Predictive maintenance management system based on big data

Abstract

The invention belongs to the technical field of predictive maintenance management of a big data technology, and particularly relates to a predictive maintenance management system based on big data, which comprises a data acquisition module, an intelligent preprocessing module, a self-adaptive feature fusion module, a hybrid prediction model module, a scenarized decision-making module and a full-cycle optimization module. The data acquisition module acquires real-time states, environmental parameters and full-life-cycle associated data of key components; the intelligent preprocessing module performs hierarchical processing on the data optimization data; the adaptive feature fusion module dynamically fuses data by using an improved attention mechanism in combination with a component fault conduction rule; the hybrid prediction model module constructs a double-layer model; the scenarized decision-making module builds the model according to scenes, and a maintenance scheme is generated through multi-objective optimization; and the full-cycle optimization module is used for reinforcing learning closed-loop iteration. According to the method, elevator fault accurate prediction, scenarized intelligent decision making and full-period self-optimization are achieved, the operation and maintenance efficiency and reliability are improved, and elevator maintenance is promoted to be converted from passive response to active prediction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of predictive maintenance management of big data technology, and particularly relates to a predictive maintenance management system based on big data. BACKGROUND

[0002] With the acceleration of urbanization and the popularization of high-rise buildings, elevators have become a key vertical transportation tool, and their operation reliability is directly related to personnel safety and travel efficiency.

[0003] However, the existing elevator maintenance system relies on single sensor data for abnormality judgment, and can only identify that "the elevator has an abnormality", but cannot accurately locate the faulty component. For example, when the vibration value is detected to be excessive, it is difficult to distinguish whether the abnormality is caused by the bearing wear of the traction machine, the loosening of the guide rail, or the tension of the steel wire rope, and the maintenance personnel need to check one by one on site, and the average fault positioning time is more than 2 hours, which seriously prolongs the downtime. At the same time, the maintenance decision lacks scene and data support: the traditional maintenance plan adopts a fixed cycle mode, without considering the difference in elevator use intensity, resulting in the coexistence of "over-maintenance" and "insufficient maintenance".

[0004] The Chinese invention patent with the application number CN202411642426.7 discloses an elevator early warning method based on multi-modal feature fusion in a big data environment, comprising the following steps: step S1: collecting video feature data, voiceprint feature data and operation feature data in the elevator operation process; step S2: video feature extraction, using a video feature extraction model to extract a video feature vector of a key video frame from the collected video feature data; step S3: voiceprint feature extraction, using a mel-frequency cepstral coefficient to extract a voiceprint feature vector from the collected voiceprint feature data. The present application solves the problems of low recognition rate and high false alarm rate of single modal data in traditional methods by fusing video feature data, voiceprint feature data and operation feature data, improves the accuracy, reliability and robustness of early warning, and can realize accurate prediction of elevator operation risk, thereby effectively improving the quality and safety of the elevator through preventive maintenance measures.

[0005] However, the above-mentioned invention can warn about abnormalities through multi-modal features, but does not explicitly associate specific faulty components and does not form component-level positioning logic. In terms of maintenance decision, the early warning result does not consider the downtime cost difference of the elevator use scene, and lacks dynamic planning of maintenance priority and time window. SUMMARY

[0006] In view of the above-mentioned deficiencies in the prior art, the present application provides a predictive maintenance management system based on big data to solve the problems in the background art.

