Dynamic equipment management service system and method for deep learning

Abstract

The invention discloses a dynamic equipment management service system and method for deep learning. The method comprises the following steps: S1, collecting real-time operation data of equipment and preprocessing the real-time operation data; s2, constructing a long-short-term memory network model, and outputting comprehensive early warning indexes of the equipment; s3, constructing an optimization problem based on the comprehensive early warning index of the equipment, and setting an objective function and constraint conditions; s4, performing global search on the optimization problem by adopting an improved dragon fly algorithm, and generating a plurality of resource scheduling and maintenance candidate schemes; s5, performing local optimization on the candidate scheme by using a simulated annealing algorithm, and determining an optimal resource scheduling and maintenance decision scheme; and S6, feeding back the real-time prediction result and the optimal decision result to form closed-loop control. Through fusion of deep learning prediction, the improved dragon fly algorithm and the simulated annealing algorithm, real-time closed-loop control of accurate early warning of the equipment state and dynamic resource scheduling decision is realized, so that the failure rate is effectively reduced, and the equipment management efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and particularly to a dynamic device management service system and method for deep learning. Background Art

[0002] At present, the fields of industrial automation and intelligent manufacturing are developing rapidly, and the efficient and stable operation of equipment puts forward higher requirements for the production efficiency and economic benefits of enterprises. In order to realize the intelligent and efficient management of equipment, advanced technologies such as the Internet of Things, big data, and deep learning have been widely adopted in the prior art to realize the real-time monitoring and prediction of equipment status. However, in the actual application process, due to the variety of equipment types, complex operating environments, and diverse data characteristics, traditional equipment management methods often face problems such as untimely data processing, insufficient prediction model accuracy, and lagging resource scheduling decisions. Existing systems mostly rely on fixed rules or simple statistical methods for equipment fault prediction and maintenance scheduling, and their reaction speed to the operating status of equipment in a dynamic environment is slow, and it is difficult to cope with the non-linear and random problems of equipment status changes. This makes it impossible for equipment to obtain effective maintenance in a timely manner in the event of sudden failures or abnormal fluctuations, resulting in production interruptions, resource waste, and increased safety risks.

[0003] In recent years, with the development of deep learning technology, models such as long short-term memory networks (LSTM) have been introduced into the field of equipment status prediction. Utilizing their advantages in processing time series data, they predict the future status, abnormal probability, and failure risk of equipment, and use this as the basis for maintenance decisions. Although such methods have improved the prediction accuracy to a certain extent, a single deep learning model usually cannot directly convert the prediction results into an efficient maintenance scheduling scheme, because equipment management not only involves data prediction, but also involves how to perform real-time resource scheduling and fault prevention on a large number of equipment in a dynamically changing environment. On the other hand, optimization algorithms such as genetic algorithms, particle swarm optimization, and simulated annealing have also been used to solve equipment scheduling and maintenance decision problems, but these traditional optimization methods have deficiencies in global search ability and local convergence speed, are easily trapped in local optimal solutions or have a slow convergence speed, and it is difficult to maintain high accuracy and real-time response ability when facing complex, multi-objective optimization problems.

[0004] In addition, in recent years, animal-inspired optimization algorithms have received attention due to their strong global search ability and flexible swarm intelligence mechanism. As an emerging swarm intelligence optimization algorithm, the dragonfly algorithm demonstrates good performance in dealing with nonlinear, multi-modal, and high-dimensional optimization problems by simulating swarm behavior. In the prior art, some studies have attempted to combine the dragonfly algorithm with other optimization methods to achieve a balance between global search and local optimization. However, the traditional dragonfly algorithm still has defects in practical applications, such as being prone to falling into local optima during the search process, having fixed parameter settings, and lacking an adaptive adjustment mechanism. In response to these deficiencies, some improvement schemes propose introducing individual memory behavior and an adaptive weight adjustment strategy into the dragonfly algorithm, as well as combining the simulated annealing algorithm to perform local optimization on candidate solutions. However, these improvement schemes still have problems such as insufficiently tight feedback between the deep learning prediction results and the optimization algorithm, and the optimization process not fully considering the dynamic matching between the actual operating state of the device and the prediction indicators, resulting in the application effect in actual device management not reaching the ideal level.

[0005] Therefore, how to provide a dynamic device management service system and method for deep learning is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] An object of the present invention is to propose a dynamic device management service system and method for deep learning. The present invention makes full use of deep learning technology, swarm intelligence optimization algorithms, and the simulated annealing algorithm, and details the closed-loop dynamic decision-making process of real-time data acquisition, deep prediction, global candidate solution generation, and local fine optimization of device data. The method first uses a deep learning model to train the historical data of the device to accurately predict the future state, abnormal probability, and failure risk of the device; then, an improved dragonfly algorithm is used to perform a global search on the optimization problem constructed based on the prediction output to generate multiple candidate resource scheduling and maintenance decision-making schemes; subsequently, by introducing a dual individual memory behavior, adaptive weight adjustment, and local perturbation mechanism, the simulated annealing algorithm is used to perform fine optimization on the candidate solutions to achieve rapid convergence to the local optimal solution; finally, the search process is adjusted through a dynamic temperature drop strategy to achieve real-time closed-loop feedback between prediction and decision-making. This method has the advantages of high real-time performance, strong adaptability, high optimization accuracy, and reduced device failure risk, thus providing an intelligent and efficient solution for device maintenance and resource scheduling.

