An operation and maintenance work order management method and system based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based operation and maintenance work order management method and system, relating to the fields of artificial intelligence and operation and maintenance management technology. The method includes constructing a CNN model, optimizing the policy network and value network using the PPO algorithm, optimizing model parameters through multi-task learning and gradient updating, outputting device failure probabilities and recommended actions, generating JSON work orders, and using a quantum decision tree based on the classical GBDT framework for split gain calculation. The quantum Monte Carlo method is used to optimize the loss function, iteratively construct a quantum decision tree, and output candidate scheduling solutions. The method uses a CNN model combined with the PPO algorithm to optimize work order generation, enhancing the accuracy and timeliness of fault prediction and action recommendations. It also introduces quantum machine learning into work order scheduling modeling, and improves the multi-objective adaptability of scheduling through the quantum decision tree and quantum Monte Carlo optimization mechanism.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and operation and maintenance management technology, and in particular to an operation and maintenance work order management method and system based on artificial intelligence. Background Art

[0002] With the rapid advancement of artificial intelligence, big data, and the Internet of Things (IoT) technologies, equipment operation and maintenance management has become increasingly complex, especially in some high-risk industries. The operating status of valves directly affects the safety and stability of the entire production system. Timely maintenance and fault prediction of valves are particularly important. Traditional equipment operation and maintenance management usually relies on manual inspections and manual recording of work orders, which is inefficient and prone to omissions or errors. Equipment operation and maintenance is gradually moving towards intelligence and automation. Through sensor monitoring, data collection and analysis, combined with machine learning and deep learning technologies, the system can monitor equipment status in real time, automatically generate fault work orders, and make predictions.

[0003] However, there are still some problems with the existing operation and maintenance work order management. The existing model cannot fully extract the deep correlation features between valve images, voice, text and time series data, resulting in limited fault prediction accuracy. It fails to combine quantum computing for optimization and fails to effectively utilize the advantages of parallel search of quantum bits and quantum state probability distribution, which limits the further improvement of scheduling accuracy and algorithm convergence speed. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an operation and maintenance work order management method and system based on artificial intelligence to solve the problem that the existing model cannot fully extract the deep correlation features between valve images, voice, text and time series data, resulting in limited fault prediction accuracy, failure to combine quantum computing for optimization, and failure to effectively utilize the advantages of parallel search of quantum bits and quantum state probability distribution, which limits the further improvement of scheduling accuracy and algorithm convergence speed.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides an operation and maintenance work order management method based on artificial intelligence, which includes collecting multimodal data, generating multimodal work order data packets, constructing a CNN model, using the PPO algorithm to optimize the policy network and the value network, optimizing the model parameters through multi-task learning and gradient updating, outputting the device failure probability and recommended actions, generating a JSON work order, using the IBM Qiskit framework to convert the quantum state, based on the classical GBDT framework, using the quantum decision tree to perform split gain calculation, and optimizing the loss function through the quantum Monte Carlo method, iteratively constructing the quantum decision tree, outputting candidate scheduling plans, performing a small perturbation on the current optimization objective function value through the simulated annealing algorithm, outputting the optimal scheduling plan, generating task instructions through the IMS, executing them, and storing, collecting and analyzing the generated multimodal data.

[0008] As a preferred solution of the artificial intelligence-based operation and maintenance work order management method of the present invention, wherein: the multimodal data is collected, a multimodal work order data packet is generated, a CNN model is constructed, the policy network and the value network are optimized using the PPO algorithm, model parameters are optimized through multi-task learning and gradient updating, the equipment failure probability and recommended action are output, and a JSON work order is generated, including:

[0009] The multimodal data includes valve pressure value, valve image and voice data;

[0010] Build a CNN model, collect historical multimodal data, use EfficientNet-B0 to extract valve image features, and generate valve image feature vectors through global average pooling;

[0011] The valve pressure value features are extracted through time series analysis technology to generate valve pressure value feature vectors. DistilBERT is used to encode log text to generate text feature vectors.

[0012] The valve image feature vector, text feature vector, and valve pressure value feature vector are fused through the attention mechanism to generate a multimodal feature vector. The vector is passed through a fully connected layer to output the number of fault categories, and through a softmax layer to output the probability of the fault category.

