Pump truck hydraulic system preventive maintenance method and system based on historical maintenance data
Patent Information
- Application Number
- CN202610970078.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-01
- Publication Date
- 2026-09-15
AI Technical Summary
然而，现有方案大多采用离线训练、固定部署的方式，缺乏对现场实际维修执行结果的反馈吸收能力，导致模型在实际应用中随着环境变化或泵车液压系统个体差异而性能衰减；同时，不同操作人员或管理者的维修经验偏好难以被有效融入决策模型，使得自动化生成的维修方案与现场实际需求之间存在偏差，降低了人机协同效率与方案的可接受度
本发明通过神经网络构建维修决策模型，以健康状态特征向量为输入自动输出最优的维修时机与动作，显著提升了维修决策的科学性与实时性，避免了传统定期维修造成的资源浪费和事后维修带来的非计划停机，引入闭环协同迭代机制，将每次执行结果与审批校验后的决策序列作为增量数据反馈至模型进行在线微调，使模型能够持续适应泵车液压系统个体差异和工况变化，有效克服了传统静态模型性能衰退的问题，通过偏好协同迭代模块，利用多个审批校验后的决策序列构建偏好监督数据对模型进行偏好微调，使模型能够学习并融合不同操作人员的维修经验与偏好，提高了决策方案的人机协同接受度和实用性，结合数据遗忘机制控制样本规模，既保证了模型对新知识的快速吸收，又防止了旧数据的干扰，维持了模型的高效稳定运行。综上，本发明实现了泵车液压系统维修决策的智能化、自适应化和个性化，提升了泵车液压系统可用性和维修经济性。
Smart Images
Figure CN122760041A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hydraulic system maintenance technology, specifically to a preventive maintenance method and system for pump truck hydraulic systems based on historical maintenance data. Background Technology
[0002] As the core power and control system of a concrete pump truck's hydraulic system, the pump truck's hydraulic system operates in harsh environments with drastic load fluctuations. It operates under high pressure and high impact conditions for extended periods, making key components such as the main oil pump, multi-way valves, and cylinders prone to progressive degradation failures such as wear, leakage, and jamming. Traditional maintenance strategies rely primarily on planned periodic maintenance or reactive fault repair. The former often leads to over-maintenance and wasted downtime, while the latter causes unplanned shutdowns due to sudden failures, severely impacting construction efficiency and the lifespan of the pump truck's hydraulic system. With the development of Industrial Internet of Things (IIoT) technology, some companies have begun to explore fault prediction methods based on sensor monitoring data and historical maintenance work orders. For example, they use threshold alarms or statistical regression models to assess component health. However, these methods generally suffer from limitations such as single feature extraction, difficulty in dynamically adapting to changes in operating conditions, and the generated maintenance solutions are mostly static suggestions, failing to provide optimal maintenance timing and action combinations based on real-time operating status.
[0003] In recent years, deep learning and reinforcement learning technologies have been gradually applied in the field of predictive maintenance of pump truck hydraulic systems. Existing research has attempted to construct degradation trend prediction models or maintenance strategy optimization models based on neural networks. However, most existing solutions employ offline training and fixed deployment, lacking the ability to absorb feedback from actual on-site maintenance execution results. This leads to performance degradation of the models in practical applications due to environmental changes or individual differences in the pump truck hydraulic system. Simultaneously, the maintenance experience preferences of different operators or managers are difficult to effectively integrate into the decision-making model, resulting in discrepancies between automatically generated maintenance plans and actual on-site needs, reducing human-machine collaboration efficiency and the acceptability of the plans. Therefore, there is an urgent need for an intelligent maintenance decision-making method that can continuously learn, dynamically iterate, and incorporate human preferences. Summary of the Invention
[0004] To address the aforementioned technical problems, this technical solution provides a preventative maintenance method and system for pump truck hydraulic systems based on historical maintenance data, thus solving at least one of the technical problems mentioned in the background section.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Preventive maintenance methods for pump truck hydraulic systems based on historical maintenance data include: Collect historical maintenance work order data, pump truck hydraulic system operating parameters and sensor monitoring data, extract degradation characteristics of key components, and construct a state space characterizing the health status of the hydraulic system; A maintenance decision model is constructed based on neural networks. The model takes the current health status feature vector of the hydraulic system as input and outputs a preventive maintenance decision sequence that includes maintenance timing and maintenance actions. The preventive maintenance decision sequence generated by the maintenance decision model is output to the human-computer interaction terminal for approval and verification. After approval and verification, the preventive maintenance decision sequence is output to the execution terminal to execute the preventive maintenance action. Record the execution results of this preventive maintenance action, and feed the execution results of this preventive maintenance action and the approved and verified preventive maintenance decision sequence as incremental data to the maintenance decision model for online fine-tuning, so as to complete the closed-loop collaborative iteration of the maintenance decision model; Based on multiple approved and verified preventive maintenance decision sequences, decision preference supervision data is constructed. The maintenance decision model is then fine-tuned using the decision preference supervision data to complete the collaborative iteration of the maintenance decision model's preferences.
