Feature alignment-based device operation and maintenance plan matching method and system

Abstract

The application relates to the technical field of equipment intelligent operation and maintenance, in particular to an equipment operation and maintenance plan matching method and system based on feature alignment; the method comprises the following steps: generating an equipment high-dimensional feature vector through multi-source data acquisition and abnormality processing, and encoding a historical operation and maintenance task into a comparable vector, and realizing preliminary alignment of heterogeneous features through nonlinear weighted mapping and mean deviation correction; fusing text semantics and numerical parameters by using contrastive learning and a large language model to form a unified equipment feature vector, calculating a multi-dimensional matching score of the equipment and the task, and screening candidate operation and maintenance tasks; dynamically triggering maintenance based on a state drift coefficient, constructing an operation and maintenance correlation graph to analyze task dependence, resource conflict and risk propagation, and generating an adaptive operation and maintenance plan through a comprehensive optimization function; continuously tracking a state recovery track after execution, reversely correcting the feature vector, and mining implicit operation and maintenance rules, and migrating high-trust experience to similar equipment groups. The application forms a self-learning and self-evolution operation and maintenance optimization mechanism.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment operation and maintenance technology, specifically to a method and system for matching equipment operation and maintenance plans based on feature alignment. Background Technology

[0002] With the increasing complexity and automation of industrial equipment, the field of equipment operation and maintenance management has accumulated massive amounts of multi-source heterogeneous data, including real-time operating status, historical operation and maintenance records, equipment text descriptions, and environmental parameters. The development of IoT, big data, and artificial intelligence technologies has provided a technological foundation for fully mining the value of this data, making intelligent equipment operation and maintenance plan matching possible. In practical applications, different equipment types, different manufacturer descriptions, and non-standardized text information make it difficult to efficiently align equipment characteristics with operation and maintenance task characteristics in the same semantic space, which places higher demands on matching accuracy and generalization ability. Equipment operating status often exhibits a gradual change pattern, and early abnormal signals under fluctuations in operating conditions need to be effectively perceived.

[0003] Chinese invention patent CN120198106B discloses an intelligent operation and maintenance method that integrates multimodal data and active learning. This method includes: deploying a hierarchical IoT device network in the target operation and maintenance area; collecting heterogeneous operation and maintenance datasets and generating an operation and maintenance feature dataset using a unified building operation and maintenance service framework with multiple MCPs working collaboratively; inputting the operation and maintenance feature dataset and user feedback information into a deep learning intent recognition model for intent recognition and demand analysis to obtain structured user demand data and operation and maintenance task priority ranking; performing real-time assessment of equipment operating status based on the operation and maintenance feature dataset and structured user demand data to obtain fault risk warning data; and conducting autonomous decision analysis through an agent-based artificial intelligence engine to generate equipment maintenance plans and resource scheduling plans.

[0004] There are complex relationships among operation and maintenance tasks, such as process dependence, resource sharing, and potential risk propagation. The rationality of task collaborative scheduling directly affects maintenance efficiency and cost. At the same time, the efficient reuse and continuous evolution of operation and maintenance experience are of great significance for improving the quality of long-term decision-making. Therefore, building a device operation and maintenance plan matching method that can achieve heterogeneous feature alignment, dynamic state perception, multi-task collaborative optimization, and self-learning ability has become an important direction for current technological development. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing a method and system for matching equipment operation and maintenance plans based on feature alignment.

[0006] The technical solution of this invention: a device operation and maintenance plan matching method based on feature equalization, comprising the following specific implementation steps: S1. Collect equipment operating status and environmental data and perform anomaly processing to form a high-dimensional feature vector. Encode historical operation and maintenance tasks into comparable vectors. Project them to a unified space through a nonlinear weighted mapping function, perform mean deviation correction, complete the initial alignment of heterogeneous features, and output the equipment feature vector. S2. Explicitly trained alignment is used to fuse text semantic features and numerical parameter features. Contrastive learning is used to aggregate similar devices and distinguish different types of devices. The device descriptions are classified by a large language model and fused with the explicitly trained features in a weighted manner. Multidimensional matching scores are generated based on a unified feature vector. A candidate maintenance task list is generated by filtering the scores through a threshold. S3. Jointly analyze the real-time operating status of equipment, historical degradation trends, operation and maintenance task correlations and resource constraints. By constructing a state drift perception mechanism, an operation and maintenance path correlation mechanism, a risk propagation suppression mechanism and a dynamic scheduling correction mechanism, intelligent optimization and adaptive generation of operation and maintenance plans are completed. S4. Continuously track the status recovery, equipment operation trajectory and long-term fault change trend during the operation and maintenance process. By analyzing the differences in equipment status before and after operation and maintenance, reverse the original equipment feature vector, risk assessment results and operation and maintenance strategy, uncover hidden operation and maintenance patterns, and transfer high-reliability operation and maintenance experience to similar equipment groups.

[0007] Preferably, step S1 specifically includes: Real-time collection of multi-dimensional equipment operation data and environmental parameters is achieved through IoT sensor networks, programmable logic controller interfaces, and equipment management systems. The equipment operation data includes temperature, vibration, load, current, voltage, power consumption, and the equipment's working cycle. Environmental parameters include humidity, ambient vibration frequency, and voltage fluctuation. Preliminary cleaning was performed using sliding window mean filtering, median filtering, and threshold-based anomaly removal methods to generate preliminary feature vectors for each device. Historical operation and maintenance records, maintenance logs, task execution time, fault type, maintenance time and required resource information are encoded to form operation and maintenance task feature vectors. Discrete features are one-hot encoded and continuous features are normalized. Then, cluster analysis is used to enhance the representativeness of the task vectors. The feature mapping function projects the device feature vector and the operation and maintenance task feature vector into a unified high-dimensional feature space, and sets the feature importance coefficient based on historical experience to weight different features, so that the features that have a greater impact on the matching results have higher weights. A preliminary consistency assessment is performed on the mapped device feature vector and the operation and maintenance task feature vector in a high-dimensional space. Weighted correction is then performed by calculating the deviation of the feature mean and standard deviation, and the device feature vector is output.

