Smart park equipment health state intelligent evaluation method based on multiple modes, electronic equipment and medium

Abstract

The invention discloses a multi-mode-based smart park equipment health state intelligent assessment method, electronic equipment and a medium. The method comprises the following steps: acquiring multi-source data from an equipment real-time monitoring platform; preprocessing the multi-source data; the preprocessed data are used, a multi-modal prediction model is adopted to carry out training of equipment comprehensive operation state evaluation on multi-source data, and a plurality of modal output results are fused based on a model feature weighting strategy so as to generate a preliminary equipment state evaluation result; and combining a preliminary equipment state evaluation result generated by the multi-modal prediction model with expert knowledge in the expert-level decision knowledge base and the equipment historical maintenance record after data preprocessing in a multi-source data weight ratio mode, taking the combined result as an input feature set of a decision model to train the decision model, and generating an equipment health state evaluation result. According to the method, the equipment operation state health assessment efficiency is improved, and the adaptability and timeliness of the model to a dynamic environment are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence algorithms and machine learning, and in particular to a method, electronic device and medium for intelligently and automatically evaluating the comprehensive health status of various types of equipment in a smart park. The method aims to improve the efficiency, timeliness and adaptability of equipment status evaluation in a smart park by using artificial intelligence algorithms and machine learning models. Background Art

[0002] With the gradual opening up of smart parks and the increasing degree of integration, the various software and hardware equipment areas in the current smart park environment are becoming more complex, diversified, and integrated, and the overall health status assessment of smart park equipment has become increasingly complex. Traditional equipment health status assessment usually relies on regular equipment inspections, regular maintenance, and abnormal monitoring, and relies on manual operation. There are problems such as low efficiency, untimely abnormal handling, and difficulty in adapting to the fast-changing real-time operating environment of equipment. In the existing technical solutions, there are methods that use neural network models to analyze equipment runtime data and expert knowledge to predict the health status of equipment, but there is a lack of consideration of the various factors that affect equipment operation, real-time feedback of equipment operation information, and full utilization of a comprehensive expert knowledge base, resulting in low efficiency, lack of accuracy and adaptability in the overall evaluation of park equipment.

[0003] In order to solve the above problems, a multimodal intelligent assessment method for the health status of smart park equipment, electronic equipment and medium are proposed. The multimodal prediction model and decision-making model are combined to integrate equipment runtime data, expert experience and historical maintenance records of equipment. The health status of equipment is automatically estimated, and equipment inspection, maintenance work plans and work task orders are automatically generated. The method automatically adapts to environmental changes during equipment operation, thereby improving the efficiency, flexibility and adaptability of equipment status assessment in smart parks. Summary of the invention

[0004] The present invention proposes a multimodal intelligent evaluation method, electronic device and medium for the health status of smart park equipment. The method generates health evaluation results by learning from multi-source data during equipment operation, and combines expert knowledge base and equipment historical maintenance records, uses decision models to dynamically adjust and optimize weight configuration, generates decision results, and finally uses decision results to autonomously generate work plans and work task orders, thereby improving the efficiency, adaptability and timeliness of smart park equipment status evaluation.

[0005] A multi-modal intelligent evaluation method for the health status of smart park equipment includes the following steps:

[0006] Data collection: Acquire multi-source data from the real-time equipment monitoring platform, including equipment electrical characteristic data, equipment status data during operation, environmental data, and external factors; the environmental data includes air temperature and humidity, and the external factors include equipment historical maintenance records, expert experience and knowledge, and industry technical standards;

[0007] Data preprocessing: preprocess the collected multi-source data, including classification and cleaning, missing value processing, outlier processing, data smoothing, standardization and normalization. Based on data classification and correlation analysis, filter out the features that have a significant impact on the equipment operation status from the multi-source data and reduce redundant features;

[0008] Multimodal prediction model training and data weighted fusion: Use preprocessed data and adopt a multimodal prediction model to train the comprehensive operation status assessment of equipment for multi-source data. Each prediction model and multi-source data within the multimodal prediction model have a one-to-one mapping relationship. Each modal model generates the health status prediction result and weight of the data source according to the corresponding data source. Then, the output results of multiple modes of the multimodal prediction model are fused based on the model feature weighting strategy to generate a preliminary equipment status assessment result.