[0007] A dynamic device management service method for deep learning according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect the real-time operating data of the device and preprocess the real-time operating data;

[0009] S2. Use the preprocessed real-time operation data to construct a long short-term memory network model, train the long short-term memory network model, and perform online real-time prediction to output the comprehensive early warning index of the device;

[0010] S3. Construct an optimization problem based on the comprehensive early warning index of the device, including setting the current state of the device, the comprehensive early warning index, the maintenance cost, and the failure risk as the objective function and constraints;

[0011] S4. Use the improved dragonfly algorithm to perform a global search on the optimization problem, initialize multiple candidate solutions, and generate multiple resource scheduling and maintenance candidate solutions through information sharing and position update;

[0012] S5. Use the generated resource scheduling and maintenance candidate solutions as the initial solution, and use the simulated annealing algorithm to perform local optimization on the initial solution. Through introducing perturbations, judging the acceptance probability, and setting the temperature reduction strategy, make fine adjustments to determine the optimal resource scheduling and maintenance decision-making solution as the final execution solution of the system;

[0013] S6. Feed back the real-time prediction result and the optimal resource scheduling and maintenance decision result to the deep learning prediction module to form a closed-loop control.

[0014] Optionally, the S2 specifically includes:

[0015] S21. Obtain the preprocessed real-time operation data, store the real-time operation data in the data buffer, and construct a data input stream in the time series format;

[0016] S22. Use the preprocessed real-time operation data to construct a long short-term memory network model. The long short-term memory network model includes an input layer, multiple hidden layers, and an output layer, where the hidden layer adjusts the information flow through a gating mechanism;

[0017] S23. Configure the model structure and initialize the parameters of the long short-term memory network model, set the number of neurons, the type of activation function, the learning rate, the loss function, and the optimization algorithm, and initialize the weight parameters and bias parameters;

[0018] S24. Perform offline training on the long short-term memory network model, perform multiple rounds of iterative training based on historical operation data, optimize the parameters of the long short-term memory network model by minimizing the loss function error, and adjust the weight matrix using the batch gradient descent strategy;

[0019] S25. Deploy the trained long short-term memory network model to the online prediction system, receive the real-time input data stream, process the input data at each time step step by step, calculate the running state of the device at the current moment in combination with historical information, and generate short-term and long-term prediction results;

[0020] S26. Output the comprehensive warning indicators of the device based on the generated short-term and long-term prediction results. The comprehensive warning indicators include the future state of the device, the probability of abnormality, and the failure risk value.

[0021] Optionally, the S3 specifically includes:

[0022] S31. Obtain the output comprehensive warning indicators of the device, including the future state of the device, the probability of abnormality, and the failure risk value, and combine the current operating state information of the device to form an input data set for optimization modeling.

[0023] S32. Analyze the operating mode and failure characteristics of the device, and set an optimization problem, where the optimization problem includes reducing the probability of device failure and reducing maintenance costs.

[0024] S33. Extract the operating constraint conditions of the device, including the upper limit of the operating capacity of the device, the maintenance cycle limit, the resource allocation constraint, and the failure impact range.

[0025] S34. Establish an optimization objective function, and define the objective function of the device management optimization problem according to the current operating state information of the device and the set optimization problem.

[0026] S35. Construct the constraint conditions of the optimization problem, and combine the operating constraint conditions of the device to set the resource scheduling constraint, the device maintenance constraint, and the operating safety requirements.

[0027] S36. Complete the construction of the optimization problem, and input the objective function and the constraint conditions into the optimization solving module.

[0028] Optionally, the S4 specifically includes:

[0029] S41. Based on the above optimization problem, set a fitness function f(X) to minimize the device maintenance cost, the device failure risk, and the deviation between the decision-making and the comprehensive warning indicators:

[0030] f(X) = αC(X) + βR(X) + γ||I - X||;

[0031] Where X represents the decision variable, C(X) represents the device maintenance cost, R(X) represents the device failure risk, I represents the comprehensive warning indicator of the device, ||I - X|| is the deviation measure, and α, β, and γ are weight coefficients.

[0032] S42. Construct an initial population P, set the population size N, and randomly initialize N solutions X i , where each solution X i represents a feasible resource scheduling and maintenance candidate solution, and at the same time record the historical optimal solution of each individual and the optimal solution in the most recent m iterations

[0033] S43. Introduce double individual memory behavior and define the individual memory behavior item P i as:

[0034]

[0035] Among them, λ1 and λ2 are individual memory influence factors. Introduce the sixth behavior. The purpose of the individual memory behavior is to make full use of the historical experience accumulated by each individual in the search process, so as to better guide the individual to move towards its historical best or recent best position. By recording and using the optimal solutions obtained by the individual in the past search process, this behavior can help the individual overcome the limitation of local optimum, accelerate convergence, and achieve a better balance between global search and local search. Compared with the original five behaviors of attraction, alignment, aggregation, environmental perception, and avoidance, the individual memory behavior provides an additional information channel for each individual, enabling the individual to refer to its own historical successful experience for local fine optimization while collaborating in the group, thus improving the search efficiency and overall optimization accuracy. In this way, the sixth behavior complements the other five behaviors and acts together on the position update process of the individual, enhancing the adaptability and robustness of the entire algorithm in a dynamic environment;

[0036] S44. According to the six behavior rules of the dragonfly algorithm, calculate the update direction of each individual, and define the individual velocity as:

[0037]