[0008] Preferably, step S2 specifically includes: A pre-trained encoder is used to extract semantic vectors from the device description text. The numerical parameters collected by the device are input into a multilayer perceptron to generate numerical feature vectors. The semantic vectors and numerical feature vectors are weighted and fused to form a preliminary multimodal feature vector. By optimizing feature alignment through comparative learning loss function, features of the same type of device are similar in vector space, while features of different types of devices are clearly distinguishable, thus achieving basic feature alignment. For non-standard device descriptions that are difficult to cover through explicit training, the original text description fragments of the device are input into the large language model. The model is guided by prompts and engineering to output standardized device types and key parameter key-value pairs, generating intelligent semantic feature vectors. The intelligent semantic feature vector output by the large language model is fused with the preliminary multimodal feature vector according to weights to form the final unified device feature vector; The system calculates a comprehensive matching score between devices and maintenance tasks based on a unified feature vector, including feature similarity, historical maintenance similarity, and operating condition parameter conformity, and allows dynamic adjustment of weights. Candidate maintenance tasks are selected based on multi-dimensional scoring thresholds, tasks below the threshold are removed, a list of candidate maintenance tasks is generated, and the scores and feature correspondences are recorded.

[0009] Preferably, the state drift sensing mechanism in step S3 specifically includes: State sequence data within a continuous time window is extracted from the equipment real-time monitoring system, including temperature fluctuation value, vibration amplitude, current imbalance rate, load fluctuation rate, start-stop frequency and energy consumption offset rate. The above multi-dimensional parameters are fused according to the feature weights corresponding to the equipment type to generate a comprehensive equipment state sequence. The system calls upon the equipment's historical stable operating condition database to extract the average baseline state of the equipment during normal operation and calculates the equipment state drift coefficient. It focuses on analyzing the degree of deviation of the equipment state from its own historical stable state in order to identify early progressive anomalies. Combining the comprehensive matching score of equipment and maintenance tasks generated in step S2, the maintenance trigger intensity is further generated. When the equipment is highly matched with a certain maintenance task and the state drift continues to increase, the maintenance priority will be increased in advance even if the equipment has not yet reached the traditional fault threshold.

[0010] Preferably, the operation and maintenance path association mechanism in step S3 specifically includes: An operation and maintenance association graph is constructed based on historical operation and maintenance work orders, equipment process flow and personnel scheduling records. This graph includes a set of operation and maintenance task nodes and a set of association edges between tasks. A detailed analysis of the relationships between different operation and maintenance tasks is conducted, including process dependencies, downtime coupling, resource conflicts, and environmental mutual exclusion. The correlation between computational tasks is determined by a combination of factors including process dependence, resource overlap, and temporal coupling. The system automatically identifies conflicts and collaborations between tasks based on their correlation and impact. When two tasks have a high degree of downtime coupling, the system automatically merges them into the same maintenance window. When multiple tasks compete for the same critical testing equipment, the system automatically adjusts the execution order. When there are process dependencies between tasks, the system automatically rearranges the execution path, thus completing multi-task collaborative scheduling and reasonable resource allocation.

[0011] Preferably, the risk transmission suppression mechanism in step S3 specifically includes: Construct equipment risk propagation factors, conduct joint analysis on equipment state drift, fault propagation trend and historical fault cumulative impact, and assess the degree of equipment risk propagation in the time dimension. Among them, fault propagation trend is used to analyze the coupling wear that may be caused by abnormal equipment vibration, the current fluctuation that may be induced by the decline in motor insulation, and the abnormal temperature rise of the entire equipment group that may be caused by the decline in cooling fan efficiency. A comprehensive optimization function is constructed for different candidate operation and maintenance solutions. This comprehensive optimization function combines the operation and maintenance trigger intensity, plan execution cost and risk propagation value to dynamically optimize the solution, giving priority to the maintenance solution that can effectively suppress risk spread and consume resources reasonably, so as to improve the scientific and economical nature of maintenance decision-making.

[0012] Preferably, the dynamic scheduling correction mechanism in step S3 specifically includes: After the optimal operation and maintenance plan is selected, real-time data such as equipment status changes, environmental parameter changes, load changes, spare parts inventory changes, operation and maintenance personnel availability, and process production plan changes are continuously received; the plan correction factor is calculated, which is generated by combining the equipment status change, load change, and environmental change with correction weights. Automatically perform lubrication and maintenance in advance when the equipment load suddenly increases; automatically increase insulation detection when the ambient humidity rises abnormally; and automatically postpone low-priority tasks when spare parts inventory is insufficient. The generated operation and maintenance plan is dynamically revised based on the planned revision amount. When the equipment status, operating conditions or resource conditions change, the maintenance priority, execution time and task sequence are automatically adjusted, and finally a dynamic operation and maintenance plan queue is generated. Each plan node includes operation and maintenance task type, execution time, risk level, required resources, scheduling priority and dynamic revision flag.

[0013] Preferably, step S4 specifically includes: After the operation and maintenance plan is executed, the equipment operating status, maintenance operation records, and short-cycle stability data are continuously collected. The collected content includes the actual start and end time of the operation and maintenance task, the arrival time of the operation and maintenance personnel, the actual execution sequence, the model and batch of the parts replaced during the operation and maintenance process, the type of lubricant or consumables, temporary abnormal records during the operation and maintenance process, the results of the trial operation after maintenance, and the short-cycle operation fluctuations after maintenance. At the same time, the status change data after maintenance is continuously collected from the equipment monitoring system, including vibration change curves, temperature rise and fall curves, current fluctuation changes, power recovery status, energy consumption change trends, start-stop stability changes, and load recovery speed. Based on the continuous status data after maintenance, a device recovery status sequence is constructed, and the recovery stability coefficient is used to analyze whether the device has truly recovered to a healthy operating state. If the device operates continuously and stably after maintenance, the credibility of this maintenance strategy is gradually increased. Based on the difference between the actual recovery effect after maintenance and the original prediction result, the equipment feature prediction error is calculated, the key feature dimensions that cause the error are analyzed, and the unified feature vector of the equipment is dynamically corrected so that the equipment feature expression can be continuously adjusted with the changes in real working conditions.