[0009] Decision analysis: The preliminary equipment status assessment results generated by the multimodal prediction model are combined with the expert knowledge in the expert-level decision knowledge base and the historical equipment maintenance records after data preprocessing in a multi-source data weight ratio as the input feature set of the decision model to generate the final equipment health status assessment results; when training the decision model, the weight addition process uses an adaptive genetic algorithm to dynamically adjust and optimize the relative weights of historical maintenance records, expert knowledge and preliminary equipment status assessment results based on the characteristics of the input multi-source data to generate the final equipment health diagnosis results to enhance the system's ability to cope with complex scenarios;

[0010] Equipment maintenance task plan generation: Import the generated equipment health diagnosis results into the equipment real-time monitoring system, and independently generate equipment maintenance and inspection work plans based on the equipment health diagnosis results, configure work task orders, and monitor the task orders and work plan status in real time;

[0011] Real-time equipment monitoring and information feedback: Through real-time monitoring of equipment and sensors to monitor the dynamic changes of equipment operation data, the inspection and maintenance feedback data are extracted and integrated into expert knowledge and experience and input into the expert-level decision-making knowledge base. At the same time, the inspection and maintenance feedback data are input into the multimodal prediction model and decision-making model again after data preprocessing for model training and optimization of model parameters.

[0012] Furthermore, the decision model includes a random forest model or a decision tree model, supports multi-source data fusion decision design, and the data combination process uses an adaptive genetic algorithm to adjust the model's multi-source data weights, model parameters and decision tree parameters. The final decision result is based on the adaptive weighted fusion of multi-source data to accurately determine the decision result.

[0013] Furthermore, the expert-level decision-making knowledge base includes the expert historical technical experience of the equipment, historical violation events, normal operating specifications, technical equipment knowledge and abnormal event handling experience, and can be dynamically updated and expanded.

[0014] Furthermore, the equipment maintenance task plan is generated based on short-term scheduled tasks, heartbeat mechanisms or message queues, which can dynamically and autonomously monitor and obtain the final decision results of the decision model, and autonomously configure the decision results into maintenance and inspection task plans and task orders.

[0015] Furthermore, the real-time monitoring and information feedback of the equipment also includes obtaining equipment operation data in real time from the equipment real-time monitoring platform, and adjusting the weight allocation strategy in the multimodal prediction model and decision model based on the equipment operation data to quickly respond to changes in equipment operation parameters.

[0016] Furthermore, the multimodal prediction model training specifically includes:

[0017] 3.1 Model construction and parameter initialization: Use Pytorch and mmseg related libraries to build the initial multimodal prediction model, define the classification cross entropy loss CCE, mean square error loss MSE, combination loss, CL and initialize the model parameters, where the learning rate Lr is 0.1, the weight decay L is 0.002, the number of iterations epoch is 50, and other parameters are randomized;

[0018] 3.2 Model training: The preprocessed training set data is input into the multimodal prediction model, and the loss is optimized by the gradient descent SGD algorithm. During the training process, the multimodal prediction model autonomously processes the features of each category, calculates the feature weights, and obtains the preliminary equipment status assessment results based on the weighted fusion of features;

[0019] 3.3 Model tuning: Use Adam optimizer, cross validation and grid search to optimize the training process and adjust the model's hyperparameters, including learning rate, decay parameter, and momentum; use Matlib to draw CCE, CMSE, and CL loss function change curves in real time during training to prepare for subsequent optimization of model training strategies;

[0020] 3.4 Model performance verification: Evaluate model performance on the validation set, record prediction errors and important features, and retrain or fine-tune the model as needed to reduce the impact of parameter randomization;

[0021] 3.5 Generalization performance verification: Use the test set to evaluate the model performance, record the classification cross entropy, mean square error loss, and combination loss of each iteration, and draw a curve chart, and compare it with the corresponding indicator data curve of the training set.