[0013] Define the state as the concatenation of multimodal feature vectors and failure probabilities, and define the discrete action space as no action, repair, and replacement;

[0014] Use a feedforward neural network to define the policy network, input the state, pass through two fully connected layers, output the action probability distribution of the current state, and set the action selected according to the current state as ;

[0015] Use MLP to define the value network, input the multimodal feature vector, pass it through two fully connected layers, and output the value estimate of the current state;

[0016] Set the reward function based on the value network, calculate the reward value, optimize the policy network and value network through the PPO algorithm, calculate the good or bad value of the fault category to the current state and the policy probability ratio between the current time t and time t-1;

[0017] Calculate the policy gradient value;

[0018] Set each fault category as a task and use the Adam optimizer to update the parameters of the policy network and value network through policy gradient values and value estimation. Update the parameters of the CNN model by calculating the gradient value of the loss function. Use the rule of thumb to set the maximum number of iterations and stop iteration when the maximum number of iterations is reached.

[0019] Multimodal data is input into the optimized CNN model, the device failure probability is output through Softmax, and the recommended action is output through the policy network. The output results and priority scores of the CNN model are packaged into JSON messages, combined with the multimodal work order data packets and integrated into JSON format task units through Schema standardization.

[0020] As a preferred solution of the artificial intelligence-based operation and maintenance work order management method of the present invention, the method uses the IBM Qiskit framework to convert quantum states, uses a quantum decision tree to calculate split gains based on the classical GBDT framework, optimizes the loss function through the quantum Monte Carlo method, iteratively constructs a quantum decision tree, and outputs candidate scheduling solutions, including:

[0021] Query current technician data through the API interface, including technician skill level and technician availability;

[0022] The task unit and technician data are converted into a unified time format using timestamp standardization, time synchronization is performed using a temporal convolutional network, and time synchronization feature vectors are generated using a cross-joining method.

[0023] The IBM Qiskit framework is used to convert the time matching features of each pair of technicians and task units in the time synchronization feature vector into quantum states through amplitude encoding.

[0024] Assign a quantum bit to each technician-task unit pair to generate a quantum state dataset;

[0025] Based on the classic GBDT gradient boosting framework, the first quantum decision tree is initialized, the average matching degree of all samples is used as the initial prediction value, and the initial residual is calculated;

[0026] Apply parameterized quantum gates to the quantum state dataset, calculate the splitting gain for each technician-task unit pair, extract the splitting gain using quantum measurements, and set the number of measurements according to the Qiskit defaults.

[0027] Select the technician-task unit pair with the largest splitting gain as the splitting point, and output the splitting point and the quantum state data set corresponding to the splitting point;

[0028] Define the mean square error loss function, use the quantum Monte Carlo method to encode the initial residual into a residual quantum state, simulate random walks through quantum circuits, adjust the expected value of the mean square error loss function, and use the COBYLA optimizer to optimize the quantum circuit parameters. Stop the optimization when the expected value reaches the minimum.

[0029] Based on the gradient improvement idea of classic GBDT, multiple quantum decision trees are iteratively constructed. The predicted value of the tree and the residual input Tree, weighted accumulation of the outputs of all trees, output of the final prediction value, and convert the final prediction value into matching probability;

[0030] Sort the matching probabilities in descending order, select the top E pairs, and generate chromosomes through encoding. Each chromosome is a complete scheduling solution.

[0031] Based on empirical analysis, the weights of response time score, resource utilization score, and GBDT prediction score are set, and a multi-objective fitness function is defined.

[0032] Calculate the fitness value of each chromosome, screen chromosomes with high probability values through roulette, generate the next generation of chromosomes through crossover mutation, repeat the iterative operation and stop when the fitness value converges, output the chromosome of the last iteration, and format it as a list of candidate scheduling solutions.

[0033] As a preferred solution of the artificial intelligence-based operation and maintenance work order management method of the present invention, the collected multimodal data is first preprocessed, and the generation of the multimodal work order data packet includes:

[0034] Build a digital twin model, input preprocessed valve pressure values, calculate the current stress value, and predict the stress value at the future time t+1 using a linear regression formula. Use statistical methods to set a safety threshold, filter out predicted stress values at the future time t+1 that are less than the safety threshold, and mark them as abnormal.

[0035] The valve image, text, and current stress value are processed through the Transformer encoder and weighted fusion calculation to generate a comprehensive feature vector;

[0036] Initialize the Q table, set the action to generate a work order or not generate a work order, and calculate the reward value of the current state;

[0037] The Q value is updated according to the Q-learning algorithm, and the Q value of the generated work order is greater than the Q value of the non-generated work order. The comprehensive feature vector, location, timestamp and predicted status are extracted to form the work order data packet, which is serialized in JSON format and output as a multimodal work order data packet.