[0014] Preferably, the experience transfer in step S4 specifically includes: The highly reliable operation and maintenance experience accumulated during long-term operation is precipitated into dynamic knowledge nodes. These knowledge nodes include high-risk state combinations, preferred maintenance paths, fault propagation chains, environmentally sensitive characteristics, high-return maintenance strategies, and stable operation modes corresponding to specific working conditions. Calculate the migration similarity between different device groups, which is based on the feature center vectors of different device groups; When the migration similarity exceeds a set threshold, high-yield maintenance strategies are automatically migrated to the target equipment group, enabling cross-equipment operation and maintenance experience sharing. This allows the anti-corrosion maintenance strategy of a certain type of circulating pump in a high-humidity environment to be migrated to a cooling pump with a similar structure, and the vibration suppression scheme of a high-load fan to be migrated to equipment with a similar power level, thereby building an operation and maintenance knowledge network.

[0015] The technical solution of this invention: A device operation and maintenance plan matching system based on feature equalization, comprising: Memory; processor; A computer program stored in the memory and capable of running on the processor; The processor executes a computer program to implement the above-mentioned device maintenance plan matching method based on feature equalization.

[0016] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention designs a method and system for matching equipment operation and maintenance plans based on feature alignment. Through a two-level feature alignment mechanism, it maps equipment operating status and historical operation and maintenance tasks to a unified high-dimensional space. By integrating contrastive learning and the semantic understanding capabilities of large language models, it significantly improves the matching accuracy and generalization ability between non-standardized equipment descriptions and heterogeneous data, enabling the system to effectively handle diverse industrial equipment types and operation and maintenance record formats. Based on the equipment's own historical stability benchmark, a state drift perception mechanism is constructed, which can identify early progressive anomalies and dynamically trigger maintenance before the fault reaches the traditional alarm threshold, thereby reducing the risk of sudden downtime and extending the equipment's healthy operating cycle. By constructing an operation and maintenance correlation graph, the system automatically analyzes the process dependencies and downtime between tasks. By addressing coupling, resource conflicts, and environmental mutual exclusion, the system enables multi-task collaborative scheduling and rational resource allocation, reducing maintenance conflicts and waiting time, and improving overall operation and maintenance efficiency. A risk propagation suppression mechanism is introduced to assess the spread of equipment degradation trends over time and within the equipment network, prioritizing maintenance solutions that can block risk propagation and consume reasonable resources, thus enhancing the scientific and economical nature of maintenance decisions. The system possesses closed-loop self-learning and knowledge transfer capabilities, continuously tracking the post-maintenance state recovery trajectory, reversing feature representations, uncovering implicit operation and maintenance patterns, and automatically migrating highly reliable experience to similar equipment groups. This allows the system to continuously evolve with operational data, maintaining high-efficiency operation and maintenance plan matching performance over the long term, reducing reliance on manual intervention, and achieving intelligent and sustainable equipment operation and maintenance management. Attached Figure Description

[0017] Figure 1 This is a flowchart of a device operation and maintenance plan matching method based on feature equalization proposed in this invention. Detailed Implementation

[0018] Example 1, as Figure 1 As shown, the device operation and maintenance plan matching method based on feature equalization proposed in this invention includes the following specific implementation steps: S1. Collect equipment operating status and environmental data, perform anomaly processing to form high-dimensional feature vectors, encode historical maintenance tasks into comparable vectors, project them onto a unified space using a non-linear weighted mapping function, and perform mean deviation correction to achieve initial alignment of heterogeneous features. Through multi-source equipment data collection, historical maintenance task feature construction, high-dimensional feature unified mapping, and initial consistency correction, a quantitative representation of equipment status and maintenance tasks in a unified feature space is achieved. The specific implementation process is as follows: S11. Collect multi-dimensional device operation data and environmental parameters through IoT sensors and device management interfaces. Perform noise filtering and anomaly removal on the raw data to form a high-quality preliminary feature vector, specifically: The system collects real-time equipment operation data through IoT sensor networks, PLC (Programmable Logic Controller) interfaces, and equipment management systems, including multi-dimensional parameters such as temperature, vibration, load, current, voltage, power consumption, and equipment duty cycle. At the same time, it monitors the environmental conditions of the equipment and collects external factors that affect the equipment's operating status, such as humidity, ambient vibration frequency, and voltage fluctuations. Preliminary cleaning was performed using sliding window mean filtering, median filtering, and threshold-based anomaly removal methods. After processing, a preliminary feature vector was generated for each device. ; In the formula, This represents the feature vector of the i-th device; This represents the j-th operating parameter value of the i-th device, after normalization; n represents the total number of feature dimensions collected. S12. Encode historical operation and maintenance records, task logs, and fault information into operation and maintenance feature vectors. Perform unified normalization on discrete and continuous features, and enhance the representativeness of task vectors through cluster analysis. Specifically: Information such as historical operation and maintenance records, maintenance logs, task execution time, fault type, maintenance time, and required resources are encoded to form an operation and maintenance task feature vector. : ; In the formula, This represents the feature vector of the k-th operation and maintenance task; The j-th characteristic metric of the task is represented, including maintenance type (routine / emergency), average time, required personnel or tools, historical failure frequency, etc.; p represents the total number of operation and maintenance characteristic dimensions. It should be noted that during the construction process, discrete features (such as task type) are one-hot encoded, and continuous features (such as time consumption and failure frequency) are normalized. To further improve the subsequent matching accuracy, cluster analysis is performed on historical task data to group similar tasks, reduce the impact of noise on feature vectors, and enhance the representativeness of the vectors. S13. Project device features and maintenance task features onto a unified high-dimensional space using a feature mapping function, and combine weighted processing to enhance the influence of key features, achieving initial alignment, specifically: After completing the construction of equipment and operation and maintenance characteristics, the feature mapping function is used. and Device characteristics Characteristics of operation and maintenance tasks Mapping to a unified high-dimensional feature space achieves initial alignment, and feature importance coefficients are set based on historical experience. Different features are weighted, so that the features that have a greater impact on the matching results have higher weights: ; ; In the formula, This represents the mapped device feature vector; This represents the mapped feature vector of the operation and maintenance task; and The feature mapping function is represented by nonlinear normalization or weighted mapping to make features of different dimensions comparable in the same space. S14. Perform preliminary consistency assessment and mean deviation weighted correction on the mapped features to eliminate distribution differences between device and task features and improve comparability in high-dimensional space. Specifically: After mapping is complete, the mapped device feature vectors are processed. Mapped operation and maintenance task feature vector A preliminary consistency assessment is performed in a high-dimensional space by calculating the deviation of the characteristic mean and standard deviation: ; In the formula, This represents the corrected device feature vector; and These represent the mean vectors of device and task features, respectively. This indicates the correction weighting coefficient, which is automatically adjusted based on historical matching results. By using weighted correction, feature distribution bias is eliminated, improving the comparability of features in a unified space. At the same time, historical execution results are taken into account, so that the correction process takes into account both data statistical regularity and actual operation and maintenance effect.