[0022] Furthermore, the update of the multi-source data weight set is expressed as:

[0023]

[0024] Among them, k represents the number of multi-source data categories, Z i Represents the multi-source data input features, ω Z,i and ω Z+1,i Respectively represent Z i The current iteration weight and the next iteration weight, Lr epoch Represents the current iterative learning rate, Lr epoch The update strategy is defined as:

[0025] Lr epoch+1 =Lr epoch -L·epoch

[0026] The classification cross entropy loss function CCE is defined as follows:

[0027]

[0028] Among them, y i is the true label, i.e. the true state of the device, y i ` is the probability distribution predicted by the model, C is the number of task categories;

[0029] The mean square error loss function MSE is defined as follows:

[0030]

[0031] Among them, N is the number of class samples, y i is the true label (actual state), y i ` is the predicted value;

[0032] The combined loss function CL is defined as follows:

[0033] CL=αMSE+(1-α)CCE

[0034] Among them, α is a weight hyperparameter, which is used to adjust the weight ratio of the prediction task and the classification task.

[0035] Furthermore, the decision analysis specifically includes:

[0036] 4.1 Initialize weights: Set the weight of the multimodal prediction model prediction value to ω L, the weight of the predicted value of the expert decision knowledge base is ω E , equipment maintenance record weight ω W , the corresponding initialization weights are ω L =0.4,ω E =0.4,ω W =0.2;

[0037] 4.2 Setting the decision-making system: The prediction model generates results based on the average strategy to obtain the decision results, combining the learning rate η, momentum α, weight decay strategy and cross entropy regression, and using the adaptive genetic algorithm to dynamically optimize the expert decision knowledge weight ω E , prediction model output result weight ω L and maintenance record weight ω W , the weight update is expressed as:

[0038]

[0039] where k∈(L,W,E);

[0040] 4.3 Dynamic weight adjustment: When the system obtains new device feedback data, it calculates the effect of the current decision and uses the decision system to adjust the weight distribution; it uses grid search to traverse the weight combination, evaluates the combination weight based on cross-validation, and obtains the global optimal weight combination; during the search process, the balanced mode is used to automatically adjust the unbalanced weight of the samples;

[0041] 4.4 Decision result output: The final decision is generated by using the adaptive feature weighted fusion method based on the adjusted weights. The calculation of the final decision combines the predicted value of the multimodal prediction model, the recommended value of the expert decision knowledge base, and the historical maintenance record data. The calculation formula is:

[0042] y final =ω L ·y L +ω E ·y E +ω W ·y W

[0043] The final decision results are imported into the monitoring platform for display. The monitoring platform generates work plans and work orders based on the decision results and designates executors. During the execution of the work order, the equipment parameters are recorded in real time and fed back to the decision model and prediction model for real-time health status prediction and decision adjustment.

[0044] An electronic device includes: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the described method.

[0045] A computer readable medium stores a computer program, which implements the method when executed by a processor.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. Combine the multimodal prediction model with multi-source data during equipment operation to automatically generate comprehensive evaluation results of equipment health status without manual intervention, thereby improving evaluation efficiency and accuracy.

[0048] 2. Combining comprehensive evaluation results, expert experience and historical equipment maintenance records, using decision-making models and adaptive genetic algorithms to dynamically adjust and optimize weights and parameter allocations, autonomously generate global optimal decisions, and enhance decision accuracy and adaptability.