[0007] In order to solve the above technical problems, the present application adopts the following technical solutions:

[0008] A predictive maintenance management system based on big data, comprising, which are sequentially data-connected:

[0009] a data acquisition module, which acquires multi-dimensional operation data of the elevator through a distributed sensor network;

[0010] an intelligent preprocessing module, which performs hierarchical processing on the multi-dimensional operation data to obtain optimized data, including: performing noise filtering and feature extraction on real-time state data; performing format standardization on structured data; marking abnormal data and triggering a priority processing mechanism;

[0011] an adaptive feature fusion module, which realizes dynamic fusion of the optimized data by using an improved attention mechanism, allocates initial feature weights based on component fault influence degree, dynamically adjusts the weight allocation strategy in combination with real-time early warning accuracy, and establishes component correlation rules to form a comprehensive feature data set;

[0012] a hybrid prediction model module, which constructs a double-layer prediction model that fuses physical constraints and data-driven based on the comprehensive feature data set: including a time series prediction layer and a fault location layer; the time series prediction layer uses an LSTM time series network with mechanical parameter constraints, introduces key component design parameters as boundary conditions, and predicts the remaining useful life; the fault location layer optimizes the classification algorithm in combination with elevator mechanical structure knowledge to realize accurate identification of fault type, risk level and specific components;

[0013] a scenario-based decision module, which constructs a dynamic maintenance decision system based on the comprehensive feature data set: establishes a differentiated cost evaluation model according to use scenarios, combines the real-time state of maintenance resources, and generates a maintenance priority ranking and time window scheme that adapts to the scenario through a multi-objective optimization algorithm;

[0014] a full-cycle optimization module, which continuously iterates system performance through a closed-loop feedback mechanism, records maintenance execution results and actual fault data, integrates common model elevators to share cases, reversely optimizes feature weights and model parameters using reinforcement learning algorithm, improves the prediction ability of newly input elevators and rare faults through transfer learning, and forms a full-cycle optimization closed loop of data, model, decision and feedback.

[0015] Further, the multi-dimensional operation data includes real-time state data of the traction machine, door machine system, guide rail, steel wire rope and control cabinet, environmental parameters and full life cycle associated data; the environmental parameters include car interior temperature and humidity, load change data; the full life cycle associated data includes control system operation log, historical fault maintenance record, same model elevator fault case and installation and debugging parameters.

[0016] Further, the real-time state data includes traction machine vibration spectrum, door machine system voiceprint signal, guide rail vibration acceleration, steel wire rope tension, control cabinet current and voltage; the load change data is specifically real-time load value, load fluctuation amplitude and daily average load peak value.

[0017] Further, the noise filtering performed by the intelligent preprocessing module on the real-time state data adopts a dynamic threshold adaptive mechanism: for the traction machine vibration signal, the 10-500Hz frequency band is taken, and for the door machine system voiceprint signal, the 200-2000Hz frequency band is taken, the standard deviation and the mean value of the signal are calculated in a 3-minute sliding window, the filter threshold is generated by taking the mean value of the signal plus twice the standard deviation, and the pulse interference exceeding the threshold is smoothed by a 5-point sliding average method.

[0018] Further, the feature extraction of the intelligent preprocessing module includes: performing Fourier frequency domain conversion on the high-frequency vibration signal of the traction machine, generating a 128-dimensional MFCC feature vector for the door machine system voiceprint signal, and extracting the joint features of the time domain peak value and the frequency domain main frequency value of the guide rail vibration acceleration signal; the priority processing mechanism includes re-acquiring and verifying abnormal data, and increasing the weight of the data in feature extraction.

[0019] Further, the way of adjusting the attention weight by combining the improved attention mechanism with the component correlation rule is: based on the historical fault co-occurrence data, the probability that component i failure is transmitted to component j is calculated and N ij is the number of simultaneous failures of component j after the failure of component i, N i is the total number of failures of component i, and a threshold value p th is set; if p ij >= p th , it is determined that components i and j are strongly correlated, and a rule is established, and the features of the strongly correlated components are given an additional weight according to Domega ij = p ij * omega base * gamma, omega base takes 0.1 to 0.3, and gamma is a correction coefficient, gamma = 1.2 for strongly correlated component features, and otherwise gamma = 0.8; based on the component failure influence degree, the initial weight is obtained by the analytic hierarchy process, and then the dynamic weight is iteratively optimized according to , wherein w i (t) is the weight of the i-th type of data at time t, k is an adjustment coefficient, and takes a value between 0.5 and 1.5, A i (t) is the early warning accuracy of the i-th type of data at time t, A avg (t) is the average early warning accuracy of the system at time t, and the weight iteration optimization is completed every 24 hours, and the comprehensive feature data set is output by fusing the initial weight, the dynamic weight and the additional weight.

