Industrial equipment energy efficiency evaluation and maintenance decision-making method and system based on reinforcement learning

Abstract

The invention provides an industrial equipment energy efficiency evaluation and maintenance decision-making method and system based on reinforcement learning, and relates to the technical field of reinforcement learning, and the method comprises the steps: collecting equipment operation data and extracting features, constructing a maintenance time sequence decision-making model based on deep reinforcement learning, and achieving the model optimization through the combination of transfer learning; and generating a maintenance decision scheme according to the energy efficiency index analysis and the time sequence causal relationship, executing maintenance, and recording process data for online learning and updating of the model. The energy efficiency management level of industrial equipment is improved, the service life of the equipment is prolonged, and the operation and maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to deep learning technology, and in particular to a method and system for industrial equipment energy efficiency evaluation and maintenance decision-making based on reinforcement learning. Background Art

[0002] Traditional industrial equipment maintenance relies primarily on regular inspections and preventive maintenance strategies, typically performed at fixed intervals. These strategies are unable to dynamically adjust based on the equipment's actual operating status. Furthermore, equipment energy efficiency assessments are often independent of maintenance decisions, lacking an integrated perspective across the entire lifecycle, leading to wasted resources and inefficient energy consumption. Furthermore, existing maintenance decision-making systems, which are mostly based on simple rules and thresholds, struggle to handle the uncertainty and complexity of equipment operation and are unable to effectively address the dynamic nature of equipment performance over time.

[0003] Currently, the following major issues exist in industrial equipment maintenance decision-making and energy efficiency management: Traditional maintenance decision-making methods lack the ability to deeply mine data from the entire equipment lifecycle, failing to fully leverage the patterns of equipment performance degradation and energy efficiency changes implicit in massive operational data, resulting in insufficiently targeted maintenance decisions. Existing maintenance strategies are typically static and cannot be adjusted in real time based on environmental changes and equipment status. They lack adaptive learning capabilities and cannot achieve dynamically optimized maintenance decisions. A lack of mechanisms for knowledge sharing and experience transfer between different types of equipment requires that maintenance models for each device be trained from scratch, making it difficult to leverage maintenance experience from existing equipment to guide maintenance decisions for new equipment. This results in inefficient learning and wasted resources. Summary of the Invention

[0004] The embodiments of the present invention provide a method and system for industrial equipment energy efficiency evaluation and maintenance decision-making based on reinforcement learning, which can solve the problems in the existing technology.

[0005] A first aspect of an embodiment of the present invention provides an industrial equipment energy efficiency assessment and maintenance decision-making method based on reinforcement learning, comprising:

[0006] Collecting real-time operating data of industrial equipment as a first feature data set; performing feature extraction and principal component analysis on the first feature data set to obtain a second feature data set;

[0007] A maintenance time series decision model is constructed based on a deep reinforcement learning algorithm, time series features are extracted from the first feature data set and the second feature data set to generate state representations, and maintenance decision actions are determined according to a preset maintenance action space;

[0008] Using a transfer learning method to perform online optimization and update of the maintenance timing decision model, migrating the multi-source domain knowledge feature matrix to the target device domain through feature mapping, and updating the network parameters of the maintenance timing decision model;

[0009] Based on the action decision results output by the maintenance time sequence decision model, the equipment energy efficiency indicators are analyzed, and a maintenance decision plan is generated based on the time sequence causal relationship optimization solution;

[0010] Execute the maintenance decision plan, record the time, operation and resource data during the maintenance operation to form a maintenance process record, collect the first feature data set and the second feature data set after the maintenance is completed, use the maintenance process record and the feature data set after maintenance as training samples, input them into the maintenance timing decision model for online learning, and update the model parameters.

[0011] In an optional embodiment,

[0012] Performing feature extraction and principal component analysis on the first feature data set to obtain a second feature data set, including:

[0013] Performing data preprocessing on the first feature data set, calculating an anomaly score based on the distribution density of each data point in the first feature data set, and determining that a data point with a distribution density less than a preset threshold is an anomaly data point;

[0014] Select multiple data before and after the abnormal data point to build a sliding window, calculate the median within the sliding window to replace the abnormal data, and generate preprocessed feature data;

[0015] Performing correlation analysis on the preprocessed feature data, calculating a time series correlation coefficient matrix of the feature data in a time dimension, and calculating a spatial correlation coefficient matrix of the feature data in a spatial dimension;

[0016] Superimposing the temporal correlation coefficient matrix and the spatial correlation coefficient matrix according to preset weights to generate a multidimensional feature correlation matrix;

[0017] Calculating feature contribution weights based on the multidimensional feature correlation matrix, dividing the preprocessed feature data into key feature groups and secondary feature groups according to the feature contribution weights, extracting principal feature vectors from the key feature groups using a principal component analysis method, and extracting local feature vectors from the secondary feature groups using a local linear embedding method;

[0018] The main feature vector and the local feature vector are adaptively weighted fused to obtain a fused feature vector, and the fused feature vector is used as a second feature data set.

[0019] In an optional embodiment,

[0020] A maintenance time series decision model is constructed based on a deep reinforcement learning algorithm, time series features are extracted from the first feature data set and the second feature data set to generate a state representation, and a maintenance decision action is determined according to a preset maintenance action space, including:

[0021] The first feature data set and the second feature data set are segmented and aligned based on timestamps to construct an optimal time segmentation sequence, and the device state time series feature vector is extracted in each time series segment to generate a time series feature sequence;

[0022] A hierarchical maintenance action space is constructed based on the temporal feature sequence, a Markov decision process is used to model the correspondence between the temporal feature sequence and the maintenance action, and an action transfer constraint matrix is constructed;

[0023] A bidirectional long short-term memory network is used to encode the temporal feature sequence and then extract an initial temporal state representation, which is input into a multi-head attention layer to fuse state information of different time scales to obtain a temporal state representation;

[0024] A maintenance time series decision model is constructed based on the time series state representation; the time series state representation is input into a multi-layer time series convolutional network, multi-scale features are extracted through dynamic void ratio, and feature fusion is performed using bidirectional residual connection to obtain enhanced time series features;

[0025] The enhanced temporal features are respectively input into the value network and the advantage network to obtain the state value function and the action advantage function. The state value function and the action advantage function are combined to obtain the temporal value score of each maintenance action. The maintenance decision action is selected in combination with the action transfer constraint matrix.

[0026] In an optional embodiment,

[0027] The temporal state representation is input into a multi-layer temporal convolutional network, multi-scale features are extracted through dynamic hole rate, and bidirectional residual connection is used for feature fusion to obtain enhanced temporal features, including:

[0028] Performing multi-scale decomposition on the temporal state representation and extracting features of different frequency bands, performing importance evaluation on the features of each frequency band, dynamically adjusting feature weights based on the importance evaluation results, and reconstructing the weighted feature components through inverse transformation to obtain an enhanced temporal state representation;

[0029] The enhanced temporal state representation is input into a multi-layer temporal convolutional network. An exponentially increasing dilation rate sequence is set in each layer of the network. A dilated convolution operation is used to extract local temporal features in multiple receptive fields. A bidirectional residual connection path is constructed between adjacent convolutional layers. The local temporal features extracted from the previous layer are adaptively selected and fused with the features of the current layer through a gating mechanism to obtain multi-scale local temporal features.

[0030] Constructing a query matrix, a key matrix, and a value matrix for the multi-scale local temporal features, performing a scaled dot product operation on the query matrix and the key matrix to calculate an attention score, and performing masking on the attention score to obtain a feature weight;

[0031] The sinusoidal position encoding method is used to add temporal position information to the features, and the feature weights after adding the position information are multiplied by the value matrix to obtain weighted features;

[0032] The weighted features are input into a feedforward neural network to construct a jump connection with the original time series state representation to generate enhanced time series features.

[0033] In an optional embodiment,

[0034] The maintenance timing decision model is optimized and updated online using a transfer learning method. The multi-source domain knowledge feature matrix is transferred to the target device domain through feature mapping, and the network parameters of the maintenance timing decision model are updated, including:

[0035] Obtain the maintenance sequence decision model of the target device and the historical maintenance data of other devices, use the multi-view spectral clustering method to construct a device similarity matrix for the historical maintenance data, and calculate the maintenance pattern similarity matrix in combination with the dynamic time warping algorithm;

[0036] Adaptively fusing the device similarity matrix and the maintenance mode similarity matrix to generate a multi-source domain knowledge feature matrix;

[0037] A variational autoencoder is used to probabilistically encode the multi-source domain knowledge feature matrix. The encoded feature sequence is input into a bidirectional gated recurrent unit network to extract long-range temporal dependency features. A multi-head self-attention mechanism is combined to capture the dynamic correlation between features. The source domain conditional feature distribution is generated through residual connections and layer normalization.

