Internet of Things log analysis system for elevator fault diagnosis

Abstract

The invention relates to the technical field of internet-of-things log analysis, and discloses an internet-of-things log analysis system for elevator fault diagnosis, which comprises a log acquisition and feature extraction module configured to perform real-time acquisition and preliminary anomaly detection on operation logs of a plurality of elevators and extract anomaly codes and time sequence features; the early warning threshold value dynamic regulation and control module dynamically adjusts a fault early warning threshold value by analyzing the frequency and the time sequence mode of the abnormal codes so as to reduce the false alarm rate; and the diagnosis model adaptive regulation and control module is used for adaptively adjusting diagnosis model parameters based on the similarity clustering result of the similar historical fault logs of the plurality of elevators. In the invention, aiming at the problem that the false alarm rate is too high due to the fact that a fixed early warning threshold value cannot adapt to the abnormal frequency fluctuation of an individual elevator in the prior art, the occurrence frequency of a specific abnormal code of the current equipment is calculated in real time, and is compared with a normal average value obtained based on historical data of multiple equipment.

Description

Technical Field

[0001] This invention relates to the field of IoT log analysis technology, specifically to an IoT log analysis system for elevator fault diagnosis. Background Technology

[0002] As an indispensable vertical transportation tool in modern buildings, the safety and reliability of elevators are of paramount importance. With the popularization of Internet of Things (IoT) technology, elevator fault diagnosis systems have gradually become intelligent and networked. These systems collect real-time operation log data through sensors and controllers deployed in various parts of the elevator, and use data analysis technology to monitor the elevator status and provide fault warnings.

[0003] Existing elevator fault diagnosis systems based on IoT logs still have many problems in practical applications. First, the diagnostic models in existing systems have poor generalization ability across different elevator models or operating environments, which limits their effectiveness in large-scale deployment. Second, network state fluctuations can easily cause data loss at edge nodes, affecting data integrity. Based on this, the present invention designs an IoT log analysis system for elevator fault diagnosis to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to provide an IoT log analysis system for elevator fault diagnosis, which solves the problems of poor generalization ability and data loss at edge nodes in the background technology.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] An IoT log analysis system for elevator fault diagnosis includes: a log collection and feature extraction module configured to collect and perform preliminary anomaly detection on the operation logs of multiple elevators in real time, and extract anomaly codes and time series features; a dynamic adjustment module for warning thresholds, which dynamically adjusts the fault warning threshold by analyzing the frequency and time series patterns of the anomaly codes to reduce the false alarm rate; an adaptive adjustment module for the diagnostic model, which adaptively adjusts the parameters of the diagnostic model based on the similarity clustering results of historical fault logs of multiple elevators of the same type, thereby improving the cross-device generalization capability; an intelligent adjustment module for data sampling frequency, which intelligently adjusts the data sampling frequency by sensing the real-time log upload response delay and network status fluctuations to reduce data loss at edge nodes; a data quality assessment and repair module, which performs integrity verification and noise detection on the collected raw operation logs, and initiates a data repair program when the data quality score is lower than a preset threshold; and a diagnostic result visualization and interpretation module, which transforms the output results of the fault diagnosis model into a visualized fault probability distribution map and a ranking of key feature contributions, and provides explanatory explanations for the diagnostic decisions.

[0007] Preferably, the early warning threshold dynamic adjustment module performs the following operations:

[0008] Obtain the frequency distribution P of historical error codes for multiple elevators. i , where P i This indicates the average number of occurrences of exception code i within a preset statistical period;

[0009] Calculate the number of times F occurs for exception code i in the current elevator within the same statistical period. current ;

[0010] The warning threshold θ is determined and adjusted based on the following formula:

[0011] like ,but ;otherwise ;

[0012] Among them, avg(P i ) represents the historical average occurrence count of exception code i across all elevators, α is the preset sensitivity coefficient, β is the preset adjustment factor, and θ is the preset sensitivity factor. old The warning threshold before adjustment, θ new The adjusted warning threshold;

[0013] Based on the adjusted threshold θ new Classify and identify real-time abnormal events.