[0038] Among them, represents the velocity of individual i at the t-th iteration, S i represents the attraction behavior item, A i represents the alignment behavior item, C i represents the aggregation behavior item, E i represents the environmental perception behavior item, M i represents the avoidance behavior item, P i represents the double individual memory behavior item, w is the inertia weight, represents the velocity of individual i at the (t + 1)-th iteration, S is the adjustment coefficient of the attraction behavior, A is the adjustment coefficient of the alignment behavior, C is the adjustment coefficient of the aggregation behavior, E is the adjustment coefficient of the environmental perception behavior, M is the adjustment coefficient of the avoidance behavior;

[0039] S45. Use the updated velocity and the comprehensive early warning index I to perform individual position update:

[0040]

[0041] Among them, γ is the convergence factor:

[0042]

[0043] Among them, γ max and γ min are the initial and minimum convergence factors respectively, δ max is the maximum feedback coefficient, T is the maximum number of iterations, t is the current number of iterations, represents the updated position of the i-th individual in the (t + 1)-th iteration, represents the position of the i-th individual in the t-th iteration;

[0044] S46. Calculate the fitness value f(X i ) for all individuals X i ), and record the current global optimal solution X * ;

[0045] S47. When the population diversity decreases or the global optimal solution has not been updated continuously for multiple times, a local perturbation mechanism is applied to some individuals, and the position is updated by combining adaptive Gaussian perturbation and differential perturbation:

[0046] X i new = X i + σ·N(0, 1)+ η·ρ·(X * - X i );

[0047] Among them, σ is the basic perturbation intensity, N(0, 1) is the standard normal distribution with a mean of 0 and a variance of 1, η is the differential perturbation factor, ρ is a random number taken from the interval [0, 1], X * is the current global optimal solution, and X i new is the updated position;

[0048] S48. When the maximum number of iterations is reached or the convergence error meets a predetermined threshold, output the resource scheduling and maintenance candidate solution X * , otherwise return to S43 to continue the iteration.

[0049] Optionally, the specific steps of S5 are as follows:

[0050] S51. Take the obtained resource scheduling and maintenance candidate solution X i as the initial solution;

[0051] S52. Set the initial temperature of the simulated annealing algorithm

[0052] S53. For each candidate solution X i generate a new candidate solution X′ i by using adaptive perturbation:

[0053] X′ i = X i + η·ΔX + κ·(I - X i );

[0054] Where η is the basic perturbation factor, ΔX is the basic perturbation amount, κ is the dynamic feedback factor, and I is the comprehensive early warning index;

[0055] S54. Calculate the fitness difference Δf between the current candidate solution X i and the newly generated candidate solution X′ i :

[0056] Δf = f(X′ i ) - f(X i );

[0057] Where f(X) is the fitness function described in S41 of claim 4;

[0058] S55. Determine whether to accept the new candidate solution X′ i , first calculate the basic acceptance probability according to the fitness difference Δf, and combine the comprehensive early warning index to evaluate the matching degree between the candidate solution and the predicted trend, and adaptively adjust the basic acceptance probability, and finally calculate the acceptance probability P a :

[0059] Ψ = exp(ζ·(‖I - X i ‖ - ‖I - X′ i ‖));

[0060]

[0061] Where T a is the current temperature, Φ is the dynamic scaling factor, λ is the feedback adjustment coefficient, ζ is the dynamic improvement coefficient, Ψ is the dynamic improvement factor, exp is the exponential function, min is the minimum operation, T a is the current temperature, generate a random number r, if r ≤ P a then update the candidate solution X i = X′ i ;

[0062] S56. Adopt an adaptive temperature drop strategy, and dynamically adjust the cooling rate according to the change of the candidate solution fitness and the deviation degree between the candidate solution and the comprehensive early warning index I:

[0063]

[0064] Where θ is the temperature drop coefficient, μ is the dynamic adjustment factor, X * is the current global optimal solution, ‖I - X *‖ represents the deviation degree between the comprehensive early warning index and the current global optimal solution. Repeat S53 to S56 until the local optimal condition is satisfied or the predetermined number of iterations is reached, and output the candidate solution after local optimization as the final resource scheduling and maintenance decision plan.

[0065] A dynamic device management service system for deep learning according to an embodiment of the present invention includes the following modules:

[0066] The data acquisition and preprocessing module is used to collect device data in real time, clean, normalize and extract features from the device data;

[0067] The deep learning prediction module is used to train and online predict the preprocessed device data by using the long short-term memory network, and output the comprehensive early warning index;

[0068] The global search module is used to generate a candidate resource scheduling and maintenance decision plan according to the comprehensive early warning index and the optimization problem construction;

[0069] The local optimization module is used to perturb the candidate resource scheduling and maintenance decision plan, judge the acceptance probability and perform adaptive temperature drop;

[0070] The closed-loop feedback and decision adjustment module is used to real-time feedback the deep learning prediction results and optimization process information, and dynamically adjust the resource scheduling and maintenance decision;

[0071] The user interaction and monitoring module is used to display the device status, prediction results and decision plans, and support manual intervention and system monitoring.

[0072] The beneficial effects of the present invention are:

[0073] The device dynamic management optimization method proposed by the present invention realizes real-time and accurate prediction of the device operation state and efficient optimization of resource scheduling and maintenance decisions through the organic combination of deep learning prediction, improved dragonfly algorithm and simulated annealing algorithm. This method fully considers the dynamic changes of the device state and the complexity of the environment in all aspects of data acquisition, deep prediction and global search, so that the entire optimization process can form a real-time closed-loop feedback mechanism, ensuring that the system can quickly adjust the scheduling strategy in the face of sudden anomalies and device state fluctuations, reduce the failure risk and achieve preventive maintenance.