[0038] Calculating a target domain feature distribution according to the maintenance sequence decision model of the target device, calculating a difference measure between the target domain feature distribution and the source domain condition feature distribution based on the bulldozer distance, performing a weighted combination of the difference measure and the multi-source domain knowledge feature matrix to generate a source domain weight allocation vector;

[0039] The source domain conditional feature distribution is input into the feature encoder, and feature mapping is performed using multi-scale convolution and dual attention mechanism combined with the source domain weight allocation vector. A weighted optimization objective is constructed to update the parameters of the maintenance timing decision model.

[0040] In an optional embodiment,

[0041] Inputting the source domain conditional feature distribution into a feature encoder, adopting multi-scale convolution and dual attention mechanism combined with the source domain weight allocation vector for feature mapping, and constructing a weighted optimization objective to update the parameters of the maintenance timing decision model, including:

[0042] Inputting the source domain conditional feature distribution into a multi-branch feature encoder to extract multi-granularity feature representation, adaptively aggregating the multi-granularity feature representation to generate a source domain feature representation;

[0043] Calculating spatial attention weights and channel attention weights for the source domain feature representation, performing tensor multiplication on the spatial attention weights and the channel attention weights to obtain a joint attention map, and generating multi-scale attention weights based on the long-range dependency of the joint attention map;

[0044] Inputting the source domain feature representation and the multi-scale attention weight cascade into a coarse-grained generative network to obtain an initial target domain feature, inputting the initial target domain feature and the source domain weight allocation vector into a fine-grained generative network for feature refinement to generate a final target domain feature;

[0045] Calculating the discriminant scores of the final target domain features at different feature levels, constructing a multi-scale adversarial loss based on the discriminant scores, and performing adaptive temperature adjustment on the source domain feature representation at the corresponding feature level to obtain a softened knowledge representation;

[0046] The distillation loss of the final target domain features and the softened knowledge representation at each level is calculated, the multi-scale adversarial loss and the distillation loss at each level are weightedly combined according to the source domain weight distribution vector, and the parameters of the feature transfer network are updated through an alternating optimization algorithm.

[0047] In an optional embodiment,

[0048] Based on the action decision results output by the maintenance time sequence decision model, the equipment energy efficiency indicators are analyzed, and a maintenance decision plan is generated based on the time sequence causal relationship optimization solution, including:

[0049] Extracting a device energy efficiency indicator from the second feature data set, performing multi-scale sampling on the device energy efficiency indicator to extract a time-frequency feature matrix, and constructing a multimodal probability distribution of the device energy efficiency indicator based on the time-frequency feature matrix;

[0050] Inputting the action decision result and the multimodal probability distribution into a multi-layer prediction unit to generate a probability prediction sequence of device energy efficiency changes;

[0051] Performing importance sampling on the probability prediction sequence to obtain multiple groups of energy efficiency prediction samples, estimating energy efficiency evaluation indicators at different confidence levels for the multiple groups of energy efficiency prediction samples; constructing a multi-criteria comprehensive evaluation model for the energy efficiency evaluation indicators to generate a maintenance benefit evaluation vector;

[0052] A graphical structure model is used to establish a temporal causal relationship between the action decision results and the probability prediction sequence. The intervention effect of each maintenance action is calculated through a counterfactual framework. The direct and indirect effects corresponding to the intervention effect are quantified based on structural equations. The maintenance benefit evaluation vector is adaptively fused with the direct and indirect effects to obtain an action importance score.

[0053] According to the action importance score, priority weights are assigned to the action decision results to obtain weighted decision results. Combined with the probability prediction sequence, the weighted decision results are used to construct an optimization objective considering timing constraints to solve the optimal maintenance timing and generate a maintenance decision plan.

[0054] A second aspect of an embodiment of the present invention provides an industrial equipment energy efficiency evaluation and maintenance decision-making system based on reinforcement learning, including:

[0055] The first unit is configured to collect real-time operating data of industrial equipment as a first feature data set; perform feature extraction and principal component analysis on the first feature data set to obtain a second feature data set;

[0056] The second unit is configured to construct a maintenance time series decision model based on a deep reinforcement learning algorithm, extract time series features from the first feature data set and the second feature data set and generate a state representation, and determine a maintenance decision action according to a preset maintenance action space;

[0057] The third unit is used to perform online optimization and update of the maintenance timing decision model using a transfer learning method, migrate the multi-source domain knowledge feature matrix to the target device domain through feature mapping, and update the network parameters of the maintenance timing decision model;

[0058] The fourth unit is used to analyze the equipment energy efficiency indicators based on the action decision results output by the maintenance time sequence decision model, and generate a maintenance decision plan based on the time sequence causal relationship optimization solution;

[0059] The fifth unit is used to execute the maintenance decision plan, record the time, operation and resource data during the maintenance operation to form a maintenance process record, collect the first feature data set and the second feature data set after the maintenance is completed, use the maintenance process record and the feature data set after maintenance as training samples, input them into the maintenance timing decision model for online learning, and update the model parameters.

[0060] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0061] processor;

[0062] a memory for storing processor-executable instructions;

[0063] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0064] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0065] By collecting real-time operating data and combining feature extraction and principal component analysis techniques, the present invention can effectively obtain the key features of the equipment's operating status, reduce data dimensions, improve data processing efficiency and feature expression capabilities, and provide more accurate data support for subsequent decision-making. The maintenance timing decision model constructed based on deep reinforcement learning can capture timing features and state change patterns from the equipment's historical operating data, and through interactive learning of optimal decision-making strategies with the environment, it realizes the intelligence and automation of maintenance decisions, significantly improving the timeliness and pertinence of equipment maintenance. The decision model is optimized and updated online using a transfer learning method, enabling the system to migrate existing equipment maintenance experience to new equipment, effectively solving data sparsity and cold start problems. At the same time, the model is continuously optimized through comprehensive analysis of maintenance process records and feature data, realizing closed-loop optimization of energy efficiency evaluation and maintenance decisions throughout the equipment's life cycle, thereby improving equipment operating efficiency and service life. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 Schematic diagram of a process for industrial equipment energy efficiency evaluation and maintenance decision-making based on reinforcement learning according to an embodiment of the present invention;

[0067] Figure 2 This is the attention weight distribution heat map of the present invention;

[0068] Figure 3 Generate a flow chart for maintenance decision-making scheme based on temporal causal relationships. DETAILED DESCRIPTION

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. 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 shall fall within the scope of protection of the present invention.

[0070] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0071] Figure 1 FIG. 1 is a flow chart of an industrial equipment energy efficiency evaluation and maintenance decision-making method based on reinforcement learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0072] Collecting real-time operating data of industrial equipment as a first feature data set; performing feature extraction and principal component analysis on the first feature data set to obtain a second feature data set;

[0073] A maintenance time series decision model is constructed based on a deep reinforcement learning algorithm, time series features are extracted from the first feature data set and the second feature data set to generate state representations, and maintenance decision actions are determined according to a preset maintenance action space;

[0074] Using a transfer learning method to perform online optimization and update of the maintenance timing decision model, migrating the multi-source domain knowledge feature matrix to the target device domain through feature mapping, and updating the network parameters of the maintenance timing decision model;

[0075] Based on the action decision results output by the maintenance time sequence decision model, the equipment energy efficiency indicators are analyzed, and a maintenance decision plan is generated based on the time sequence causal relationship optimization solution;

[0076] Execute the maintenance decision plan, record the time, operation and resource data during the maintenance operation to form a maintenance process record, collect the first feature data set and the second feature data set after the maintenance is completed, use the maintenance process record and the feature data set after maintenance as training samples, input them into the maintenance timing decision model for online learning, and update the model parameters.