[0020] Further, the key component design parameters in the time sequence prediction layer include traction machine design fatigue life, door machine motor rated power, guide rail safety vibration threshold, steel wire rope rated tension, control cabinet voltage fluctuation range;The improved random forest algorithm fusing elevator mechanical topology knowledge is adopted in the fault positioning layer, and the correlation graph of fault-component-feature is constructed: the mapping relationship between the traction machine bearing wear and the specific frequency band vibration, the association rule between the door machine transmission mechanism jam and the voiceprint signal mutation are taken as priori knowledge, the priori knowledge is quantified as the decision tree splitting constraint condition, the decision tree splitting criterion based on the mechanical topology correlation degree is adopted, and the fault type is output simultaneously with the specific fault component and position.

[0021] Further, the dynamic maintenance decision system establishes a downtime cost sensitive evaluation model according to the use scene, including:

[0022] Downtime loss coefficient: distinguish commercial, residential and medical scenes, and set differential coefficients according to the downtime loss level of the scene;

[0023] Fault risk coefficient: the calculation formula is R=P·S·W, wherein P is the fault occurrence probability, S is the consequence severity, and W is the component safety weight;

[0024] Maintenance resource cost coefficient: the calculation formula is C=α·C_d+β·C_s, wherein α and β are cost weight coefficients, and α+β=1, C_d is distance cost, and C_s is inventory cost;

[0025] The multi-objective optimization algorithm calculates the maintenance priority through the formula Score=ω1·R_r+ω2·C_s+ω3·S_a, wherein ω1, ω2 and ω3 are target weight coefficients, and ω1+ω2+ω3=1, and the target weight coefficients are dynamically assigned according to the elevator safety requirement and cost sensitivity;R_r is the standardized value corresponding to the fault risk reduction rate;C_s is the standardized value corresponding to the maintenance cost saving rate;S_a is the standardized value corresponding to the scene adaptation degree.

[0026] Further, the full-cycle optimization module adopts the reinforcement learning framework of Q-learning, constructs a state-action-reward closed loop, takes feature weight and model parameter as state space, takes prediction accuracy improvement value as core reward and introduces false alarm rate constraint, and updates the system through iteration.

[0027] Compared with the prior art, the present application has the following beneficial effects:

[0028] 1. Constructing a double-layer prediction model of component fault conduction association rule and mechanical topology knowledge fusion, accurately associating component life cycle parameters and residual life in time series prediction layer, and realizing accurate mapping of fault type to "component-position" in fault positioning layer relying on mechanical topology correlation decision tree, so as to improve the accuracy of fault component positioning and solve the problem of disconnection between abnormal early warning and specific fault component.

[0029] 2. The scene decision module first creates a "shutdown cost sensitive + multi-objective dynamic optimization" mechanism, distinguishes the shutdown loss coefficients of commercial, medical and other scenes, dynamically plans maintenance priority and time window combined with the real-time state of maintenance resources, shortens the fault response time in medical scenes, and reduces unnecessary shutdown loss in commercial scenes, realizes the full-link breakthrough from "abnormal early warning" to "component-level accurate positioning + scene dynamic maintenance decision", and promotes the predictive maintenance of the elevator from "finding abnormalities" to "accurate tracing and intelligent decision-making". BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The structure diagram of the predictive maintenance management system based on big data. DETAILED DESCRIPTION

[0031] In order to enable those skilled in the art to better understand the present application, the technical solutions of the present application are further described below in conjunction with the drawings and examples.