[0074] In addition, the present invention introduces an improved dragonfly algorithm, which is innovative in adopting dual individual memory behavior, adaptive weight adjustment, and local perturbation mechanism, effectively solving the problems that traditional optimization algorithms are prone to falling into local optima and having slow convergence speed, and significantly improving the global search and local fine optimization capabilities. The synergistic effect of the simulated annealing algorithm and the improved dragonfly algorithm enables the candidate solutions to better fit the comprehensive early warning indicators output by the deep learning prediction module while meeting the real-time requirements, thus achieving the balanced optimization among maintenance costs, failure risks, and prediction matching degrees.

[0075] In summary, the device dynamic management optimization method of the present invention not only has high real-time performance and adaptive ability, but also can continuously provide efficient and accurate resource scheduling and maintenance decisions in a changing industrial environment, effectively reducing the device failure rate and maintenance costs, improving the overall stability and efficiency of the system operation, and providing a solution with significant economic benefits and technological advancement for the fields of industrial automation and intelligent manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0077] Figure 1 is a flowchart of a dynamic device management service method for deep learning proposed by the present invention;

[0078] Figure 2 is a schematic structural diagram of a dynamic device management service system for deep learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0079] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0080] Refer to Figure 1 , a dynamic device management service method for deep learning, includes the following steps:

[0081] S1. Collect the real-time operation data of the device, and preprocess the real-time operation data;

[0082] S2. Use the preprocessed real-time operation data to construct a long short-term memory network model, perform long short-term memory network model training and online real-time prediction, and output the comprehensive early warning indicators of the device;

[0083] S3. Optimize the construction of comprehensive warning indicators for the device, including setting the current state of the device, comprehensive warning indicators, maintenance costs, and failure risks as the objective function and constraints;

[0084] S4. Use an improved dragonfly algorithm to perform a global search on the optimization problem, initialize multiple candidate solutions, and generate multiple resource scheduling and maintenance candidate solutions through information sharing and position updates;

[0085] S5. Take the generated resource scheduling and maintenance candidate solutions as the initial solution, use the simulated annealing algorithm to perform local optimization on the initial solution, and make fine adjustments by introducing perturbations, judging the acceptance probability, and setting the temperature drop strategy to determine the optimal resource scheduling and maintenance decision-making solution as the final execution solution of the system;

[0086] S6. Feed the real-time prediction results and the optimal resource scheduling and maintenance decision results back to the deep learning prediction module to form a closed-loop control.

[0087] In this embodiment, the S2 specifically includes:

[0088] S21. Obtain the preprocessed real-time operation data, store the real-time operation data in the data buffer, and construct a data input stream in the time series format;

[0089] S22. Use the preprocessed real-time operation data to construct a long short-term memory network model. The long short-term memory network model includes an input layer, multiple hidden layers, and an output layer, where the hidden layers regulate the information flow through a gating mechanism;

[0090] S23. Configure the model structure and initialize the parameters of the long short-term memory network model, set the number of neurons, the type of activation function, the learning rate, the loss function, and the optimization algorithm, and initialize the weight parameters and bias parameters;

[0091] S24. Perform offline training on the long short-term memory network model, perform multiple rounds of iterative training based on historical operation data, optimize the parameters of the long short-term memory network model by minimizing the loss function error, and adjust the weight matrix using the batch gradient descent strategy;

[0092] S25. Deploy the trained long short-term memory network model to the online prediction system, receive the real-time input data stream, process the input data at each time step step by step, calculate the current operating state of the device in combination with historical information, and generate short-term and long-term prediction results;

[0093] S26. Based on the generated short-term and long-term prediction results, output the comprehensive warning indicators of the device. The comprehensive warning indicators include the future state of the device, the abnormal probability, and the failure risk value.

[0094] In this embodiment, S3 specifically includes:

[0095] S31. Obtain the comprehensive early warning indicators of the output device, including the future state of the device, the abnormal probability, and the failure risk value, and combine the current operating status information of the device to form a data input set for optimization modeling;

[0096] S32. Analyze the operating mode and failure characteristics of the device, and set optimization problems, including reducing the probability of device failures and reducing maintenance costs;

[0097] S33. Extract the operating constraint conditions of the device, including the upper limit of the operating capacity of the device, the maintenance cycle limit, the resource allocation constraint, and the failure impact range;

[0098] S34. Establish an optimization objective function, and define the objective function of the device management optimization problem according to the current operating status information of the device and the set optimization problems;

[0099] S35. Construct the constraint conditions of the optimization problem, and combine the operating constraint conditions of the device to set the resource scheduling constraint, the device maintenance constraint, and the operating safety requirements;

[0100] S36. Complete the construction of the optimization problem, and input the objective function and the constraint conditions into the optimization solution module.