[0049] 3. Through real-time monitoring and information feedback, multimodal prediction models and decision-making models can self-learn and optimize, continuously improving the adaptability and scientificity of comprehensive equipment health status assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of a multi-modal intelligent evaluation method for health status of smart park equipment according to an embodiment of the present invention;

[0051] Figure 2 It is a diagram of the overall implementation steps of the present invention;

[0052] Figure 3 It is a system structure diagram of the present invention;

[0053] Figure 4 This is a flow chart of the decision model training of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] See also Figure 1-3 The embodiment of the present invention provides a multi-modal intelligent evaluation method for health status of smart park equipment, comprising the following steps:

[0056] Step 1. Data Collection

[0057] Access multi-source data from the real-time equipment monitoring platform, including equipment electrical characteristic data, equipment status data during operation, environmental data and external factors. Environmental data includes air temperature and humidity. External factors include equipment historical maintenance records, expert experience and industry knowledge and technical standards.

[0058] Step 2. Data preprocessing

[0059] The collected data is cleaned, outliers are processed, missing values ​​are filled, data is classified, and normalized. Common features of the data are extracted to construct a classification data set, including a training set, a validation set, and a test set, with a size ratio of 8:1:1. Based on data classification and correlation analysis, features that have a significant impact on the operating status of the equipment are screened out, redundant features are cut, the amount of calculation is reduced, and the calculation efficiency is improved. Finally, the data file is exported according to the classification ratio of the data set.

[0060] Step 3. Multimodal prediction model training and data weighted fusion

[0061] The training steps of the multimodal prediction model are shown in the figure below. The preprocessed data is used as the model input, and the multimodal prediction model is used to train the corresponding multi-source data. During the training process, gradient descent and optimizer are used to assist model training. The Matlib library is used to draw the training loss curve, and the training changes are monitored in real time to facilitate the adjustment of the training strategy later. The multimodal prediction model training in step 3 specifically includes:

[0062] 3.1 Model construction and parameter initialization: Use Pytorch and mmseg related libraries to build the initial multimodal prediction model, define the classification cross entropy loss (CCE), mean square error loss (MSE), combined losses (CL) and initialize the model parameters. The learning rate Lr is 0.1, the weight decay L is 0.002, the number of iterations epoch is 50, and other parameters are randomized.

[0063] The update of multi-source data weight set can be expressed as:

[0064]

[0065] Among them, k represents the number of multi-source data categories, Z i Represents the multi-source data input features, ω Z,i and ω Z+1,i Respectively represent Z i The current iteration weight and the next iteration weight, Lr epoch Represents the current iterative learning rate, Lr epoch The update strategy is defined as:

[0066] Lr epoch+1 =Lr epoch-L·epoch

[0067] The categorical cross entropy loss function (CCE) is defined as follows:

[0068]

[0069] Among them, y i is the true label, i.e. the true state of the device, y i ` is the probability distribution predicted by the model, and C is the number of task categories. Categorical cross entropy loss (CCE) is an indicator that measures the difference between the probability distribution predicted by the model and the actual label. It encourages the model to increase the probability of the correct category and reduce the probability of the wrong category. It is used in the model for health status decision classification tasks. The smaller the CCE value, the smaller the difference between the predicted value and the true value, and the higher the model prediction accuracy.

[0070] The mean square error loss function (MSE) is defined as follows:

[0071]

[0072] Among them, N is the number of class samples, y i is the true label (actual state), y i ` is the predicted value. The mean square error loss (MSE) is an indicator to measure the difference between the predicted value and the true value. Based on the maximum reward and punishment mechanism, a higher penalty weight is given to larger errors. It is used in the model to compare the actual health status of the equipment with the health status predicted by the model. The smaller the MSE value, the closer the model prediction result is to the true value and the higher the accuracy.

[0073] The combined loss function (CL) is defined as follows:

[0074] CL=αMSE+(1-α)CCE

[0075] α is a weight hyperparameter. It is used to adjust the weight ratio of prediction task and classification task. This model involves two major tasks: health status prediction and problem classification evaluation. It is necessary to evaluate the overall fitting effect of the model for multi-source data, but it focuses more on equipment health status prediction, so α is initialized to 0.6.