[0032] Among them, the drawings are only used for illustrative description, and the representation is only a schematic diagram, not a physical diagram, and cannot be understood as a limitation on the patent; in order to better illustrate the embodiments of the present application, some components of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some known structures and their descriptions in the drawings may be omitted.

[0033] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the terms "upper", "lower", "left", "right", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationship in the drawings are only used for illustrative description, and cannot be understood as a limitation on the patent, for those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0034] In the description of the present application, unless otherwise explicitly specified and limited, if the term "connection" or the like appears to indicate the connection relationship between components, the term should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two components or the interaction relationship between two components. For those skilled in the art, the specific meaning of the above-mentioned terms in the present application can be understood according to the specific circumstances.

[0035] As shown in the present application. Figure 1 Specifically, a predictive maintenance management system based on big data includes a data acquisition module, an intelligent preprocessing module, an adaptive feature fusion module, a hybrid prediction model module, a scenario decision module, and a full-cycle optimization module connected in turn.

[0036] In this embodiment, in the data acquisition module, the system collects multi-dimensional operation data of the elevator through a distributed sensor network. Sensors are installed at key parts of the elevator such as the traction machine, the door machine system, the guide rail, the steel wire rope, and the control cabinet. For example, a vibration sensor is installed on the traction machine to obtain vibration spectrum data, which can reflect the running stability of the traction machine. Abnormal vibration may indicate bearing wear and other faults; an acoustic fingerprint sensor is installed on the door machine system to collect acoustic fingerprint signals. Changes in acoustic fingerprints can indicate whether the door machine transmission mechanism is stuck; a vibration acceleration sensor is arranged on the guide rail to monitor the guide rail vibration acceleration and ensure smooth operation of the elevator; a tension sensor is used to measure the tension of the steel wire rope to ensure safe operation of the steel wire rope; and a voltage and current sensor is used to collect control cabinet current and voltage data to understand the working state of the control cabinet.

[0037] At the same time, the temperature and humidity sensor data in the car are collected to understand the temperature and humidity conditions of the elevator operating environment. Extreme temperature and humidity may affect the performance of elevator electronic components. Load change data is collected by a weighing sensor to obtain real-time load values, load fluctuation amplitudes, and daily average load peak values for analyzing the elevator load situation. In terms of full life cycle related data, the system is connected with the elevator control system to obtain control system operation logs containing elevator operating status, fault records, etc.; historical fault maintenance records are collected to record the time of each fault occurrence, fault phenomena, maintenance measures, etc. to provide a basis for fault analysis; similar elevator fault cases are integrated to learn from similar elevator fault experience; installation and debugging parameters such as elevator rated speed, load, etc. are obtained as basic data for subsequent analysis.

[0038] Further, in this embodiment, the intelligent preprocessing module processes the multi-dimensional operation data in layers to obtain optimized data. This includes real-time state data processing, structured data processing, and abnormal data processing.

[0039] In real-time state data processing, the 10-500Hz frequency band is selected for the traction machine vibration signal, the standard deviation and mean value of the signal are calculated in a 3-minute sliding window, and a filter threshold is generated by adding twice the standard deviation to the signal mean value. The 5-point moving average method is used to smooth the pulse interference that exceeds the threshold to ensure the accuracy of the vibration signal. The 200-2000Hz frequency band is selected for the door machine system voiceprint signal, and the same processing is performed to ensure the purity of the voiceprint signal. Fourier frequency domain conversion is performed on the high-frequency vibration signal of the traction machine to extract the vibration frequency characteristics, and different frequencies correspond to different fault types. The door machine system voiceprint signal generates a 128-dimensional MFCC feature vector to identify the running state. The joint features of time domain peak value and frequency domain main frequency value are extracted from the rail vibration acceleration signal to fully reflect the vibration characteristics.

[0040] Secondly, for structured data such as elevator operation parameters and equipment archives, the unified data format includes type, length, and coding method, etc. For example, the elevator running time data from different sources is unified to the "year-month-day hour: minute: second" format to ensure data consistency and compatibility.