[0014] Preferably, the diagnostic model adaptive adjustment module performs the following operations:

[0015] Extract the feature vector V from the historical fault logs of multiple elevators, where the feature vector V is an n-dimensional real number vector;

[0016] The feature vector V is divided into C clusters using the K-means clustering algorithm;

[0017] Calculate the weights for each cluster c. , where N c N is the number of log samples in cluster c. total This is the total number of samples;

[0018] The diagnostic model parameters ω are updated based on the following formula, where ω is an n-dimensional parameter vector:

[0019] ;

[0020] Where η is the learning rate, V c V is the central feature vector of cluster c. i Let be the feature vector of sample i belonging to cluster c.

[0021] Preferably, the data sampling frequency intelligent control module performs the following operations:

[0022] Detect the real-time changes in upload latency D and network bandwidth B;

[0023] The sampling period T is determined by the following function:

[0024] ;

[0025] T min The minimum sampling interval; T max Maximum sampling interval, γ is the nonlinear adjustment coefficient, and K is the proportionality constant;

[0026] The sampling period T is allocated to each edge node device for task scheduling;

[0027] Synchronously update the packet loss rate metrics for each node.

[0028] Preferably, it also includes an event matching rule real-time control module, which records the arrival time S of each door opening signal. door and position sensor reporting time S pos ;

[0029] Calculate the time difference between the two: ;

[0030] if Then a new logical matching rule will be used. Otherwise, maintain the original matching rule R. prev ;

[0031] Where τ is the maximum allowable error range, δ is the adjustment step size, and sign(.) is the sign function used to distinguish between leading and lagging gating signals;

[0032] The subsequent event correlation analysis is performed based on the new logical rules.

[0033] Preferably, it also includes an alarm priority iteration control module, used for:

[0034] Collect feedback records R from maintenance personnel on handling logs and alarms. R takes values ​​of 0, 1, and 2, which represent ignore, medium priority, and high priority, respectively.

[0035] Construct a log feature matrix H and extract the key attribute vector for each alarm;

[0036] The priority weight of each alarm is calculated using a machine learning model:

[0037] , where σ(.) is the sigmoid activation function, W is the model weight vector, and H and W are matrix multiplications;

[0038] The alarm priority sorting list is reconstructed based on the priority weight w, and used for decision output in the next stage.

[0039] Preferably, the method for the alarm priority iterative control module to dynamically correct the alarm priority further includes:

[0040] Set the initial weight vector W initial ;

[0041] The gradient descent method is used to iteratively optimize the value of W:

[0042] Where λ is the learning rate, L(W) is the loss function designed based on the false alarm rate and actual feedback, and W new With W old These are the weight vectors before and after iterative optimization, respectively, and ∇L(W) is the gradient of the loss function L(W) with respect to the weight vector W.

[0043] After each optimization, the current training feature set is normalized.

[0044] Preferably, it also includes an interference source shielding module, configured as follows:

[0045] Interference sources during data transmission are identified and marked, and the frequency bands of the interference sources are analyzed and targeted shielding is implemented.

[0046] The interference source shielding module is communicatively connected to the log collection and feature extraction module, and receives the data transmission quality indicators reported by the module as an auxiliary judgment for identifying interference.

[0047] Preferably, the diagnostic result visualization and interpretation module is specifically configured as follows:

[0048] The failure probability distribution map is overlaid on the corresponding components of the elevator's three-dimensional structural model in the form of a heat map;

[0049] It generates a natural language explanation based on the contribution of key features, which includes the top K abnormal features that lead to an increased probability of failure and their contribution weights.

[0050] Preferably, in the data quality assessment and repair module, the data repair procedure includes methods based on time series interpolation and prediction based on historical data patterns;

[0051] The data quality assessment and repair module also marks the repaired data as estimated values ​​and inputs them into the subsequent feature extraction and diagnosis module.

[0052] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0053] 1. In this invention, the frequency of occurrence of specific abnormal codes of the current equipment is calculated in real time through the early warning threshold dynamic adjustment module, and compared with the normal average value obtained based on historical data of multiple equipment. When the frequency deviates significantly, the design enables the early warning mechanism to perceive and adapt to the specificity of different elevators or different operating stages, respond sensitively when abnormalities first appear, and remain stable under normal fluctuations. Thus, without excessive triggering of invalid alarms, it ensures the effective capture of real fault risks, significantly improving the accuracy and practicality of the early warning system.

