information. All the above-mentioned types of information related to the operation process of the digital twin model can be used as the operation status information for monitoring the digital twin model in this embodiment, so the operation status information includes:

[0058] Input stimulus information to the digital twin model;

[0059] and / or, response information when the digital twin model responds to stimulus information according to its internal operating logic.

[0060] It is understandable that the operating status information of the digital twin model is input into the multi-layer deep learning model, and the output result of the multi-layer deep learning model is exactly the monitoring result of the digital twin model. This effect is produced because the multi-layer deep learning model is pre-trained as a monitoring model for the digital twin model. The training of the multi-layer deep learning model can be referred to Figure 2 As shown, Figure 2 is a flowchart of the training process of the multi-layer deep learning model in the embodiment of the present application, Figure 2 The training process can be performed by the terminal processor device or the server.

[0061] It can be understood that in this embodiment, the target device and the digital twin model completely correspond to each other. When the target device changes, the digital twin model also needs to change accordingly. The running status information of the digital twin model completely corresponds to the multi-layer deep learning model. The multi-layer deep learning models used for different running status information of the same digital twin model or the running status information of different digital twin models are likely to be different, and targeted training is required. Therefore, the training sample data set required for training and the test sample data set required for testing of the multi-layer deep learning model with the current running status information as input both correspond to the running status information. The training sample data set and the test sample data set both include input samples and output samples. The form of the input samples is consistent with the running status information. The output samples are specifically the monitoring results corresponding to different input samples. The monitoring results may include fault status and normal status. The fault status may also include the specific cause of the fault, and the normal status may also include a specific working status.

[0062] In this embodiment, the training of the multi-layer deep learning model is performed based on a hyperparameter population, which includes multiple member individuals. The size of the hyperparameter population is the number of member individuals, and each member individual includes a set of specific candidate hyperparameters. The training process of the multi-layer deep learning model in this embodiment is essentially a process of continuously performing performance evaluation and iterative optimization based on the hyperparameter population to find a better set of hyperparameters.

[0063] Among them, hyperparameters are parameters pre-set before training of the machine model, which are used to control the algorithm's search space, learning speed, model complexity, etc. The values ​​of these parameters will not be optimized during the deep learning model training process, so they are usually pre-set by the user manually based on experience or selected through a certain strategy. The numerical value of the hyperparameter has an important impact on the performance of the machine model, which will determine the structure, training efficiency, accuracy and other performance of the machine model. Therefore, how to choose the appropriate hyperparameter is a key issue in machine model training. This embodiment uses iterative optimization of the hyperparameter population to determine a set of best-performing hyperparameters to obtain a multi-layer deep learning model with the best effect.

[0064] It is understandable that in the embodiment of the present application, a set of candidate hyperparameters included in each member individual may include multiple of the total number of hidden layers, the number of neural units in each hidden layer, the batch size, the time window size, the number of training times, and the learning rate. Among them, the total number of hidden layers and the number of neural units can be collectively referred to as the hidden layer size, which determines the network structure of the multi-layer deep learning model, the batch size refers to the number of samples used for each gradient descent update, the number of training times refers to the maximum number of training times, and the learning rate is used to control the amplitude of each step parameter update.

[0065] Specifically, the ultimate goal of this embodiment is to obtain the monitoring results of the digital twin model of the target device. Therefore, the multi-layer deep learning model in this embodiment can select a model including a multi-layer LSTM (Long Short-term Memory) structure. The model with a multi-layer LSTM structure can effectively capture and analyze complex data features in long time series, effectively capture long-term dependencies, thereby enhancing the accuracy and robustness of dynamic fault prediction, improving the accuracy of fault prediction, and having obvious advantages in predicting the timeliness and accuracy of fault time, which can reduce the occurrence of missed reports and false alarms, and ultimately improve the efficiency and safety of equipment use. Since the digital twin model can express long-term characteristics on the timeline, combined with the multi-layer LSTM structure's ability to capture long-term characteristics, it can more predictively discover potential faults.

[0066] The time window size in the candidate hyperparameters mentioned above is a hyperparameter that describes the LSTM structure. It defines the length of the time series that the LSTM network can process. The time window size can also be called sequence length or time step.

[0067] It is understandable that, according to the operating performance of different target devices, the multi-layer deep learning model can also select multi-layer models of other structures, which is not limited here. The candidate hyperparameters in each member individual can also select other different types of hyperparameters according to the specific structure of the multi-layer deep learning model.

[0068] It is understandable that for a hyperparameter population, the candidate hyperparameters of each member are of the same type, and the values ​​of the same type of candidate hyperparameters are different, but the values ​​of each type of candidate hyperparameters should be within the corresponding range. For example, the ranges of the batch size, the total number of hidden layers, the number of neural units in each hidden layer, and the learning rate can be set to [16,128], [1,10], [16,100], [0.001,0.02], respectively.

[0069] It can be understood that each member individual in the initialization hyperparameter population, that is, each group of candidate hyperparameters is initialized, and each candidate hyperparameter is assigned a value from the corresponding value range according to a specific rule or randomly. Each member individual can be regarded as a candidate solution represented by a multidimensional vector corresponding to this group of candidate hyperparameters.