[0076] 3.2 Model training: The preprocessed training set data is input into the multimodal prediction model, and the loss is optimized by the gradient descent (SGD) algorithm. During the training process, the multimodal prediction model autonomously processes the features of each category, calculates the feature weights, and obtains the preliminary equipment status assessment results based on the weighted fusion of features;

[0077] 3.3 Model tuning: First, use Adam optimizer, cross validation and grid search to optimize the training process and adjust the model's hyperparameters, such as learning rate, decay parameter, momentum, etc., to improve the model's generalization ability and prediction accuracy. Second, use Matlib to draw the CCE, CMSE, and CL loss function change curves in real time during the training process to prepare for subsequent optimization of the model training strategy.

[0078] 3.4 Model performance verification: Evaluate model performance on the validation set, record prediction errors and important features, and retrain or fine-tune the model as needed to reduce the impact of parameter randomization.

[0079] 3.5 Generalization performance verification: Use the test set to evaluate the model performance, record the classification cross entropy, mean square error loss, and combination loss of each iteration, and draw a curve chart, and compare it with the corresponding indicator data curve of the training set.

[0080] Step 4. Decision Analysis

[0081] The decision model uses the random forest model to integrate the expert knowledge base, equipment historical maintenance records and the prediction results of the multimodal prediction model to form the final optimal decision result. The decision model training process is as follows: Figure 4 During the training process, the weights between expert decision-making knowledge, equipment maintenance records and model prediction results are dynamically adjusted based on the adaptive genetic algorithm to ensure optimal decision-making under different internal and external environments.

[0082] Step 4 specifically includes:

[0083] 4.1 Initialization weight: Assume that the weight of the predicted value of the multimodal prediction model is ω L , the weight of the predicted value of the expert decision knowledge base is ω E , equipment maintenance record weight ω W The corresponding initialization weights are ω L =0.4,ω E =0.4,ω W =0.2.

[0084] 4.2 Setting the decision-making system: The prediction model generates results based on the average strategy to obtain the decision results, combining the learning rate η, momentum α, weight decay strategy and cross entropy regression, and using the adaptive genetic algorithm to dynamically optimize the expert decision knowledge weight ω E , prediction model output result weight ω L and maintenance record weight ω W , the weight update can be expressed as:

[0085]

[0086] Where k∈(L,W,E).

[0087] 4.3 Dynamic weight adjustment: When the system obtains new device feedback data, it calculates the effect of the current decision and uses the decision system to adjust the weight distribution. Use grid search to traverse the weight combination, evaluate the combination weight based on cross-validation, and obtain the global optimal weight combination. Use the balanced mode to automatically adjust the sample imbalance weight during the search process. When the device data fluctuates greatly, perform multi-point sampling on the data and mark the abnormal data, and smooth the sampled data based on the moving average method.

[0088] 4.4 Decision result output: The final decision is generated by using the adaptive feature weighted fusion method based on the adjusted weights. The calculation of the final decision combines the predicted value of the multimodal prediction model, the recommended value of the expert decision knowledge base, and the historical maintenance record data. The calculation formula is:

[0089] y final =ω L ·y L +ω E ·y E +ω W ·y W

[0090] The final decision results are imported into the monitoring platform for display. The monitoring platform generates work plans and work orders based on the decision results and designates executors. During the execution of the work order, the equipment parameters are recorded in real time and fed back to the decision model and prediction model for real-time health status prediction and decision adjustment.

[0091] 5. Generate equipment maintenance task plan

[0092] The system background automatically imports the generated equipment health diagnosis results (decision results) into the equipment real-time monitoring system, and independently generates equipment maintenance and inspection work plans based on the equipment health diagnosis results, configures work task orders, and monitors the status of task orders and work plans in real time.