[0038] As a preferred solution of the artificial intelligence-based operation and maintenance work order management method of the present invention, wherein: the simulated annealing algorithm is used to slightly perturb the current optimization objective function value and output the optimal scheduling solution, including:

[0039] Define the optimization objective function, set the initial temperature based on the analysis of the problem characteristics, and simulate the beginning stage of the annealing process;

[0040] Calculate the optimization objective function value of the current solution as the starting point of the simulated annealing process, make a small perturbation to the current solution, generate a new solution, and recalculate the new objective function value;

[0041] Determine whether to accept the new solution based on the difference in the optimization objective function, and update the accepted new solution as the current solution;

[0042] After each iteration, the initial temperature decays according to the cooling strategy, and the iteration continues. When the temperature drops to the minimum value, the iteration stops, the current solution is output, and the optimal scheduling plan is generated.

[0043] As a preferred solution of the artificial intelligence-based operation and maintenance work order management method of the present invention, the generating of task instructions through IMS and executing the same include:

[0044] The optimal scheduling plan generates task instructions through the IMS system and pushes them to the technician terminal, who then executes the task instructions.

[0045] As a preferred solution of the artificial intelligence-based operation and maintenance work order management method of the present invention, the multimodal data generated by storage, collection and analysis includes:

[0046] It refers to storing the collected multimodal data and the optimal scheduling plan generated by analysis in a database. The database marks the stored data with a timestamp and uploads the stored data to the cloud for backup.

[0047] In a second aspect, the present invention provides an operation and maintenance work order management system based on artificial intelligence, comprising:

[0048] The collection and decision-making module is used to collect multimodal data, generate multimodal work order data packets, build a CNN model, use the PPO algorithm to optimize the policy network and value network, optimize model parameters through multi-task learning and gradient updating, output device failure probability and recommended actions, and generate JSON work orders;

[0049] The scheduling optimization module uses the IBM Qiskit framework to convert quantum states. Based on the classical GBDT framework, it uses quantum decision trees to calculate split gains, optimizes the loss function through the quantum Monte Carlo method, iteratively constructs quantum decision trees, and outputs candidate scheduling solutions.

[0050] The global adjustment module is used to perform small perturbations on the current optimization objective function value through the simulated annealing algorithm and output the optimal scheduling solution;

[0051] The execution and storage module is used to generate task instructions through IMS, execute them, and store the multimodal data generated by collection and analysis.

[0052] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the artificial intelligence-based operation and maintenance work order management method as described in the first aspect of the present invention.

[0053] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the artificial intelligence-based operation and maintenance work order management method as described in the first aspect of the present invention.

[0054] The beneficial effects of the present invention are as follows: the present invention collects multimodal data, generates multimodal work order data packets, constructs a CNN model, uses the PPO algorithm to optimize the policy network and value network, optimizes model parameters through multi-task learning and gradient updating, outputs device failure probability and recommended actions, generates JSON work orders, uses the IBM Qiskit framework to convert quantum states, uses a quantum decision tree based on the classical GBDT framework to calculate splitting gains, optimizes the loss function through the quantum Monte Carlo method, iteratively constructs a quantum decision tree, and outputs candidate scheduling solutions; the accuracy and timeliness of fault prediction and action recommendation are enhanced, and the multi-objective adaptability of scheduling is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0056] Figure 1 This is a flowchart of the operation and maintenance work order management method based on artificial intelligence in Example 1.

[0057] Figure 2This is a schematic diagram of the artificial intelligence-based operation and maintenance work order management system in Example 1.

[0058] Figure 3 This is a flowchart for generating a multimodal feature vector in Example 1.

[0059] Figure 4 This is a flowchart for outputting the optimal scheduling solution in Example 1. DETAILED DESCRIPTION

[0060] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0061] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0062] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0063] Example 1, with reference to Figures 1 to 4 , which is the first embodiment of the present invention, provides an operation and maintenance work order management method based on artificial intelligence, comprising the following steps:

[0064] S1. Collect multimodal data, generate multimodal work order data packets, build a CNN model, use the PPO algorithm to optimize the policy network and value network, optimize model parameters through multi-task learning and gradient update, output device failure probability and recommended actions, and generate JSON work orders;

[0065] Specifically, collect multimodal data and generate multimodal work order data packets, including:

[0066] The smart sensor includes a pressure sensor, an infrared camera, a voice recognition module and a microphone;

[0067] The multimodal data includes valve pressure value, valve image and voice data;

[0068] The preprocessing includes converting the voice data into text using a voice recognition module, normalizing the valve pressure value using linear normalization, and gray-scaling and scaling the valve image using a weighted average method and linear double interpolation respectively;

[0069] Build a digital twin model, including input layer, model layer, simulation layer, analysis layer and output layer;

[0070] Collect historical valve pressure values, perform normalization, and train the digital twin model;

[0071] Input the pre-processed valve pressure value, calculate the current stress value, and predict the stress value at the future time t+1 through the linear regression formula. The output formula is:

[0072] ,

[0073] ,

[0074] in, Current time The stress value, is the normalized valve pressure value, is the valve area (provided by the manufacturer), For time The predicted value of ΔW is the stress change rate in the historical time t (obtained based on the intelligent sensor collecting historical valve stress values and calculating the average change rate);

[0075] A safety threshold is set by statistical methods, and stress values predicted for the future time t+1 that are less than the safety threshold are screened and marked as abnormal.