[0101] In this embodiment, S4 specifically includes:

[0102] S41. Based on the above optimization problems, set a fitness function f(X) to minimize the device maintenance cost, the device failure risk, and the deviation between the decision-making and the comprehensive early warning indicators:

[0103] f(X) = αC(X) + βR(X) + γ||I - X||;

[0104] Where X represents the decision variable, C(X) represents the device maintenance cost, R(X) represents the device failure risk, I represents the comprehensive early warning indicator of the device, ||I - X|| is the deviation metric, and α, β, and γ are weight coefficients;

[0105] S42. Construct an initial population P, set the population size N, and randomly initialize N solutions X i , where each solution X i represents a feasible resource scheduling and maintenance candidate solution, and at the same time record the historical optimal solution of each individual and the optimal solution in the most recent m iterations

[0106] S43. Introduce a dual individual memory behavior, and define the individual memory behavior item P i as:

[0107]

[0108] Among them, λ1 and λ2 are individual memory influence factors. A sixth behavior, the individual memory behavior, aims to make full use of the historical experience accumulated by each individual during the search process, so as to better guide the individual to move towards its historical best or recent best position. By recording and utilizing the optimal solutions obtained by the individual during past search processes, this behavior can help the individual overcome the limitation of local optimum, accelerate convergence, and achieve a better balance between global search and local search. Compared with the original five behaviors of attraction, alignment, aggregation, environmental perception, and avoidance, the individual memory behavior provides an additional information channel for each individual, enabling the individual to refer to its own historical successful experience for local fine-tuning while collaborating in the group, thereby improving the search efficiency and overall optimization accuracy. In this way, the sixth behavior complements the other five behaviors and jointly acts on the position update process of the individual, enhancing the adaptability and robustness of the entire algorithm in a dynamic environment;

[0109] S44. According to the six behavior rules of the dragonfly algorithm, calculate the update direction of each individual, and define the individual velocity as:

[0110]

[0111] Among them, represents the velocity of individual i at the t-th iteration, S i represents the attraction behavior term, A i represents the alignment behavior term, C i represents the aggregation behavior term, E i represents the environmental perception behavior term, M i represents the avoidance behavior term, P i represents the dual individual memory behavior term, w is the inertia weight, represents the velocity of individual i at the (t + 1)-th iteration, S is the adjustment coefficient of the attraction behavior, A is the adjustment coefficient of the alignment behavior, C is the adjustment coefficient of the aggregation behavior, E is the adjustment coefficient of the environmental perception behavior, M is the adjustment coefficient of the avoidance behavior;

[0112] S45. Use the updated velocity and the comprehensive warning index I to perform individual position update:

[0113]

[0114] Among them, γ is the convergence factor:

[0115]

[0116] Among them, γ max and γmin are the initial and minimum convergence factors, δ max is the maximum feedback coefficient, T is the maximum number of iterations, t is the current number of iterations, represents the updated position of the i-th individual in the (t + 1)-th iteration, represents the position of the i-th individual in the t-th iteration;

[0117] S46. For all individuals X i calculate the fitness value f(X i ), and record the current global optimal solution X * ;

[0118] S47. When the population diversity decreases or the global optimal solution has not been updated continuously for multiple times, a local perturbation mechanism is applied to some individuals, and the position is updated by combining adaptive Gaussian perturbation and differential perturbation:

[0119] X i new = X i + σ·N(0, 1)+ η·ρ·(X * - X i );

[0120] where σ is the basic perturbation intensity, N(0, 1) is the standard normal distribution with a mean of 0 and a variance of 1, η is the differential perturbation factor, ρ is a random number taken from the interval [0, 1], X * is the current global optimal solution, X i new is the updated position;

[0121] S48. When the maximum number of iterations is reached or the convergence error meets the predetermined threshold, output the resource scheduling and maintenance candidate solution X * , otherwise return to S43 to continue the iteration.

[0122] In this embodiment, the S5 specifically includes:

[0123] S51. Take the obtained resource scheduling and maintenance candidate solution X i as the initial solution;

[0124] S52. Set the initial temperature of the simulated annealing algorithm

[0125] S53. For each candidate solution X i generate a new candidate solution X' by using adaptive perturbation i :

[0126] X' i = X i + η·ΔX + κ·(I - X i));

[0127] Among them, η is the basic perturbation factor, ΔX is the basic perturbation amount, κ is the dynamic feedback factor, and I is the comprehensive early warning index;

[0128] S54. Calculate the fitness difference Δf between the current candidate solution X i and the newly generated candidate solution X′ i :

[0129] Δf = f(X′ i ) - f(X i );

[0130] Among them, f(X) is the fitness function described in S41 of claim 4;

[0131] S55. Determine whether to accept the new candidate solution X′ i , first calculate the basic acceptance probability according to the fitness difference Δf, and combine the comprehensive early warning index to evaluate the matching degree between the candidate solution and the predicted trend, and adaptively adjust the basic acceptance probability. Finally, calculate the acceptance probability P a :

[0132] Ψ = exp(ζ · (‖I - X i ‖ - ‖I - X′ i ‖));

[0133]

[0134] Among them, T a is the current temperature, Φ is the dynamic scaling factor, λ is the feedback adjustment coefficient, ζ is the dynamic improvement coefficient, Ψ is the dynamic improvement factor, exp is the exponential function, min is the minimum value operation, T a is the current temperature, generate a random number r. If r ≤ P a , then update the candidate solution X i = X′ i ;

[0135] S56. Adopt an adaptive temperature drop strategy, and dynamically adjust the cooling rate according to the change of the candidate solution fitness and the deviation degree between the candidate solution and the comprehensive early warning index I:

[0136]

[0137] Among them, θ is the temperature drop coefficient, μ is the dynamic adjustment factor, X * is the current global optimal solution, ‖I - X * ‖ represents the deviation degree between the comprehensive early warning index and the current global optimal solution. Repeat S53 to S56 until the local optimal condition is met or the predetermined number of iterations is reached, and output the locally optimized candidate solution as the final resource scheduling and maintenance decision plan.