[0054] 2. In this invention, the diagnostic model adaptive control module uses unsupervised clustering to extract common fault pattern clusters from the historical fault data of multiple elevators; the universality of each cluster is characterized by calculating the weight of each cluster, and the model parameters are driven to be updated towards more representative cluster center features by using this weight as a guide.

[0055] 3. In this invention, an adaptive sampling period decision function is constructed using a data sampling frequency intelligent control module, taking network upload latency and bandwidth as inputs. When the network is congested, the system intelligently extends the sampling interval and reduces transmission attempts, thereby reducing the risk of data loss due to continuous packet loss; when the network is unobstructed, the interval is shortened to obtain richer data. This negative feedback adjustment mechanism, which is linked to the network status in real time, provides dynamic protection for data integrity in unreliable network environments, laying a reliable data foundation for subsequent analysis. Attached Figure Description

[0056] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0057] Figure 2 This is a flowchart of the adaptive control process of the diagnostic model of the present invention;

[0058] Figure 3 This is a flowchart of the alarm priority iteration control process of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Example 1;

[0061] Please see Figures 1-3 An IoT log analysis system for elevator fault diagnosis includes:

[0062] The log collection and feature extraction module is configured to collect and perform preliminary anomaly detection on the operation logs of multiple elevators in real time, and extract the anomaly code and time series features.

[0063] The early warning threshold dynamic adjustment module dynamically adjusts the fault early warning threshold by analyzing the frequency and time series patterns of abnormal codes in order to reduce the false alarm rate.

[0064] The diagnostic model adaptive adjustment module adjusts the diagnostic model parameters adaptively based on the similarity clustering results of historical fault logs of multiple elevators of the same type, thereby improving the cross-device generalization ability.

[0065] The intelligent data sampling frequency control module intelligently adjusts the data sampling frequency by sensing the real-time log upload response latency and network status fluctuations, so as to reduce data loss at edge nodes.

[0066] The data quality assessment and repair module is used to perform integrity verification and noise detection on the collected raw operation logs, and to start the data repair program when the data quality score is lower than the preset threshold.

[0067] The diagnostic results visualization and interpretation module is used to transform the output of the fault diagnosis model into a visualized fault probability distribution map and a ranking of the contribution of key features, and to provide explanatory explanations for the diagnostic decisions.

[0068] The early warning threshold dynamic adjustment module performs the following operations:

[0069] Obtain the frequency distribution P of historical error codes for multiple elevators. i , where P i This indicates the average number of occurrences of exception code i within a preset statistical period;

[0070] Calculate the number of times F occurs for exception code i in the current elevator within the same statistical period. current ;

[0071] The warning threshold θ is determined and adjusted based on the following formula:

[0072] like ,but ;otherwise ;

[0073] Among them, avg(P i ) represents the historical average occurrence count of exception code i across all elevators, α is the preset sensitivity coefficient, β is the preset adjustment factor, and θ is the preset sensitivity factor. old The warning threshold before adjustment, θ new The adjusted warning threshold;

[0074] Based on the adjusted threshold θ new Classify and identify real-time abnormal events.

[0075] The diagnostic model adaptive control module performs the following operations:

[0076] Extract the feature vector V from the historical fault logs of multiple elevators. The feature vector V is an n-dimensional real number vector.

[0077] The feature vector V is divided into C clusters using the K-means clustering algorithm;

[0078] Calculate the weights for each cluster c. , where N c N is the number of log samples in cluster c. total This is the total number of samples;

[0079] The diagnostic model parameters ω are updated based on the following formula, where ω is an n-dimensional parameter vector:

[0080] ;

[0081] Where η is the learning rate, V c V is the central feature vector of cluster c. i Let ω be the feature vector of sample i belonging to cluster c. This update rule adjusts the model parameter ω towards the weighted average of the cluster centers, enhancing the learning of common fault modes.