[0093] 6. Real-time equipment monitoring and information feedback

[0094] The monitoring platform monitors the changes in the internal and external environmental parameters of the equipment in real time, dynamically adjusts the policy configuration based on the feedback data, and re-optimizes the adjusted policy in the prediction model and decision model. After the work order is executed, it is automatically written into the historical records and the experience is written into the expert decision knowledge base, and the latest data feedback is continuously obtained through the real-time monitoring platform. Based on the feedback data, the system continuously learns and optimizes the prediction model and decision model to form a self-optimization mechanism.

[0095] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A multi-modal intelligent evaluation method for the health status of smart park equipment, characterized in that: The following steps are involved: Data collection: Acquire multi-source data from the real-time equipment monitoring platform, including equipment electrical characteristic data, equipment status data during operation, environmental data, and external factors; the environmental data includes air temperature and humidity, and the external factors include equipment historical maintenance records, expert experience and knowledge, and industry technical standards; Data preprocessing: preprocess the collected multi-source data, including classification and cleaning, missing value processing, outlier processing, data smoothing, standardization and normalization. Based on data classification and correlation analysis, filter out the features that have a significant impact on the equipment operation status from the multi-source data and reduce redundant features; Multimodal prediction model training and data weighted fusion: Use preprocessed data and adopt a multimodal prediction model to train the comprehensive operation status assessment of equipment for multi-source data. Each prediction model and multi-source data within the multimodal prediction model have a one-to-one mapping relationship. Each modal model generates the health status prediction result and weight of the data source according to the corresponding data source. Then, the output results of multiple modes of the multimodal prediction model are fused based on the model feature weighting strategy to generate a preliminary equipment status assessment result. Decision analysis: The preliminary equipment status assessment results generated by the multimodal prediction model are combined with the expert knowledge in the expert-level decision knowledge base and the historical equipment maintenance records after data preprocessing in a multi-source data weight ratio as the input feature set of the decision model to generate the final equipment health status assessment results; when training the decision model, the weight addition process uses an adaptive genetic algorithm to dynamically adjust and optimize the relative weights of historical maintenance records, expert knowledge and preliminary equipment status assessment results based on the characteristics of the input multi-source data to generate the final equipment health diagnosis results to enhance the system's ability to cope with complex scenarios; Equipment maintenance task plan generation: Import the generated equipment health diagnosis results into the equipment real-time monitoring system, and independently generate equipment maintenance and inspection work plans based on the equipment health diagnosis results, configure work task orders, and monitor the task orders and work plan status in real time; Real-time equipment monitoring and information feedback: Through real-time monitoring of equipment and sensors to monitor the dynamic changes of equipment operation data, the inspection and maintenance feedback data are extracted and integrated into expert knowledge and experience and input into the expert-level decision-making knowledge base. At the same time, the inspection and maintenance feedback data are input into the multimodal prediction model and decision-making model again after data preprocessing for model training and optimization of model parameters.

2. The method according to claim 1, characterized in that The decision model includes a random forest model or a decision tree model, supports multi-source data fusion decision design, and the data combination process uses an adaptive genetic algorithm to adjust the model multi-source data weights, model parameters and decision tree parameters. The final decision result is based on the adaptive weighted fusion of multi-source data to accurately determine the decision result.

3. The method according to claim 1, characterized in that The expert-level decision-making knowledge base includes the expert historical technical experience of the equipment, historical violation events, normal operating specifications, technical equipment knowledge and abnormal event handling experience, and can be dynamically updated and expanded.

4. The method according to claim 1, characterized in that: The equipment maintenance task plan is generated based on short-term scheduled tasks, heartbeat mechanisms or message queues, which can dynamically and autonomously monitor and obtain the final decision results of the decision model, and autonomously configure the decision results into maintenance and inspection task plans and task orders.

5. The method according to claim 1, characterized in that The real-time monitoring and information feedback of the equipment also includes obtaining equipment operation data from the equipment real-time monitoring platform in real time, and adjusting the weight allocation strategy in the multimodal prediction model and the decision model based on the equipment operation data to quickly respond to changes in equipment operation parameters.