[0006] Preferably, the process of collecting historical maintenance work order data, operating parameters, and sensor monitoring data of the pump truck hydraulic system, extracting degradation characteristics of key components, and constructing a state space characterizing the health status of the hydraulic system specifically includes: For each key component in the pump truck hydraulic system, extract at least one typical maintenance event and the pump truck hydraulic system operating characteristics within the window period before the typical maintenance event occurs based on its historical maintenance work order data. Based on the operating parameters of the pump truck hydraulic system and the sensor monitoring data, the real-time operating characteristics of the pump truck hydraulic system are extracted. The real-time operating characteristics of the pump truck hydraulic system are matched with the operating characteristics of the pump truck hydraulic system within the window period before all typical maintenance events of each key component. If the match is successful, the key component is marked with the typical maintenance event; otherwise, no response is made, and the health status feature vector of each key component is obtained. By summing the health status feature vectors of all key components, the state space of the hydraulic system's health status is obtained.
[0007] Preferably, the initial training steps of the maintenance decision model are as follows: Analyze historical maintenance work order data to obtain the health status feature vector of the hydraulic system corresponding to each historical maintenance work order, including maintenance timing, maintenance action sequence, and maintenance result evaluation as model sample data; A neural network structure consisting of an input layer, an information processing layer, and an output layer is constructed. The information processing layer consists of multiple fully connected layers and includes a preference calculation branch. In the initial training step, the system's preference supervision data is zero. The input layer receives the health status feature vector of the hydraulic system, and the output layer outputs the optimal maintenance time and maintenance action. A loss function based on weighted cross-entropy and preference loss is constructed. With the goal of minimizing the loss function, the complex mapping relationship between multiple fully connected layers of the information processing layer is trained through model sample data to obtain the maintenance decision model in the initial state.
[0008] Preferably, the closed-loop collaborative iterative process of the maintenance decision model is as follows: Design an incremental learning data space, in which the execution result of each preventive maintenance action and the preventive maintenance decision sequence after approval and verification are written into the incremental learning data space. When the data in the incremental learning data space reaches the preset incremental training threshold, the data in the incremental learning data space is added to the model sample data of the maintenance decision model to learn and optimize the complex mapping relationship between multiple fully connected layers of the information processing layer, thus completing a closed-loop collaborative iteration.
[0009] Preferably, the closed-loop collaborative iterative process includes a data forgetting mechanism, which specifically includes: Design an incremental iteration threshold. When the number of times the data in the incremental learning data space added to the model sample data is less than or equal to the incremental iteration threshold, the data in the incremental learning data space is directly added to the model sample data to learn and optimize the complex mapping relationship between multiple fully connected layers of the information processing layer. When the number of times data in the incremental learning data space added to the model sample data exceeds the incremental iteration threshold, the data in the incremental learning data space that was added to the model sample data the longest time is removed from the model sample data, and the data in the current incremental learning data space is added to the model sample data to learn and optimize the complex mapping relationship between multiple fully connected layers of the information processing layer.
[0010] Preferably, the preference collaborative iteration of the maintenance decision model specifically includes: A preference analysis data space is constructed based on multiple approved and verified preventive maintenance decision sequences. Preference supervision values regarding maintenance timing and maintenance actions are extracted based on the preference analysis data space. The preference supervision values of maintenance timing and maintenance actions are introduced as preference supervision data into the loss function based on weighted cross-entropy and preference loss to obtain the updated loss function; Based on the model sample data, with the goal of minimizing the updated loss function, the complex mapping relationship between multiple fully connected layers of the information processing layer is trained to complete the preference collaborative iteration.