[0138] Reference Figure 2 , a dynamic device management service system for deep learning, includes the following modules:

[0139] Data acquisition and preprocessing module, used to collect device data in real time, clean, normalize and extract features from the device data;

[0140] Deep learning prediction module, used to train and online predict the preprocessed device data using the long short-term memory network, and output comprehensive warning indicators;

[0141] Global search module, used to generate candidate resource scheduling and maintenance decision-making schemes according to the comprehensive warning indicators and the construction of optimization problems;

[0142] Local optimization module, used to perturb the candidate resource scheduling and maintenance decision-making schemes, judge the acceptance probability and perform adaptive temperature drop;

[0143] Closed-loop feedback and decision adjustment module, used to feedback the deep learning prediction results and optimization process information in real time, and dynamically adjust the resource scheduling and maintenance decisions;

[0144] User interaction and monitoring module, used to display the device status, prediction results and decision-making schemes, and support manual intervention and system monitoring.

[0145] Example 1:

[0146] To verify the feasibility of the present invention in implementation, the present invention is applied to a large manufacturing enterprise. The stability and efficiency of the equipment have always been the key factors affecting production efficiency and product quality. The enterprise has multiple key equipment such as numerically controlled machine tools and hydraulic supports. During the long-term operation of these equipment, failures often occur due to inaccurate device status prediction and lagging maintenance scheduling response, which in turn lead to production downtime and increased maintenance costs. To effectively solve the above problems, the present invention proposes a device dynamic management optimization method based on deep learning prediction and the combination of improved dragonfly algorithm and simulated annealing algorithm, and has been applied and verified in the actual production environment of the enterprise.

[0147] In practical applications, enterprises first deploy data collection and preprocessing modules on each key device, and use advanced Internet of Things sensors to collect multi-dimensional data such as temperature, vibration, power consumption, and operation duration of the device in real time. After the collected data is subjected to data cleaning, normalization processing, and feature extraction, it is input into the deep learning prediction module. This module uses a long short-term memory network to train and online predict the historical data of the device, and can accurately output comprehensive warning indicators such as the future state of the device, abnormal probability, and failure risk. After three consecutive weeks of data collection and model tuning, the accuracy rate of the prediction module reached over 92%, significantly improving the perception ability of potential device failure risks.

[0148] With the support of the deep learning prediction module, the system further constructs an optimization problem with the equipment maintenance cost, failure risk, and prediction index deviation as the objectives. By improving the global search ability of the dragonfly algorithm, multiple candidate resource scheduling and maintenance decision-making schemes are generated. The improved dragonfly algorithm adds a dual individual memory behavior and an adaptive dynamic weight adjustment mechanism on the basis of the traditional algorithm, effectively overcoming the problem of being easily trapped in local optimality in the traditional global search process. Subsequently, the system introduces the simulated annealing algorithm to locally optimize the candidate solutions, and adopts an adaptive perturbation and dynamic temperature drop strategy to finely adjust the resource scheduling scheme, so that the optimization result can better fit the comprehensive warning indicators output by the deep learning prediction module, realizing seamless docking between maintenance scheduling and prediction.

[0149] In the actual application process of this enterprise, through the closed-loop dynamic decision-making system of the present invention, the system can capture the changes in the device state in real time and automatically adjust the maintenance scheduling scheme according to the latest prediction results, achieving the purpose of preventive maintenance. After one month of continuous system operation, the enterprise has detailed records and statistics on the operation status of key devices. The data shows that significant changes have occurred in the operation of the devices and maintenance decisions before and after the system is put into use. Previously, due to untimely device status monitoring and unscientific maintenance scheduling, the failure rate of each device per month of the enterprise was about 8%, the cumulative maintenance downtime reached 120 hours, and the maintenance cost was as high as 50,000 yuan. After adopting this system, the failure rate has dropped to about 2%, the maintenance downtime has been reduced to 40 hours, the maintenance cost has been reduced to 30,000 yuan, and the production efficiency of the device has increased by nearly 15%, and the energy consumption has been reduced by about 10%. This significant improvement not only ensures the continuity and safety of production, but also greatly reduces the operating costs of the enterprise.

[0150] Table 1 Comparison data table of maintenance decision optimization effects

[0151]

[0152] The above table details the changes in the key indicators of the equipment before and after the system implementation, visually reflecting the significant effect of the present invention in optimizing maintenance decisions from the data. First of all, the table shows that the equipment failure rate has decreased from 8% before the system implementation to 2% after the implementation, with a reduction rate of 75%. This indicates that through the closed-loop decision-making of introducing deep learning prediction and optimization algorithms, the probability of equipment failure has decreased significantly, significantly improving the reliability and safety of equipment operation. The reduction of the equipment failure rate not only reduces production interruptions caused by equipment failures but also helps enterprises achieve preventive maintenance, thus reducing high later maintenance costs and downtime losses.

[0153] Secondly, in terms of maintenance downtime, the data shows that the average monthly downtime of the equipment before the system implementation was 120 hours, while after the optimization of the present invention, the downtime has been reduced to 40 hours, a decrease of approximately 66.7%. This significant improvement indicates that the system can timely adjust the maintenance scheduling plan when abnormal signs of the equipment appear and take measures in advance through the early warning mechanism, thus reducing the long-term downtime caused by equipment failure maintenance. This is of great significance for continuous production and maintaining production efficiency because short-term downtime can not only reduce production losses but also improve equipment utilization and production capacity stability.