[0076] Output digital twin status data, including current stress value, predicted stress value and abnormality mark;

[0077] The valve image, text, and current stress value are processed through the Transformer encoder and weighted fusion calculation to generate a comprehensive feature vector;

[0078] Initialize the Q table, set the action to generate a work order or not generate a work order, and calculate the reward value of the current state. The formula is:

[0079] J ,

[0080] Among them, J is the reward value of the current state, is the comprehensive feature vector The strength value of (calculated by L2 norm), To predict the stress severity value (obtained by comparing the current stress value with the maximum stress value), and The comprehensive feature vectors are The weight of the strength value and the weight of the predicted stress severity value (both weights are obtained through the entropy weight method);

[0081] The Q value is updated according to the Q-learning algorithm, and the Q value of the generated work order is greater than the Q value of the non-generated work order. The comprehensive feature vector, location, timestamp and predicted status are extracted to form the work order data packet, which is serialized in JSON format and output as a multimodal work order data packet.

[0082] By collecting multimodal data, the comprehensiveness of work order analysis is improved, avoiding the limitations of a single data source. The digital twin model is used to calculate current stress values and predict future stress values, avoiding losses caused by excessive maintenance or sudden failures. By generating a comprehensive feature vector, a better status representation is provided for subsequent use.

[0083] Furthermore, we built a CNN model, used the PPO algorithm to optimize the policy network and value network, optimized the model parameters through multi-task learning and gradient update, output the device failure probability and recommended actions, and generated a JSON work order, including:

[0084] Build a CNN model, including input layer, convolution layer, fully connected layer and output layer;

[0085] Collect historical multimodal data, input it into the CNN model, extract valve image features using EfficientNet-B0, and perform global average pooling on the extracted valve image features to generate a valve image feature vector;

[0086] Extract the characteristics of valve pressure value through time series analysis technology and generate valve pressure value characteristic vector;

[0087] Use DistilBERT to encode log text and generate text feature vectors;

[0088] The valve image feature vector, text feature vector, and valve pressure value feature vector are fused through the attention mechanism to generate a multimodal feature vector. The vector is passed through a fully connected layer to output the number of fault categories, and through a softmax layer to output the probability of the fault category.

[0089] Define the state as the concatenation of multimodal feature vectors and failure probabilities, and define the discrete action space as no action, repair, and replacement;

[0090] Use a feedforward neural network to define the policy network, input the state, pass through two fully connected layers, output the action probability distribution of the current state, and set the action selected according to the current state as ;

[0091] Use MLP to define the value network, input the multimodal feature vector, pass it through two fully connected layers, and output the value estimate of the current state;

[0092] Set the reward function based on the value network and calculate the reward value. The formula is:

[0093] ,

[0094] in, Current time The reward value, is the discount factor (historical training verification), Current time Status The estimated value of is the immediate reward value (set according to the discrete action space);

[0095] Generate a priority score by normalizing the reward value;

[0096] The PPO algorithm is used to optimize the strategy network and value network, and calculate the good or bad value of the fault category to the current state. The formula is:

[0097] ,

[0098] in, Current time The good or bad value of

[0099] Calculate the current time t and time -1 strategy probability ratio , the formula is:

[0100] ,

[0101] in, Current time The probability distribution of the strategy, For time -1 strategy probability distribution;

[0102] Calculate the policy gradient value, the formula is:

[0103] ,

[0104] in, is the policy gradient value, is the current time t and time -1 strategy probability ratio, is a constant;

[0105] Set each fault category as a task and use the Adam optimizer to update the parameters of the policy network and value network through policy gradient values and value estimation. Update the parameters of the CNN model by calculating the gradient value of the loss function. Use the rule of thumb to set the maximum number of iterations and stop iteration when the maximum number of iterations is reached.

[0106] Input multimodal data into the optimized CNN model, output the device failure probability through Softmax, and output the recommended action through the policy network;

[0107] The output results and priority scores of the CNN model are packaged into JSON messages, combined with multimodal work order data packets and integrated into JSON format task units through Schema standardization.