[0011] Preferably, when extracting the preference supervision values for maintenance timing and maintenance actions based on the preference analysis data space, an adjusted preventive maintenance decision sequence in the preference analysis data space is identified as a preference-emphasized sequence, and the weight of the preference-emphasized sequence is increased when extracting the preference supervision values for maintenance timing and maintenance actions.
[0012] Furthermore, a preventive maintenance system for pump truck hydraulic systems based on historical maintenance data is proposed to implement the aforementioned preventive maintenance method for pump truck hydraulic systems based on historical maintenance data, including: The data acquisition module is used to collect historical maintenance work order data, pump truck hydraulic system operating parameters and sensor monitoring data, extract degradation characteristics of key components, and construct a state space characterizing the health status of the hydraulic system. The maintenance decision module, built on a neural network, takes the current health status feature vector of the hydraulic system as input and outputs a preventive maintenance decision sequence that includes maintenance timing and maintenance actions. The human-computer interaction approval module is used to receive and display the preventive maintenance decision sequence output by the maintenance decision module for operators to approve and verify, and output the approved and verified preventive maintenance decision sequence to the execution end; The closed-loop collaborative iteration module is used to record the execution result of this preventive maintenance action, and feeds the execution result and the approved and verified preventive maintenance decision sequence as incremental data to the maintenance decision module for online fine-tuning; The preference collaborative iteration module is used to construct decision preference supervision data based on multiple approved and verified preventive maintenance decision sequences, and to fine-tune the maintenance decision module with the decision preference supervision data.
[0013] Optionally, the closed-loop collaborative iteration module includes: The incremental data storage unit is used to store the execution results of each preventive maintenance action and the preventive maintenance decision sequence after approval and verification. An incremental training triggering unit is used to determine whether the amount of data in the incremental data storage unit has reached a preset incremental training threshold. If so, the maintenance decision module is triggered to perform incremental learning. The data forgetting control unit is used to determine whether to remove the oldest incremental data from the model sample data based on a preset incremental iteration threshold when the maintenance decision module performs incremental learning, so as to control the size of the model sample data.
[0014] Optionally, the preference collaborative iteration module includes: The preference analysis unit is used to construct a preference analysis data space based on multiple approved and verified preventive maintenance decision sequences, and extract preference supervision values for maintenance timing and maintenance actions from it. At the same time, it identifies the manually adjusted preventive maintenance decision sequences as preference emphasis sequences and increases the weight of these preference emphasis sequences. The loss function update unit is used to introduce the preference supervision value extracted by the preference analysis unit into the loss function of the maintenance decision module to generate an updated loss function for the maintenance decision module to perform preference fine-tuning training.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a maintenance decision-making model using a neural network. Taking the health status feature vector as input, it automatically outputs the optimal maintenance timing and actions, significantly improving the scientific rigor and real-time performance of maintenance decisions. It avoids resource waste caused by traditional periodic maintenance and unplanned downtime resulting from reactive maintenance. A closed-loop collaborative iteration mechanism is introduced, feeding the results of each execution and the approved decision sequence as incremental data back to the model for online fine-tuning. This allows the model to continuously adapt to individual differences and changing operating conditions in the pump truck hydraulic system, effectively overcoming the performance degradation problem of traditional static models. Through a preference collaborative iteration module, preference supervision data is constructed using multiple approved decision sequences to fine-tune the model's preferences. This enables the model to learn and integrate the maintenance experience and preferences of different operators, improving the human-machine collaborative acceptance and practicality of the decision scheme. Combined with a data forgetting mechanism to control the sample size, it ensures the model's rapid absorption of new knowledge while preventing interference from old data, maintaining the model's efficient and stable operation. In summary, this invention achieves intelligent, adaptive, and personalized maintenance decisions for pump truck hydraulic systems, improving the availability and maintenance economy of pump truck hydraulic systems. Attached Figure Description
[0016] Figure 1 This is a flowchart of the preventive maintenance method for pump truck hydraulic systems based on historical maintenance data proposed in this solution; Figure 2 This is a flowchart of the method for obtaining the state space of the hydraulic system health state proposed in this scheme; Figure 3 This is a flowchart of the closed-loop collaborative iterative method for the maintenance decision-making model proposed in this scheme. Detailed Implementation
[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.