[0077] In an optional embodiment, feature extraction and principal component analysis are performed on the first feature data set to obtain a second feature data set, including:

[0078] Performing data preprocessing on the first feature data set, calculating an anomaly score based on the distribution density of each data point in the first feature data set, and determining that a data point with a distribution density less than a preset threshold is an anomaly data point;

[0079] Select multiple data before and after the abnormal data point to build a sliding window, calculate the median within the sliding window to replace the abnormal data, and generate preprocessed feature data;

[0080] Performing correlation analysis on the preprocessed feature data, calculating a time series correlation coefficient matrix of the feature data in a time dimension, and calculating a spatial correlation coefficient matrix of the feature data in a spatial dimension;

[0081] Superimposing the temporal correlation coefficient matrix and the spatial correlation coefficient matrix according to preset weights to generate a multidimensional feature correlation matrix;

[0082] Calculating feature contribution weights based on the multidimensional feature correlation matrix, dividing the preprocessed feature data into key feature groups and secondary feature groups according to the feature contribution weights, extracting principal feature vectors from the key feature groups using a principal component analysis method, and extracting local feature vectors from the secondary feature groups using a local linear embedding method;

[0083] The main feature vector and the local feature vector are adaptively weighted fused to obtain a fused feature vector, and the fused feature vector is used as a second feature data set.

[0084] Exemplarily, data preprocessing is performed on the first feature data set. The system receives a first feature data set containing multidimensional features, such as an industrial equipment operating parameter data set, including time series data of ten dimensions such as temperature, pressure, vibration, and current, and each dimension contains 1000 sampling points. The system calculates the anomaly score based on the distribution density of each data point. The specific implementation method is: calculate the average distance between each data point and its neighboring data points, and convert the distance into a standardized score. In practical applications, the K nearest neighbor algorithm can be selected to calculate the distribution density, and the K value can be set to 10. When the distribution density of a data point is less than a preset threshold (for example, 0.05), the system marks it as an abnormal data point.

[0085] For detected anomalous data points, the system constructs a sliding window for correction. For example, if the 150th temperature point is detected as abnormal (85°C, while the normal range is 30-45°C), the system selects the five data points before and after this point to construct a sliding window of size 11. The median within the window (e.g., 38°C) is calculated and used to replace the anomalous value. After processing all anomalous points, a preprocessed feature dataset is generated.

[0086] Correlation analysis is performed on the preprocessed feature data. In the time dimension, the autocorrelation of each feature over time and the cross-correlations between different features are calculated to form a time series correlation coefficient matrix. For example, analyzing the time-delayed correlation between temperature and pressure may reveal that a temperature rise is followed by a pressure rise 10 minutes after the temperature rises, with a correlation coefficient of 0.78. In the spatial dimension, the system analyzes the correlation between different features at the same time point to form a spatial correlation coefficient matrix. For example, the spatial correlation coefficient between temperature and current is 0.65, indicating that the two parameters tend to change in a similar manner.

[0087] The system superimposes the two correlation coefficient matrices according to preset weights to generate a multidimensional feature correlation matrix. In a specific implementation, the temporal correlation coefficient weight can be set to 0.6, and the spatial correlation coefficient weight can be set to 0.4. For example, if the temporal correlation coefficient of temperature and pressure is 0.78 and the spatial correlation coefficient is 0.55, the superimposed correlation is 0.78 × 0.6 + 0.55 × 0.4 = 0.688.

[0088] Based on the multi-dimensional feature correlation matrix, the system calculates the contribution weight of each feature. The specific method is to calculate the sum of the correlations between each feature and all other features and perform normalization. For example, the sum of the correlations between the temperature feature and other features is 6.5, ranking second among all features, and the normalized contribution weight is 0.23. The system divides the features into key feature groups and secondary feature groups based on the contribution weights. Features with a contribution weight exceeding 0.15 (such as temperature, pressure, and current) are classified into the key feature group, and the remaining features are classified into the secondary feature group.

[0089] For key feature groups, the system uses principal component analysis to extract principal eigenvectors. Specifically, this involves calculating the covariance matrix of the key features, solving for their eigenvalues and eigenvectors, and selecting the eigenvectors with the largest contributions to form the principal eigenvectors. For example, from the three key features of temperature, pressure, and current, two principal components were extracted, explaining 75% and 20% of the variance in the original data, respectively.

[0090] For secondary feature groups, the system uses a local linear embedding method to extract local eigenvectors. First, the K nearest neighbors of each data point are determined (K can be set to 12). A local reconstruction weight matrix is then constructed. Finally, the characteristic equation is solved to obtain the local eigenvectors. For example, a local eigenvector extracted from secondary features such as vibration and noise can explain 95% of the local structural information of the original secondary feature.

[0091] The fusion weights are dynamically adjusted based on the proportion of data variance explained by the main eigenvector and the local eigenvectors. In this example, the weight of the main eigenvector is set to 0.85, and the weight of the local eigenvector is set to 0.15. The fusion process is achieved through weighted concatenation, resulting in a fused feature vector of length 3. The first two dimensions come from the main eigenvector, and the third dimension comes from the local eigenvector. The values of each dimension have been adjusted based on the weights. This fused feature vector serves as the second feature dataset for subsequent data analysis and processing.

[0092] The present invention detects outliers by calculating distribution density and uses sliding windows to correct them, effectively improving data quality; combines time series and spatial correlation analysis to construct a multidimensional feature matrix and accurately identify the correlation between features; uses principal component analysis and local linear embedding methods to process key features and secondary features respectively, and performs adaptive weight fusion, which retains key information and local structural features while reducing data dimensions, improves the accuracy and computational efficiency of feature extraction, and lays a solid foundation for subsequent data analysis.

[0093] In an optional embodiment, a maintenance time series decision model is constructed based on a deep reinforcement learning algorithm, time series features are extracted from the first feature data set and the second feature data set to generate a state representation, and a maintenance decision action is determined according to a preset maintenance action space, including:

[0094] The first feature data set and the second feature data set are segmented and aligned based on timestamps to construct an optimal time segmentation sequence, and the device state time series feature vector is extracted in each time series segment to generate a time series feature sequence;

[0095] A hierarchical maintenance action space is constructed based on the temporal feature sequence, a Markov decision process is used to model the correspondence between the temporal feature sequence and the maintenance action, and an action transfer constraint matrix is constructed;

[0096] A bidirectional long short-term memory network is used to encode the temporal feature sequence and then extract an initial temporal state representation, which is input into a multi-head attention layer to fuse state information of different time scales to obtain a temporal state representation;

[0097] A maintenance time series decision model is constructed based on the time series state representation; the time series state representation is input into a multi-layer time series convolutional network, multi-scale features are extracted through dynamic void ratio, and feature fusion is performed using bidirectional residual connection to obtain enhanced time series features;

[0098] The enhanced temporal features are respectively input into the value network and the advantage network to obtain the state value function and the action advantage function. The state value function and the action advantage function are combined to obtain the temporal value score of each maintenance action. The maintenance decision action is selected in combination with the action transfer constraint matrix.

[0099] Exemplarily, the first feature dataset and the second feature dataset are segmented and aligned based on timestamps to construct an optimal time segmentation sequence. The two datasets are sorted according to timestamps, and then a dynamic time warping algorithm is used to calculate the time alignment relationship between the two datasets. For example, for a certain device, the sensor data in the first feature dataset is collected at a frequency of once per minute, while the maintenance records in the second feature dataset are unevenly distributed. In this case, the dynamic time warping algorithm is used to divide the two datasets into several time segments, each of which is 6 hours long, to ensure that the data in each time segment is complete and has temporal continuity.

[0100] Within each time series segment, the device status time series feature vector is extracted to generate a time series feature sequence. For sensor data, statistical features (mean, variance, peak value, valley value, and rate of change) are calculated. For text-based maintenance records, the TF-IDF method is used to extract keywords and convert them into vector representations. For a wind turbine, for example, within a 6-hour time segment, statistical features of 10 types of sensor data, such as wind speed, power generation, and bearing temperature, are extracted to form a 50-dimensional feature vector. This is then combined with the 30-dimensional text feature vector extracted from the maintenance record to ultimately generate an 80-dimensional time series feature vector.

[0101] A hierarchical maintenance action space is constructed based on temporal feature sequences. Maintenance action types are summarized based on historical maintenance records, forming a hierarchical action space containing specific maintenance actions. A Markov decision process is used to model the correspondence between temporal feature sequences and maintenance actions, and an action transition constraint matrix is constructed. The transition probabilities between different maintenance actions are calculated based on the order of maintenance actions in historical data, forming a 7×7 action transition constraint matrix.