[0108] By building a CNN model, the accuracy of fault detection is improved. Feature extraction based on EfficientNet-B0, time series, and DistilBERT encoded log text enhances model robustness, improves the comprehensiveness of fault diagnosis, and improves classification accuracy. Multimodal feature fusion based on the attention mechanism improves the model's sensitivity to key fault signals. The reinforcement learning architecture of the policy network and value network reduces manual intervention and avoids short-sighted decision-making. The PPO algorithm is optimized to improve training stability, realize intelligent sorting of work orders, and optimize resource allocation.

[0109] S2. Use the IBM Qiskit framework to convert quantum states. Based on the classic GBDT framework, use a quantum decision tree to calculate split gains, optimize the loss function using the quantum Monte Carlo method, iteratively construct a quantum decision tree, and output candidate scheduling solutions, including:

[0110] Query current technician data through the API interface, including technician skill level and technician availability;

[0111] The task unit and technician data are converted into a unified time format using timestamp standardization, time synchronization is performed using a temporal convolutional network, and time synchronization feature vectors are generated using a cross-joining method.

[0112] The IBM Qiskit framework is used to convert the time matching features of each pair of technicians and task units in the time synchronization feature vector into quantum states through amplitude encoding.

[0113] Each technician-task unit pair is assigned a qubit, generating a quantum state dataset representing a superposition of so-called features;

[0114] Based on the gradient boosting framework of the classic GBDT, the first quantum decision tree is initialized, the average matching degree of all samples is used as the initial prediction value, and the initial residual is calculated. The formula is:

[0115] ,

[0116] ,

[0117] in, is the initial prediction value, For the The true matching degree of samples (obtained from historical experimental data), is the total number of samples, the samples are historical technician-task unit pairs, For the The initial residuals of samples;

[0118] Apply parameterized quantum gates to the quantum state dataset and calculate the splitting gain for each technician-task unit pair using the formula:

[0119] ,

[0120] in, For the The split gain of the feature, is the current predicted value;

[0121] Use quantum measurement to extract the splitting gain, setting the number of measurements according to the Qiskit default value;

[0122] Select the technician-task unit pair with the largest splitting gain as the splitting point, and output the splitting point and the quantum state data set corresponding to the splitting point;

[0123] Define the mean square error loss function and use the quantum Monte Carlo method to encode the initial residual into the residual quantum state. The formula is:

[0124] ,

[0125] ,

[0126] Among them, L is the expected value of the mean square error loss function, The encoded residual quantum state;

[0127] Through quantum circuit simulation of random walk, the expected value of the mean square error loss function is adjusted, and the quantum circuit parameters are optimized using the COBYLA optimizer. The optimization is stopped when the expected value reaches the minimum.

[0128] Based on the gradient boosting idea of classic GBDT, multiple quantum decision trees are iteratively constructed. The prediction values and residuals of g-1 trees are input into the g-th tree. The outputs of all trees are weighted and accumulated to output the final prediction value. The formula is:

[0129] ,

[0130] in, is the learning rate (set according to cross-validation), is the sum of the predicted values of g-1 trees, is the fitted value of the residual of the g-th tree;

[0131] The final prediction value is converted into a matching probability using the following formula:

[0132] ,

[0133] in, For technicians and task units The matching probability, For technicians and task units The final predicted value of

[0134] Sort the matching probabilities in descending order, select the top E pairs (set according to specific needs), and generate chromosomes through encoding. Each chromosome is a complete scheduling plan;

[0135] The weights of response time score, resource utilization score and GBDT prediction score are set based on empirical analysis, respectively. 、 as well as , define the multi-objective fitness function, the formula is:

[0136] ,

[0137] ,

[0138] ,

[0139] ,

[0140] in, For the The multi-objective fitness function value of chromosomes, is the number of task units, is the actual response time (preset value), is the estimated start time (based on technician availability), is the duration of the task unit, is the matching degree between the task unit and the technician, is the matching probability predicted by GBDT, is the response time score, is the resource utilization score, For GBDT prediction score, For the chromosome;

[0141] Calculate the fitness value of each chromosome using the formula:

[0142] ,

[0143] in, is the sum of the multi-objective fitness function values of all chromosomes, For the The fitness value of a chromosome, that is, the probability of being selected;

[0144] The chromosomes with high probability values are selected through roulette, and the next generation of chromosomes are generated through crossover mutation. The iterative operation is repeated and stops when the fitness value converges (if there is no new highest record of the best fitness value after consecutive iterations, the algorithm has converged). The chromosomes of the last iteration are output and formatted as a list of candidate scheduling solutions.