[0018] Example 1: Reference Figure 1As shown, the preventive maintenance method for the hydraulic system of a pump truck based on historical maintenance data includes: Collect historical maintenance work order data, pump truck hydraulic system operating parameters and sensor monitoring data, extract degradation characteristics of key components, and construct a state space characterizing the health status of the hydraulic system; Specifically, the state space characterizing the health state of a hydraulic system is composed of the degradation characteristics of key components within the hydraulic system, referring to... Figure 2 As shown, the specific steps to obtain the state space of the hydraulic system's health state are as follows: For each key component in the pump truck's hydraulic system, extract at least one typical maintenance event and the pump truck's hydraulic system operating characteristics within the window period preceding the typical maintenance event based on its historical maintenance work order data. , ,in, The pump truck hydraulic system operating characteristics during the window period before the j-th typical maintenance event of the i-th critical component. Let k be the operating parameter value of the pump truck hydraulic system for the type k during the window period before the occurrence of the type j typical maintenance event of the type i critical component. This refers to the total number of operating parameters in the pump truck's hydraulic system. Real-time operating characteristics of the pump truck hydraulic system are extracted based on operating parameters and sensor monitoring data. , ,in, This refers to the real-time operating characteristics of the pump truck's hydraulic system. This represents the k-th type of pump truck hydraulic system operating parameter value in the real-time operating characteristics of the pump truck hydraulic system. The real-time operating characteristics of the pump truck's hydraulic system are matched with the operating characteristics of the pump truck's hydraulic system during the window period preceding all typical maintenance events for each critical component. The matching method involves calculation. and The Euclidean distance between them, if and If the Euclidean distance between the components is less than a threshold, the match passes, and the j-th typical maintenance event label is added to the i-th critical component; otherwise, no response is made, and the health status feature vector of each critical component is obtained. , , Let be the health status feature vector of the i-th critical component. Let be the labeling feature value of the i-th critical component with respect to the j-th typical maintenance event. If the i-th critical component has a label for the j-th typical maintenance event, then... ,otherwise, , which represents the total number of typical maintenance events occurring in critical components; By summing the health status feature vectors of all key components, the state space of the hydraulic system's health status is obtained.
[0019] A maintenance decision model is constructed based on neural networks. The model takes the current health status feature vector of the hydraulic system as input and outputs a preventive maintenance decision sequence that includes maintenance timing and maintenance actions. The maintenance decision model is specifically a neural network structure consisting of an input layer, an information processing layer, and an output layer. The information processing layer consists of multiple fully connected layers and is designed with a preference calculation branch. In the initial training step, the system's preference supervision data is zero. The input layer receives the health status feature vector of the hydraulic system; The output layer outputs the optimal repair timing and repair actions; The initial model sample data was obtained by analyzing historical maintenance work order data and obtaining the health status feature vector of the hydraulic system corresponding to each historical maintenance work order, including maintenance timing, maintenance action sequence, and maintenance result evaluation as model sample data. The loss function of the maintenance decision model is constructed based on weighted cross-entropy and preference loss, and is as follows: in, The weighted cross-entropy is calculated using the following formula: The amount of data in the model sample data. The maintenance actions taken for the model sample data of the eth item. The timing of maintenance for the model sample data of the eth example. Let e be the health state feature vector of the hydraulic system corresponding to the e-th model sample data. The score for evaluating the repair results of the e-th model sample data. In model parameters The model then performs maintenance. The probability, In model parameters The model then performs maintenance. The probability of; The formula for calculating preference loss is as follows: This represents the total number of maintenance actions. This represents the total number of types of maintenance opportunities. For maintenance actions Preference supervision value, For the timing of maintenance The preference supervision value, in the initial training mode, has no interaction preference, therefore it can make ; , To emphasize weight, satisfy The value is customized based on the emphasis on human-computer interaction experience. If the system operation requires adjustments based on the experience of the staff, then increase it. .