[0082] The data sampling frequency intelligent control module performs the following operations:

[0083] Detect the real-time changes in upload latency D and network bandwidth B;

[0084] The sampling period T is determined by the following function:

[0085] ;

[0086] T min The minimum sampling interval; T max Maximum sampling interval, γ is the nonlinear adjustment coefficient, and K is the proportionality constant;

[0087] The sampling period T is allocated to each edge node device for task scheduling;

[0088] Synchronously update the packet loss rate metrics for each node.

[0089] The working principle of this invention is as follows: The log collection and feature extraction module first collects the operation logs generated by sensors and controllers deployed in multiple elevators in real time. It then identifies and extracts abnormal codes and their corresponding timestamp sequences by parsing the log content. This module uses a sliding window method to segment the time-series data and extracts multi-dimensional features, including mean, variance, and frequency domain features, forming a standardized feature vector that can be used for subsequent analysis. This step achieves preliminary anomaly detection and structured representation of the elevator's operating status, providing a high-quality data foundation for subsequent intelligent diagnosis.

[0090] The early warning threshold dynamic control module starts working, first obtaining the frequency distribution P of historical abnormal codes for multiple elevators. i This represents the average number of occurrences of exception code i within a preset statistical period (e.g., 24 hours), reflecting the frequency of this exception code. The module calculates in real time the number of occurrences F of a specific exception code i for the current elevator within the same period. current By comparing F current Compared with the historical average avg (P i The ratio of avg(P) is used to dynamically determine whether the warning threshold θ needs to be adjusted. i The threshold θ is obtained by averaging the occurrences of all elevator anomaly codes i in the historical database. This design allows the warning threshold to adaptively adjust based on the deviation of the current equipment anomaly frequency from the group's normal range, effectively solving the problem of excessively high false alarm rates caused by fixed thresholds being unsuitable for individual or time-period differences. new It was immediately applied to classify and identify abnormal events that flow in in real time, thereby achieving a more precise and personalized early warning mechanism.

[0091] The adaptive adjustment module of the diagnostic model first extracts the feature vector V (an n-dimensional real vector) from the historical fault logs of multiple elevators. Then, it uses an unsupervised K-means clustering algorithm to divide all feature vectors V into C clusters, each representing a fault mode with similar characteristics. The module updates the core parameter vector ω of the diagnostic model based on a formula. The core idea of ​​this formula is to use the weights of each fault mode cluster to calculate the average deviation of all samples from their respective cluster centers, and use this deviation as the direction for parameter updates. The model parameters then converge towards more common fault mode features, i.e., those with higher weights, while also considering the internal differences between various modes. In this way, the diagnostic model can absorb common fault knowledge from multiple elevators and adaptively adjust its parameters, thus maintaining a high fault identification accuracy even when facing new elevator models or elevators operating in different environments, solving the problem of poor model generalization ability.

[0092] Clustering reveals common fault patterns across devices, and their weights are used to emphasize the learning of common and important patterns. Parameter updates drive the model parameters ω to adjust towards a weighted average of the central features of each common pattern. This allows the model to absorb and reinforce general knowledge mined from multi-device data, rather than simply fitting the specific noise of individual devices. When faced with a new, unseen elevator, the model, with its parameters now biased towards general fault feature representations, can better understand and diagnose its logs, significantly improving its cross-device generalization ability.

[0093] To address data loss at edge nodes caused by network instability, the intelligent data sampling frequency control module implements a dynamic strategy. The module continuously monitors two key network metrics: the data upload latency D from the edge node to the central server and the current available network bandwidth B. Based on this real-time information, it calculates and dynamically adjusts the sampling period T using the following function. When network congestion occurs (D increases) or bandwidth is insufficient (B decreases), the fraction value increases significantly, leading to a longer calculated suggested period T. The system intelligently extends the sampling interval, reducing the number of data packets attempted to be uploaded per unit time, thereby effectively reducing the risk of data loss due to network packet loss or timeouts. When the network is unobstructed (D decreases, B increases), the fraction value decreases, the suggested period T shortens, and the system increases the sampling frequency to capture richer details of device operation.

[0094] The calculated optimal sampling period T is immediately sent to each edge node for execution, and the packet loss rate of each node is updated synchronously for monitoring. This adaptive negative feedback adjustment mechanism based on real-time network status perception provides intelligent protection for data integrity in unreliable IoT environments.