[0145] By using the IBM Qiskit framework for quantum state conversion, quantum circuit simulation random walk to adjust the expected value of the loss function, COBYLA to optimize the quantum circuit parameters, efficiently exploring the optimization space, reducing computing time, avoiding the local optimality problem of gradient descent optimizers, improving the global optimality of the scheduling scheme, reducing the number of iterations required for classical optimization, and reducing computing costs, a quantum decision tree is constructed based on the GBDT framework for multi-objective scheduling optimization, efficiently processing high-dimensional features, improving the accuracy of matching predictions, and improving the adaptability of scheduling schemes. Chromosomes are generated by encoding, fitness values are calculated, and candidate scheduling schemes are screened for crossover and mutation output. This avoids scheduling imbalances caused by a single optimization objective, improves the robustness of the scheduling scheme, reduces manual intervention, and improves operation and maintenance efficiency.

[0146] S3. Use the simulated annealing algorithm to slightly perturb the current optimization objective function value and output the optimal scheduling solution, including:

[0147] The CRITIC method is used to calculate the weights of the expected repair time, resource utilization, and quality stability, and the optimization objective function is defined as follows:

[0148] ,

[0149] Among them, Z is the optimization objective function, as well as is the weight of expected repair time, resource utilization and quality stability, is the estimated repair time (obtained from historical repair data), is the resource utilization rate (the ratio of the number of currently used technicians to the total number of technicians), is quality stability (obtained from historical repair data);

[0150] The initial temperature is set based on the analysis of the problem characteristics to simulate the beginning of the annealing process;

[0151] Calculate the optimization objective function value of the current solution as the starting point of the simulated annealing process;

[0152] Make a small perturbation to the current solution, generate a new solution, and recalculate the new objective function value;

[0153] The difference between the optimization objective function determines whether to accept the new solution. If the difference is less than 0, the new solution is accepted. If the difference is greater than 0, the acceptance probability is calculated. The formula is:

[0154] ,

[0155] in, For an acceptable probability, To optimize the difference of the objective function, is the initial temperature;

[0156] Generate a uniform random number. If the acceptance probability is less than the uniform random number, accept the new solution; otherwise, keep the current solution.

[0157] Update the accepted new solution to the current solution;

[0158] After each iteration, the initial temperature decays according to the cooling strategy, and the formula is:

[0159] ,

[0160] in, is the temperature after decay, is the initial temperature, is the temperature reduction coefficient;

[0161] Based on the current solution and temperature decay, the iteration continues. When the temperature drops to the minimum value, the iteration stops, the current solution is output, and the optimal scheduling plan is generated.

[0162] The high-temperature exploration and low-temperature convergence characteristics of the quasi-annealing algorithm are used to avoid the local optimality problem of traditional optimization methods. Small perturbations and exponential cooling strategies reduce computational overhead while ensuring the superiority of the solution. The weight setting and probability acceptance mechanism based on historical data improve the practicality and robustness of the solution.

[0163] S4, generating task instructions through IMS, executing them, storing, collecting and analyzing the generated multimodal data;

[0164] Specifically, task instructions are generated and executed through IMS, including:

[0165] The optimal scheduling plan generates task instructions through the IMS system and pushes them to the technician terminal, who then executes the task instructions.

[0166] Task instructions are generated through the IMS system and pushed to the technician's terminal, reducing the delay of traditional manual dispatching, improving execution accuracy, and enhancing the transparency of operation and maintenance management.

[0167] Furthermore, the multimodal data generated by storage, collection and analysis includes:

[0168] The collected multimodal data and the optimal scheduling plan generated by the analysis are stored in the database, the database marks the stored data with a timestamp, and the stored data is uploaded to the cloud for backup.

[0169] By uploading scheduling plans and execution status to the central database in real time, data consistency and traceability are improved, and decision-making deviations caused by data dispersion are reduced.

[0170] This embodiment also provides an artificial intelligence-based operation and maintenance work order management system, including:

[0171] The collection and decision-making module is used to collect multimodal data, generate multimodal work order data packets, build a CNN model, use the PPO algorithm to optimize the policy network and value network, optimize model parameters through multi-task learning and gradient updating, output device failure probability and recommended actions, and generate JSON work orders;

[0172] The scheduling optimization module uses the IBM Qiskit framework to convert quantum states. Based on the classical GBDT framework, it uses quantum decision trees to calculate split gains, optimizes the loss function through the quantum Monte Carlo method, iteratively constructs quantum decision trees, and outputs candidate scheduling solutions.

[0173] The global adjustment module is used to perform small perturbations on the current optimization objective function value through the simulated annealing algorithm and output the optimal scheduling solution;

[0174] The execution and storage module is used to generate task instructions through IMS, execute them, and store the multimodal data generated by collection and analysis.

[0175] This embodiment also provides a computer device, which is suitable for the case of an operation and maintenance work order management method based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the operation and maintenance work order management method based on artificial intelligence proposed in the above embodiment.