[0070] After that, the optimization search operation starts to be executed cyclically. In each round of optimization search operation, a candidate model is first generated for each member individual. The candidate model is trained with the hyperparameters and training sample data set corresponding to the member individual; then each candidate model is tested based on the test sample data set, and its fitness value is determined according to the test results. The smaller the fitness value, the higher the accuracy of the candidate model, so as to determine the optimal candidate model of this round of cycle; finally, the member individuals are updated, that is, a set of candidate hyperparameters for each member individual is updated. The overall performance of the member individual is used as the basis for the update. The update measures are usually to adjust the values ​​of the candidate hyperparameters, that is, the position of the candidate solution. The update target can also be set to make each candidate solution close to the optimal solution.

[0071] It can be understood that the training action when generating a candidate model for each individual member is the action of adjusting the parameters in the model after building the network structure of the model based on the hyperparameters with known values ​​of the individual member. The parameters here refer to the weight parameters and bias values ​​of each neural unit, that is, parameter training; and the training process of the multi-layer deep learning model referred to in this embodiment is carried out under the premise that the hyperparameters are unknown, and the main content is to determine, update, and optimize each set of hyperparameters, and finally determine an optimal hyperparameter. The parameter training when generating the candidate model above is a small link in updating each set of hyperparameters in the training process of the multi-layer deep learning model in this embodiment.

[0072] The optimization search operation is executed in a loop until the iteration end condition is met. The iteration end condition usually includes but is not limited to whether the maximum number of cycles is reached and whether one or more model training indicators are met. Model training indicators such as preset fitness values, determination coefficients, etc. are used to evaluate the performance of the model.

[0073] When the iteration end condition is met, the optimal candidate model determined in this round of loop can be used as a multi-layer deep learning model that has completed training and is used to execute the monitoring method of this embodiment.

[0074] It is understandable that the optimal candidate model determined by the optimization search operation in this embodiment is determined by the optimal hyperparameters. The optimal candidate model has a very ideal fitness value, so it has a very high recognition accuracy and better prediction effect. Conventional model training in the prior art is usually pre-trained by manually inputting fixed hyperparameters based on experience. The hyperparameters will not change again, and the numerical quality of the hyperparameters themselves determines the performance of the model. Compared with the prior art, this embodiment does not set fixed hyperparameters. A hyperparameter population is used in the training process. After multiple optimization search operations on each member of the hyperparameter population, the hyperparameters are optimized and updated, and the optimal candidate model finally determined has a more ideal performance.

[0075] The method of the embodiment of the present application utilizes a multi-layer deep learning model to analyze the operating status information in the digital twin model to obtain the monitoring results of the digital twin model. The training of the multi-layer deep learning model is implemented by a cyclic iteration of the optimization search operation based on the hyperparameter population. Compared with the scheme of parameter training after taking empirical values ​​of the hyperparameters, the multi-layer deep learning model that has been trained in the present application has higher recognition accuracy and better prediction effect. The multi-layer deep learning model is combined with the digital simulation characteristics of the digital twin model to predictably analyze nonlinear and complex mode equipment failures and more accurately and deeply discover potential problems in operation.

[0076] In some embodiments, the optimization search operation can be implemented according to the idea of ​​the Harris Hawk optimization algorithm, in which each member individual acts as a hunter, and the prey approaches the optimal solution with iteration, and the hunter's update goal is to get close to the prey; each hunter in the hyperparameter population is represented by a hunter position, and the hunter position is specifically a candidate solution represented by a multidimensional vector of this set of candidate hyperparameters; there is usually only one prey, and the relevant parameters of the prey include prey position, prey escape energy and prey capture probability, in which the prey position and the hunter position are multidimensional vectors of the same dimension, usually corresponding to the current optimal solution, the prey escape energy and the prey capture probability affect the possible changes in the prey position in the next iteration, the prey escape energy refers to the change in the prey escape position at each iteration, usually decreases with the increase in the number of iterations, the prey capture probability refers to the possibility of the prey being captured at each iteration, usually a random amount. All hunters in the hyperparameter population will decide the hunting strategy based on the prey position, prey escape energy and prey capture probability, that is, how to update the hunter position in the next iteration. It can be understood that the hunting strategy for updating the hunter position is also the update strategy for individual members.

[0077] Therefore, the optimization search operation is cyclically performed until the iteration end condition is met, and before the current optimal candidate model is used as the multi-layer deep learning model, the following steps are also included:

[0078] Initialize the prey escape energy and prey capture probability corresponding to the hyperparameter population;

[0079] In the optimization search operation, the process of updating each member individual based on all current member individuals includes:

[0080] Update the prey position of the hyperparameter population based on the current best candidate model;

[0081] Based on the current number of operations of the optimal search operation, the current prey escape energy is updated;

[0082] Determine the current update strategy based on the current prey escape energy and prey capture probability;

[0083] Based on the prey position and all current member individuals, each member individual is updated according to the current update strategy.