[0095] Example 2;

[0096] Please see Figures 1-3 In this embodiment of the invention, an event matching rule real-time control module is also included, which records the arrival time S of each door opening signal. door and position sensor reporting time S pos ;

[0097] Calculate the time difference between the two: ;

[0098] if Then a new logical matching rule will be used. Otherwise, maintain the original matching rule R. prev ;

[0099] Where τ is the maximum allowable error range, δ is the adjustment step size, and sign(.) is the sign function used to distinguish between leading and lagging gating signals;

[0100] The subsequent event correlation analysis is performed based on the new logical rules.

[0101] It also includes an alarm priority iteration and control module, used for:

[0102] Collect feedback records R from maintenance personnel on handling logs and alarms. R takes values ​​of 0, 1, and 2, which represent ignore, medium priority, and high priority, respectively.

[0103] Construct a log feature matrix H and extract the key attribute vector for each alarm;

[0104] The priority weight of each alarm is calculated using a machine learning model:

[0105] , where σ(.) is the sigmoid activation function, which maps the linear output to the (0,1) interval. The closer the obtained w value is to 1, the higher the priority; W is the model weight vector, and H and W are matrix multiplication;

[0106] The alarm priority sorting list is reconstructed based on the priority weight w, and used for decision output in the next stage.

[0107] The alarm priority iteration control module also includes the following methods for dynamically correcting alarm priorities:

[0108] Set the initial weight vector W initial ;

[0109] The gradient descent method is used to iteratively optimize the value of W:

[0110] Where λ is the learning rate, L(W) is the loss function designed based on the false alarm rate and actual feedback, and W new With W old Let W be the weight vector before and after iterative optimization, and ∇L(W) be the gradient of the loss function L(W) with respect to the weight vector W. This formula is specifically designed to simultaneously consider the difference between the priority weights w predicted by the model and the actual feedback R from the operations personnel, and incorporates a penalty term for the false positive rate.

[0111] This function comprehensively considers the difference between the model's prediction results and the actual feedback R from operations and maintenance personnel, while also taking into account factors such as the false alarm rate.

[0112] After each optimization, the current training feature set is normalized.

[0113] The working principle of this invention is as follows: The real-time event matching rule control module is activated, specifically designed to resolve the logical judgment misalignment problem caused by asynchronous sensor signal timing, which is common in the collaborative judgment between door systems and position systems. The module executes the following process: continuously recording the arrival time of each elevator door opening command and the time when the position sensor reports the exact arrival time of the car, and calculating the time difference between the two.

[0114] When the absolute value of the measured timing deviation exceeds the tolerance τ, it indicates an abnormal shift in the synchronization relationship between the gating and position signals. The system does not simply issue an alarm; instead, it adjusts the original event matching logic rule R based on the direction of the deviation and the preset step size δ. prev Targeted fine-tuning is performed. If the gate signal continues to lead, the rule is adjusted to increase delay compensation. This mechanism enables online dynamic calibration and compensation at the software level for systematic timing errors caused by hardware response differences and minor communication delays. Based on the updated rule R... new Performing event correlation analysis can significantly improve the accuracy of logical judgments in asynchronous signal situations, effectively avoiding misjudgments of critical events such as door status or floor leveling position.

[0115] The experiential knowledge of operations and maintenance personnel is formalized through feedback data R, and a machine learning model is used to automatically learn the priority mapping relationship from historical alarm features H and feedback R. Through continuous gradient descent optimization, the model continuously fine-tunes its weights W, making the predicted priority w increasingly closer to the actual urgency of operations and maintenance needs. This achieves adaptive iterative evolution of the alarm ranking strategy, effectively distinguishing between alarm "signals" and "noise," ensuring that critical fault information is not missed in the flood of alarms, and greatly improving operations and maintenance response efficiency and system security.

[0116] The data quality assessment and repair module and the diagnostic result visualization and interpretation module continuously function in both the backend and frontend. The former performs integrity checks and noise detection, and intelligently selects methods such as time series interpolation or historical pattern prediction to repair data when the quality score is too low. At the same time, it marks the repaired data as "estimated values" to maintain data traceability and transparency, ensuring the data quality of the input model. The latter maps the failure probability output by the model onto the elevator's 3D model in the form of a heat map and generates a natural language explanation that includes a ranking of the contribution of key features, such as "Traction machine overheating risk (85%): 1. The motor winding temperature continues to rise (contribution weight 0.45); 2. Current harmonics increase (contribution weight 0.32)", which greatly improves the understandability, credibility, and efficiency of operation and maintenance decision support of the diagnostic results.