[0019] S2. Based on the device feature vector output in step S1, textual semantic features and numerical parameter features are fused using explicit training alignment. Contrastive learning is used to aggregate similar devices and distinguish different types. Intelligent emergence is achieved through a large language model to classify non-standardized or rare device descriptions, which are then weighted and fused with the explicitly trained features. A multi-dimensional matching score is generated based on the unified feature vector, integrating type matching, historical maintenance similarity, and operating condition parameter fit. A candidate maintenance task list is generated by filtering through a scoring threshold. Through two-level feature alignment and multi-dimensional scoring, intelligent matching between devices and maintenance tasks is achieved. The specific implementation process is as follows: S21. Extract the device text description and numerical parameters into semantic vectors and numerical vectors respectively, fuse them to generate a preliminary multimodal feature vector, and optimize through comparative learning to make the features of similar devices similar in vector space and the features of different types of devices clearly distinguishable, thereby achieving basic feature alignment. Specifically: Semantic vectors are extracted from device description text (such as model number, function description, manufacturer information) using a pre-trained encoder (BERT or RoBERTa). For example, descriptions from different manufacturers such as "single-stage centrifugal pump", "C-type centrifugal pump", and "centrifugal transfer pump" can all be encoded into semantically similar vector representations. The numerical parameters collected by the equipment (such as temperature, flow rate, power, efficiency, pressure, etc.) are input into a multilayer sensor (MLP) to generate numerical feature vectors. ; semantic vectors With numerical eigenvectors Weighted fusion forms a preliminary multimodal feature vector : ; In the formula, The text semantic feature vector corresponding to the i-th device is obtained by encoding the device description text through a pre-trained language model and is used to represent semantic information such as device name, model, purpose, structural form, and functional description. This represents the feature vector of the numerical parameters of the i-th device; This represents the unified feature vector of the device after the first level of explicit training, i.e., the preliminary multimodal feature vector; This represents the text feature fusion weight, used to measure the contribution of text semantic features to the overall device features; This represents the weighting of numerical parameters, used to measure the contribution of equipment operating parameters to the overall characteristics; Feature alignment is optimized by using a contrastive learning loss function (InfoNCE): ; In the formula, This represents the value of the contrastive learning loss function; Represents a positive sample vector (device of the same type as i); represents the negative sample vector (device of a different type than i); K represents the total number of negative samples; This represents a temperature coefficient used to adjust the sharpness of the similarity distribution. ; The feature similarity function is represented by cosine similarity in this embodiment; S22. For non-standard device descriptions that are difficult to cover through explicit training, input them into a large language model to generate standardized type and parameter vectors, and then weightedly fuse them with explicit features to form a unified feature vector, specifically: Input the original text description fragment of the device into a large language model (such as GPT, deepseek, etc.), and use the Prompt project to guide the model to output standardized device type and key parameter key-value pairs; By leveraging the inherent knowledge and contextual understanding capabilities of large models, reasonable classifications can be generated even for combinations of device models that have never been seen before. For example, "Single-stage centrifugal pump, 5 m³ / min, 20 mH₂O" is correctly classified as "centrifugal pump, flow rate 5, head 20". Output features of the large model With the initial multimodal feature vector Weighted fusion: ; In the formula, Represents the intelligent semantic feature vector generated by the large language model; This represents the final unified device feature vector; Indicates the feature fusion coefficient; S23. Based on a unified feature vector, calculate a multi-dimensional matching score between equipment and maintenance tasks, including feature similarity, historical maintenance similarity, and operating condition parameter conformity. The weights can be dynamically adjusted to achieve a quantitative evaluation of the matching quality, providing a basis for candidate task selection. Specifically: After completing the unified feature vector Subsequently, a multi-dimensional matching score is performed on the equipment and maintenance tasks to provide a quantitative basis for matching maintenance plans. ; In the formula, This represents the overall matching score between device i and maintenance task k; This represents the feature vector of the k-th operation and maintenance task; This represents the historical maintenance similarity function, used to measure the degree of similarity between the current equipment and the historically repaired equipment in terms of maintenance behavior; The function representing the fit between operating parameters is used to determine the degree of compatibility between the current equipment operating status and the target maintenance plan; This represents the type feature matching weight, used to control the proportion of influence of device type similarity in the total score; This represents the weight of historical operation and maintenance similarity, used to control the degree of influence of historical experience on the matching of operation and maintenance plans; This indicates the weight of the operating condition matching degree, which is used to reflect the importance of real-time operating status to plan matching; S24. Filter candidate maintenance tasks based on multi-dimensional scoring thresholds, generate a candidate list, and record the correspondence between scores and features. Specifically: A candidate maintenance task list is generated based on the score. Remove those below the threshold Task: ; In the formula, This represents the set of candidate maintenance tasks for device i. This indicates the threshold for selecting candidate tasks.

