New energy equipment management system based on artificial intelligence vision
Patent Information
- Application Number
- CN202510540075.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-27
- Publication Date
- 2025-08-15
Smart Images
Figure CN120492887A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of new energy equipment management, and specifically to a new energy equipment management system based on artificial intelligence vision. Background Art
[0002] With the rapid development of new energy technologies, new energy equipment such as wind turbines, solar panels, and energy storage systems are becoming increasingly prominent in the energy supply system. The efficient and stable operation of these devices is crucial for ensuring energy supply and promoting sustainable development. However, new energy equipment faces numerous operational and maintenance challenges due to its technical complexity, volatile operating environments, and the natural wear and tear associated with long-term operation.
[0003] Traditional new energy equipment management systems primarily rely on threshold-based fault alarm mechanisms, which present numerous limitations. Traditional methods often rely solely on sensor data to monitor equipment status, lacking intuitive visual information and making it difficult to fully and accurately reflect the equipment's actual operating conditions. This is particularly true for issues that cannot be directly detected by sensors, such as cosmetic damage and abnormal operating behavior. Traditional management systems typically only issue alarms after equipment failures occur, lacking effective fault prediction methods. This often forces maintenance personnel to take action only after equipment failures have already occurred or performance has significantly degraded, increasing repair costs and downtime.
[0004] Traditional maintenance decisions are primarily based on empirical rules and fixed cycles, lacking specificity and flexibility. This one-size-fits-all approach often leads to over- or under-maintenance, wasting resources and potentially impacting equipment operation. Traditional management systems often take a long time to detect and respond to changes in equipment status. This is primarily due to the cumbersome process of data collection, processing, and analysis, coupled with a lack of real-time data feedback mechanisms. Summary of the Invention
[0005] The purpose of the present invention is to provide a new energy equipment management system based on artificial intelligence vision to solve the problems raised in the above background technology.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a new energy equipment management system based on artificial intelligence vision, the system comprising:
[0007] The equipment monitoring unit builds an equipment status monitoring model to monitor the operating status of new energy equipment in real time. Its input data includes:
[0008] Real-time video data of new energy equipment, including the equipment's appearance, operating movements, and environmental background;
[0009] Device sensor data, including device temperature, humidity, vibration frequency, and electrical parameters;
[0010] The output data is: a characteristic description of the current operating status of the equipment, including the integrity and operating efficiency of each component of the equipment;
[0011] The fault prediction unit builds a fault prediction model to predict possible equipment failures in the future. Its input data is the characteristic description of the equipment's current operating status output by the equipment status monitoring model; its output data is the possible future equipment failure type, failure probability, and expected failure time.
[0012] The maintenance decision unit builds an equipment maintenance decision model to determine the maintenance strategy for new energy equipment. Its input data is the predicted value of the equipment's future state output by the equipment status monitoring module, and its output data is the preliminary maintenance instructions.
[0013] The instruction optimization module builds an optimization adjustment model to optimize the preliminary maintenance instructions output by the maintenance decision module based on the difference between the equipment's predicted status output by the equipment status monitoring module and the equipment's actual operating status obtained through real-time sensor monitoring, and outputs the optimized final maintenance instructions;
[0014] The system management center is used to combine the historical operation data of new energy equipment, actual equipment image data and actual working environment data, and generate preliminary maintenance instructions for new energy equipment through the equipment status monitoring module and maintenance decision module to adjust equipment operation. It continuously monitors the actual operating status of new energy equipment and compares it with the prediction results of the equipment status monitoring module. If significant deviations are found, the optimization adjustment model will be immediately triggered to recalculate the final maintenance instructions.
[0015] Preferably, the equipment status monitoring model adopts a combined architecture of a convolutional neural network and a recurrent neural network. The convolutional neural network part consists of an input layer, multiple convolutional layers, a pooling layer, and a fully connected layer, which is used to process real-time video data and extract equipment appearance and action features; the recurrent neural network part contains an LSTM unit, which is used to process the time series of equipment sensor data and capture the dynamic changes in the equipment's operating status; the model output layer integrates the outputs of the convolutional neural network and the recurrent neural network to generate a feature description of the equipment's current operating status.