[0117] Example 3;

[0118] Please see Figures 1-3 In this embodiment of the invention, an interference source shielding module is also included, configured as follows:

[0119] Interference sources during data transmission are identified and marked, and their frequency bands are analyzed and targeted shielding is implemented. The interference source shielding module communicates with the log collection and feature extraction module, receiving data transmission quality indicators reported by the module as auxiliary indicators for interference identification.

[0120] The specific configuration of the diagnostic results visualization and interpretation module is as follows:

[0121] The failure probability distribution map is overlaid on the corresponding components of the elevator's 3D structural model in the form of a heat map; and a natural language explanation is generated based on the contribution of key features, including the top K abnormal features that lead to an increase in failure probability and their contribution weights.

[0122] In the data quality assessment and repair module, the data repair procedure includes methods based on time series interpolation or prediction based on historical data patterns; the data quality assessment and repair module also marks the repaired data as estimated values ​​and inputs them into the subsequent feature extraction and diagnosis module.

[0123] The working principle of this invention embodiment is as follows: An interference source shielding module is added to this embodiment. This module is communicatively connected to the log acquisition and feature extraction module, receiving real-time data transmission quality indicators such as signal strength, bit error rate, and signal-to-noise ratio. The module incorporates spectrum analysis and pattern recognition algorithms, enabling it to actively scan the data transmission channel, identify periodic or sudden electromagnetic interference sources, and mark their respective frequency bands. Once the interference source and its main affected frequency bands are confirmed, the module initiates targeted shielding strategies, such as dynamically adjusting the communication frequency through software-defined radio (SDR) technology or activating hardware filters to filter specific frequency bands. By actively identifying and shielding interference sources, this module improves the reliability and integrity of log data transmission from the source of data acquisition, providing a cleaner and more stable data stream for subsequent analysis, with particularly significant effects in complex electromagnetic environments, such as industrial scenarios.

[0124] Secondly, this embodiment specifies and enhances the functionality of the diagnostic result visualization and interpretation module. This module is configured not only to generate a fault probability heatmap overlaid on the 3D model, but also to perform in-depth feature attribution analysis. It extracts the top K most critical features affecting the final decision from the diagnostic model and accurately calculates the contribution weight of each feature to the current fault probability. Based on this, the module automatically generates a structured natural language report. The report clearly lists the primary factors leading to the increased fault probability, such as: "Fault diagnosis points to the risk of 'unexpected car movement,' and the main abnormal features and their contribution are as follows: 1. The time difference between the landing door locking signal and the car position signal exceeds the limit (contribution: 35%); 2. The leveling accuracy has consistently exceeded the standard in the last 10 operations (contribution: 28%); 3. Abnormal fluctuations in the brake force sensor readings (contribution: 22%)." This highly interpretable output allows maintenance personnel not only to know "where might be broken," but also to understand "why the system judged it to be broken," greatly enhancing the trustworthiness and operability of the diagnostic results.

[0125] Finally, when missing or anomaly data is detected, the repair program intelligently selects a repair strategy based on the pattern and context of the missing data: for short-term, randomly missing time series data points, time series-based interpolation methods, such as linear interpolation or cubic spline interpolation, are used to smoothly estimate the missing value using the valid data points before and after; for data with obvious periodicity or trend but with anomalies over a longer period, methods based on historical data pattern prediction are used, such as using a trained seasonal autoregressive integral moving average (SARIMA) model or a long short-term memory (LSTM) network to predict the current missing or anomaly value based on the data patterns of the same period and working conditions in history.

[0126] Only this labeled data is input into the subsequent feature extraction and diagnostic modules, maintaining the transparency of the data chain. Downstream modules can identify and process this estimated data as appropriate, thereby making full use of existing information while clarifying the source of data uncertainty. This avoids the bias that may be introduced by correcting data being mistaken for completely accurate observations, further improving the reliability of the entire analysis system.