[0102] A bidirectional long short-term memory network is used to encode the temporal feature sequence and extract the initial temporal state representation. A bidirectional LSTM network with 128 hidden units is constructed. The input is an 80-dimensional temporal feature vector sequence, and the output is a 256-dimensional initial state representation vector. The initial temporal state representation is input into a multi-head attention layer to fuse state information at different time scales. The resulting temporal state representation is then constructed using a multi-head attention layer with 8 attention heads. Each attention head has an output dimension of 32, for a total output dimension of 256. This achieves attention and fusion of features at different time scales (e.g., short-term, medium-term, and long-term).

[0103] A maintenance time series decision model is constructed based on time series state representation. The time series state representation is input into a multi-layer time series convolutional network, and multi-scale features are extracted using dynamic dilation ratio. Bidirectional residual connections are used for feature fusion to obtain enhanced time series features.

[0104] The enhanced time series features are fed into the value network and advantage network, respectively, to generate the state value function and action advantage function. The value network consists of two fully connected layers, with 128 hidden nodes and an output dimension of 1. The advantage network also consists of two fully connected layers, with 128 hidden nodes and an output dimension of 7 (corresponding to the seven maintenance actions).

[0105] The state value function and the action advantage function are combined to obtain the temporal value score of each maintenance action. This score is then used in conjunction with the action transition constraint matrix to select the maintenance decision action. For each maintenance action, its Q-value (total value score) is calculated as the sum of the state value plus the action's advantage value minus the average of all action advantages. This is then adjusted based on the action transition constraint matrix to reduce the probability of unreasonable action sequences. Ultimately, the action with the highest Q-value is selected as the final maintenance decision.

[0106] The present invention uses a dynamic time warping algorithm to achieve precise alignment and segmentation of data sets and effectively extract temporal features. It adopts a bidirectional LSTM and multi-head attention mechanism to realize multi-scale state representation, enhancing the model's ability to capture temporal features. It combines temporal convolution with dynamic void rate and bidirectional residual connection to improve the receptive field of feature extraction, accurately evaluates the value of maintenance actions through value networks and advantage networks, and uses action transfer constraint matrices to ensure the rationality of decision sequences, significantly improving the accuracy and reliability of maintenance decisions.

[0107] In an optional embodiment, the temporal state representation is input into a multi-layer temporal convolutional network, multi-scale features are extracted by dynamic hole rate, and feature fusion is performed using bidirectional residual connection to obtain enhanced temporal features, including:

[0108] Performing multi-scale decomposition on the temporal state representation and extracting features of different frequency bands, performing importance evaluation on the features of each frequency band, dynamically adjusting feature weights based on the importance evaluation results, and reconstructing the weighted feature components through inverse transformation to obtain an enhanced temporal state representation;

[0109] The enhanced temporal state representation is input into a multi-layer temporal convolutional network. An exponentially increasing dilation rate sequence is set in each layer of the network. A dilated convolution operation is used to extract local temporal features in multiple receptive fields. A bidirectional residual connection path is constructed between adjacent convolutional layers. The local temporal features extracted from the previous layer are adaptively selected and fused with the features of the current layer through a gating mechanism to obtain multi-scale local temporal features.

[0110] Constructing a query matrix, a key matrix, and a value matrix for the multi-scale local temporal features, performing a scaled dot product operation on the query matrix and the key matrix to calculate an attention score, and performing masking on the attention score to obtain a feature weight;

[0111] The sinusoidal position encoding method is used to add temporal position information to the features, and the feature weights after adding the position information are multiplied by the value matrix to obtain weighted features;

[0112] The weighted features are input into a feedforward neural network to construct a jump connection with the original time series state representation to generate enhanced time series features.

[0113] Exemplarily, the time series state representation is subjected to multi-scale decomposition and features of different frequency bands are extracted. Specifically, a wavelet transform is used to perform multi-scale decomposition on the input time series state representation, decomposing the original time series signal into sub-signals of different frequency bands. For example, a time series state representation with a length of 1024 can be decomposed into four frequency bands, corresponding to sub-signals with frequency ranges of 0-16 Hz, 16-32 Hz, 32-64 Hz, and 64-128 Hz, respectively.

[0114] The importance of each frequency band feature is assessed. By calculating the energy distribution of each frequency band's sub-signal, the frequency components that contribute most to the target task are identified. For example, the variance or entropy of each frequency band's sub-signal is calculated as an importance indicator. A larger variance or higher entropy indicates that the frequency band contains more effective information. In practical applications, assume that the importance scores of the four frequency bands are 0.8, 0.5, 0.3, and 0.1, respectively. A soft attention mechanism is used to assign different weight coefficients to each frequency band feature, with higher weights given to frequency bands with higher importance. For example, based on the above scores, the weights of the four frequency bands can be set to 0.47, 0.29, 0.18, and 0.06, respectively, to ensure that the sum of the weights is 1. The weighted feature components are reconstructed using an inverse wavelet transform to obtain an enhanced temporal state representation.

[0115] The enhanced temporal state representation is input into a multi-layer temporal convolutional network. In this embodiment, the network consists of 8 layers of temporal convolutional layers, each layer contains 64 convolution kernels, and the convolution kernel size is 3.

[0116] An exponentially increasing sequence of dilation rates is set in each network layer. The dilation rate refers to the sampling interval during the convolution operation; a larger dilation rate results in a larger receptive field. Specifically, from the first to the eighth layers, the dilation rates are set to 1, 2, 4, 8, 16, 32, 64, and 128, respectively, enabling the network to capture multi-scale features from local details to global patterns. For example, for the third convolution layer, the dilation rate is 4, meaning that when performing the convolution operation, a sample point is selected for every three elements, effectively expanding the receptive field.

[0117] Dilated convolution is used to extract local temporal features with multiple receptive fields. For input feature X, convolution is performed using a dilated convolution kernel W with a dilation ratio d. The output feature Y is calculated by multiplying the input element at the corresponding position by the convolution kernel weight. For example, for the second convolution layer with a dilation ratio of 2 and a convolution kernel size of 3, the actual receptive field covers 5 elements in the input sequence (the original element plus 2 left and right spacing elements).

[0118] A bidirectional residual connection pathway is constructed between adjacent convolutional layers. The forward path transmits the output features of layer i to layer i+1, and the backward path feeds the features of layer i+1 back to layer i after undergoing a 1×1 convolution. For example, the output feature F3 of layer 3 is directly transmitted to layer 4 via the forward path, while the feature F4 of layer 4 is fused with F3 via the backward path after undergoing a 1×1 convolution.

[0119] The gating mechanism enables adaptive feature selection and fusion. A sigmoid activation function is used to generate the gating signal G, which controls the fusion ratio of features at different levels. In practice, the value of G ranges from 0 to 1. When the quality of a feature Fi at a certain location is higher, the G value at that location is closer to 1, and vice versa.

[0120] Construct query, key, and value matrices for multi-scale local temporal features. Assuming the feature dimension is 64, the feature F is projected into query matrix Q, key matrix K, and value matrix V through three different linear transformations, all of which have dimensions [sequence length, 64].

[0121] The attention score is calculated by performing a scaled dot product operation on the query matrix and the key matrix. Specifically, the transpose of QK is calculated and divided by the scaling factor 8 (the square root of 64) to obtain the attention score matrix A. For example, for an input with a sequence length of 100, the dimension of the attention matrix A is [100, 100], and A[i, j] represents the attention weight of position i to position j.

[0122] The attention scores are masked to obtain feature weights. To ensure temporal causality, an upper triangular mask matrix is used to mask future information after the current moment. This is done by setting the attention scores at the corresponding positions to negative infinity. A softmax function is then applied to the masked attention scores to obtain the weight matrix W. For example, for position i, only W[i, j] with j ≤ i has a nonzero value, and the sum of all weights is 1.

[0123] Sine positional encoding is used to add temporal position information to features. For each position and feature dimension in the sequence, the positional encoding value is calculated using the sine and cosine functions, respectively, taking into account the proportional relationship between position and dimension. The calculated positional encoding value is added to the original feature, allowing the model to effectively identify and distinguish feature information at different time steps. Weighted features are obtained by performing tensor multiplication of the feature weights after adding positional information with the value matrix. Specifically, the matrix product of W and V is calculated to obtain the contextual feature C, with the dimension [sequence length, 64]. For example, the contextual feature C[i] at position i is the weighted sum of all features V[j] at positions j ≤ i, with the weight W[i, j].