6. The method according to claim 1, characterized in that The multimodal prediction model training specifically includes: 3.1 Model construction and parameter initialization: Use Pytorch and mmseg related libraries to build the initial multimodal prediction model, define the classification cross entropy loss CCE, mean square error loss MSE, combination loss, CL and initialize the model parameters, where the learning rate Lr is 0.1, the weight decay L is 0.002, the number of iterations epoch is 50, and other parameters are randomized; 3.2 Model training: The preprocessed training set data is input into the multimodal prediction model, and the loss is optimized by the gradient descent SGD algorithm. During the training process, the multimodal prediction model autonomously processes the features of each category, calculates the feature weights, and obtains the preliminary equipment status assessment results based on the weighted fusion of features; 3.3 Model tuning: Use Adam optimizer, cross validation and grid search to optimize the training process and adjust the model's hyperparameters, including learning rate, decay parameter, and momentum; use Matlib to draw CCE, CMSE, and CL loss function change curves in real time during training to prepare for subsequent optimization of model training strategies; 3.4 Model performance verification: Evaluate model performance on the validation set, record prediction errors and important features, and retrain or fine-tune the model as needed to reduce the impact of parameter randomization; 3.5 Generalization performance verification: Use the test set to evaluate the model performance, record the classification cross entropy, mean square error loss, and combination loss of each iteration, and draw a curve chart, and compare it with the corresponding indicator data curve of the training set.

7. The method according to claim 6, characterized in that in, The update of multi-source data weight set is expressed as: Among them, k represents the number of multi-source data categories, Z i Represents the multi-source data input features, ω Z,i and ω Z+1,i Respectively represent Z i The current iteration weight and the next iteration weight, Lr epoch Represents the current iterative learning rate, Lr epoch The update strategy is defined as: The classification cross entropy loss function CCE is defined as follows: Among them, y i is the true label, i.e. the true state of the device, y i ` is the probability distribution predicted by the model, C is the number of task categories; The mean square error loss function MSE is defined as follows: Among them, N is the number of class samples, y i is the true label (actual state), y i ` is the predicted value; The combined loss function CL is defined as follows: CL=αMSE+(1-α)CCE Among them, α is a weight hyperparameter, which is used to adjust the weight ratio of the prediction task and the classification task.

8. The method according to claim 7, characterized in that The decision analysis specifically includes: 4.1 Initialize weights: Set the weight of the multimodal prediction model prediction value to ω L , the weight of the predicted value of the expert decision knowledge base is ω E , equipment maintenance record weight ω W , the corresponding initialization weights are ω L =0.4,ω E =0.4,ω W =0.2; 4.2 Setting the decision-making system: The prediction model generates results based on the average strategy to obtain the decision results, combining the learning rate η, momentum α, weight decay strategy and cross entropy regression, and using the adaptive genetic algorithm to dynamically optimize the expert decision knowledge weight ω E , prediction model output weight ω L and maintenance record weight ω W , the weight update is expressed as: where k∈(L,W,E); 4.3 Dynamic weight adjustment: When the system obtains new device feedback data, it calculates the effect of the current decision and uses the decision system to adjust the weight distribution; it uses grid search to traverse the weight combination, evaluates the combination weight based on cross-validation, and obtains the global optimal weight combination; during the search process, the balanced mode is used to automatically adjust the unbalanced weight of the samples; 4.4 Decision result output: The final decision is generated by using the adaptive feature weighted fusion method based on the adjusted weights. The calculation of the final decision combines the predicted value of the multimodal prediction model, the recommended value of the expert decision knowledge base, and the historical maintenance record data. The calculation formula is: y final =ω L ·y L +ω E ·y E +ω W ·y W The final decision results are imported into the monitoring platform for display. The monitoring platform generates work plans and work orders based on the decision results and designates executors. During the execution of the work order, the equipment parameters are recorded in real time and fed back to the decision model and prediction model for real-time health status prediction and decision adjustment.

9. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 8.

10. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images