[0020] During training, with the goal of minimizing the loss function, the complex mapping relationship between multiple fully connected layers of the information processing layer is trained through model sample data. The information processing layer outputs the probability distribution of each maintenance action and maintenance timing based on the health status feature vector of the hydraulic system. The output layer takes the maintenance action and maintenance timing with the highest probability as the optimal maintenance timing and maintenance action output based on the probability distribution of each maintenance action and maintenance timing output by the information processing layer. The preventive maintenance decision sequence generated by the maintenance decision model is output to the human-computer interaction terminal for approval and verification. After approval and verification, the preventive maintenance decision sequence is output to the execution terminal to execute the preventive maintenance action. This step is a preliminary data collection step to obtain the experience preferences of the staff. The preventive maintenance decision sequence, which includes maintenance actions and timing, is generated by the maintenance decision model. After being approved and verified by the staff through the human-computer interaction terminal, it represents the staff's preference attributes for the maintenance actions and timing to be taken. Subsequent analysis is based on these preference attributes to obtain decision preference supervision data. Record the execution results of this preventive maintenance action, and feed the execution results of this preventive maintenance action and the approved and verified preventive maintenance decision sequence as incremental data to the maintenance decision model for online fine-tuning, so as to complete the closed-loop collaborative iteration of the maintenance decision model; Reference Figure 3 As shown, the closed-loop collaborative iteration of the maintenance decision-making model specifically includes: Design an incremental learning data space, in which the execution result of each preventive maintenance action and the preventive maintenance decision sequence after approval and verification are written into the incremental learning data space. When the data in the incremental learning data space reaches the preset incremental training threshold, the data in the incremental learning data space is added to the model sample data of the maintenance decision model to learn and optimize the complex mapping relationship between multiple fully connected layers of the information processing layer, and complete a closed-loop collaborative iteration. To control the sample size and prevent the model from overfitting to the running data and forgetting the initial training experience, a data forgetting mechanism is adopted in the closed-loop collaborative iteration process. The data forgetting mechanism is as follows: Design an incremental iteration threshold. When the number of times the data in the incremental learning data space added to the model sample data is less than or equal to the incremental iteration threshold, the data in the incremental learning data space is directly added to the model sample data to learn and optimize the complex mapping relationship between multiple fully connected layers of the information processing layer. When the number of times data in the incremental learning data space added to the model sample data exceeds the incremental iteration threshold, the data in the incremental learning data space that was added to the model sample data the longest time is removed from the model sample data, and the data in the current incremental learning data space is added to the model sample data to learn and optimize the complex mapping relationship between multiple fully connected layers of the information processing layer.
[0021] By designing an incremental learning data space and a closed-loop collaborative iteration mechanism, the execution results of each preventive maintenance operation and the decision sequence after approval and verification are fed back to the maintenance decision model in real time as incremental data for online fine-tuning. This allows the model to continuously absorb actual field experience and dynamically adapt to individual equipment differences and changes in operating conditions, effectively overcoming the problem of performance degradation of traditional static models over time. At the same time, a data forgetting mechanism is introduced. By setting an incremental iteration threshold, the earliest batch of historical incremental data is automatically removed while ensuring sufficient learning of new knowledge. This precisely controls the model sample size and prevents the model from forgetting the general experience of the initial training stage due to excessive focus on recent data. On the basis of maintaining model stability, this achieves efficient and sustainable self-evolution, significantly improving the long-term accuracy of maintenance decisions.
[0022] Based on multiple approved and verified preventive maintenance decision sequences, decision preference supervision data is constructed. The maintenance decision model is then fine-tuned using the decision preference supervision data to complete the collaborative iteration of the maintenance decision model's preferences.
[0023] Specifically, the preference-based collaborative iteration of the maintenance decision-making model includes: A preference analysis data space is constructed based on multiple approved and verified preventive maintenance decision sequences. Preference supervision values regarding maintenance timing and maintenance actions are extracted based on the preference analysis data space. The preference supervision values of maintenance timing and maintenance actions are introduced as preference supervision data into the loss function based on weighted cross-entropy and preference loss to obtain the updated loss function; Based on the model sample data, with the goal of minimizing the updated loss function, the complex mapping relationship between multiple fully connected layers of the information processing layer is trained to complete the preference collaborative iteration.