[0016] Preferably, the fault prediction model consists of an input layer, multiple restricted Boltzmann machine (RBM) layers, a fully connected layer, and an output layer; the input layer receives the feature description of the current operating status of the device output by the device status monitoring model; the RBM layer is used to learn high-order correlations between features in an unsupervised manner; the fully connected layer maps the output of the RBM layer to the fault prediction space; and the output layer outputs possible future fault information of the device.
[0017] Preferably, the training step of the fault prediction model includes:
[0018] Step 1: Construct the network structure of the model, including determining the number of RBM layers and the number of nodes in each layer;
[0019] Step 2: Collect training data, including the characteristic description of the current operating status of the equipment output by the equipment status monitoring model, and the corresponding actual equipment fault records as labels;
[0020] Step 3: Initialize the parameters of the DBN model and set the hyperparameters for pre-training and fine-tuning;
[0021] Step 4: First, pre-train the RBM layer through unsupervised learning, and then fine-tune the entire DBN model in a supervised manner to enable the model to accurately predict possible future equipment failures;
[0022] Step 5: After training is completed, evaluate the model performance and select the model with the best performance as the final fault prediction model.
[0023] Preferably, in the fault prediction model, the training of the restricted Boltzmann machine (RBM) layer adopts a contrastive divergence algorithm, which is performed by the following steps:
[0024] S1: Initialize the weights and biases of RBM;
[0025] S2: For a given input data, calculate the activation probability of the hidden layer and sample the hidden layer state;
[0026] S3: Calculate the reconstruction probability of the visible layer based on the hidden layer state, and sample to obtain the reconstructed visible layer state;
[0027] S4: Using the difference between the reconstructed visible layer state and the original input data, the weights and biases of the RBM are updated by gradient descent.
[0028] S5: Repeat steps S2 to S4 until the RBM model parameters converge, completing the pre-training of the RBM layer.
[0029] Preferably, the maintenance decision module uses a deep reinforcement learning (DDPG) algorithm to construct an equipment maintenance decision model.
[0030] Preferably, the method for constructing the equipment maintenance decision model includes:
[0031] Define the state space, action space, and reward function of the DDPG algorithm. The state space is the future state prediction output by the device state prediction model, the action space is the possible maintenance instructions, and the reward function is set based on the difference between the predicted device state output by the device state monitoring module and the actual device operating state obtained through real-time sensor monitoring.
[0032] Preferably, the instruction optimization module uses a random forest RF algorithm to construct an optimization adjustment model.
[0033] Preferably, the training step of the optimization adjustment model includes:
[0034] Collect training data, including the difference between the predicted state output by the equipment condition monitoring module and the actual operating state obtained by real-time sensor monitoring. Also, collect expert-annotated optimal maintenance instructions for a given difference between the predicted and actual states as labels for the training data.
[0035] Build a random forest RF algorithm model and set RF parameters, including the number of trees, maximum depth, and minimum number of sample splits;
[0036] The random forest RF algorithm model is trained using training data and the bagging method is used to improve the model performance so that the model can accurately predict the optimal maintenance instructions under given input differences.
[0037] After the training is completed, the optimized and adjusted model is obtained.
[0038] Preferably, the method for determining whether the deviation between the actual operating state and the predicted state is significant is: calculating the mean square error between the actual operating state and the predicted state, setting an error threshold, comparing the calculated mean square error with the set threshold, and if the mean square error exceeds the set threshold, it is determined to be a significant deviation.
[0039] Compared with the prior art, the present invention has the following beneficial effects:
[0040] This system integrates real-time video data with equipment sensor data to build a comprehensive equipment status monitoring model. This not only addresses the shortcomings of traditional monitoring methods that rely solely on sensor data, but also intuitively captures visual information such as damage to the equipment's appearance and abnormal operating behavior, thereby more comprehensively and accurately reflecting the equipment's actual operating status, greatly improving the accuracy and efficiency of monitoring. The system's built-in fault prediction model can predict the type, probability, and expected time of future equipment failures based on the characteristic description of the equipment's current operating status. This function enables maintenance personnel to take preventive measures before failures occur, effectively reducing maintenance costs and downtime, and improving equipment reliability and availability.