[0020] S3. Based on the final unified device feature vector and candidate maintenance task set output in step S2, a joint analysis is performed on the real-time operating status of the device, historical degradation trends, maintenance task correlations, and resource constraints. By constructing a state drift perception mechanism, a maintenance path correlation mechanism, a risk propagation suppression mechanism, and a dynamic scheduling correction mechanism, intelligent optimization and adaptive generation of maintenance plans are achieved. The specific implementation process is as follows: S31. By collecting the operating parameters of the equipment within a continuous time window, a time series of equipment status is constructed. Combined with the baseline status of the equipment during its historical stable operating phases, the equipment status drift coefficient is calculated. This coefficient, along with the equipment and maintenance task matching score generated in step S2, generates the maintenance trigger intensity, thereby achieving dynamic maintenance triggering based on real-world operating condition changes. Specifically: The real-time equipment monitoring system extracts state sequence data within a continuous time window, including but not limited to: temperature fluctuation value, vibration amplitude, current imbalance rate, load fluctuation rate, start-stop frequency, and energy consumption deviation rate. The above multidimensional parameters are fused according to the feature weights corresponding to the equipment type to generate a comprehensive equipment status sequence: ; In the formula, Represents the time series of the state of device i; This represents the overall state value of the device at time t; t is the sampling time. Access the equipment's historical stable operating condition database to extract the baseline average value of the equipment during normal operation. And calculate the state drift coefficient: ; In the formula, The device state drift coefficient is represented by T; the analysis window length is represented by T. The analysis focuses on the degree of deviation of the equipment status from its historical stable state, thus enabling the identification of early, gradual anomalies. For example, if a centrifugal pump's temperature does not exceed the alarm threshold, but its vibration frequency and current fluctuations have been continuously deviating from the historical stable range, the system will determine that the equipment has a potential deterioration trend. Combined with the comprehensive matching score generated in step S2 Further generate the operation and maintenance trigger strength: ; In the formula, This indicates the trigger strength of device i for maintenance task k; and Indicates the adjustment weight; It should be noted that when the equipment is highly matched with a certain maintenance task but its operating status remains stable, high-level maintenance will not be triggered immediately; conversely, when the state drift continues to increase, the system will raise the maintenance priority in advance even if the equipment has not yet reached the traditional fault threshold. S32. Based on historical maintenance work orders, equipment process flows, and resource scheduling records, an maintenance relationship graph is constructed. This graph analyzes the process dependencies, downtime coupling relationships, resource conflict relationships, and environmental mutual exclusion relationships among different maintenance tasks. The correlation impact degree between tasks is calculated, and task conflicts and collaboration relationships are automatically identified based on this degree. The execution order of maintenance tasks is dynamically adjusted to achieve multi-task collaborative scheduling and reasonable resource allocation. Specifically: After generating the operation and maintenance trigger strength, the next step is to further analyze the correlation between different operation and maintenance tasks; Based on historical maintenance work orders, equipment process flows, and personnel scheduling records, an maintenance relationship graph is constructed: ; In the formula, V represents the set of operation and maintenance task nodes; V represents the set of associated edges between tasks; A detailed analysis of the relationships between tasks is conducted, including but not limited to: Process dependencies, such as "lubrication maintenance" must be performed before "high-speed no-load testing"; Shutdown coupling relationships, such as "bearing replacement" and "coupling calibration" both require equipment shutdown, can be performed together; Resource conflicts, such as two maintenance tasks requiring the same technician or the same testing equipment at the same time; Environmental incompatibilities, such as the inability to perform insulation testing and load impulse testing simultaneously under high-temperature conditions; Calculate the degree of correlation between tasks: ; In the formula, This indicates the degree of correlation and influence between task u and task v; Indicates the degree of dependence on the manufacturing process; Indicates the degree of resource overlap; Indicates the degree of time coupling; , and Indicates the adjustment coefficient; It should be further explained that the system automatically identifies conflicts and collaborations between tasks based on their correlation and impact. For example, if two tasks have a high degree of downtime coupling, the system will automatically merge them into the same maintenance window; if multiple tasks compete for the same critical testing equipment, the system will automatically adjust the execution order; if there are process dependencies between tasks, the system will automatically rearrange the execution paths. S33. By constructing equipment risk propagation factors, a joint analysis is conducted on equipment state drift, fault propagation trends, and the cumulative impact of historical faults to assess the degree of equipment risk propagation over time. A comprehensive optimization function is constructed by combining maintenance trigger intensity, planned execution cost, and risk propagation value to dynamically optimize candidate maintenance solutions, prioritizing those that effectively suppress risk propagation and have reasonable resource consumption. Specifically: After completing the operation and maintenance correlation analysis, we further consider the propagation effect of equipment risks in the time dimension and equipment network, and dynamically optimize the candidate operation and maintenance solutions. For example: abnormal vibration of a pump body may further lead to wear of the coupling; decreased insulation of a motor may further induce current fluctuations; decreased efficiency of a cooling fan may cause abnormal temperature rise of the entire equipment group. Introducing a risk propagation analysis mechanism to construct equipment risk propagation factors: ; In the formula, This represents the risk propagation value of the device at time t; Indicates the trend of fault propagation; This represents the cumulative impact value of historical faults; , and Indicates the weighting coefficient; A comprehensive optimization function is constructed for different candidate operation and maintenance solutions, prioritizing the solution that can effectively suppress risk spread while consuming reasonable resources. ; In the formula, This represents the optimal comprehensive value of the operation and maintenance scheme k; Indicates the cost of plan execution; , and Indicates the adjustment coefficient; S34. After completing the optimal operation and maintenance plan, continuously receive real-time data such as equipment operating status, environmental parameters, load changes, spare parts inventory, and personnel scheduling, and calculate the plan correction amount. Dynamically revise the generated operation and maintenance plan. When equipment status, operating conditions, or resource conditions change, automatically adjust the maintenance priority, execution time, and task sequence, and finally generate an operation and maintenance plan queue, specifically as follows: After the optimal operation and maintenance plan is selected, a dynamic correction mechanism is established so that the operation and maintenance plan can be adjusted in real time as the on-site conditions change. In practice, the system continuously receives the following data: real-time changes in equipment status; changes in environmental parameters; changes in load; changes in spare parts inventory; the availability of maintenance personnel; and changes in process production plans. Calculate the plan correction factor: ; In the formula, Indicates the planned adjustment amount; Indicates the amount of change in equipment status; Indicates the amount of load change; Indicates the amount of environmental change; , and Indicates the adjusted weights; For example: when the equipment load suddenly increases, automatic lubrication and maintenance are performed in advance; when the ambient humidity rises abnormally, insulation detection is automatically added; when spare parts inventory is insufficient, low-priority tasks are automatically postponed. The final output is a dynamic operation and maintenance plan queue: ; Each planned node includes, but is not limited to: operation and maintenance task type; execution time; risk level; required resources; scheduling priority; and dynamic correction flag.