[0124] The weighted features are input into a feedforward neural network and skip connections are constructed with the original temporal state representation to generate enhanced temporal features. The feedforward network consists of two fully connected layers, with a ReLU activation function used in the middle to expand the feature dimension from 64 to 128 and then map it back to 64. Finally, the network output is added to the original features using a residual connection to obtain the final enhanced temporal features. In specific implementation, if the original feature dimension is 64, the parameter matrices of the two fully connected layers of the feedforward network are [64, 128] and [128, 64] respectively. When processing features with a sequence length of 100, the final output enhanced feature dimension is [100, 64].

[0125] Figure 2 This is the attention weight distribution heat map of the present invention, which shows the weight distribution characteristics of the attention mechanism of the present invention in time series data processing: the attention distribution is highly focused, the weight values are mainly concentrated on specific time steps, and the attention weights of most key positions reach high values between 0.84 and 0.93; the weights on the diagonal are generally high, indicating that the features of the current time step have a strong attention to themselves, such as the weight of 0.93 at position t2 and the weight of 0.89 at position t3. The attention distribution presents a block-like clustering structure rather than a uniform dispersion, indicating that the model can adaptively identify key time points in time series data and their interrelationships. The low-value area (light-colored part) in the weight matrix indicates that the model has successfully filtered out the interference information of irrelevant time steps, thereby improving the accuracy of feature extraction. Overall, the attention mechanism of the present invention achieves accurate capture of key dependencies in time series data, and has stronger feature selection capabilities and pattern recognition efficiency than traditional methods.

[0126] The present invention uses wavelet transform to achieve multi-scale decomposition and frequency band importance assessment, effectively extracting key temporal features; adopts a multi-layer convolutional network with exponentially increasing void ratio and bidirectional residual connection to expand the receptive field range and enhance the feature extraction capability; combines the query-key-value attention mechanism and sinusoidal position encoding to accurately capture long-range temporal dependencies, and realizes feature adaptive selection and fusion through a gating mechanism, significantly improving the expressiveness and robustness of temporal features, and providing high-quality feature representation for subsequent maintenance decisions.

[0127] In an optional embodiment, the maintenance timing decision model is optimized and updated online using a transfer learning method, and the multi-source domain knowledge feature matrix is transferred to the target device domain through feature mapping to update the network parameters of the maintenance timing decision model, including:

[0128] Obtain the maintenance sequence decision model of the target device and the historical maintenance data of other devices, use the multi-view spectral clustering method to construct a device similarity matrix for the historical maintenance data, and calculate the maintenance pattern similarity matrix in combination with the dynamic time warping algorithm;

[0129] Adaptively fusing the device similarity matrix and the maintenance mode similarity matrix to generate a multi-source domain knowledge feature matrix;

[0130] A variational autoencoder is used to probabilistically encode the multi-source domain knowledge feature matrix. The encoded feature sequence is input into a bidirectional gated recurrent unit network to extract long-range temporal dependency features. A multi-head self-attention mechanism is combined to capture the dynamic correlation between features. The source domain conditional feature distribution is generated through residual connections and layer normalization.

[0131] Calculating a target domain feature distribution according to the maintenance sequence decision model of the target device, calculating a difference measure between the target domain feature distribution and the source domain condition feature distribution based on the bulldozer distance, performing a weighted combination of the difference measure and the multi-source domain knowledge feature matrix to generate a source domain weight allocation vector;

[0132] The source domain conditional feature distribution is input into the feature encoder, and feature mapping is performed using multi-scale convolution and dual attention mechanism combined with the source domain weight allocation vector. A weighted optimization objective is constructed to update the parameters of the maintenance timing decision model.

[0133] Exemplarily, the maintenance sequence decision model of the target device and the historical maintenance data of other devices are obtained. The multi-perspective spectral clustering method is used to construct the device similarity matrix for the acquired historical maintenance data. Specifically, feature extraction is performed on the device data from three perspectives: device parameters, fault types, and maintenance cycles, and a similarity sub-matrix is generated for each perspective. From the device parameter perspective, the statistical characteristics (mean, variance, peak, etc.) of the device operating parameters are extracted to construct the parameter similarity sub-matrix; from the fault type perspective, the proportion of the same fault type between different devices is counted as the fault similarity sub-matrix; from the maintenance cycle perspective, the distribution difference of the equipment maintenance interval time is calculated as the maintenance cycle similarity sub-matrix. The three sub-matrices are weighted and fused to obtain a 10×10 device similarity matrix S.

[0134] At the same time, the dynamic time warping algorithm is combined to calculate the maintenance pattern similarity matrix. The maintenance activity sequence of each device is regarded as a time series. The dynamic time warping algorithm is applied to the maintenance sequences of any two devices to calculate the time series distance. Taking the first and second devices as examples, the maintenance activity time series M1 = [t 11 ,t 12 ,...,t1n] and M2=[t 21 ,t 22 ,...,t2m], the optimal alignment path is solved by dynamic programming, and the temporal distance D(M1,M2) = 218.5 hours is obtained. After normalizing the temporal distance calculation results of all device pairs, a 10×10 maintenance mode similarity matrix P is constructed.

[0135] The device similarity matrix S and the maintenance mode similarity matrix P are adaptively fused to generate the multi-source domain knowledge feature matrix K. The fusion process uses a dynamic weighting mechanism to automatically adjust the weights of the two matrices based on the current state of the target device. For example, when the target device has just completed maintenance, the weight of the maintenance mode similarity matrix is increased (α = 0.7); when the device is in normal operation, the weight of the device similarity matrix is increased (α = 0.6). The specific fusion formula is the weighted sum of the two matrices. For example, when α = 0.6, K = 0.6 × S + 0.4 × P, resulting in a 10 × 10 multi-source domain knowledge feature matrix.

[0136] A variational autoencoder is used to probabilistically encode the multi-source domain knowledge feature matrix K. The variational autoencoder consists of an encoder and a decoder. The encoder, consisting of a three-layer fully connected network (10×10 input layer, 128 hidden layers, and a 64-dimensional latent space), maps K to the mean vector μ and variance vector σ of the latent space. Reparameterization techniques are then used to generate the latent variable z, with a dimension of 64. The decoder, also consisting of a three-layer fully connected network, reconstructs z into the original feature matrix.

[0137] The encoded feature sequence z is fed into a bidirectional gated recurrent unit network to extract long-range temporal dependency features. The network consists of a two-layer bidirectional GRU structure with 128 hidden units. The input is the 64-dimensional latent variable z generated by the variational autoencoder. Temporal information is captured through forward and backward propagation, and the output is a 256-dimensional feature vector h.

[0138] Combining a multi-head self-attention mechanism to capture dynamic correlations between features, an 8-head attention mechanism is employed. The feature vector h from the previous step is split into eight subspaces, with attention weights calculated independently for each subspace. For example, the first attention head maps the input into three vectors: query, key, and value. Similarity is calculated using dot products, and the softmax function is applied to obtain normalized weights. Weighted aggregation is then performed to generate enhanced features. The output features of the eight heads are concatenated and linearly transformed to produce a 256-dimensional attention feature a. The source domain conditional feature distribution Fs is generated through residual connections and layer normalization. h and a are added to form a residual connection, and layer normalization is performed to obtain the source domain conditional feature distribution Fs, maintaining a dimensionality of 256. The target domain feature distribution Ft is calculated based on the maintenance time series decision model for the target device. The same feature extraction process is applied to the target device's historical data to obtain the target domain feature distribution Ft, also a 256-dimensional vector.

[0139] The difference metric d between the target domain feature distribution Ft and the source domain conditional feature distribution Fs is calculated based on the bulldozer distance. This measure, which measures the distance between two distributions by solving an optimal transfer problem, yields a result of d = 0.385. The difference metric d is weightedly combined with the multi-source domain knowledge feature matrix K to generate the source domain weight assignment vector w = [0.12, 0.15, 0.06, 0.09, 0.18, 0.11, 0.08, 0.14, 0.07], representing the knowledge transfer weights of the nine source domain devices.

[0140] The source domain conditional feature distribution Fs is fed into the feature encoder, and feature mapping is performed using a multi-scale convolution and dual attention mechanism combined with the source domain weight distribution vector w. A weighted optimization objective is constructed to update the parameters of the maintenance temporal decision model.

[0141] The present invention uses multi-view spectral clustering and dynamic time warping algorithms to accurately calculate the similarity between devices and construct a multi-source domain knowledge feature matrix; uses variational autoencoders to implement feature probability encoding, and combines bidirectional gated recurrent units and multi-head attention mechanisms to capture long-range dependencies; measures inter-domain differences based on the bulldozer distance and generates weight allocation vectors, and implements feature transfer mapping through multi-scale convolution and dual attention mechanisms, significantly improving the model's generalization ability and maintenance decision accuracy on target devices.