[0041] By building an equipment maintenance decision-making model, this system can generate targeted maintenance instructions based on the equipment's actual operating status and predicted data. Compared to traditional maintenance methods based on empirical rules and fixed cycles, this intelligent maintenance decision-making is more flexible and accurate, effectively avoiding the problems of over-maintenance or under-maintenance, and improving resource utilization and maintenance efficiency. The system management center continuously monitors the actual operating status of new energy equipment and compares it with the predicted results of the equipment status monitoring model. Once a significant deviation is found, the optimization and adjustment model is immediately triggered to recalculate the final maintenance instructions. This real-time response and dynamic adjustment capability ensures the timeliness and effectiveness of maintenance strategies, further improving the operational stability and safety of the equipment.
[0042] Through intelligent and precise equipment management and maintenance, this system effectively extends the service life of new energy equipment, reduces operation and maintenance costs, and improves energy efficiency. This not only contributes to the economic efficiency and reliability of new energy equipment, but also promotes the sustainable development of the new energy industry, which is of great significance to promoting energy transformation and addressing climate change. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 This is a working principle diagram of the new energy equipment management system of the present invention;
[0044] Figure 2 This is the training flow chart of the fault prediction model;
[0045] Figure 3 This is a flowchart of the construction and working principle of the equipment maintenance decision model based on the DDPG algorithm. DETAILED DESCRIPTION
[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.
[0047] See also Figure 1-3 The present invention provides a technical solution: a new energy equipment management system based on artificial intelligence vision, the system comprising:
[0048] Equipment Monitoring Unit: This unit builds an equipment status monitoring model that receives two types of input data: real-time video data from new energy equipment, captured by cameras installed on or near the equipment, including the equipment's appearance, operating behavior, and environmental context; and sensor data, such as temperature, humidity, vibration frequency, and electrical parameters, collected in real time by various built-in sensors. The equipment status monitoring model processes and analyzes this data and outputs a characteristic description of the equipment's current operating status, including the health of its components, its operating efficiency, and any signs of anomalies.
[0049] Fault Prediction Unit: Builds a fault prediction model whose input data is the characteristic description of the device's current operating status output by the device condition monitoring model. The fault prediction model uses machine learning or deep learning algorithms to analyze the device's historical operating data and current status characteristics to predict the type, probability, and expected time of future faults.
[0050] Maintenance Decision Unit: This unit builds an equipment maintenance decision model. Its input data is the predicted future equipment status output by the equipment status monitoring module. Based on this prediction, the maintenance decision model combines factors such as the equipment's maintenance history, maintenance costs, and equipment importance to determine the maintenance strategy for the new energy equipment and output preliminary maintenance instructions.
[0051] The Instruction Optimization Module builds an optimization and adjustment model. This model receives as input the difference between the predicted equipment status output by the Equipment Condition Monitoring Module and the actual equipment operating status obtained through real-time sensor monitoring. Based on this difference, the optimization and adjustment model optimizes the preliminary maintenance instructions output by the Maintenance Decision Module to more accurately reflect the actual maintenance needs of the equipment and outputs the optimized final maintenance instructions.
[0052] System Management Center: The System Management Center is responsible for integrating historical operating data, actual device image data, and actual operating environment data for new energy equipment. Using the Equipment Status Monitoring Module and the Maintenance Decision Module, the System Management Center generates preliminary maintenance instructions for new energy equipment to adjust equipment operation. The System Management Center continuously monitors the actual operating status of new energy equipment and compares it with the predictions from the Equipment Status Monitoring Module. If significant deviations are detected, the optimization and adjustment model is immediately triggered to recalculate and output the final maintenance instructions.
[0053] The present invention will be further described below in conjunction with Examples 1 to 4:
[0054] Example 1:
[0055] In the device status monitoring unit, the present invention uses a combination of a convolutional neural network (CNN) and a recurrent neural network (specifically, an RNN containing LSTM units) to construct a device status monitoring model. The network structure of the device status monitoring model includes:
[0056] Input layer: This layer is designed to receive real-time video data streams from new energy devices. Video data is in frames, and each frame contains information about the device's appearance and movement details.