[0084] It is understandable that if the prey escape energy and the prey capture probability are different, the update basis and the target are different, and it is necessary to select an update strategy with different speeds. In the specific optimization search operation, the process of determining the current update strategy based on the current prey escape energy and the prey capture probability may include:

[0085] Determine the absolute value of the current prey escape energy as the corresponding current energy absolute value;

[0086] Determine whether the absolute value of the current energy is less than the first energy;

[0087] If not, determine that the current update strategy is the exploration strategy;

[0088] If so, determine whether the absolute value of the current energy is less than the second energy, and determine whether the probability of prey capture is less than a preset probability; the first energy is greater than the second energy and greater than zero;

[0089] When the absolute value of the energy is not less than the second energy and the prey capture probability is not less than the preset probability, determining that the current update strategy is a gentle siege strategy;

[0090] When the absolute value of the energy is less than the second energy and the prey capture probability is not less than the preset probability, the current update strategy is determined to be a fierce siege strategy;

[0091] When the absolute value of the energy is not less than the second energy and the prey capture probability is less than the preset probability, the current update strategy is determined to be a cumulative speed diving gentle siege strategy;

[0092] When the absolute value of the energy is less than the second energy and the prey capture probability is less than the preset probability, the current update strategy is determined to be a cumulative dive-style fierce siege strategy;

[0093] The specific exploration strategy is: updating the current member individual based on any member individual, or updating the current member individual based on the average position of all member individuals and the prey position;

[0094] The gentle siege strategy is specifically as follows: based on the current prey escape energy, the current member individual, the prey position and the preset escape coefficient corresponding to the prey position, the current member individual is updated;

[0095] The fierce siege strategy is as follows: based on the current member individual and the prey position, update the current member individual;

[0096] The specific strategy of the gradual dive-type gentle siege is as follows: based on the current prey escape energy, the current member individual, the prey position, the preset escape coefficient and the preset constant, the current member individual is updated;

[0097] The specific strategy of the accelerated dive and fierce siege is as follows: based on the current prey escape energy, average position, prey position, preset escape coefficient and preset constant, the current member individuals are updated.

[0098] Therefore, see Figure 3 As shown in the flowchart, the training process of the multi-layer deep learning model may specifically include:

[0099] Obtain a training sample data set and a test sample data set;

[0100] Initialize each member of the hyperparameter population;

[0101] Initialize the prey escape energy and prey capture probability corresponding to the hyperparameter population;

[0102] Cyclic optimization search operation, the optimization search operation includes:

[0103] Based on the training sample data set and each member individual, a candidate model corresponding to each member individual is generated;

[0104] Determine the fitness value of each candidate model based on the test sample data set, and take the candidate model with the smallest fitness value as the current optimal candidate model;

[0105] Update the prey position of the hyperparameter population based on the current best candidate model;

[0106] Based on the current number of operations of the optimal search operation, the current prey escape energy is updated;

[0107] Determine the current update strategy based on the current prey escape energy and prey capture probability;

[0108] Based on the prey position and all current member individuals, update each member individual according to the current update strategy;

[0109] Until the iteration end condition is met, the optimization search operation ends and the current optimal candidate model is used as the multi-layer deep learning model.

[0110] It can be understood that the training sample data set is recorded as Ztrain and the test sample data set is recorded as Ztest. Both sample data sets include multiple sample data, and each sample data includes an input sample and a corresponding actual output sample.

[0111] The total number of member individuals in the hyperparameter population and the iteration end condition are pre-set. For example, the total number can be set to 20, and the maximum number of cycles in the iteration end condition is 100; the i-th member individual can be expressed as X i , i is 1-N, N is the total number of member individuals, X i During initialization, the value assigned can be determined from the value range of each hyperparameter according to the assignment strategy;

[0112] The prey escape energy is represented by E, and the initial value is recorded as E0, taking a random value in the interval [-1,1];

[0113] The probability of prey capture is denoted as u, and the initialization value is a random value in the interval [0,1].

[0114] Based on individual member X i Initially generate candidate model M i The network structure of i and training sample data set Ztrain for candidate model Mi The parameters are trained to obtain the complete candidate model M i .

[0115] Calculate the candidate model M based on the test sample dataset Ztest i The fitness value of the test sample data set Ztest is specifically input into the candidate model M i , get the corresponding predicted output sample, calculate the fitness based on the actual output sample and the predicted output sample in the test sample data set, and get the fitness value. The formula for fitness calculation is as follows:

[0116]

[0117] Among them, e MAE is the fitness value, m is the total number of sample data in the test sample data set Ztest, and j are the predicted output sample and actual output sample of the j-th sample data respectively.

[0118] After obtaining the fitness values ​​of all candidate models, the candidate model with the smallest fitness value is determined as the optimal candidate model for this iteration, and the prey position is updated based on the corresponding member individuals. The prey position is recorded as X rabbit (t), where t is the current operation number of the optimization search operation, also known as the number of iterations, and t is a positive integer. In fact, the update of the prey position in each iteration here is directly updated to the member individual of the current optimal candidate model.

[0119] Update the prey escape energy based on the current number of operations, which can be updated according to the following formula:

[0120]

[0121] Where t and T are the current operation number and the maximum loop number respectively.

[0122] The current update strategy is determined based on the prey escape energy E and the prey capture probability u. Specifically, the absolute value of the prey escape energy E is determined as the current energy absolute value |E|. The update strategy is determined based on the energy relationship between |E| and the first energy and the second energy, and the probability relationship between the prey capture probability and the preset probability. Considering the values ​​of E and u, the first energy is usually 1, the second energy is usually 0.5, and the preset probability is usually 0.5. Therefore, there is:

[0123] If |E|≥1, the update strategy is the exploration strategy;

[0124] If 0.5≤|E|<1 and u≥0.5, the update strategy is the gentle siege strategy;

[0125] If |E|<0.5 and u≥0.5, the update strategy is a fierce siege strategy;

[0126] If 0.5≤|E|<1 and u<0.5, the update strategy is a gentle siege strategy with cumulative dive;

[0127] If |E|<0.5 and u<0.5, the update strategy is a cumulative dive and fierce siege strategy.