[0024] Among them, since the preventive maintenance decision sequence after personnel adjustment can better reflect the staff's preferences for maintenance actions and maintenance events, when extracting the preference supervision values for maintenance timing and maintenance actions based on the preference analysis data space, the adjusted preventive maintenance decision sequence in the preference analysis data space is identified as the preference-emphasized sequence, and the weight of the preference-emphasized sequence is increased when extracting the preference supervision values for maintenance timing and maintenance actions. Specifically, the method for obtaining the supervised values for maintenance timing and maintenance actions is as follows: The preference emphasis sequence includes maintenance actions. The number of sequences, Maintenance actions are included in the non-preference-focused sequence. The number of sequences, The preference emphasis sequence includes maintenance timing. The number of sequences, Maintenance timing is included in the non-preference-focused sequence. The number of sequences, To favor the total number of sequences, The total number of non-preference-based sequences. The weight is set to a value greater than 1, and in some preferred embodiments the weight is set to 1.5. By constructing a preference analysis data space, preference supervision values regarding maintenance timing and actions are extracted from multiple approved and verified preventive maintenance decision sequences. These values are then introduced as preference supervision data into the joint loss function of weighted cross-entropy and preference loss. This allows the maintenance decision model to further learn from and integrate the maintenance preferences of on-site personnel based on historical experience. For decision sequences that have been manually adjusted, their proportion in the calculation of preference supervision values is increased by setting emphasis weights. This enables the model to more accurately capture the subjective tendencies of personnel regarding specific maintenance timings and actions, thereby automatically outputting solutions that better fit actual operating habits in subsequent decisions. This improves the efficiency of human-machine collaboration and the acceptability of decision solutions, and enhances the integration between model output and on-site experience.
[0025] Example 2: This embodiment proposes a preventive maintenance system for pump truck hydraulic systems based on historical maintenance data, including: The data acquisition module is used to collect historical maintenance work order data, pump truck hydraulic system operating parameters and sensor monitoring data, extract degradation characteristics of key components, and construct a state space characterizing the health status of the hydraulic system. The maintenance decision module, built on a neural network, takes the current health status feature vector of the hydraulic system as input and outputs a preventive maintenance decision sequence that includes maintenance timing and maintenance actions. The human-computer interaction approval module is used to receive and display the preventive maintenance decision sequence output by the maintenance decision module, for operators to approve and verify, and output the approved and verified preventive maintenance decision sequence to the execution end; The closed-loop collaborative iteration module is used to record the execution results of this preventive maintenance action, and feeds the execution results and the approved and verified preventive maintenance decision sequence back to the maintenance decision module as incremental data for online fine-tuning. The preference collaborative iteration module is used to construct decision preference supervision data based on multiple approved and verified preventive maintenance decision sequences, and to fine-tune the maintenance decision module with this decision preference supervision data.
[0026] The closed-loop collaborative iteration module includes: The incremental data storage unit is used to store the execution results of each preventive maintenance action and the preventive maintenance decision sequence after approval and verification. The incremental training trigger unit is used to determine whether the amount of data in the incremental data storage unit has reached the preset incremental training threshold. If so, it triggers the maintenance decision module to perform incremental learning. The data forgetting control unit is used to determine whether to remove the oldest incremental data from the model sample data based on a preset incremental iteration threshold when the maintenance decision module is performing incremental learning, so as to control the size of the model sample data.
[0027] The preference collaborative iteration module includes: The preference analysis unit is used to construct a preference analysis data space based on multiple approved and verified preventive maintenance decision sequences, and extract preference supervision values for maintenance timing and maintenance actions from it. At the same time, it identifies the manually adjusted preventive maintenance decision sequences as preference emphasis sequences and increases the weight of these preference emphasis sequences. The loss function update unit is used to introduce the preference supervision values extracted by the preference analysis unit into the loss function of the maintenance decision module, and generate an updated loss function for the maintenance decision module to perform preference fine-tuning training.
[0028] In summary, the advantages of this invention are as follows: By constructing a maintenance decision-making model through a neural network, the optimal maintenance timing and actions are automatically output using health status feature vectors as input, significantly improving the scientific and real-time nature of maintenance decisions. This avoids resource waste caused by traditional periodic maintenance and unplanned downtime resulting from post-maintenance. The introduction of a closed-loop collaborative iteration mechanism feeds the results of each execution and the approved decision sequence as incremental data back to the model for online fine-tuning, enabling the model to continuously adapt to individual differences and changes in operating conditions within the pump truck hydraulic system. This effectively overcomes the performance degradation problem of traditional static models. Through a preference collaborative iteration module, preference supervision data is constructed using multiple approved decision sequences to fine-tune the model's preferences, allowing the model to learn and integrate the maintenance experience and preferences of different operators. This improves the human-machine collaborative acceptance and practicality of the decision-making scheme. The combination of a data forgetting mechanism to control the sample size ensures both rapid absorption of new knowledge and prevents interference from old data, maintaining the model's efficient and stable operation. In conclusion, this invention achieves intelligent, adaptive, and personalized maintenance decisions for pump truck hydraulic systems, improving the availability and maintenance economy of pump truck hydraulic systems.