[0057] Multiple convolutional layers: Convolutional layers convolve the input image with a series of convolution kernels (or filters) to extract spatial features from the image. These features include the shape, edges, and texture of the device. A nonlinear activation function (such as ReLU) is typically applied after each convolutional layer to enhance the model's expressiveness.
[0058] Pooling layer: The pooling layer downsamples the output of the convolutional layer, reducing the spatial dimension of the data while retaining important feature information. This helps reduce the amount of computation and prevents overfitting.
[0059] Fully connected layer: The fully connected layer flattens the output of the pooling layer into a one-dimensional vector and further extracts and integrates features through a series of fully connected operations. These features are then used to generate a description of the device's appearance and motion.
[0060] LSTM units are used to process time series of device sensor data. This data, including temperature, humidity, vibration frequency, and electrical parameters, changes over time, reflecting the dynamic characteristics of the device's operating status. Through its internal forget gate, input gate, and output gate mechanism, LSTM units effectively capture and memorize long-term dependencies in time series. This enables the model to accurately describe the changing trends of device operating status over time.
[0061] The model output layer integrates the outputs of the convolutional neural network and the recurrent neural network. The convolutional neural network output provides the device's appearance and motion characteristics, while the recurrent neural network output reflects the dynamic changes in the device's operating status. Through appropriate integration mechanisms (such as fully connected layers and weighted sums), these two outputs are fused together to generate a feature description of the device's current operating status. This feature description includes the integrity of each device component, operating efficiency, and signs of abnormalities.
[0062] Specific example: Condition monitoring of wind turbines
[0063] Real-time video data processing: A camera mounted on a wind turbine captures real-time video data. The convolutional neural network processes each frame of video, extracting visual features such as the shape, color, and texture of the wind turbine blades, as well as motion characteristics such as the blades' rotational speed and angle.
[0064] Sensor data processing: Meanwhile, the wind turbine's sensors collect real-time sensor data such as temperature and vibration frequency. This data is fed into the recurrent neural network. The LSTM unit processes this time series data to capture dynamic changes in the wind turbine's operating status, such as temperature fluctuations and abnormal increases in vibration frequency.
[0065] Feature Description Generation: The model output layer integrates the outputs of the convolutional neural network and the recurrent neural network to generate a feature description of the wind turbine's current operating status. These feature descriptions can be used for subsequent tasks such as fault prediction and maintenance decision-making, providing strong support for the safe operation of the wind turbine.
[0066] Traditional monitoring methods for new energy equipment rely primarily on sensor data and lack intuitive visual information. This makes it difficult to detect issues such as damage to the equipment's exterior or abnormal operation. By combining convolutional neural networks with recurrent neural networks, the present invention can simultaneously process real-time video data and sensor data, providing comprehensive equipment status monitoring. Traditional monitoring methods often struggle to accurately capture dynamic changes in equipment operating status. By introducing LSTM units, the present invention can effectively process time series data, capture long-term dependencies in equipment operating status, and improve the accuracy of status monitoring.
[0067] Example 2:
[0068] This paper proposes a fault prediction model based on a deep belief network (DBN). The model consists of an input layer, multiple restricted Boltzmann machine (RBM) layers, a fully connected layer, and an output layer. It is used to predict possible future failures of new energy equipment. The following is a detailed implementation of the fault prediction model:
[0069] The network structure of the fault prediction model includes:
[0070] Input layer: Receives the characteristic description of the current operating status of the device output by the device status monitoring model. These characteristic descriptions include information such as the device's appearance, action characteristics, and dynamic changes in the operating status.
[0071] Multiple RBM layers: RBM layers are used to learn high-order correlations between features in an unsupervised manner. By stacking multiple RBM layers, the model is able to extract and represent higher-level, more abstract features of the input data layer by layer.
[0072] Fully connected layer: Maps the output of the RBM layer to the fault prediction space. Through a series of fully connected operations, the fully connected layer integrates the high-order features extracted by the RBM layer to provide a powerful feature representation for fault prediction.
[0073] Output layer: Outputs information about possible future device failures. This layer typically uses a softmax function or other classifier to map the output of the fully connected layer to a specific fault type or probability.