[0128] Among them, the exploration strategy is usually in the global search stage. The specific exploration strategy is: update the current member individual based on any member individual, or update the current member individual based on the average position of all member individuals and the prey position; the specific scheme selected when updating the member individual can be determined according to a random variable q with a value in the interval [0,1]. For example, when q≥0.5, the current member individual is updated based on any member individual, and when q<0.5, the current member individual is updated based on the average position of all member individuals and the prey position. The update formula of the exploration strategy for the i-th member individual is as follows:

[0129]

[0130] Where X i (t) represents the member individuals that have not been updated under the current number of operations, X i (t+1) represents the updated member individual, X rd (t) is any member individual in the hyperparameter population that has not been updated under the current number of operations. r1, r2, r3, r4, and q are all random numbers in the interval [0,1]. ub and lb represent the upper and lower limits of the search space. X m (t) represents the average position of the hyperparameter population that has not been updated under the current number of operations, and the expression is as follows;

[0131]

[0132] It can be understood that when |E| is not less than 1, the exploration strategy of the global search stage is continued until |E| changes to less than 1, and the update strategy enters the local search stage from the global search stage. The local search stage includes a gentle siege strategy, a fierce siege strategy, a cumulative dive gentle siege strategy, and a cumulative dive fierce siege strategy. The current update strategy is determined according to the prey escape energy E and the prey capture probability u.

[0133] Among them, the gentle siege strategy corresponds to 0.5≤|E|<1 and u≥0.5. The prey has enough energy to try to escape but the probability of capture is high. The gentle siege strategy is specifically as follows: based on the current prey escape energy, the current member individual, the prey position and the preset escape coefficient corresponding to the prey position, the current member individual is updated; specifically, the i-th member individual in the gentle siege strategy can be updated according to the following formula:

[0134] X i (t+1)=(X rabbit (t)-X i (t))-E|JX rabbit (t)-X i (t)|;

[0135] Where X rabbit (t)-X i (t) is the distance between the i-th member and the prey position, J is the preset escape coefficient when the prey escapes, J = 2(1-r5), r5 is a random number in the interval [0,1].

[0136] Among them, the fierce siege strategy corresponds to |E|<0.5 and u≥0.5, the prey has low energy and poor escape ability, at this time the member individuals will attack quickly, the fierce siege strategy is specifically: based on the current member individual and the prey position, update the current member individual; specifically, the i-th member individual in the fierce siege strategy can be updated according to the following formula:

[0137] X i (t+1)=X rabbit (t)-E|X rabbit (t)-X i (t)|.

[0138] Among them, the cumulative speed diving gentle siege strategy corresponds to 0.5≤|E|<1 and u<0.5. The prey has sufficient escape ability. The individual members need to form a gradual and rapid gentle siege before attacking. Specifically, based on the current prey escape energy, the current individual member, the prey position, the preset escape coefficient and the preset constant, the current individual member is updated; the preset constant can be based on the mechanism of Levy flight. Specifically, the i-th individual member in the cumulative speed diving gentle siege strategy can be updated according to the following formula:

[0139]

[0140] Where F() is the fitness function, S×LF(D) is a preset constant, where D is the dimension of the problem, S is a random vector of size 1×D, and LF is the Levy flight function, which can be obtained using the following formula:

[0141]

[0142] where μ and v are random values ​​in the interval (0,1), and β is a default constant set to 1.5.

[0143] Among them, the cumulative dive fierce siege strategy corresponds to |E|<0.5 and u<0.5. The prey has low escape energy, but still has a chance to escape. In this case, the member individuals try to reduce the distance between them and the average position of the target prey. Specifically, based on the current prey escape energy, average position, prey position, preset escape coefficient and preset constant, the current member individual is updated. Specifically, the i-th member individual in the cumulative dive fierce siege strategy can be updated according to the following formula:

[0144]

[0145] This embodiment uses different strategies in different search stages to find the best solution.

[0146] Select the update strategy according to different situations and update each member individual, then judge whether the iteration end condition is met. If not, perform the complete optimization search operation again. If it is met, end the optimization search operation and use the current optimal candidate model as the multi-layer deep learning model to be used by the monitoring method. The training of the multi-layer deep learning model is completed.

[0147] It is understandable that only using the above update strategy is prone to fall into the problem of local optimality, so the problem of local optimality can be solved by updating the prey position. The update action of the prey position is performed after updating the prey position based on the current optimal candidate model and before updating each update member individual according to the current update strategy. The specific update method can be Cauchy mutation or adaptive adjustment.

[0148] Taking adaptive adjustment as an example, specifically, in the optimization search operation, based on the prey position and all current member individuals, before updating each member individual according to the current update strategy, it also includes:

[0149] Based on the current operation number of the optimization search operation, determine the corresponding current adaptive weight;

[0150] Based on the current adaptive weight, update the current prey position.