[0041] At the same time, when abnormal data is detected, the system automatically marks and triggers the priority processing mechanism, re-collects and verifies the abnormal data, reads the sensor data again or compares and confirms the authenticity from other data sources. If the abnormality is confirmed, the weight of the data in feature extraction is increased, and attention is focused on to discover potential fault hidden dangers.

[0042] In this embodiment, the adaptive feature fusion module uses an improved attention mechanism to realize dynamic data fusion.

[0043] Based on the historical fault co-occurrence data of the elevator throughout its life cycle, a dynamic component fault transmission network is constructed to calculate the component fault transmission probability p ij , which is the probability of component i fault transmission to component j. The calculation formula is:

[0044] where N ij is the number of component j faults occurring at the same time after component i fault, and N i is the total number of component i faults. The system sets the fault transmission strength threshold p th , with a range of 0.2 to 0.4. When p ij ≥ p th , it is determined that there is a strong correlation transmission relationship between components i and j, and the corresponding component correlation rule is established.

[0045] The initial weight w The allocation is based on the impact of component failures, quantitatively assessed from three dimensions: safety risk level, maintenance cost, and downtime loss. The safety risk level is assigned a score of 1-5 based on the degree of threat the failure poses to personnel safety (higher threats result in higher scores). Maintenance cost is based on the average historical maintenance cost for that component (in yuan). Downtime loss is calculated based on the average daily economic loss caused by the failure (in yuan / day). A judgment matrix is ​​established using the Analytic Hierarchy Process (AHP), and the initial weights of each component are calculated using normalization. Satisfy the formula

[0046] Dynamic weight w i (t) The accuracy of the early warning output based on the fault location layer is iteratively optimized using the following formula:

[0047]

[0048] Among them, w i (t) represents the weight of the i-th type of data at time t, and k is an adjustment coefficient, ranging from 0.5 to 1.5. In elevators with high failure rates or high safety levels (such as medical elevators), k is automatically adjusted to ≥1.2; A i (t) represents the real-time early warning accuracy of the i-th type of data, A avg (t) represents the system's average early warning accuracy, ω i (t-1) represents the historical weights at time t-1, introduced by introducing initial weights. As a regularization term, it prevents the weights from deviating excessively from the physical priors. The iteration cycle is fixed at once every 24 hours to ensure that the weights are dynamically updated according to the elevator's operating status.

[0049] For component features with strong correlation and transmission relationships, the system uses the formula Δω ij =p ij · ω base ·γ assigns dynamic additional attention weights Δω ij , where ω base The base additional weight has a value between 0.1 and 0.3, with a default value of ≥0.2 for medical elevator scenarios; γ is the correlation strength correction coefficient, when ρ ij ≥ρ th γ is 1.2 when the condition is met, and γ is 0.8 otherwise. Initial weights for fusion. Dynamic iterative weight w i (t) and additional attention weights Δω ij This forms a comprehensive feature dataset, providing multi-dimensional feature inputs for the hybrid prediction model module.

[0050] In the present embodiment, the hybrid prediction model module constructs a double-layer prediction model based on the comprehensive feature dataset. The time series prediction layer adopts an LSTM time series network with mechanical parameter constraints, introduces key component design parameters such as elevator design fatigue life, door machine motor rated power, guide rail safety vibration threshold, steel wire rope rated tension, control cabinet voltage fluctuation range as boundary conditions, and inputs the LSTM network with elevator operation time series data, learns the relationship between time series characteristics and mechanical parameters in historical data, and predicts the remaining service life of key components, such as predicting the time that the traction machine can still work normally according to the historical vibration and temperature data of the traction machine and the design fatigue life parameter. The fault positioning layer adopts an improved random forest algorithm that integrates elevator mechanical topology knowledge, constructs a fault-component-feature correlation graph, embeds correlation rules such as the correlation between traction machine bearing wear and specific frequency band vibration, and the correlation between door machine transmission mechanism jamming and voiceprint signal mutation as prior knowledge, quantifies the prior knowledge as decision tree splitting constraint conditions, adopts a decision tree splitting criterion based on mechanical topology correlation degree, and after inputting the comprehensive feature dataset, the decision tree splits according to the feature data and prior knowledge, and outputs the fault type, risk level and specific component location, such as when the voiceprint signal specific frequency band mutation occurs, the transmission mechanism component jamming is judged and the fault component is located by combining the door machine mechanical topology structure and the prior knowledge.