[0154] In addition, the maintenance cost has also been significantly controlled. The average maintenance cost before implementation was approximately 50,000 yuan per month, while after optimization, it has dropped to 30,000 yuan per month, a decrease of approximately 40%. The reduction of the maintenance cost is mainly due to the accurate prediction and dynamic scheduling of the equipment status by the system, making the maintenance work more scientific and reasonable, avoiding unnecessary over-maintenance and emergency repairs, thus reducing cost expenditures. At the same time, the system optimizes resource allocation through the comprehensive analysis of various equipment operation data, enabling the enterprise to obtain higher economic benefits in overall operation.

[0155] Furthermore, the table also reflects the changes in production efficiency and energy consumption. The production efficiency per hour has increased from 1000 pieces to 1150 pieces, an increase of approximately 15%, while the daily energy consumption has decreased from 500 kWh to 450 kWh, a decrease of 10%. These data further prove that after optimizing equipment scheduling and maintenance decisions by the method of the present invention, it can not only improve the operation efficiency and production capacity of the equipment but also show significant advantages in energy conservation and consumption reduction. The increase in production efficiency and the reduction of energy consumption have a positive effect on reducing the production cost of enterprises and enhancing market competitiveness.

[0156] In summary, through the analysis of the above data table, it can be seen that the device dynamic management optimization method of the present invention has achieved remarkable improvement effects in multiple key indicators. The reduction of the failure rate, the decrease of the maintenance downtime, the control of the maintenance cost, the improvement of the production efficiency, and the reduction of the energy consumption all indicate that this system has high practicability and economic benefits in practical applications. These data not only verify the technical advantages of the present invention in optimizing the equipment maintenance scheduling decision, but also provide strong technical support and decision-making basis for enterprises in the process of realizing intelligent manufacturing and industrial automation.

[0157] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent replacements or changes, should be covered within the protection scope of the present invention.

Claims

1. A dynamic device management service system and method for deep learning, characterized in that It includes the following steps: S1. Collect the real-time operation data of the device and preprocess the real-time operation data; S2. Use the preprocessed real-time operation data to construct a long short-term memory network model, conduct training and online real-time prediction of the long short-term memory network model, and output the comprehensive warning index of the device; S3. Construct an optimization problem based on the comprehensive warning index of the device, including setting the current state of the device, the comprehensive warning index, the maintenance cost, and the failure risk as the objective function and constraint conditions; S4. Use the improved dragonfly algorithm to globally search for the optimization problem, initialize multiple candidate solutions, and generate multiple resource scheduling and maintenance candidate solutions through information sharing and position update; S5. Take the generated resource scheduling and maintenance candidate solutions as the initial solution, use the simulated annealing algorithm to locally optimize the initial solution, and conduct detailed adjustment by introducing perturbations, judging the acceptance probability, and setting the temperature drop strategy to determine the optimal resource scheduling and maintenance decision-making solution as the final execution solution of the system; S6. Feed back the real-time prediction result and the optimal resource scheduling and maintenance decision result to the deep learning prediction module to form a closed-loop control.

2. The dynamic device management service method for deep learning according to claim 1, wherein The specific content of S2 includes: S21. Obtain the preprocessed real-time operation data, store the real-time operation data in the data buffer, and construct a data input stream in the time series format; S22. Use the preprocessed real-time operation data to construct a long short-term memory network model. The long short-term memory network model includes an input layer, multiple hidden layers, and an output layer, where the hidden layer adjusts the information flow through a gating mechanism; S23. Configure the model structure and initialize the parameters of the long short-term memory network model, set the number of neurons, the type of activation function, the learning rate, the loss function, and the optimization algorithm, and initialize the weight parameters and bias parameters; S24. Conduct offline training on the long short-term memory network model, conduct multiple rounds of iterative training based on the historical operation data, optimize the parameters of the long short-term memory network model by minimizing the loss function error, and adjust the weight matrix using the batch gradient descent strategy; S25. Deploy the trained long short-term memory network model to the online prediction system, receive the real-time input data stream, process the input data at each time step step by step, calculate the operation state of the device at the current moment in combination with the historical information, and generate short-term and long-term prediction results; S26. Based on the generated short-term and long-term prediction results, output the comprehensive warning index of the device. The comprehensive warning index includes the future state of the device, the abnormal probability, and the failure risk value.

3. The dynamic device management service method for deep learning according to claim 1, characterized in that, The specific content of S3 includes: S31. Obtain the output comprehensive warning index of the device, including the future state of the device, the abnormal probability, and the failure risk value, and combine the current operation state information of the device to form an input data set for optimization modeling; S32. Analyze the operation mode and failure characteristics of the device, and set the optimization problem. The optimization problem includes reducing the probability of device failure and reducing the maintenance cost; S33. Extract the operation constraint conditions of the device, including the upper limit of the operation capacity of the device, the maintenance cycle limit, the resource allocation constraint, and the failure impact range; S34. Establish an optimization objective function, and define the objective function of the equipment management optimization problem according to the current operating status information of the equipment and the set optimization problem. S35. Construct the constraint conditions of the optimization problem, and set the resource scheduling constraints, equipment maintenance constraints, and operation safety requirements in combination with the operation constraints of the equipment. S36. Complete the construction of the optimization problem, and input the objective function and constraint conditions into the optimization solving module.