[0151] The adaptive weight can be determined according to the following formula:

[0152]

[0153] The update of the prey position can be performed according to the following formula:

[0154] X' rabbit (t) = ω(t) × X rabbit (t);

[0155] Where ω(t) is the adaptive weight, X' rabbit (t) is the updated prey position corresponding to this adaptive weight.

[0156] When the adaptive weight is large, the training of the multi-layer deep learning model will use relatively more time for global search, and when the adaptive weight is small, the local search time will be more; the prey position represents the optimal solution for the current number of operations, and its update determines the performance of search optimization. Therefore, in order to improve the performance of local search, the neighborhood of the prey position can be updated using the adaptive weight to find the optimal solution, so that the search optimization in this embodiment has a better local mining effect. This optimization method can be used in the local search stage.

[0157] Taking Cauchy mutation as an example, specifically, in the optimization search operation, based on the prey position and all current member individuals, before updating each member individual according to the current update strategy, it also includes:

[0158] Determine the current amount of variation based on the Cauchy distribution;

[0159] Use the current mutation amount to update the current prey position.

[0160] Specifically, the standard distribution function formula of the Cauchy distribution is as follows:

[0161]

[0162] The current variation determined based on the Cauchy distribution can be recorded as Cauchy(0,1), and the process of updating the prey position includes:

[0163] X” rabbit (t) = X rabbit (t)+X rabbit (t)×Cauchy(0,1);

[0164] X” rabbit (t) is the updated prey position corresponding to this Cauchy distribution.

[0165] The peak value of the Cauchy function is relatively small. After the Cauchy mutation, this embodiment can search for the global optimal value more and use relatively less time to mine the local interval. In addition, the Cauchy function drops relatively slowly from the peak to both sides. After the position is updated through the Cauchy mutation, the constraint of the member individual on the local extreme point decreases, and the local extreme point can be jumped out. This optimization method can be used in the global search stage.

[0166] In some embodiments, the purpose of setting up the digital twin model is to simulate the physical device. The data receiving and processing process and model of the physical device can be used as the basis for constructing the digital twin model. Therefore, the process of obtaining the digital twin model can specifically include:

[0167] Establish a data perception process to collect component or board status information from the test endpoints reserved by the components or boards of the measurement and control equipment;

[0168] Data preprocessing: Since the collected current, voltage, displacement and other information are mostly analog data, the collected information is converted into A / D;

[0169] The pre-processed status data is collected through the internal port to the data acquisition device, which usually consists of a controller and supporting working software;

[0170] The data collection equipment is synchronized according to time, and the collected multi-channel data is packaged and fused according to the agreed protocol to form a network data packet;

[0171] Obtain multi-source real-time data and multi-source historical data from measurement and control equipment, integrate and store these data, and build a real database;

[0172] Establish a three-dimensional model corresponding to the physical measurement and control equipment: accurately measure the actual size of the physical model of the measurement and control equipment, and establish a refined three-dimensional model in proportion; when establishing the three-dimensional model, it is necessary to retain the physical characteristics and operating status of the actual equipment as much as possible; the model material and texture of the three-dimensional model are modeled using high-definition photos of the acquisition equipment; the three-dimensional model can simulate the normal state and fault state of the equipment, and highlight the fault state and recovery state; after the three-dimensional model is established, each three-dimensional model has an information association interface for associating with behavior-driven data to achieve visualization of information flow and behavior logic.

[0173] Construct a logical behavior model for the assembled digital three-dimensional model: For measurement and control equipment, the working principle of the equipment, the internal information flow, and the logical relationship between system components can be regarded as solidified quantitative after the equipment is manufactured. Constructing a behavior model is to digitally express the working principle of the equipment, the internal information flow, and the logical relationship between system components; the construction of the initial model is achieved by software programming; the collection of digital three-dimensional models and their logical behavior models is the digital twin model in this embodiment.

[0174] The running status information of the digital twin model is generated based on the operation of the digital twin model. When the digital twin model is running, the physical acquisition device sends the packaged data to the remote digital device through the network. After receiving the data packet, the digital device needs to parse the data packet and reversely parse the data according to the previous packaging protocol of the acquisition device. After parsing, different types of channel data are classified according to the data identifier and sent to the corresponding driver mapping module to drive the corresponding digital twin model on the digital device.

[0175] Taking the valve positioner device in the measurement and control equipment as an example, the monitoring of the digital twin model of the valve positioner device is implemented based on the specific multi-layer deep learning model of the valve positioner device. The training process of the multi-layer deep learning model here depends on the training sample data set and the test sample data set corresponding to the device. The sample data collection process in these two sample data sets includes:

[0176] Sampling parameters such as the control current I, power supply voltage V, and ambient temperature T of the device, with a sampling period of 10s, and using the sampled control current I, power supply voltage V, and ambient temperature T as excitation information for the digital twin model of the input valve positioner device, so that the digital twin model is based on the input action, and then the response information of the digital twin model, such as displacement S, is obtained, and at the same time, the device state at each sampling is determined to be normal or faulty;

[0177] A sample database is constructed based on the current excitation information, response information and device status. In the sample database, the input sample of each sample data includes the control current I, the power supply voltage V, the ambient temperature T and the displacement S, and the actual output sample is the device status.