[0074] The training steps of the fault prediction model are as follows:
[0075] Step 1: Determine the number of RBM layers and the number of nodes in each layer. This is usually determined based on the complexity of the equipment and the requirements for fault prediction.
[0076] Step 2: Collect the characteristic description of the current operating status of the equipment output by the equipment status monitoring model and the corresponding actual equipment fault records as labels. This data is used to train the fault prediction model.
[0077] Step 3: Initialize the weights and biases of the RBM layer, and set the hyperparameters for pre-training and fine-tuning, such as learning rate, number of iterations, etc.
[0078] Step 4: First, pre-train the RBM layer through unsupervised learning. Using the contrastive divergence algorithm, pre-train the RBM layer layer by layer to learn the feature representation of the input data. Then, fine-tune the entire DBN model using a supervised approach. Based on the pre-training, fine-tune the model using labeled training data to ensure that it can accurately predict possible future equipment failures.
[0079] Step 5: After training is complete, use the test dataset to evaluate the model performance. The model with the best performance is selected as the final fault prediction model.
[0080] The training steps of the RBM layer include:
[0081] S1: Initialize RBM weights and biases: Randomly initialize RBM weights and biases as the starting point for training.
[0082] S2: Calculate the activation probability of the hidden layer and sample it: For a given input data, calculate the activation probability of the hidden layer. Then, sample the hidden layer state according to the activation probability to obtain the binary representation of the hidden layer.
[0083] S3: Calculate the reconstruction probability of the visible layer and sample: Based on the hidden layer state, calculate the reconstruction probability of the visible layer. Then, sample the visible layer state according to the reconstruction probability to obtain the reconstructed visible layer state.
[0084] S4: Update the RBM weights and biases: Using the difference between the reconstructed visible layer state and the original input data, the RBM weights and biases are updated via gradient descent. This step aims to minimize the reconstruction error and improve the model's ability to represent the input data.
[0085] S5: Repeat steps S2 to S4 until the RBM model parameters converge. The convergence criterion may be that the reconstruction error reaches a preset threshold, or the number of iterations reaches a preset upper limit.
[0086] Taking the fault prediction of wind turbines as an example, the fault prediction model proposed in the present invention is applied to predict faults of wind turbines. First, the feature description of the current operating status of the wind turbine is obtained through the equipment status monitoring model. These feature descriptions are input into the fault prediction model as input data. The model learns the high-order correlation between features through multiple RBM layers and maps the output of the RBM layer to the fault prediction space through a fully connected layer. The output layer then outputs possible future fault information of the wind turbine, such as gearbox failure, generator failure, etc. This information can be used to formulate maintenance plans in advance to reduce the possibility of failures and repair costs.
[0087] Problem Solved: The fault prediction model proposed in this paper addresses the problems of traditional fault prediction methods, which rely on manual feature extraction and are unable to automatically learn high-order correlations between features. By introducing multiple RBM layers and unsupervised learning, the model can automatically extract and represent higher-level, more abstract features of the input data. Pre-training the RBM layers using the contrastive divergence algorithm improves the model's training efficiency and accuracy. The supervised fine-tuning process enables the model to accurately predict possible future equipment failures, providing strong support for equipment maintenance and management.
[0088] Example 3:
[0089] The equipment maintenance decision module uses the deep reinforcement learning DDPG (Deep Deterministic Policy Gradient) algorithm to build an equipment maintenance decision model. The following is a detailed construction method of the equipment maintenance decision model:
[0090] The DDPG algorithm is a combination of deep learning and reinforcement learning, suitable for problems in continuous action spaces. In equipment maintenance decisions, maintenance instructions are often continuous or refinable, making the DDPG algorithm a suitable choice.
[0091] The key elements that define the DDPG algorithm include:
[0092] State space: The state space is defined as the future state predictions output by the equipment state prediction model. These predictions include information such as the equipment's future operating state, possible failure modes, and performance degradation trends. The equipment state prediction model can capture changes in the equipment's state over a period of time, providing forward-looking information for maintenance decisions.
[0093] Action Space: The action space represents possible maintenance instructions. These instructions can include scheduled maintenance, troubleshooting, component replacement, and adjusting operating parameters. The design of the action space should take into account the actual maintenance needs and operability of the equipment to ensure the effectiveness and feasibility of maintenance instructions.