[0142] In an optional embodiment, the source domain conditional feature distribution is input into a feature encoder, multi-scale convolution and dual attention mechanism are used in combination with the source domain weight distribution vector for feature mapping, and a weighted optimization objective is constructed to update the parameters of the maintenance time series decision model, including:

[0143] Inputting the source domain conditional feature distribution into a multi-branch feature encoder to extract multi-granularity feature representation, adaptively aggregating the multi-granularity feature representation to generate a source domain feature representation;

[0144] Calculating spatial attention weights and channel attention weights for the source domain feature representation, performing tensor multiplication on the spatial attention weights and the channel attention weights to obtain a joint attention map, and generating multi-scale attention weights based on the long-range dependency of the joint attention map;

[0145] Inputting the source domain feature representation and the multi-scale attention weight cascade into a coarse-grained generative network to obtain an initial target domain feature, inputting the initial target domain feature and the source domain weight allocation vector into a fine-grained generative network for feature refinement to generate a final target domain feature;

[0146] Calculating the discriminant scores of the final target domain features at different feature levels, constructing a multi-scale adversarial loss based on the discriminant scores, and performing adaptive temperature adjustment on the source domain feature representation at the corresponding feature level to obtain a softened knowledge representation;

[0147] The distillation loss of the final target domain features and the softened knowledge representation at each level is calculated, the multi-scale adversarial loss and the distillation loss at each level are weightedly combined according to the source domain weight distribution vector, and the parameters of the feature transfer network are updated through an alternating optimization algorithm.

[0148] For example, the source domain conditional feature distribution is input into a multi-branch feature encoder to extract multi-granularity feature representations. In actual implementation, the multi-branch feature encoder contains three parallel branches: (1) the first branch uses three convolutional layers with a convolution kernel size of 3×3 and a stride of 1, and each layer is followed by batch normalization and ReLU activation function; (2) the second branch uses two convolutional layers with a convolution kernel size of 5×5 and a stride of 1; (3) the third branch uses one convolutional layer with a convolution kernel size of 7×7 and a stride of 1. These three branches extract feature information of different granularities from the source domain input features. For example, for image features with an input dimension of 64×64×3, the three branches output feature maps of 32×32×64, 32×32×128, and 32×32×256, respectively. The multi-granularity feature representations are adaptively aggregated to generate the source domain feature representation. During implementation, a channel attention mechanism is used to weight the importance of each branch feature. First, global average pooling is performed on the features of each branch to obtain a channel-dimensional feature vector. Then, weight coefficients for each channel are calculated through two fully connected layers. Finally, the weighted features are concatenated and fused through a 1×1 convolution to generate the source domain feature representation. In this example, the fused source domain feature representation has a dimension of 32×32×256.

[0149] Spatial attention weights and channel attention weights are calculated separately for the source domain feature representation. The spatial attention weights are calculated by performing the maximum and average calculations on the source domain feature representation in the channel dimension to obtain two spatial feature maps. These two feature maps are concatenated and passed through a 7×7 convolutional layer, followed by a sigmoid function to obtain a spatial attention map of size 32×32×1. The channel attention weights are calculated by performing global average pooling and max pooling on the source domain feature representation in the spatial dimension to obtain two channel feature vectors. These two vectors are passed through a fully connected layer with shared weights, and finally fused and passed through a sigmoid function to obtain a channel attention vector of size 1×1×256. The spatial attention weights are tensor-multiplied with the channel attention weights to obtain a joint attention map. Specifically, the channel attention vector is first spatially replicated and expanded to 32×32×256, and then element-wise multiplied with the spatial attention map (expanded to 32×32×256) to obtain the initial joint attention map. Subsequently, long-range dependencies are captured through non-local operations to generate multi-scale attention weights. In the implementation process, the initial joint attention map is passed through the convolution layers of three convolution kernels: 1×1, 3×3, and 5×5 to obtain attention features of three scales respectively. These features are then concatenated and passed through 1×1 convolution to obtain the final multi-scale attention weights with a dimension of 32×32×256.

[0150] The source domain feature representation and the multi-scale attention weight are cascaded and input into the coarse-grained generative network. The source domain feature representation and the multi-scale attention weight are first connected in the channel dimension to obtain a cascade feature with a dimension of 32×32×512, and then the initial target domain feature is generated through a coarse-grained generative network consisting of a residual block and two transposed convolutional layers, with a dimension of 64×64×128. The initial target domain feature and the source domain weight allocation vector are input into the fine-grained generative network for feature refinement. In an embodiment, the source domain weight allocation vector has a dimension of 128, which is first converted into a spatial feature map through a fully connected layer and a reshaping operation, connected with the initial target domain feature, and then the final target domain feature is generated through a fine-grained generative network consisting of two residual blocks and a convolutional layer, with a dimension of 64×64×64.

[0151] The final target domain features are used to calculate the discriminant scores at different feature levels. A multi-scale discriminator network is used, consisting of three sub-discriminators of different scales. These discriminate features at the original size, 1 / 2 size, and 1 / 4 size, and output the corresponding discriminant scores. For example, for a 64×64×64 target domain feature, the three discriminators output discriminant score maps of 64×64, 32×32, and 16×16, respectively.

[0152] A multi-scale adversarial loss is constructed based on the discriminant score. For each scale of the discriminant score, the mean squared error between it and the true label is calculated, and then weighted according to the corresponding weight value in the source domain weight allocation vector to obtain the weighted adversarial loss. At the same time, the source domain feature representation is adaptively temperature-adjusted at the corresponding feature level to obtain a softened knowledge representation. In the specific implementation, the statistics on the channel dimension are first calculated for each layer of features, and then the temperature parameter T (usually between 1 and 10) is determined based on these statistics. The original feature is divided by T and then passed through the Softmax function to obtain the softened knowledge representation.

[0153] The distillation loss of the final target domain features and the softened knowledge representation is calculated at each level. A weighted combination of the multi-scale adversarial loss and the distillation loss at each level is performed based on the source domain weight distribution vector. In practice, the weight coefficients are typically set to 0.5 and 0.5, and the parameters of the feature transfer network are updated using an alternating optimization algorithm. The optimization process uses the Adam optimizer with a learning rate of 0.0002, a batch size of 16, and 50,000 training iterations. This approach effectively transfers source domain knowledge to the target domain, enhancing the model's performance on the target domain.

[0154] The present invention extracts multi-granularity feature representations through a multi-branch feature encoder and generates a joint attention map by combining spatial and channel attention mechanisms. It adopts a coarse and fine two-level generative network and source domain weight distribution to achieve precise feature transfer. It constructs a weighted optimization objective based on multi-scale adversarial loss and hierarchical knowledge distillation loss, and dynamically updates the feature transfer network through an alternating optimization algorithm, which significantly improves the model's feature expression ability and transfer learning effect in the target domain, and provides effective guarantees for the online optimization of maintenance decision models.

[0155] In an optional embodiment, based on the action decision results output by the maintenance time sequence decision model, the equipment energy efficiency index is analyzed, and a maintenance decision plan is generated based on the time sequence causal relationship optimization solution, including:

[0156] Extracting a device energy efficiency indicator from the second feature data set, performing multi-scale sampling on the device energy efficiency indicator to extract a time-frequency feature matrix, and constructing a multimodal probability distribution of the device energy efficiency indicator based on the time-frequency feature matrix;

[0157] Inputting the action decision result and the multimodal probability distribution into a multi-layer prediction unit to generate a probability prediction sequence of device energy efficiency changes;

[0158] Performing importance sampling on the probability prediction sequence to obtain multiple groups of energy efficiency prediction samples, estimating energy efficiency evaluation indicators at different confidence levels for the multiple groups of energy efficiency prediction samples; constructing a multi-criteria comprehensive evaluation model for the energy efficiency evaluation indicators to generate a maintenance benefit evaluation vector;

[0159] A graphical structure model is used to establish a temporal causal relationship between the action decision results and the probability prediction sequence. The intervention effect of each maintenance action is calculated through a counterfactual framework. The direct and indirect effects corresponding to the intervention effect are quantified based on structural equations. The maintenance benefit evaluation vector is adaptively fused with the direct and indirect effects to obtain an action importance score.