[0178] According to the proportion, all sample data are divided into training sample data set and test sample data set.

[0179] For example, in a specific sampling, after statistics, there are 75,800 sample data in the sample database with normal equipment status, accounting for 93%; there are 5,430 sample data with equipment status failure, accounting for 7%. 90% of the sample data can be used as a training sample data set, and the remaining 10% can be used as a test sample data set.

[0180] It can be understood that the performance evaluation of the diagnostic results of the model can be achieved through fitness and determination coefficient. The smaller the fitness value, the better the prediction effect, and the closer the determination coefficient is to 1, the higher the identification accuracy.

[0181] The fitness is calculated according to the following formula:

[0182]

[0183] The coefficient of determination is calculated according to the following formula:

[0184]

[0185] Among them, e MAE is the fitness value, R 2 is the determination coefficient, m is the total number of sample data in the test sample data set, and j are the predicted output sample and actual output sample of the j-th sample data, respectively. is the average value of the sample data in the test sample data set.

[0186] In order to compare the accuracy of the model in this embodiment with the conventional model, this embodiment compares multiple trainings with different ideas:

[0187] Idea 1: Use the LSTM model for fault prediction. According to human experience, set the batch size of the LSTM model to 16, the total number of hidden layers (i.e., the number of hidden layers) to 1, the number of hidden layer neural units (i.e., the number of hidden layer neurons) to 32, and the learning rate to 0.01. Use this prediction model to predict faults based on the test sample data set and obtain the final diagnosis result.

[0188] Idea 2: Using the training process of the multi-layer deep learning model in this embodiment, but excluding the Cauchy mutation or adaptive adjustment of the prey position, the optimal solution of the hyperparameters is: batch size is 32, number of hidden layers is 2, number of hidden neurons is 20, and learning rate is 0.012. The multi-layer deep learning model corresponding to the optimal solution is used to predict faults based on the historical data set to obtain the final diagnosis result.

[0189] Idea 3: Using the training process of the multi-layer deep learning model in this embodiment, and adding Cauchy mutation or adaptive adjustment of the prey position, the optimal solution of the hyperparameters is finally obtained: batch size is 32, number of hidden layers is 3, number of hidden neurons is 64, and learning rate is 0.006. The multi-layer deep learning model corresponding to the optimal solution is used to predict faults based on the historical data set to obtain the final diagnosis result.

[0190] The hyperparameters of the learning models generated by these three ideas and the final performance evaluation can be shown in Table 1 below:

[0191] Table 1 Hyperparameters and performance evaluation of three ideas

[0192] Ideas Batch size Number of hidden layers Number of neurons in the hidden layer Learning Rate Fitness Coefficient of determination Idea 1 16 1 32 0.01 0.257 0.832 Idea 2 32 2 20 0.012 0.098 0.912 Idea 3 32 3 64 0.006 0.012 0.981

[0193] Obviously, Ideas 2 and 3 use the training of the multi-layer deep learning model in this embodiment, and the prediction accuracy is significantly improved compared with the hyperparameters empirically set in Idea 1. Idea 3 is more accurate than Idea 2.

[0194] Therefore, the multi-layer deep learning model in this embodiment is used to monitor the operating status information of the digital twin model, which can more accurately and predictively detect potential faults and complex faults.

[0195] In all the above optional solutions, the selection basis of the specific strategy, the updating means, and the numerical settings of each parameter can be adjusted according to the actual situation and user needs, and are not limited here.

[0196] All the above optional technical solutions can be combined in any way to form optional embodiments of the present application, and will not be described in detail here. It should be understood that the size of the sequence number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0197] The following is an embodiment of the device of the present application, which can be used to execute the embodiment of the method of the present application. For details not disclosed in the embodiment of the device of the present application, please refer to the embodiment of the method of the present application.

[0198] Figure 4 It is a schematic diagram of a monitoring device based on a digital twin model provided in an embodiment of the present application.

[0199] like Figure 4 As shown, the monitoring device comprises:

[0200] An acquisition module 401 is used to acquire a digital twin model of a target device and also to acquire operating status information of the digital twin model;

[0201] A monitoring module 402 is used to input the operating status information into the trained multi-layer deep learning model to obtain the output result of the multi-layer deep learning model as the monitoring result of the digital twin model;

[0202] Among them, the training process of the multi-layer deep learning model includes:

[0203] Obtain a training sample data set and a test sample data set;

[0204] Initialize each member individual in the hyperparameter population, each member individual includes a set of candidate hyperparameters;

[0205] The optimization search operation is repeated until the iteration end condition is met, and the current optimal candidate model is used as the multi-layer deep learning model;

[0206] Among them, the optimization search operation includes:

[0207] Based on the training sample data set and each member individual, a candidate model corresponding to each member individual is generated;

[0208] Determine the fitness value of each candidate model based on the test sample data set, and take the candidate model with the smallest fitness value as the current optimal candidate model;

[0209] Update each member based on all current member individuals.

[0210] The device of the embodiment of the present application utilizes a multi-layer deep learning model to analyze the operating status information in the digital twin model to obtain the monitoring results of the digital twin model. The training of the multi-layer deep learning model is implemented by a cyclic iteration of the optimization search operation based on the hyperparameter population. Compared with the training scheme based on empirical values, the multi-layer deep learning model that has been trained in the present application has higher recognition accuracy and better prediction effect. The multi-layer deep learning model is combined with the digital simulation characteristics of the digital twin model to predictably analyze nonlinear and complex mode equipment failures and more accurately and deeply discover potential problems in operation.