[0094] Reward function: The reward function is based on the difference between the predicted equipment state output by the equipment state monitoring module and the actual equipment operating state obtained through real-time sensor monitoring. When the difference between the predicted and actual states is small, a positive reward is given, indicating that the maintenance decision is effective; when the difference is large, a negative reward is given, indicating that the maintenance decision needs to be adjusted. The design of the reward function should take into account factors such as equipment operational safety, performance stability, and maintenance costs to guide the model to make the optimal maintenance decision.
[0095] The steps to construct the equipment maintenance decision model include:
[0096] Model initialization: Initialize the DDPG algorithm's network structure, including the actor network and the critic network. The actor network is responsible for generating maintenance instructions, while the critic network is responsible for evaluating the quality of maintenance instructions.
[0097] Data collection and preprocessing: This process collects the predicted device status output by the device status monitoring module and the actual device operating status data obtained from real-time sensor monitoring. This data is preprocessed, such as through denoising and normalization, to improve model training effectiveness.
[0098] Model training: The DDPG algorithm is trained using the collected data. During training, the actor network generates maintenance instructions based on the current state, and the critic network evaluates these instructions and provides rewards. Through continuous iterative training, the actor network gradually learns to generate optimal maintenance instructions, and the critic network gradually learns to accurately evaluate the quality of maintenance instructions.
[0099] Model testing and optimization: Use the test dataset to test the trained model and evaluate the accuracy and effectiveness of its maintenance decisions. Based on the test results, optimize the model, such as adjusting the network structure and reward function, to improve model performance.
[0100] Taking wind turbine maintenance decisions as an example, the equipment maintenance decision model proposed in this invention is applied to make maintenance decisions for wind turbines. The state space includes predictions of the wind turbine's future wind speed, power output, and health status of the gearbox and generator. The action space includes maintenance instructions such as adjusting blade angles, inspecting and replacing worn parts, and adjusting the lubrication system. The reward function is set based on factors such as the difference between the wind turbine's actual power output and predicted power output, and the frequency and severity of faults. Through the trained equipment maintenance decision model, optimal maintenance instructions can be generated in real time to ensure the stable operation of the wind turbine and extend its service life.
[0101] The equipment maintenance decision model proposed in this paper addresses the problems of traditional maintenance decision-making methods, which rely on empirical rules and are unable to adapt to changes in equipment status. Using the deep reinforcement learning (DDPG) algorithm, the model automatically learns the mapping between equipment status and maintenance instructions and makes optimal maintenance decisions based on real-time equipment status. This improves the efficiency and accuracy of equipment maintenance, reduces maintenance costs, and provides a strong guarantee for reliable equipment operation.
[0102] Example 4:
[0103] This paper proposes an instruction optimization module based on the Random Forest (RF) algorithm for constructing an optimized adjustment model for equipment maintenance. Random Forest is an ensemble learning method that improves model accuracy and stability by constructing multiple decision trees and integrating their outputs. The Random Forest algorithm is a suitable choice for optimizing equipment maintenance instructions because it can handle complex nonlinear relationships and effectively avoid overfitting.
[0104] Training data collection for the optimization model: The difference between the predicted state output by the equipment condition monitoring module and the actual operating state obtained through real-time sensor monitoring is collected. This difference data reflects the accuracy of the equipment condition prediction and the changes in the actual operating state. Furthermore, expert-annotated optimal maintenance instructions are collected for each given difference between the predicted and actual states. These optimal maintenance instructions, derived from expert experience and equipment maintenance knowledge, serve as labels for the training data.
[0105] Building a random forest RF algorithm model includes:
[0106] Model initialization: Build a random forest (RF) algorithm model and set RF parameters, including the number of trees, maximum depth, and minimum number of sample splits. These parameters affect model performance and accuracy and need to be adjusted based on the specific problem and data characteristics.
[0107] Parameter Setting Explanation: Determines the number of decision trees in the random forest. More trees improve model stability but also increase computational cost. Limits the maximum depth of the decision tree to prevent overfitting. Shallower trees may not capture complex patterns, while deeper trees may lead to overfitting. Determines the minimum number of samples required for a node split to ensure reliable splits.