[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. A method for preventive maintenance of a pump truck hydraulic system based on historical repair data, characterized by, include: Collect historical maintenance work order data, pump truck hydraulic system operating parameters and sensor monitoring data, extract degradation characteristics of key components, and construct a state space characterizing the health status of the hydraulic system; A maintenance decision model is constructed based on neural networks. The model takes the current health status feature vector of the hydraulic system as input and outputs a preventive maintenance decision sequence that includes maintenance timing and maintenance actions. The preventive maintenance decision sequence generated by the maintenance decision model is output to the human-computer interaction terminal for approval and verification. After approval and verification, the preventive maintenance decision sequence is output to the execution terminal to execute the preventive maintenance action. Record the execution results of this preventive maintenance action, and feed the execution results of this preventive maintenance action and the approved and verified preventive maintenance decision sequence as incremental data to the maintenance decision model for online fine-tuning, so as to complete the closed-loop collaborative iteration of the maintenance decision model; Based on multiple approved and verified preventive maintenance decision sequences, decision preference supervision data is constructed. The maintenance decision model is then fine-tuned using the decision preference supervision data to complete the collaborative iteration of the maintenance decision model's preferences.
2. The method of claim 1, wherein, The process of collecting historical maintenance work order data, operating parameters, and sensor monitoring data of the pump truck hydraulic system, extracting degradation characteristics of key components, and constructing a state space characterizing the health status of the hydraulic system specifically includes: For each key component in the pump truck hydraulic system, extract at least one typical maintenance event and the pump truck hydraulic system operating characteristics within the window period before the typical maintenance event occurs based on its historical maintenance work order data. Based on the operating parameters of the pump truck hydraulic system and the sensor monitoring data, the real-time operating characteristics of the pump truck hydraulic system are extracted. The real-time operating characteristics of the pump truck hydraulic system are matched with the operating characteristics of the pump truck hydraulic system within the window period before all typical maintenance events of each key component. If the match is successful, the key component is marked with the typical maintenance event; otherwise, no response is made, and the health status feature vector of each key component is obtained. By summing the health status feature vectors of all key components, the state space of the hydraulic system's health status is obtained.
3. The method of claim 2, wherein, The initial training steps for the maintenance decision model are as follows: Analyze historical maintenance work order data to obtain the health status feature vector of the hydraulic system corresponding to each historical maintenance work order, including maintenance timing, maintenance action sequence, and maintenance result evaluation as model sample data; A neural network structure consisting of an input layer, an information processing layer, and an output layer is constructed. The information processing layer consists of multiple fully connected layers and includes a preference calculation branch. In the initial training step, the system's preference supervision data is zero. The input layer receives the health status feature vector of the hydraulic system, and the output layer outputs the optimal maintenance time and maintenance action. A loss function based on weighted cross-entropy and preference loss is constructed. With the goal of minimizing the loss function, the complex mapping relationship between multiple fully connected layers of the information processing layer is trained through model sample data to obtain the maintenance decision model in the initial state.
4. The preventive maintenance method for pump truck hydraulic systems based on historical maintenance data according to claim 3, characterized in that, The closed-loop collaborative iterative process of the maintenance decision-making model is as follows: Design an incremental learning data space, in which the execution result of each preventive maintenance action and the preventive maintenance decision sequence after approval and verification are written into the incremental learning data space. When the data in the incremental learning data space reaches the preset incremental training threshold, the data in the incremental learning data space is added to the model sample data of the maintenance decision model to learn and optimize the complex mapping relationship between multiple fully connected layers of the information processing layer, thus completing a closed-loop collaborative iteration.