[0021] S4. Continuously track the state recovery status, equipment operation trajectory, and long-term fault change trends during the operation and maintenance process. By analyzing the differences in equipment state before and after operation and maintenance, reverse-correct the original equipment feature vectors, risk assessment results, and operation and maintenance strategies. Further explore the implicit operation and maintenance patterns in long-term operation to form a sustainable evolutionary operation and maintenance knowledge system. Transfer highly reliable operation and maintenance experience to similar equipment groups to realize the sharing and reuse of operation and maintenance strategies among different equipment, thereby building a dynamic operation and maintenance optimization mechanism with self-learning, self-calibration, and self-evolution capabilities. The specific implementation process is as follows: S41. After the operation and maintenance plan is executed, continuously collect equipment operating status, maintenance operation records, and short-cycle stability data to construct an equipment status recovery trajectory. Analyze whether the equipment has truly recovered to a healthy operating state using the recovery stability coefficient. Specifically: After generating the dynamic operation and maintenance plan queue P, the entire process of actual execution of the operation and maintenance tasks is continuously tracked, and the state changes of the equipment before and after maintenance are perceived in real time to determine whether the current maintenance action has truly improved the equipment's operating status. Collect traditional work order execution results and simultaneously record fine-grained execution trajectory information during the maintenance process, including but not limited to: the actual start and end time of the maintenance task; the arrival time of maintenance personnel; the actual execution sequence; the model and batch of components replaced during the maintenance process; the type of lubricant or consumables; temporary anomaly records during the maintenance process; the results of the trial run after maintenance; and the short-term operational fluctuations after maintenance. At the same time, the equipment monitoring system continuously collects data on changes in status after maintenance, including but not limited to: vibration change curves after maintenance; temperature rise and fall curves; current fluctuations; power recovery status; energy consumption trends; start-stop stability changes; and load recovery speed. Subsequently, a device recovery state sequence is constructed based on the continuous status data after maintenance: ; In the formula, This represents the recovery status trajectory of device i after maintenance is completed; This represents the recovery state value of the equipment at time t. The state value is generated by combining the vibration recovery rate, temperature rise recovery rate, load stability rate, and energy consumption recovery rate. Further analysis of the sustained stability of the recovery trajectory was conducted, and the recovery stability coefficient was calculated: ; In the formula, Indicates the equipment's recovery stability coefficient; This represents the baseline value for the device's health status; N represents the length of the recovery analysis window. For example, if the vibration value of a centrifugal pump drops immediately after the bearing is replaced, but the vibration fluctuates frequently in the following hours, the maintenance will not be immediately deemed effective. Instead, the observation period will be extended. If the equipment continues to operate stably, the credibility of the maintenance strategy will be gradually increased. S42. Based on the difference between the actual recovery effect after maintenance and the original prediction result, calculate the equipment feature prediction error and dynamically correct the unified feature vector of the equipment so that the equipment feature representation can be continuously adjusted according to changes in actual operating conditions. Specifically: After obtaining the actual recovery effect after equipment maintenance, the deviation relationship between the "original feature judgment" and the "actual maintenance result" is further analyzed, and the equipment feature representation formed in steps S1 and S2 is dynamically reversed. The predicted maintenance results before equipment maintenance are compared with the actual recovery results after maintenance, and the feature prediction error is calculated: ; In the formula, This represents the feature prediction error of device i; This represents the system's predicted maintenance performance value; This indicates the actual recovery effect value; Analyze the key characteristic dimensions that lead to the error, such as: whether the load fluctuation characteristic has too low a weight; whether the environmental humidity characteristic is not fully reflected; whether a certain type of fault is more sensitive to a specific parameter; whether a certain manufacturer's equipment has special operating patterns. After identifying key deviations, the unified feature vector of the device is dynamically corrected: ; In the formula, This represents the corrected device feature vector; Indicates the feature offset correction amount; Indicates the characteristic calibration coefficient; For example, if the recovery effect of a certain type of motor is significantly worse after maintenance in a high humidity environment, the system will gradually increase the weight of "humidity-related characteristics"; if a certain type of fan is stable under low load conditions for a long time, but the failure frequency increases sharply during high load periods, the system will enhance the influence of "load fluctuation characteristics". S43. Perform correlation analysis on a large amount of historical operation and maintenance records, fault recurrence, equipment recovery speed, and resource consumption to automatically uncover hidden operation and maintenance patterns that are difficult to detect using traditional manual rules. Based on long-term stability and fault suppression effectiveness, generate new, optimized operation and maintenance strategies, specifically: After completing the equipment feature correction, a deep correlation analysis is further conducted on a large number of historical operation and maintenance results to automatically identify hidden operation and maintenance patterns and transform them into new operation and maintenance strategies. Specifically, the following long-term historical information is analyzed jointly: recovery speed corresponding to different maintenance schemes; fault recurrence cycle; maintenance stability under different environmental conditions; maintenance cost changes under different operating conditions; relationship between operation and maintenance resource input and return; long-term performance of the same equipment under different maintenance paths. Construct the strategy reward function: ; In the formula, This represents the overall benefit value of operation and maintenance strategy k; This indicates a long-term, stable increase in value; Indicates the effectiveness of fault suppression; Indicates the cost of resource consumption; , and Indicates the weighting coefficient; Continuously compare the revenue values ​​corresponding to different maintenance paths and automatically generate high-return maintenance patterns: For example, if it is found that the subsequent stability of a certain type of fan is significantly higher when "high-speed testing is delayed for 8 to 12 hours after lubrication maintenance is completed" than when it is tested immediately, then this maintenance path will be automatically selected as the preferred strategy. For example, if it is found that the lifespan of a certain type of cooling pump is significantly extended when a combination of "high-frequency inspection + low-intensity maintenance" is used during the high humidity period in summer, the priority of this strategy will be automatically increased under similar operating conditions. S44. High-reliability operation and maintenance experience accumulated over long-term operation is precipitated as dynamic knowledge nodes, and optimized maintenance strategies are migrated to similar equipment groups based on equipment feature similarity, achieving cross-device operation and maintenance experience sharing. Specifically: After completing the strategy optimization, the highly reliable operation and maintenance experience is precipitated as dynamic knowledge nodes and migrated to similar device groups, so that effective operation and maintenance experience can be shared among different devices, and an operation and maintenance knowledge network is built. Knowledge nodes include: high-risk state combinations; optimal maintenance paths; fault propagation chains; environmentally sensitive characteristics; high-yield maintenance strategies; and stable operating modes corresponding to specific working conditions. Calculate migration similarity between different device groups: ; In the formula, This represents the migration similarity between device group a and device group b; and Represents the feature center vectors of different device groups; When the similarity exceeds the set threshold, the high-yield maintenance strategy will be automatically migrated to the target device group. For example, the anti-corrosion maintenance strategy for a certain type of circulating pump in a high-humidity environment can be applied to a cooling pump with a similar structure; the vibration suppression scheme for a high-load fan can be applied to equipment with a similar power level.