[0211] In an exemplary embodiment, the optimization search operation is cyclically performed until an iteration end condition is satisfied, and before the current optimal candidate model is used as the multi-layer deep learning model, the method further includes:

[0212] Initialize the prey escape energy and prey capture probability corresponding to the hyperparameter population;

[0213] In the optimization search operation, the process of updating each member individual based on all current member individuals includes:

[0214] Update the prey position of the hyperparameter population based on the current best candidate model;

[0215] Based on the current number of operations of the optimal search operation, the current prey escape energy is updated;

[0216] Determine the current update strategy based on the current prey escape energy and prey capture probability;

[0217] Based on the prey position and all current member individuals, each member individual is updated according to the current update strategy.

[0218] In an exemplary embodiment, in the optimization search operation, the process of determining the current update strategy based on the current prey escape energy and the prey capture probability includes:

[0219] Determine the absolute value of the current prey escape energy as the corresponding current energy absolute value;

[0220] Determine whether the absolute value of the current energy is less than the first energy;

[0221] If not, determine that the current update strategy is the exploration strategy;

[0222] If so, determine whether the absolute value of the current energy is less than the second energy, and determine whether the probability of prey capture is less than a preset probability; the first energy is greater than the second energy and greater than zero;

[0223] When the absolute value of the energy is not less than the second energy and the prey capture probability is not less than the preset probability, determining that the current update strategy is a gentle siege strategy;

[0224] When the absolute value of the energy is less than the second energy and the prey capture probability is not less than the preset probability, the current update strategy is determined to be a fierce siege strategy;

[0225] When the absolute value of the energy is not less than the second energy and the prey capture probability is less than the preset probability, the current update strategy is determined to be a cumulative speed diving gentle siege strategy;

[0226] When the absolute value of the energy is less than the second energy and the prey capture probability is less than the preset probability, the current update strategy is determined to be a cumulative dive-style fierce siege strategy;

[0227] The specific exploration strategy is: updating the current member individual based on any member individual, or updating the current member individual based on the average position of all member individuals and the prey position;

[0228] The gentle siege strategy is specifically as follows: based on the current prey escape energy, the current member individual, the prey position and the preset escape coefficient corresponding to the prey position, the current member individual is updated;

[0229] The fierce siege strategy is as follows: based on the current member individual and the prey position, update the current member individual;

[0230] The specific strategy of the gradual dive-type gentle siege is as follows: based on the current prey escape energy, the current member individual, the prey position, the preset escape coefficient and the preset constant, the current member individual is updated;

[0231] The specific strategy of the accelerated dive and fierce siege is as follows: based on the current prey escape energy, average position, prey position, preset escape coefficient and preset constant, the current member individuals are updated.

[0232] In an exemplary embodiment, in the optimization search operation, based on the prey position and all current member individuals, before updating each member individual according to the current update strategy, it also includes:

[0233] Based on the current operation number of the optimization search operation, determine the corresponding current adaptive weight;

[0234] Based on the current adaptive weight, update the current prey position.

[0235] In an exemplary embodiment, in the optimization search operation, based on the prey position and all current member individuals, before updating each member individual according to the current update strategy, it also includes:

[0236] Determine the current amount of variation based on the Cauchy distribution;

[0237] Use the current mutation amount to update the current prey position.

[0238] In an exemplary embodiment, the candidate hyperparameters included in each member include multiple ones of the total number of hidden layers, the number of neural units in each hidden layer, the batch size, the time window size, the number of training times and the learning rate.

[0239] In an exemplary embodiment, the operating status information includes:

[0240] Input stimulus information to the digital twin model;

[0241] and / or, response information when the digital twin model responds to stimulus information according to its internal operating logic.

[0242] Figure 5 Schematic diagram of an electronic device 5 provided in an embodiment of the present application. Figure 5 As shown, the electronic device 5 of this embodiment includes: a processor 501, a memory 502, and a computer program 503 stored in the memory 502 and executable on the processor 501. When the processor 501 executes the computer program 503, the steps in the above-mentioned method embodiments are implemented. Alternatively, when the processor 501 executes the computer program 503, the functions of the modules / units in the above-mentioned device embodiments are implemented.

[0243] The electronic device 5 may be a desktop computer, a notebook, a PDA, a cloud server, or other electronic device. The electronic device 5 may include, but is not limited to, a processor 501 and a memory 502. Those skilled in the art will appreciate that Figure 5 The electronic device 5 is merely an example and does not limit the electronic device 5 , and may include more or less components than those shown in the figure, or different components.

[0244] The processor 501 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0245] The memory 502 may be an internal storage unit of the electronic device 5, for example, a hard disk or memory of the electronic device 5. The memory 502 may also be an external storage device of the electronic device 5, for example, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 5. The memory 502 may also include both an internal storage unit of the electronic device 5 and an external storage device. The memory 502 is used to store computer programs and other programs and data required by the electronic device.

[0246] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units.