[0176] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0177] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the artificial intelligence-based operation and maintenance work order management method proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, disk or optical disk.

[0178] In summary, the present invention collects multimodal data, generates multimodal work order data packets, constructs a CNN model, uses the PPO algorithm to optimize the policy network and value network, optimizes model parameters through multi-task learning and gradient updating, outputs device failure probability and recommended actions, generates JSON work orders, uses the IBM Qiskit framework for quantum state conversion, uses a quantum decision tree based on the classical GBDT framework for split gain calculation, optimizes the loss function through the quantum Monte Carlo method, iteratively constructs a quantum decision tree, and outputs candidate scheduling solutions; thus, the accuracy and timeliness of fault prediction and action recommendation are enhanced, and the multi-objective adaptability of scheduling is improved.

[0179] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An artificial intelligence-based operation and maintenance work order management method, characterized by: include, Collect multimodal data, generate multimodal work order data packets, build a CNN model, use the PPO algorithm to optimize the policy network and value network, optimize model parameters through multi-task learning and gradient update, output device failure probability and recommended actions, and generate JSON work orders; Use the IBM Qiskit framework to convert quantum states. Based on the classic GBDT framework, use quantum decision trees to calculate split gains, optimize the loss function through the quantum Monte Carlo method, iteratively construct quantum decision trees, and output candidate scheduling solutions. Use the simulated annealing algorithm to slightly perturb the current optimization objective function value and output the optimal scheduling solution; Generate task instructions through IMS, execute them, store, collect and analyze the generated multimodal data; The multimodal data includes valve pressure value, valve image and voice data; Build a CNN model, collect historical multimodal data, use EfficientNet-B0 to extract valve image features, and generate valve image feature vectors through global average pooling; The valve pressure value features are extracted through time series analysis technology to generate valve pressure value feature vectors. DistilBERT is used to encode log text to generate text feature vectors. The valve image feature vector, text feature vector, and valve pressure value feature vector are fused through the attention mechanism to generate a multimodal feature vector; Set the reward function based on the value network, calculate the reward value, optimize the policy network and value network through the PPO algorithm, calculate the good or bad value of the fault category to the current state and the policy probability ratio between the current time t and time t-1; Apply parameterized quantum gates to the quantum state dataset, calculate the splitting gain for each technician-task unit pair, and extract the splitting gain using quantum measurements; The quantum Monte Carlo method is used to encode the initial residual into a residual quantum state, and the random walk is simulated through the quantum circuit to adjust the expected value of the mean square error loss function.

2. The artificial intelligence-based operation and maintenance work order management method according to claim 1, characterized in that: The process of collecting multimodal data, generating a multimodal work order data package, building a CNN model, optimizing the policy network and value network using the PPO algorithm, outputting device failure probability and recommended actions, and generating a JSON work order includes: Through the fully connected layer, the number of fault categories is output, and through the Softmax layer, the probability of the fault category is output; Define the state as the concatenation of multimodal feature vectors and failure probabilities, and define the discrete action space as no action, repair, and replacement; Use a feedforward neural network to define the policy network, input the state, pass through two fully connected layers, output the action probability distribution of the current state, and set the action selected according to the current state as ; Use MLP to define the value network, input the multimodal feature vector, pass it through two fully connected layers, and output the value estimate of the current state; Calculate the policy gradient value; Set each fault category as a task and use the Adam optimizer to update the parameters of the policy network and value network through policy gradient values and value estimation. Update the parameters of the CNN model by calculating the gradient value of the loss function. Use the rule of thumb to set the maximum number of iterations and stop iteration when the maximum number of iterations is reached. Multimodal data is input into the optimized CNN model, the device failure probability is output through Softmax, and the recommended action is output through the policy network. The output results and priority scores of the CNN model are packaged into JSON messages, combined with the multimodal work order data packets and integrated into JSON format task units through Schema standardization.