5. The preventive maintenance method for pump truck hydraulic systems based on historical maintenance data according to claim 4, characterized in that, The closed-loop collaborative iterative process includes a data forgetting mechanism, which specifically includes: Design an incremental iteration threshold. When the number of times the data in the incremental learning data space added to the model sample data is less than or equal to the incremental iteration threshold, the data in the incremental learning data space is directly added to the model sample data to learn and optimize the complex mapping relationship between multiple fully connected layers of the information processing layer. When the number of times data in the incremental learning data space added to the model sample data exceeds the incremental iteration threshold, the data in the incremental learning data space that was added to the model sample data the longest time is removed from the model sample data, and the data in the current incremental learning data space is added to the model sample data to learn and optimize the complex mapping relationship between multiple fully connected layers of the information processing layer.
6. The preventive maintenance method for pump truck hydraulic systems based on historical maintenance data according to claim 5, characterized in that, The preference collaborative iteration of the maintenance decision model specifically includes: A preference analysis data space is constructed based on multiple approved and verified preventive maintenance decision sequences. Preference supervision values regarding maintenance timing and maintenance actions are extracted based on the preference analysis data space. The preference supervision values of maintenance timing and maintenance actions are introduced as preference supervision data into the loss function based on weighted cross-entropy and preference loss to obtain the updated loss function; Based on the model sample data, with the goal of minimizing the updated loss function, the complex mapping relationship between multiple fully connected layers of the information processing layer is trained to complete the preference collaborative iteration.
7. The preventive maintenance method for pump truck hydraulic systems based on historical maintenance data according to claim 6, characterized in that, When extracting preference supervision values for maintenance timing and maintenance actions based on the preference analysis data space, an adjusted preventive maintenance decision sequence in the preference analysis data space is identified as a preference-emphasized sequence. When extracting preference supervision values for maintenance timing and maintenance actions, the weight of the preference-emphasized sequence is increased.
8. A preventive maintenance system for pump truck hydraulic systems based on historical maintenance data, characterized in that: A method for implementing the preventive maintenance method for a pump truck hydraulic system based on historical maintenance data as described in any one of claims 1-7 includes: The data acquisition module is used to collect historical maintenance work order data, pump truck hydraulic system operating parameters and sensor monitoring data, extract degradation characteristics of key components, and construct a state space characterizing the health status of the hydraulic system. The maintenance decision module, built on a neural network, takes the current health status feature vector of the hydraulic system as input and outputs a preventive maintenance decision sequence that includes maintenance timing and maintenance actions. The human-computer interaction approval module is used to receive and display the preventive maintenance decision sequence output by the maintenance decision module for operators to approve and verify, and output the approved and verified preventive maintenance decision sequence to the execution end; The closed-loop collaborative iteration module is used to record the execution result of this preventive maintenance action, and feeds the execution result and the approved and verified preventive maintenance decision sequence as incremental data to the maintenance decision module for online fine-tuning; The preference collaborative iteration module is used to construct decision preference supervision data based on multiple approved and verified preventive maintenance decision sequences, and to fine-tune the maintenance decision module with the decision preference supervision data.
9. The preventive maintenance system for pump truck hydraulic systems based on historical maintenance data according to claim 8, characterized in that, The closed-loop collaborative iteration module includes: The incremental data storage unit is used to store the execution results of each preventive maintenance action and the preventive maintenance decision sequence after approval and verification. An incremental training triggering unit is used to determine whether the amount of data in the incremental data storage unit has reached a preset incremental training threshold. If so, the maintenance decision module is triggered to perform incremental learning. The data forgetting control unit is used to determine whether to remove the oldest incremental data from the model sample data based on a preset incremental iteration threshold when the maintenance decision module performs incremental learning, so as to control the size of the model sample data.
10. The preventive maintenance system for pump truck hydraulic systems based on historical maintenance data according to claim 8, characterized in that, The preference collaborative iteration module includes: The preference analysis unit is used to construct a preference analysis data space based on multiple approved and verified preventive maintenance decision sequences, and extract preference supervision values for maintenance timing and maintenance actions from it. At the same time, it identifies the manually adjusted preventive maintenance decision sequences as preference emphasis sequences and increases the weight of these preference emphasis sequences. The loss function update unit is used to introduce the preference supervision value extracted by the preference analysis unit into the loss function of the maintenance decision module to generate an updated loss function for the maintenance decision module to perform preference fine-tuning training.