[0247] If the integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium, such as a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. The computer program may include computer program code, which may be in source code form, object code form, executable file or some intermediate form. The readable storage medium may include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0248] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A monitoring method based on a digital twin model, characterized in that: include: Obtain the digital twin model of the target device; Obtaining operating status information of the digital twin model; Inputting the operating status information into a multi-layer deep learning model that has completed training to obtain an output result of the multi-layer deep learning model as a monitoring result of the digital twin model; The training process of the multi-layer deep learning model includes: Obtain a training sample data set and a test sample data set; Initializing each member individual in the hyperparameter population, each member individual including a set of candidate hyperparameters; The optimization search operation is repeated until the iteration end condition is met, and the current optimal candidate model is used as the multi-layer deep learning model; The optimization search operation includes: Based on the training sample data set and each of the individual members, generating a candidate model corresponding to each of the individual members; Determine the fitness value of each candidate model based on the test sample data set, and use the candidate model with the smallest fitness value as the current optimal candidate model; Based on all the current member individuals, each member individual is updated.

2. The method according to claim 1, characterized in that The optimization search operation is cyclically performed until the iteration end condition is satisfied, and before the current optimal candidate model is used as the multi-layer deep learning model, the method further includes: Initializing the prey escape energy and prey capture probability corresponding to the hyperparameter population; In the optimization search operation, the process of updating each member individual based on all the current member individuals includes: Update the prey position of the hyperparameter population based on the current optimal candidate model; Based on the current number of operations of the optimization search operation, updating the current prey escape energy; Determining a current update strategy based on the current prey escape energy and the prey capture probability; Based on the position of the prey and all the current member individuals, each member individual is updated according to the current update strategy.

3. The method according to claim 2, characterized in that In the optimization search operation, the process of determining the current update strategy based on the current prey escape energy and the prey capture probability includes: Determining the absolute value of the current prey escape energy as the corresponding current energy absolute value; Determine whether the current absolute value of the energy is less than the first energy; If not, determine that the current update strategy is the exploration strategy; If so, determine whether the current absolute value of the energy is less than the second energy, and determine whether the prey capture probability is less than a preset probability; the first energy is greater than the second energy and greater than zero; When the energy absolute value is not less than the second energy and the prey capture probability is not less than the preset probability, determining that the current update strategy is a gentle siege strategy; When the energy absolute value is less than the second energy and the prey capture probability is not less than the preset probability, determining that the current update strategy is a fierce siege strategy; When the energy absolute value is not less than the second energy and the prey capture probability is less than the preset probability, determining that the current update strategy is a cumulative speed diving gentle siege strategy; When the energy absolute value is less than the second energy and the prey capture probability is less than the preset probability, determining that the current update strategy is a cumulative speed diving fierce siege strategy; The exploration strategy is specifically: updating the current member individual based on any of the member individuals, or updating the current member individual based on the average position of all the member individuals and the position of the prey; The gentle siege strategy specifically includes: updating the current individual member based on the current escape energy of the prey, the current individual member, the prey position and the preset escape coefficient corresponding to the prey position; The fierce siege strategy specifically includes: updating the current member individual based on the current member individual and the prey position; The cumulative speed diving gentle siege strategy specifically includes: updating the current individual member based on the current escape energy of the prey, the current individual member, the position of the prey, the preset escape coefficient and the preset constant; The cumulative dive-style fierce siege strategy is specifically: based on the current prey escape energy, the average position, the prey position, the preset escape coefficient and the preset constant, the current member individuals are updated.

4. The method according to claim 2, characterized in that: In the optimization search operation, based on the prey position and all the current member individuals, before updating each member individual according to the current update strategy, it also includes: Determining a corresponding current adaptive weight based on the current operation number of the optimization search operation; Based on the current adaptive weight, the current prey position is updated.

5. The method according to claim 2, characterized in that: In the optimization search operation, based on the prey position and all the current member individuals, before updating each member individual according to the current update strategy, it also includes: Determine the current amount of variation based on the Cauchy distribution; The current position of the prey is updated using the current variation.

6. The method according to claim 1, characterized in that The candidate hyperparameters included in each of the member individuals include multiple ones of the total number of hidden layers, the number of neural units in each hidden layer, the batch size, the time window size, the number of training times and the learning rate.

7. The method according to any one of claims 1 to 6, characterized in that The operation status information includes: Inputting stimulus information of the digital twin model; and / or, response information when the digital twin model responds to the stimulus information according to its internal operating logic.

8. A monitoring device based on a digital twin model, characterized in that: include: An acquisition module, used to acquire a digital twin model of a target device, and also used to acquire operating status information of the digital twin model; A monitoring module, used to input the operating status information into a multi-layer deep learning model that has completed training, so as to obtain an output result of the multi-layer deep learning model as a monitoring result of the digital twin model; The training process of the multi-layer deep learning model includes: Obtain a training sample data set and a test sample data set; Initializing each member individual in the hyperparameter population, each member individual including a set of candidate hyperparameters; The optimization search operation is repeated until the iteration end condition is met, and the current optimal candidate model is used as the multi-layer deep learning model; The optimization search operation includes: Based on the training sample data set and each of the individual members, generating a candidate model corresponding to each of the individual members; Determine the fitness value of each candidate model based on the test sample data set, and use the candidate model with the smallest fitness value as the current optimal candidate model; Based on all the current member individuals, each member individual is updated.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images