[0051] In this embodiment, the scenario decision module constructs a dynamic maintenance decision system based on the comprehensive feature dataset output by the adaptive feature fusion module and the fault prediction results of the hybrid prediction model module. A downtime cost sensitive evaluation model is established according to the use scenario. Commercial, residential, and medical scenarios are distinguished. The downtime of a commercial scenario affects business activities and causes significant economic losses. A higher downtime loss coefficient is set according to the business type and passenger flow. The residential scenario mainly affects the convenience of residents' life, and the coefficient is relatively low. The medical scenario uses elevators to transport patients and materials, and the reliability requirement is high, so an extremely high coefficient is set. The fault risk coefficient calculation formula is R=P×S×W, P is obtained from the prediction results of the hybrid prediction model module, S is evaluated according to the influence degree of the fault type on the operation safety and personnel safety, and W is determined according to the importance of the component in the safe operation. The maintenance resource cost coefficient calculation formula is C=α·C_d+β·C_s, where α and β are cost weight coefficients, and α+β=1. The actual maintenance resource configuration is adjusted, C_d is the distance cost, considering the distance cost of maintenance personnel and spare parts to the elevator location, and C_s is the inventory cost, including spare parts inventory management cost, etc. Through a multi-objective optimization algorithm, combined with the standardized value C_s corresponding to the fault risk reduction rate and the standardized value S_a corresponding to the scene adaptation degree, the maintenance priority is calculated using the formula Score=ω1·R_r+ω2·C_s+ω3·S_a, where ω1, ω2, and ω3 are target weight coefficients, and ω1+ω2+ω3=1. According to the elevator safety requirements and cost sensitivity, dynamic assignment is performed, such as increasing ω1 to prioritize reducing fault risk during the high-fault period of medical elevators, and adjusting ω2 to balance risk and cost in cost-sensitive places. Combined with the real-time state of maintenance resources, the maintenance priority ranking and time window scheme for the adapted scene are generated to determine the elevator maintenance time.

[0052] In this embodiment, the whole cycle optimization module adopts the Q-learning reinforcement learning framework to build a state-action-reward closed loop, takes feature weights and model parameters as state space, takes prediction accuracy improvement value as core reward and introduces false alarm rate constraint, gives positive reward to prediction accuracy improvement and negative reward to false alarm, and iteratively updates the system to make the early warning accuracy converge to a stable threshold within 3 months. During the iteration process, the maintenance execution results and actual fault data are recorded, the same type of elevator is shared, the feature weights and model parameters are optimized in reverse, the existing model knowledge is transferred to new input elevators and rare fault prediction through transfer learning technology, the prediction ability for new elevators and rare faults is improved, and a whole cycle optimization closed loop of data, model, decision, and feedback is formed to continuously improve the system predictive maintenance ability.

[0053] The above is only an embodiment of the present application, and relates to circuits and electronic components and modules, which are all prior art. Those skilled in the art can implement the present application without further description. The present application does not involve improvement of software and methods. Commonly known specific structures and characteristics in the scheme are not described in detail herein. Those skilled in the art know all common technical knowledge in the technical field of the present application before the filing date or the priority date, can know all prior art in the field, and have the ability to apply conventional experimental means before the date. Those skilled in the art can perfect and implement the present scheme based on their own ability under the guidance of the present application. Some typical known structures or known methods should not be an obstacle for those skilled in the art to implement the present application. It should be pointed out that, for those skilled in the art, a number of modifications and improvements can be made without departing from the structure of the present application. These should also be considered as the protection scope of the present application, and these will not affect the implementation effect and practicality of the patent.