[0108] The random forest RF algorithm model is trained using training data as follows:
[0109] Training process: The Random Forest (RF) algorithm model is trained using the collected difference data and optimal maintenance instructions. Bagging is used to improve model performance. This involves randomly sampling the dataset multiple times, training a decision tree for each sampling, and then combining the outputs of all decision trees.
[0110] Bagging method description: Bagging is a method for building a strong learner by combining multiple weak learners (such as decision trees). It reduces the variance of the model and improves the generalization ability of the model by randomly sampling the data set and training multiple models.
[0111] After training, the optimized model is obtained. The model is evaluated using a test dataset to verify its accuracy and reliability. Evaluation metrics can include accuracy, recall, and F1 score. Once the optimized model is obtained, it can be applied to actual equipment maintenance scenarios. The difference between the predicted state output by the equipment status monitoring module and the actual operating state obtained through real-time sensor monitoring is input into the optimized model to generate optimal maintenance instructions.
[0112] Taking the equipment maintenance of industrial production lines as an example, the optimization and adjustment model proposed in this invention is applied to make maintenance decisions for the equipment of industrial production lines. The difference data between the predicted state and the actual operating state of the industrial production line equipment, as well as the optimal maintenance instructions marked by experts, are collected. A random forest RF algorithm model is constructed and appropriate parameters are set. The model is trained using the training data, and the model performance is improved through the bagging method. After the training is completed, the optimization and adjustment model is obtained and applied to actual equipment maintenance. Based on the difference data of the equipment, the model can accurately predict the optimal maintenance instructions, such as adjusting equipment parameters, replacing worn parts, etc., to ensure the stable operation and production efficiency of the production line.
[0113] The optimization and adjustment model proposed in this paper addresses the problems of traditional equipment maintenance decision-making, such as reliance on empirical rules, inability to adapt to changes in equipment status, and a lack of scientific evidence. Using the random forest RF algorithm and bagging, the model automatically learns the mapping between equipment status differences and maintenance instructions, and makes optimal maintenance decisions based on real-time equipment status differences. This improves the efficiency and accuracy of equipment maintenance, reduces maintenance costs, and provides a strong guarantee for the stable operation and productivity of industrial production lines.
[0114] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.
[0115] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. A new energy equipment management system based on artificial intelligence vision, characterized in that: The system comprises: The equipment monitoring unit builds an equipment status monitoring model to monitor the operating status of new energy equipment in real time. Its input data includes: Real-time video data of new energy equipment, including the equipment's appearance, operating movements, and environmental background; Device sensor data, including device temperature, humidity, vibration frequency, and electrical parameters; The output data is: a characteristic description of the current operating status of the equipment, including the integrity and operating efficiency of each component of the equipment; The fault prediction unit builds a fault prediction model to predict possible equipment failures in the future. Its input data is the characteristic description of the equipment's current operating status output by the equipment status monitoring model; its output data is the possible future equipment failure type, failure probability, and expected failure time. The maintenance decision unit builds an equipment maintenance decision model to determine the maintenance strategy for new energy equipment. Its input data is the predicted value of the equipment's future state output by the equipment status monitoring module, and its output data is the preliminary maintenance instructions. The instruction optimization module builds an optimization adjustment model to optimize the preliminary maintenance instructions output by the maintenance decision module based on the difference between the equipment's predicted status output by the equipment status monitoring module and the equipment's actual operating status obtained through real-time sensor monitoring, and outputs the optimized final maintenance instructions; The system management center is used to combine the historical operation data of new energy equipment, actual equipment image data and actual working environment data, and generate preliminary maintenance instructions for new energy equipment through the equipment status monitoring module and maintenance decision module to adjust equipment operation. It continuously monitors the actual operating status of new energy equipment and compares it with the prediction results of the equipment status monitoring module. If significant deviations are found, the optimization adjustment model will be immediately triggered to recalculate the final maintenance instructions.