[0160] According to the action importance score, priority weights are assigned to the action decision results to obtain weighted decision results. Combined with the probability prediction sequence, the weighted decision results are used to construct an optimization objective considering timing constraints to solve the optimal maintenance timing and generate a maintenance decision plan.

[0161] For example, combined Figure 3The flow chart for generating a maintenance decision plan based on temporal causal relationships is explained as follows: the equipment energy efficiency index is extracted from the second feature data set, the equipment energy efficiency index is sampled at multiple scales to extract a time-frequency feature matrix, and a multimodal probability distribution of the equipment energy efficiency index is constructed based on the time-frequency feature matrix. Specifically, energy efficiency indicators including equipment energy consumption rate, conversion efficiency, and power factor are extracted from the second feature data set. Sliding window sampling is performed for each energy efficiency index using different time windows (such as 1 hour, 12 hours, 24 hours, and 7 days), and the time-frequency features of the data in each window are extracted using wavelet transform to form a feature matrix. A Gaussian mixture model is used to construct a multimodal probability distribution for the time-frequency feature matrix.

[0162] The action decision results and the multimodal probability distribution are input into a multi-layer prediction unit to generate a probability prediction sequence for equipment energy efficiency changes. Specifically, the maintenance action decision results (such as replacing a filter or cleaning a heat exchanger) output by the time series decision model are encoded as a vector, concatenated with the multimodal probability distribution parameters of the energy efficiency indicator, and input into a multi-layer prediction unit composed of a long short-term memory (LSTM) network. This prediction unit contains three layers of LSTM, each with 128 neurons, and outputs a probability distribution of equipment energy efficiency changes over the next 30 days.

[0163] Importance sampling is performed on the probabilistic prediction sequence to obtain multiple sets of energy efficiency prediction samples. Energy efficiency evaluation indicators at different confidence levels are estimated for these multiple sets of energy efficiency prediction samples. Using the Monte Carlo importance sampling method, 1,000 sets of energy efficiency change samples are extracted from the probabilistic prediction sequence. For each set of samples, energy efficiency evaluation indicators, including average energy saving rate, peak power reduction ratio, and system stability index, are estimated at 90%, 95%, and 99% confidence levels. A multi-criteria comprehensive evaluation model is constructed for these energy efficiency evaluation indicators to generate a maintenance benefit evaluation vector. The analytic hierarchy process is used to determine the weights of each energy efficiency indicator to construct the comprehensive evaluation model. For example, for a certain device, the weight of energy saving rate is 0.5, the weight of peak power reduction ratio is 0.3, and the weight of system stability is 0.2. The calculated maintenance benefit evaluation vector is [0.72, 0.65, 0.81], representing the short-term, medium-term, and long-term maintenance benefit scores, respectively.

[0164] A graphical model is used to establish a temporal causal relationship between the action decision results and the probability prediction sequence. The intervention effect of each maintenance action is calculated using a counterfactual framework. The direct and indirect impacts corresponding to the intervention effect are quantified based on structural equations. A directed acyclic graph is used to represent the causal relationship between maintenance actions and energy efficiency changes, with nodes representing maintenance actions or energy efficiency indicators and edges representing causal impacts. Using the counterfactual framework, a "do vs. do not do" comparison experiment is performed for each maintenance action to calculate its intervention effect. The maintenance benefit evaluation vector is adaptively fused with the direct and indirect impacts to obtain an action importance score. An adaptive weight fusion algorithm is used to dynamically adjust the weights of the benefit evaluation vector, direct impact, and indirect impact based on the current system state to calculate the importance score of each maintenance action.

[0165] A priority weight is assigned to each maintenance action based on the action importance score, forming a weighted decision result. An optimization objective function is constructed, taking into account maintenance resource constraints (personnel, spare parts, time windows) and timing dependency constraints (some maintenance must be performed in a specific order). A genetic algorithm is used to solve the optimization problem and obtain the optimal maintenance sequence. For example, for an industrial refrigeration system, the generated maintenance decision plan is: perform chiller inspection on the first day (importance 0.95), perform cooling tower cleaning on the third day (importance 0.87), perform water pump maintenance on the fifth day (importance 0.82), perform pipeline system inspection on the seventh day (importance 0.76), and perform terminal equipment debugging on the tenth day (importance 0.68).

[0166] The present invention extracts the time-frequency feature matrix through multi-scale sampling to construct the multimodal probability distribution of energy efficiency indicators; uses multi-layer prediction units to generate probability prediction sequences, and combines importance sampling and multi-criteria evaluation to achieve maintenance benefit quantification; adopts a graphical structure model to establish temporal causal relationships, analyzes the intervention effects of maintenance actions through a counterfactual framework, and calculates the importance of actions based on structural equations and adaptive fusion, which significantly improves the scientific nature and reliability of maintenance decision-making plans.

[0167] A second aspect of an embodiment of the present invention provides an industrial equipment energy efficiency evaluation and maintenance decision-making system based on reinforcement learning, including:

[0168] The first unit is configured to collect real-time operating data of industrial equipment as a first feature data set; perform feature extraction and principal component analysis on the first feature data set to obtain a second feature data set;

[0169] The second unit is configured to construct a maintenance time series decision model based on a deep reinforcement learning algorithm, extract time series features from the first feature data set and the second feature data set and generate a state representation, and determine a maintenance decision action according to a preset maintenance action space;

[0170] The third unit is used to perform online optimization and update of the maintenance timing decision model using a transfer learning method, migrate the multi-source domain knowledge feature matrix to the target device domain through feature mapping, and update the network parameters of the maintenance timing decision model;

[0171] The fourth unit is used to analyze the equipment energy efficiency indicators based on the action decision results output by the maintenance time sequence decision model, and generate a maintenance decision plan based on the time sequence causal relationship optimization solution;

[0172] The fifth unit is used to execute the maintenance decision plan, record the time, operation and resource data during the maintenance operation to form a maintenance process record, collect the first feature data set and the second feature data set after the maintenance is completed, use the maintenance process record and the feature data set after maintenance as training samples, input them into the maintenance timing decision model for online learning, and update the model parameters.

[0173] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0174] processor;

[0175] a memory for storing processor-executable instructions;

[0176] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0177] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0178] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An industrial equipment energy efficiency evaluation and maintenance decision-making method based on reinforcement learning, characterized by: include: Collecting real-time operating data of industrial equipment as the first feature data set; Performing feature extraction and principal component analysis on the first feature data set to obtain a second feature data set; A maintenance time series decision model is constructed based on a deep reinforcement learning algorithm, time series features are extracted from the first feature data set and the second feature data set to generate state representations, and maintenance decision actions are determined according to a preset maintenance action space; Using a transfer learning method to perform online optimization and update of the maintenance timing decision model, migrating the multi-source domain knowledge feature matrix to the target device domain through feature mapping, and updating the network parameters of the maintenance timing decision model; Based on the action decision results output by the maintenance time sequence decision model, the equipment energy efficiency indicators are analyzed, and a maintenance decision plan is generated based on the time sequence causal relationship optimization solution; Execute the maintenance decision plan, record the time, operation and resource data during the maintenance operation to form a maintenance process record, collect the first feature data set and the second feature data set after the maintenance is completed, use the maintenance process record and the feature data set after maintenance as training samples, input them into the maintenance timing decision model for online learning, and update the model parameters.

2. The method according to claim 1, characterized in that Performing feature extraction and principal component analysis on the first feature data set to obtain a second feature data set, including: Performing data preprocessing on the first feature data set, calculating an anomaly score based on the distribution density of each data point in the first feature data set, and determining that a data point with a distribution density less than a preset threshold is an anomaly data point; Select multiple data before and after the abnormal data point to build a sliding window, calculate the median within the sliding window to replace the abnormal data, and generate preprocessed feature data; Performing correlation analysis on the preprocessed feature data, calculating a time series correlation coefficient matrix of the feature data in a time dimension, and calculating a spatial correlation coefficient matrix of the feature data in a spatial dimension; Superimposing the temporal correlation coefficient matrix and the spatial correlation coefficient matrix according to preset weights to generate a multidimensional feature correlation matrix; Calculating feature contribution weights based on the multidimensional feature correlation matrix, dividing the preprocessed feature data into key feature groups and secondary feature groups according to the feature contribution weights, extracting principal feature vectors from the key feature groups using a principal component analysis method, and extracting local feature vectors from the secondary feature groups using a local linear embedding method; The main feature vector and the local feature vector are adaptively weighted fused to obtain a fused feature vector, and the fused feature vector is used as a second feature data set.