Claims

1. A predictive maintenance management system based on big data, characterized in that: Including sequential data connections: The data acquisition module collects multi-dimensional operational data of the elevator through a distributed sensor network; The intelligent preprocessing module performs hierarchical processing on the multi-dimensional operational data to obtain optimized data, including: performing noise filtering and feature extraction on real-time status data; standardizing the format of structured data; and marking abnormal data and triggering a priority processing mechanism. The adaptive feature fusion module uses an improved attention mechanism to achieve dynamic fusion of optimized data. It assigns initial feature weights based on the impact of component failures and dynamically adjusts the weight allocation strategy in conjunction with the real-time early warning accuracy. At the same time, it establishes component association rules to form a comprehensive feature dataset. The hybrid prediction model module, based on the comprehensive feature dataset, constructs a two-layer prediction model that integrates physical constraints and data-driven approaches: including a time-series prediction layer and a fault location layer; the time-series prediction layer uses an LSTM time-series network with mechanical parameter constraints, introduces key component design parameters as boundary conditions, and predicts the remaining service life; the fault location layer combines elevator mechanical structure knowledge to optimize the classification algorithm, achieving accurate identification of fault type, risk level, and specific components; The scenario-based decision-making module constructs a dynamic maintenance decision-making system based on a comprehensive feature dataset: it establishes differentiated cost assessment models according to usage scenarios, combines the real-time status of maintenance resources, and generates maintenance priority ranking and time window schemes adapted to the scenarios through multi-objective optimization algorithms; The full-cycle optimization module continuously iterates system performance through a closed-loop feedback mechanism, records maintenance execution results and actual fault data, integrates shared cases of elevators of the same model, uses reinforcement learning algorithms to optimize feature weights and model parameters in reverse, and improves the predictive ability of newly put elevators and rare faults through transfer learning, forming a full-cycle optimization closed loop of data, model, decision and feedback.

2. The predictive maintenance management system based on big data as described in claim 1, characterized in that: The multi-dimensional operational data includes real-time status data, environmental parameters, and full lifecycle related data for the traction machine, door operator system, guide rails, wire ropes, and control cabinet; the environmental parameters include temperature and humidity inside the car and load change data; the full lifecycle related data includes control system operation logs, historical fault repair records, fault cases of elevators of the same model, and installation and commissioning parameters.

3. The predictive maintenance management system based on big data as described in claim 2, characterized in that: The real-time status data includes the traction machine vibration spectrum, the gantry crane system acoustic signal, the guide rail vibration acceleration, the wire rope tension, and the control cabinet current and voltage; the load change data specifically includes the real-time load value, the load fluctuation amplitude, and the daily average load peak value.

4. The predictive maintenance management system based on big data as described in claim 1, characterized in that: The intelligent preprocessing module performs noise filtering on real-time status data using a dynamic threshold adaptive mechanism: for the traction machine vibration signal, the frequency band is taken as 10-500Hz, and for the gantry crane system acoustic signal, the frequency band is taken as 200-2000Hz. The standard deviation and mean of the signal are calculated within a 3-minute sliding window. The filtering threshold is generated by adding twice the standard deviation to the signal mean. Pulse interference exceeding the threshold is smoothed using a 5-point moving average method.

5. The predictive maintenance management system based on big data as described in claim 1, characterized in that: The feature extraction of the intelligent preprocessing module includes: performing Fourier frequency domain transformation on the high-frequency vibration signal of the traction machine, generating a 128-dimensional MFCC feature vector on the acoustic fingerprint signal of the gantry crane system, and extracting the joint features of the time domain peak value and the frequency domain main frequency value from the guide rail vibration acceleration signal; the priority processing mechanism includes re-acquiring and verifying abnormal data and increasing the weight of the data in feature extraction.