[0022] Example 2: The present invention proposes a device operation and maintenance plan matching system based on feature equalization, which is used to execute a device operation and maintenance plan matching method based on feature equalization proposed in Example 1, and includes: Memory; processor; A computer program stored in the memory and capable of running on the processor; The processor executes a computer program to implement a device maintenance plan matching method based on feature equalization as described in Embodiment 1 above.

[0023] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A feature alignment-based device operation and maintenance plan matching method, characterized in that, The specific implementation steps include the following: S1. Collect equipment operating status and environmental data and perform anomaly processing to form a high-dimensional feature vector. Encode historical operation and maintenance tasks into comparable vectors. Project them to a unified space through a nonlinear weighted mapping function, perform mean deviation correction, complete the initial alignment of heterogeneous features, and output the equipment feature vector. S2. Explicitly trained alignment is used to fuse text semantic features and numerical parameter features. Contrastive learning is used to aggregate similar devices and distinguish different types of devices. The device descriptions are classified by a large language model and fused with the explicitly trained features in a weighted manner. Multidimensional matching scores are generated based on a unified feature vector. A candidate maintenance task list is generated by filtering the scores through a threshold. S3. Jointly analyze the real-time operating status of equipment, historical degradation trends, operation and maintenance task correlations and resource constraints. By constructing a state drift perception mechanism, an operation and maintenance path correlation mechanism, a risk propagation suppression mechanism and a dynamic scheduling correction mechanism, intelligent optimization and adaptive generation of operation and maintenance plans are completed. S4. Continuously track the status recovery, equipment operation trajectory and long-term fault change trend during the operation and maintenance process. By analyzing the differences in equipment status before and after operation and maintenance, reverse the original equipment feature vector, risk assessment results and operation and maintenance strategy, uncover hidden operation and maintenance patterns, and transfer high-reliability operation and maintenance experience to similar equipment groups.

2. The device operation and maintenance plan matching method based on feature alignment according to claim 1, characterized in that, Step S1 specifically includes: Real-time collection of multi-dimensional equipment operation data and environmental parameters is achieved through IoT sensor networks, programmable logic controller interfaces, and equipment management systems. The equipment operation data includes temperature, vibration, load, current, voltage, power consumption, and the equipment's working cycle. Environmental parameters include humidity, ambient vibration frequency, and voltage fluctuation. Preliminary cleaning was performed using sliding window mean filtering, median filtering, and threshold-based anomaly removal methods to generate preliminary feature vectors for each device. Historical operation and maintenance records, maintenance logs, task execution time, fault type, maintenance time and required resource information are encoded to form operation and maintenance task feature vectors. Discrete features are one-hot encoded and continuous features are normalized. Then, cluster analysis is used to enhance the representativeness of the task vectors. The feature mapping function projects the device feature vector and the operation and maintenance task feature vector into a unified high-dimensional feature space, and sets the feature importance coefficient based on historical experience to weight different features, so that the features that have a greater impact on the matching results have higher weights. A preliminary consistency assessment is performed on the mapped device feature vector and the operation and maintenance task feature vector in a high-dimensional space. Weighted correction is then performed by calculating the deviation of the feature mean and standard deviation, and the device feature vector is output.

3. The device operation and maintenance plan matching method based on feature alignment according to claim 2, characterized in that, Step S2 specifically includes: A pre-trained encoder is used to extract semantic vectors from the device description text. The numerical parameters collected by the device are input into a multilayer perceptron to generate numerical feature vectors. The semantic vectors and numerical feature vectors are weighted and fused to form a preliminary multimodal feature vector. By optimizing feature alignment through comparative learning loss function, features of the same type of device are similar in vector space, while features of different types of devices are clearly distinguishable, thus achieving basic feature alignment. For non-standard device descriptions that are difficult to cover through explicit training, the original text description fragments of the device are input into the large language model. The model is guided by prompts and engineering to output standardized device types and key parameter key-value pairs, generating intelligent semantic feature vectors. The intelligent semantic feature vector output by the large language model is fused with the preliminary multimodal feature vector according to weights to form the final unified device feature vector; The system calculates a comprehensive matching score between devices and maintenance tasks based on a unified feature vector, including feature similarity, historical maintenance similarity, and operating condition parameter conformity, and allows dynamic adjustment of weights. Candidate maintenance tasks are selected based on multi-dimensional scoring thresholds, tasks below the threshold are removed, a list of candidate maintenance tasks is generated, and the scores and feature correspondences are recorded.