3. The method according to claim 1, characterized in that A maintenance time series decision model is constructed based on a deep reinforcement learning algorithm, time series features are extracted from the first feature data set and the second feature data set to generate a state representation, and a maintenance decision action is determined according to a preset maintenance action space, including: The first feature data set and the second feature data set are segmented and aligned based on timestamps to construct an optimal time segmentation sequence, and the device state time series feature vector is extracted in each time series segment to generate a time series feature sequence; A hierarchical maintenance action space is constructed based on the temporal feature sequence, a Markov decision process is used to model the correspondence between the temporal feature sequence and the maintenance action, and an action transfer constraint matrix is constructed; A bidirectional long short-term memory network is used to encode the temporal feature sequence and then extract an initial temporal state representation, which is input into a multi-head attention layer to fuse state information of different time scales to obtain a temporal state representation; A maintenance time series decision model is constructed based on the time series state representation; the time series state representation is input into a multi-layer time series convolutional network, multi-scale features are extracted through dynamic void ratio, and feature fusion is performed using bidirectional residual connection to obtain enhanced time series features; The enhanced temporal features are respectively input into the value network and the advantage network to obtain the state value function and the action advantage function. The state value function and the action advantage function are combined to obtain the temporal value score of each maintenance action. The maintenance decision action is selected in combination with the action transfer constraint matrix.

4. The method according to claim 3, characterized in that The temporal state representation is input into a multi-layer temporal convolutional network, multi-scale features are extracted through dynamic hole rate, and bidirectional residual connection is used for feature fusion to obtain enhanced temporal features, including: Performing multi-scale decomposition on the temporal state representation and extracting features of different frequency bands, performing importance evaluation on the features of each frequency band, dynamically adjusting feature weights based on the importance evaluation results, and reconstructing the weighted feature components through inverse transformation to obtain an enhanced temporal state representation; The enhanced temporal state representation is input into a multi-layer temporal convolutional network. An exponentially increasing dilation rate sequence is set in each layer of the network. A dilated convolution operation is used to extract local temporal features in multiple receptive fields. A bidirectional residual connection path is constructed between adjacent convolutional layers. The local temporal features extracted from the previous layer are adaptively selected and fused with the features of the current layer through a gating mechanism to obtain multi-scale local temporal features. Constructing a query matrix, a key matrix, and a value matrix for the multi-scale local temporal features, performing a scaled dot product operation on the query matrix and the key matrix to calculate an attention score, and performing masking on the attention score to obtain a feature weight; The sinusoidal position encoding method is used to add temporal position information to the features, and the feature weights after adding the position information are multiplied by the value matrix to obtain weighted features; The weighted features are input into a feedforward neural network to construct a jump connection with the original time series state representation to generate enhanced time series features.

5. The method according to claim 1, characterized in that The maintenance timing decision model is optimized and updated online using a transfer learning method. The multi-source domain knowledge feature matrix is transferred to the target device domain through feature mapping, and the network parameters of the maintenance timing decision model are updated, including: Obtain the maintenance sequence decision model of the target device and the historical maintenance data of other devices, use the multi-view spectral clustering method to construct a device similarity matrix for the historical maintenance data, and calculate the maintenance pattern similarity matrix in combination with the dynamic time warping algorithm; Adaptively fusing the device similarity matrix and the maintenance mode similarity matrix to generate a multi-source domain knowledge feature matrix; A variational autoencoder is used to probabilistically encode the multi-source domain knowledge feature matrix. The encoded feature sequence is input into a bidirectional gated recurrent unit network to extract long-range temporal dependency features. A multi-head self-attention mechanism is combined to capture the dynamic correlation between features. The source domain conditional feature distribution is generated through residual connections and layer normalization. Calculating a target domain feature distribution according to the maintenance sequence decision model of the target device, calculating a difference measure between the target domain feature distribution and the source domain condition feature distribution based on the bulldozer distance, performing a weighted combination of the difference measure and the multi-source domain knowledge feature matrix to generate a source domain weight allocation vector; The source domain conditional feature distribution is input into the feature encoder, and feature mapping is performed using multi-scale convolution and dual attention mechanism combined with the source domain weight allocation vector. A weighted optimization objective is constructed to update the parameters of the maintenance timing decision model.

6. The method according to claim 5, characterized in that Inputting the source domain conditional feature distribution into a feature encoder, adopting multi-scale convolution and dual attention mechanism combined with the source domain weight allocation vector for feature mapping, and constructing a weighted optimization objective to update the parameters of the maintenance timing decision model, including: Inputting the source domain conditional feature distribution into a multi-branch feature encoder to extract multi-granularity feature representation, adaptively aggregating the multi-granularity feature representation to generate a source domain feature representation; Calculating spatial attention weights and channel attention weights for the source domain feature representation, performing tensor multiplication on the spatial attention weights and the channel attention weights to obtain a joint attention map, and generating multi-scale attention weights based on the long-range dependency of the joint attention map; Inputting the source domain feature representation and the multi-scale attention weight cascade into a coarse-grained generative network to obtain an initial target domain feature, inputting the initial target domain feature and the source domain weight allocation vector into a fine-grained generative network for feature refinement to generate a final target domain feature; Calculating the discriminant scores of the final target domain features at different feature levels, constructing a multi-scale adversarial loss based on the discriminant scores, and performing adaptive temperature adjustment on the source domain feature representation at the corresponding feature level to obtain a softened knowledge representation; The distillation loss of the final target domain features and the softened knowledge representation at each level is calculated, the multi-scale adversarial loss and the distillation loss at each level are weightedly combined according to the source domain weight distribution vector, and the parameters of the feature transfer network are updated through an alternating optimization algorithm.

7. The method according to claim 1, characterized in that Based on the action decision results output by the maintenance time sequence decision model, the equipment energy efficiency indicators are analyzed, and a maintenance decision plan is generated based on the time sequence causal relationship optimization solution, including: Extracting a device energy efficiency indicator from the second feature data set, performing multi-scale sampling on the device energy efficiency indicator to extract a time-frequency feature matrix, and constructing a multimodal probability distribution of the device energy efficiency indicator based on the time-frequency feature matrix; Inputting the action decision result and the multimodal probability distribution into a multi-layer prediction unit to generate a probability prediction sequence of device energy efficiency changes; Performing importance sampling on the probability prediction sequence to obtain multiple groups of energy efficiency prediction samples, estimating energy efficiency evaluation indicators at different confidence levels for the multiple groups of energy efficiency prediction samples; constructing a multi-criteria comprehensive evaluation model for the energy efficiency evaluation indicators to generate a maintenance benefit evaluation vector; A graphical structure model is used to establish a temporal causal relationship between the action decision results and the probability prediction sequence. The intervention effect of each maintenance action is calculated through a counterfactual framework. The direct and indirect effects corresponding to the intervention effect are quantified based on structural equations. The maintenance benefit evaluation vector is adaptively fused with the direct and indirect effects to obtain an action importance score. According to the action importance score, priority weights are assigned to the action decision results to obtain weighted decision results. Combined with the probability prediction sequence, the weighted decision results are used to construct an optimization objective considering timing constraints to solve the optimal maintenance timing and generate a maintenance decision plan.

8. An industrial equipment energy efficiency evaluation and maintenance decision system based on reinforcement learning, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect real-time operating data of industrial equipment as a first feature data set; Performing feature extraction and principal component analysis on the first feature data set to obtain a second feature data set; The second unit is configured to construct a maintenance time series decision model based on a deep reinforcement learning algorithm, extract time series features from the first feature data set and the second feature data set and generate a state representation, and determine a maintenance decision action according to a preset maintenance action space; The third unit is used to perform online optimization and update of the maintenance timing decision model using a transfer learning method, migrate the multi-source domain knowledge feature matrix to the target device domain through feature mapping, and update the network parameters of the maintenance timing decision model; The fourth unit is used to analyze the equipment energy efficiency indicators based on the action decision results output by the maintenance time sequence decision model, and generate a maintenance decision plan based on the time sequence causal relationship optimization solution; The fifth unit is used to execute the maintenance decision plan, record the time, operation and resource data during the maintenance operation to form a maintenance process record, collect the first feature data set and the second feature data set after the maintenance is completed, use the maintenance process record and the feature data set after maintenance as training samples, input them into the maintenance timing decision model for online learning, and update the model parameters.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

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

Images