6. The predictive maintenance management system based on big data as described in claim 1, characterized in that: The improved attention mechanism, which combines component association rules to adjust attention weights, works by: statistically analyzing the probability that a fault in component i will propagate to component j based on historical fault co-occurrence data. And N ij N represents the number of simultaneous failures of component j after component i fails. i Set a threshold ρ for the total number of failures of component i. th If ρ ij ≥ρ th Determine the strong association between components i and j and establish rules accordingly. Then, apply Δω to the features of strongly associated components. ij =ρ ij ·ω base ·γ assigns additional weights, ω base The value is set to 0.1 to 0.3, with γ as a correction coefficient. For strongly correlated component characteristics, γ = 1.2; otherwise, γ = 0.

8. Based on the impact of component failure, the initial weights are obtained using the analytic hierarchy process (AHP). Based on the accuracy of the fault location layer's early warning, according to Iterative optimization of dynamic weights, where w i (t) represents the weight of the i-th class of data at time t, and k is the adjustment coefficient, which takes a value between 0.5 and 1.

5. i (t) represents the warning accuracy of the i-th type of data at time t, A avg (t) represents the average warning accuracy of the system at time t. The weight is iterated and optimized every 24 hours, and the initial weight, dynamic weight and additional weight are combined to output a comprehensive feature dataset.

7. The predictive maintenance management system based on big data as described in claim 1, characterized in that: The key component design parameters in the time-series prediction layer include the design fatigue life of the traction machine, the rated power of the door operator motor, the safety vibration threshold of the guide rail, the rated tension of the wire rope, and the voltage fluctuation range of the control cabinet. The fault location layer adopts an improved random forest algorithm that integrates elevator mechanical topology knowledge to construct a fault-component-feature association graph: embedding the mapping relationship between traction machine bearing wear and vibration in a specific frequency band, and the association rules between door operator transmission mechanism jamming and abrupt changes in acoustic signals as prior knowledge, quantifying the prior knowledge into decision tree splitting constraints, and adopting a decision tree splitting criterion based on mechanical topology association degree, simultaneously associating specific faulty components and locations when outputting the fault type.

8. The predictive maintenance management system based on big data as described in claim 1, characterized in that: The dynamic maintenance decision-making system establishes a downtime cost-sensitive assessment model based on usage scenarios, including: Downtime loss coefficient: Differentiate between commercial, residential, and medical scenarios, and set differentiated coefficients based on the downtime loss level of each scenario; Failure risk coefficient: The calculation formula is R = P·S·W, where P is the probability of failure, S is the severity of the consequences, and W is the safety weight of the component. Maintenance resource cost coefficient: The calculation formula is: C=α·C_d+β·C_s, where α and β are cost weight coefficients, and α+β=1, C_d is the distance cost, and C_s is the inventory cost; The multi-objective optimization algorithm calculates maintenance priority using the formula Score=ω1·R_r+ω2·C_s+ω3·S_a, where ω1, ω2, and ω3 are objective weight coefficients, and ω1+ω2+ω3=1, dynamically assigned based on elevator safety requirements and cost sensitivity; R_r is the standardized value corresponding to the failure risk reduction rate; C_s is the standardized value corresponding to the maintenance cost saving rate; and S_a is the standardized value corresponding to the scenario adaptability.

9. The predictive maintenance management system based on big data as described in claim 1, characterized in that: The full-cycle optimization module adopts the Q-learning reinforcement learning framework to construct a state-action-reward closed loop; Using feature weights and model parameters as the state space, with the improvement in prediction accuracy as the core reward and introducing a false alarm rate constraint, the system's early warning accuracy converges to a stable threshold within 3 months through iterative updates, achieving full-cycle self-enhancement of predictive maintenance capabilities.

Images