2. The new energy equipment management system based on artificial intelligence vision according to claim 1 is characterized in that: The equipment status monitoring model adopts a combined architecture of convolutional neural networks and recurrent neural networks. The convolutional neural network part consists of an input layer, multiple convolutional layers, a pooling layer, and a fully connected layer, which is used to process real-time video data and extract equipment appearance and action features; the recurrent neural network part contains LSTM units, which are used to process the time series of equipment sensor data and capture the dynamic changes in the equipment's operating status; the model output layer integrates the outputs of the convolutional neural network and the recurrent neural network to generate a feature description of the equipment's current operating status.
3. The new energy equipment management system based on artificial intelligence vision according to claim 1 is characterized in that: The fault prediction model consists of an input layer, multiple restricted Boltzmann machine (RBM) layers, a fully connected layer, and an output layer. The input layer receives the feature description of the current operating state of the device output by the device state monitoring model. The RBM layer is used to learn high-order correlations between features in an unsupervised manner. The fully connected layer maps the output of the RBM layer to the fault prediction space. The output layer outputs information about possible future failures of the device.
4. The new energy equipment management system based on artificial intelligence vision according to claim 3 is characterized in that: The training steps of the fault prediction model include: Step 1: Construct the network structure of the model, including determining the number of RBM layers and the number of nodes in each layer; Step 2: Collect training data, including the characteristic description of the current operating status of the equipment output by the equipment status monitoring model, and the corresponding actual equipment fault records as labels; Step 3: Initialize the parameters of the DBN model and set the hyperparameters for pre-training and fine-tuning; Step 4: First, pre-train the RBM layer through unsupervised learning, and then fine-tune the entire DBN model in a supervised manner to enable the model to accurately predict possible future equipment failures; Step 5: After training is completed, evaluate the model performance and select the model with the best performance as the final fault prediction model.
5. The new energy equipment management system based on artificial intelligence vision according to claim 4 is characterized in that: In the fault prediction model, the training of the restricted Boltzmann machine (RBM) layer adopts the contrastive divergence algorithm, which is performed through the following steps: S1: Initialize the weights and biases of RBM; S2: For a given input data, calculate the activation probability of the hidden layer and sample the hidden layer state; S3: Calculate the reconstruction probability of the visible layer based on the hidden layer state, and sample to obtain the reconstructed visible layer state; S4: Using the difference between the reconstructed visible layer state and the original input data, the weights and biases of the RBM are updated by gradient descent. S5: Repeat steps S2 to S4 until the RBM model parameters converge, completing the pre-training of the RBM layer.
6. The new energy equipment management system based on artificial intelligence vision according to claim 1 is characterized in that: The maintenance decision module uses the deep reinforcement learning (DDPG) algorithm to build an equipment maintenance decision model.
7. The new energy equipment management system based on artificial intelligence vision according to claim 6 is characterized in that: The method for constructing the equipment maintenance decision model includes: Define the state space, action space, and reward function of the DDPG algorithm. The state space is the future state prediction output by the device state prediction model, the action space is the possible maintenance instructions, and the reward function is set based on the difference between the predicted device state output by the device state monitoring module and the actual device operating state obtained through real-time sensor monitoring.
8. The new energy equipment management system based on artificial intelligence vision according to claim 1 is characterized in that: The instruction optimization module uses the random forest RF algorithm to build an optimization adjustment model.
9. The new energy equipment management system based on artificial intelligence vision according to claim 8 is characterized in that: The training steps of the optimization adjustment model include: Collect training data, including the difference between the predicted state output by the equipment condition monitoring module and the actual operating state obtained by real-time sensor monitoring. Also, collect expert-annotated optimal maintenance instructions for a given difference between the predicted and actual states as labels for the training data. Build a random forest RF algorithm model and set RF parameters, including the number of trees, maximum depth, and minimum number of sample splits; The random forest RF algorithm model is trained using training data and the bagging method is used to improve the model performance so that the model can accurately predict the optimal maintenance instructions under given input differences. After the training is completed, the optimized and adjusted model is obtained.
10. The new energy equipment management system based on artificial intelligence vision according to claim 1 is characterized in that: The method for determining whether the deviation between the actual operating state and the predicted state is significant is as follows: calculating the mean square error between the actual operating state and the predicted state, setting an error threshold, comparing the calculated mean square error with the set threshold, and if the mean square error exceeds the set threshold, it is determined to be a significant deviation.