3. The artificial intelligence-based operation and maintenance work order management method according to claim 2, characterized in that: The method uses the IBM Qiskit framework to convert quantum states, uses a quantum decision tree to calculate split gains based on the classical GBDT framework, optimizes the loss function through the quantum Monte Carlo method, iteratively constructs a quantum decision tree, and outputs candidate scheduling solutions, including: Query current technician data through the API interface, including technician skill level and technician availability; The task unit and technician data are converted into a unified time format using timestamp standardization, time synchronization is performed using a temporal convolutional network, and time synchronization feature vectors are generated using a cross-joining method. The IBM Qiskit framework is used to convert the time matching features of each pair of technicians and task units in the time synchronization feature vector into quantum states through amplitude encoding. Assign a quantum bit to each technician-task unit pair to generate a quantum state dataset; Based on the classic GBDT gradient boosting framework, the first quantum decision tree is initialized, the average matching degree of all samples is used as the initial prediction value, and the initial residual is calculated; Set the number of measurements according to Qiskit default values; Select the technician-task unit pair with the largest splitting gain as the splitting point, and output the splitting point and the quantum state data set corresponding to the splitting point; Define the mean square error loss function and use the COBYLA optimizer to optimize the quantum circuit parameters. Stop the optimization when the expected value reaches the minimum. Based on the gradient improvement idea of classic GBDT, multiple quantum decision trees are iteratively constructed, and the prediction values and residuals of g-1 trees are input into the first Tree, weighted accumulation of the outputs of all trees, output of the final prediction value, and convert the final prediction value into matching probability; Sort the matching probabilities in descending order, select the top E pairs, and generate chromosomes through encoding. Each chromosome is a complete scheduling solution. Based on empirical analysis, the weights of response time score, resource utilization score, and GBDT prediction score are set, and a multi-objective fitness function is defined. Calculate the fitness value of each chromosome, screen chromosomes with high probability values through roulette, generate the next generation of chromosomes through crossover mutation, repeat the iterative operation and stop when the fitness value converges, output the chromosome of the last iteration, and format it as a list of candidate scheduling solutions.

4. The artificial intelligence-based operation and maintenance work order management method according to claim 1, characterized in that: The collected multimodal data is preprocessed to generate a multimodal work order data packet, including: Build a digital twin model, input preprocessed valve pressure values, calculate the current stress value, and predict the stress value at the future time t+1 using a linear regression formula. Use statistical methods to set a safety threshold, filter out predicted stress values at the future time t+1 that are less than the safety threshold, and mark them as abnormal. The valve image, text, and current stress value are processed through the Transformer encoder and weighted fusion calculation to generate a comprehensive feature vector; Initialize the Q table, set the action to generate a work order or not generate a work order, and calculate the reward value of the current state; The Q value is updated according to the Q-learning algorithm, and the Q value of the generated work order is greater than the Q value of the non-generated work order. The comprehensive feature vector, location, timestamp and predicted status are extracted to form the work order data packet, which is serialized in JSON format and output as a multimodal work order data packet.

5. The artificial intelligence-based operation and maintenance work order management method according to claim 3, characterized in that: The simulated annealing algorithm is used to slightly perturb the current optimization objective function value and output the optimal scheduling solution, including: Define the optimization objective function, set the initial temperature based on the analysis of the problem characteristics, and simulate the beginning stage of the annealing process; Calculate the optimization objective function value of the current solution as the starting point of the simulated annealing process, make a small perturbation to the current solution, generate a new solution, and recalculate the new objective function value; Determine whether to accept the new solution based on the difference in the optimization objective function, and update the accepted new solution as the current solution; After each iteration, the initial temperature decays according to the cooling strategy, and the iteration continues. When the temperature drops to the minimum value, the iteration stops, the current solution is output, and the optimal scheduling plan is generated.

6. The artificial intelligence-based operation and maintenance work order management method according to claim 4, characterized in that: The generating of task instructions through the IMS and executing the same include: The optimal scheduling plan generates task instructions through the IMS system and pushes them to the technician terminal, who then executes the task instructions.

7. The artificial intelligence-based operation and maintenance work order management method according to claim 4, characterized in that: The multimodal data generated by the storage, collection and analysis includes: It refers to storing the collected multimodal data and the optimal scheduling plan generated by analysis in a database. The database marks the stored data with a timestamp and uploads the stored data to the cloud for backup.

8. An artificial intelligence-based operation and maintenance work order management system, based on the artificial intelligence-based operation and maintenance work order management method according to any one of claims 1 to 7, characterized in that: include, The collection and decision-making module is used to collect multimodal data, generate multimodal work order data packets, build a CNN model, use the PPO algorithm to optimize the policy network and value network, optimize model parameters through multi-task learning and gradient updating, output device failure probability and recommended actions, and generate JSON work orders; The scheduling optimization module uses the IBM Qiskit framework to convert quantum states. Based on the classical GBDT framework, it uses quantum decision trees to calculate split gains, optimizes the loss function through the quantum Monte Carlo method, iteratively constructs quantum decision trees, and outputs candidate scheduling solutions. The global adjustment module is used to perform small perturbations on the current optimization objective function value through the simulated annealing algorithm and output the optimal scheduling solution; The execution and storage module is used to generate task instructions through IMS, execute them, and store the multimodal data generated by collection and analysis.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the artificial intelligence-based operation and maintenance work order management method described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the artificial intelligence-based operation and maintenance work order management method according to any one of claims 1 to 7 are implemented.

Images