4. The device operation and maintenance plan matching method based on feature alignment according to claim 3, characterized in that, The state drift sensing mechanism in step S3 specifically includes: State sequence data within a continuous time window is extracted from the equipment real-time monitoring system, including temperature fluctuation value, vibration amplitude, current imbalance rate, load fluctuation rate, start-stop frequency and energy consumption offset rate. The above multi-dimensional parameters are fused according to the feature weights corresponding to the equipment type to generate a comprehensive equipment state sequence. The system calls upon the equipment's historical stable operating condition database to extract the average baseline state of the equipment during normal operation and calculates the equipment state drift coefficient. It focuses on analyzing the degree of deviation of the equipment state from its own historical stable state in order to identify early progressive anomalies. Combining the comprehensive matching score of equipment and maintenance tasks generated in step S2, the maintenance trigger intensity is further generated. When the equipment is highly matched with a certain maintenance task and the state drift continues to increase, the maintenance priority will be increased in advance even if the equipment has not yet reached the traditional fault threshold.

5. The device operation and maintenance plan matching method based on feature alignment according to claim 4, characterized in that, The operation and maintenance path association mechanism in step S3 specifically includes: An operation and maintenance association graph is constructed based on historical operation and maintenance work orders, equipment process flow and personnel scheduling records. This graph includes a set of operation and maintenance task nodes and a set of association edges between tasks. A detailed analysis of the relationships between different operation and maintenance tasks is conducted, including process dependencies, downtime coupling, resource conflicts, and environmental mutual exclusion. The correlation between computational tasks is determined by a combination of factors including process dependence, resource overlap, and temporal coupling. The system automatically identifies conflicts and collaborations between tasks based on their correlation and impact. When two tasks have a high degree of downtime coupling, the system automatically merges them into the same maintenance window. When multiple tasks compete for the same critical testing equipment, the system automatically adjusts the execution order. When there are process dependencies between tasks, the system automatically rearranges the execution path, thus completing multi-task collaborative scheduling and reasonable resource allocation.

6. The device operation and maintenance plan matching method based on feature alignment according to claim 5, characterized in that, The risk transmission suppression mechanism in step S3 specifically includes: Construct equipment risk propagation factors, conduct joint analysis on equipment state drift, fault propagation trend and historical fault cumulative impact, and assess the degree of equipment risk propagation in the time dimension. Among them, fault propagation trend is used to analyze the coupling wear that may be caused by abnormal equipment vibration, the current fluctuation that may be induced by the decline in motor insulation, and the abnormal temperature rise of the entire equipment group that may be caused by the decline in cooling fan efficiency. A comprehensive optimization function is constructed for different candidate operation and maintenance solutions. This comprehensive optimization function combines the operation and maintenance trigger intensity, plan execution cost and risk propagation value to dynamically optimize the solution, giving priority to the maintenance solution that can effectively suppress risk spread and consume resources reasonably, so as to improve the scientific and economical nature of maintenance decision-making.

7. The device operation and maintenance plan matching method based on feature alignment according to claim 6, characterized in that, The dynamic scheduling correction mechanism in step S3 specifically includes: After the optimal operation and maintenance plan is selected, real-time data such as equipment status changes, environmental parameter changes, load changes, spare parts inventory changes, operation and maintenance personnel availability, and process production plan changes are continuously received; the plan correction factor is calculated, which is generated by combining the equipment status change, load change, and environmental change with correction weights. Automatically perform lubrication and maintenance in advance when the equipment load suddenly increases; automatically increase insulation detection when the ambient humidity rises abnormally; and automatically postpone low-priority tasks when spare parts inventory is insufficient. The generated operation and maintenance plan is dynamically revised based on the planned revision amount. When the equipment status, operating conditions or resource conditions change, the maintenance priority, execution time and task sequence are automatically adjusted, and finally a dynamic operation and maintenance plan queue is generated. Each plan node includes operation and maintenance task type, execution time, risk level, required resources, scheduling priority and dynamic revision flag.

8. The device operation and maintenance plan matching method based on feature alignment according to claim 7, characterized in that, Step S4 specifically includes: After the operation and maintenance plan is executed, the equipment operating status, maintenance operation records, and short-cycle stability data are continuously collected. The collected content includes the actual start and end time of the operation and maintenance task, the arrival time of the operation and maintenance personnel, the actual execution sequence, the model and batch of the parts replaced during the operation and maintenance process, the type of lubricant or consumables, temporary abnormal records during the operation and maintenance process, the results of the trial operation after maintenance, and the short-cycle operation fluctuations after maintenance. At the same time, the status change data after maintenance is continuously collected from the equipment monitoring system, including vibration change curves, temperature rise and fall curves, current fluctuation changes, power recovery status, energy consumption change trends, start-stop stability changes, and load recovery speed. Based on the continuous status data after maintenance, a device recovery status sequence is constructed, and the recovery stability coefficient is used to analyze whether the device has truly recovered to a healthy operating state. If the device operates continuously and stably after maintenance, the credibility of this maintenance strategy is gradually increased. Based on the difference between the actual recovery effect after maintenance and the original prediction result, the equipment feature prediction error is calculated, the key feature dimensions that cause the error are analyzed, and the unified feature vector of the equipment is dynamically corrected so that the equipment feature expression can be continuously adjusted with the changes in real working conditions.

9. The device operation and maintenance plan matching method based on feature alignment according to claim 8, characterized in that, The experience transfer in step S4 specifically includes: The highly reliable operation and maintenance experience accumulated during long-term operation is precipitated into dynamic knowledge nodes. These knowledge nodes include high-risk state combinations, preferred maintenance paths, fault propagation chains, environmentally sensitive characteristics, high-return maintenance strategies, and stable operation modes corresponding to specific working conditions. Calculate the migration similarity between different device groups, which is based on the feature center vectors of different device groups; When the migration similarity exceeds a set threshold, high-yield maintenance strategies are automatically migrated to the target equipment group, enabling cross-equipment operation and maintenance experience sharing. This allows the anti-corrosion maintenance strategy of a certain type of circulating pump in a high-humidity environment to be migrated to a cooling pump with a similar structure, and the vibration suppression scheme of a high-load fan to be migrated to equipment with a similar power level, thereby building an operation and maintenance knowledge network.

10. A feature alignment-based device operation and maintenance plan matching system, characterized in that, include: Memory; processor; A computer program stored in the memory and capable of running on the processor; The processor executes a computer program to implement the device maintenance plan matching method based on feature equalization as described in any one of claims 1 to 9.

Images