4. The dynamic device management service method for deep learning according to claim 1, characterized in that, The specific steps of S4 are as follows: S41. Based on the above optimization problem, set the fitness function f(X) to minimize the equipment maintenance cost, equipment failure risk, and the deviation between the decision-making and comprehensive warning indicators: f(X) = αC(X) + βR(X) + γ||I - X||; Where X represents the decision variable, C(X) represents the equipment maintenance cost, R(X) represents the equipment failure risk, I represents the comprehensive warning indicator of the equipment, ||I - X|| is the deviation metric, and α, β, and γ are weight coefficients. S42. Construct an initial population P, set the population size N, and randomly initialize N solutions X i , where each solution X i represents a feasible candidate solution for resource scheduling and maintenance, and simultaneously records the historical optimal solution of each individual and the optimal solution in the most recent m iterations S43. Introduce dual individual memory behaviors and define the individual memory behavior item P i It is defined as: Where λ1 and λ2 are individual memory influence factors. S44. According to the six behavioral rules of the dragonfly algorithm, calculate the update direction of each individual, and define the individual velocity as: Among them, represents the velocity of individual i at the t-th iteration, S i represents the attraction behavior term, A i represents the alignment behavior term, C i represents the aggregation behavior term, E i represents the environmental perception behavior term, M i represents the avoidance behavior term, P i represents the dual individual memory behavior term, w is the inertia weight, represents the velocity of individual i at the (t + 1)-th iteration, S is the adjustment coefficient of the attraction behavior, A is the adjustment coefficient of the alignment behavior, C is the adjustment coefficient of the aggregation behavior, E is the adjustment coefficient of the environmental perception behavior, M is the adjustment coefficient of the avoidance behavior; S45. Update the individual position using the updated speed and the comprehensive warning index I: Where γ is the convergence factor: where γ max and γ min are the initial and minimum convergence factors respectively, δ max is the maximum feedback coefficient, T is the maximum number of iterations, t is the current number of iterations, represents the updated position of the i-th individual in the (t + 1)-th iteration, and represents the position of the i-th individual in the t-th iteration; S46. For all individuals X i Calculate the fitness value f(X i ), and record the current global optimal solution X * ; S47. When the population diversity decreases or the global optimal solution has not been updated continuously for multiple times, adopt a local perturbation mechanism for some individuals, and use a combination of adaptive Gaussian perturbation and differential perturbation for position update: X i new = X i + σ·N(0,1) + η·ρ·(X * - X i ); Among them, σ is the basic perturbation intensity, N(0,1) is the standard normal distribution with a mean of 0 and a variance of 1, η is the differential perturbation factor, ρ is a random number taken from the interval [0,1], X * is the current global optimal solution, X i new is the updated position; S48. When the maximum number of iterations is reached or the convergence error meets a predetermined threshold, output the resource scheduling and maintenance candidate solution X * , otherwise return to S43 to continue the iteration.

5. A dynamic device management service method for deep learning according to claim 1, characterized in that The specific steps of S5 are as follows: S51. Take the obtained resource scheduling and maintenance candidate solution X i as the initial solution; S52. Set the initial temperature of the simulated annealing algorithm S53. For each candidate solution X i Use adaptive perturbation to generate a new candidate solution X' i : X′ i = X i + η·ΔX + κ·(I - X i ) Where η is the basic perturbation factor, ΔX is the basic perturbation amount, κ is the dynamic feedback factor, and I is the comprehensive warning indicator. S54, Calculate the fitness difference Δf between the current candidate solution X i and the newly generated candidate solution X′ i : Δf = f(X′ i ) - f(X i ); Where f(X) is the fitness function described in S41 of claim 4. S55. Determine whether to accept the new candidate solution X' i , first calculate the basic acceptance probability according to the fitness difference Δf, combine the comprehensive early warning index to evaluate the matching degree between the candidate solution and the predicted trend, adaptively adjust the basic acceptance probability, and finally calculate the acceptance probability P a : Ψ = exp(ζ·(‖I - X i ‖ - ‖I - X' i ‖)); Among them, T a is the current temperature, Φ is the dynamic scaling factor, λ is the feedback adjustment coefficient, ζ is the dynamic improvement coefficient, Ψ is the dynamic improvement factor, exp is the exponential function, min is the minimum value operation, T a is the current temperature, generate a random number r, if r ≤ P a then update the candidate solution X i = X′ i ; S56. Adopt an adaptive temperature drop strategy, and dynamically adjust the cooling rate according to the change of the candidate solution fitness and the deviation degree between the candidate solution and the comprehensive warning indicator I. where θ is the temperature drop coefficient, μ is the dynamic adjustment factor, and X * is the current global optimal solution, and ‖I - X * ‖ represents the deviation between the comprehensive warning index and the current global optimal solution. Repeat S53 to S56 until the local optimal condition is met or the predetermined number of iterations is reached, and output the locally optimized candidate solution as the final resource scheduling and maintenance decision plan.

6. A dynamic device management service system for deep learning, the method for a dynamic device management service for deep learning according to any one of claims 1 to 5, characterized in that, It includes the following modules: Data acquisition and preprocessing module, which is used to collect equipment data in real time, clean, normalize, and extract features from the equipment data. Deep learning prediction module, which is used to train and online predict the preprocessed equipment data using the long short-term memory network, and output the comprehensive warning indicator. Global search module, which is used to generate candidate resource scheduling and maintenance decision-making schemes according to the comprehensive warning indicator and the optimization problem construction. Local optimization module, which is used to perform perturbation, acceptance probability judgment, and adaptive temperature drop on the candidate resource scheduling and maintenance decision-making schemes. Closed-loop feedback and decision adjustment module, which is used to real-time feedback the deep learning prediction results and optimization process information, and dynamically adjust the resource scheduling and maintenance decisions. User interaction and monitoring module, which is used to display the equipment status, prediction results, and decision-making schemes, and support manual intervention and system monitoring.

Images