[0127] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An IoT log analysis system for elevator fault diagnosis, characterized in that, include: The log collection and feature extraction module is configured to collect and perform preliminary anomaly detection on the operation logs of multiple elevators in real time, and extract the abnormal codes and time series features. The early warning threshold dynamic adjustment module dynamically adjusts the fault early warning threshold by analyzing the frequency and time series pattern of the abnormal code, so as to reduce the false alarm rate. The diagnostic model adaptive adjustment module adjusts the diagnostic model parameters adaptively based on the similarity clustering results of historical fault logs of multiple elevators of the same type, thereby improving the cross-device generalization ability. The intelligent data sampling frequency control module intelligently adjusts the data sampling frequency by sensing the real-time log upload response latency and network status fluctuations, so as to reduce data loss at edge nodes. The data quality assessment and repair module is used to perform integrity verification and noise detection on the collected raw operation logs, and to start the data repair program when the data quality score is lower than the preset threshold. The diagnostic results visualization and interpretation module is used to transform the output of the fault diagnosis model into a visualized fault probability distribution map and a ranking of the contribution of key features, and to provide explanatory explanations for the diagnostic decisions.

2. The IoT log analysis system for elevator fault diagnosis according to claim 1, characterized in that, The early warning threshold dynamic adjustment module performs the following operations: Obtain the frequency distribution P of historical error codes for multiple elevators. i , where P i This indicates the average number of occurrences of exception code i within a preset statistical period; Calculate the number of times F occurs for exception code i in the current elevator within the same statistical period. current ; The warning threshold θ is determined and adjusted based on the following formula: like ,but ;otherwise ; Among them, avg(P i ) represents the historical average number of occurrences of exception code i across all elevators, α is a preset sensitivity coefficient (α > 1), β is a preset adjustment factor (β > 1), and θ old The warning threshold before adjustment, θ new The adjusted warning threshold; Based on the adjusted threshold θ new Classify and identify real-time abnormal events.

3. The IoT log analysis system for elevator fault diagnosis according to claim 2, characterized in that, The diagnostic model adaptive adjustment module performs the following operations: Extract the feature vector V from the historical fault logs of multiple elevators, where the feature vector V is an n-dimensional real number vector; The feature vector V is divided into C clusters using the K-means clustering algorithm; Calculate the weights for each cluster c. , where N c N is the number of log samples in cluster c. total This is the total number of samples; The diagnostic model parameters ω are updated based on the following formula, where ω is an n-dimensional parameter vector: ； Where η is the learning rate, V c V is the central feature vector of cluster c. i Let be the feature vector of sample i belonging to cluster c.

4. The IoT log analysis system for elevator fault diagnosis according to claim 3, characterized in that, The intelligent data sampling frequency control module performs the following operations: Detect the real-time changes in upload latency D and network bandwidth B; The sampling period T is determined by the following function: ； T min The minimum sampling interval; T max Maximum sampling interval, γ is the nonlinear adjustment coefficient, and K is the proportionality constant; The sampling period T is allocated to each edge node device for task scheduling; Synchronously update the packet loss rate metrics for each node.

5. The IoT log analysis system for elevator fault diagnosis according to claim 4, characterized in that, It also includes an event matching rule real-time control module, which records the arrival time S of each door opening signal. door and position sensor reporting time S pos ; Calculate the time difference between the two: ; if Then a new logical matching rule will be used. Otherwise, maintain the original matching rule R. prev ; Where τ is the maximum allowable error range, δ is the adjustment step size, and sign(.) is the sign function used to distinguish between leading and lagging gating signals; The subsequent event correlation analysis is performed based on the new logical rules.

6. The IoT log analysis system for elevator fault diagnosis according to claim 5, characterized in that, It also includes an alarm priority iteration and control module, used for: Collect feedback records R from maintenance personnel on handling logs and alarms. R takes values ​​of 0, 1, and 2, which represent ignore, medium priority, and high priority, respectively. Construct a log feature matrix H and extract the key attribute vector for each alarm; The priority weight of each alarm is calculated using a machine learning model: , where σ(.) is the sigmoid activation function, W is the model weight vector, and H and W are matrix multiplications; The alarm priority sorting list is reconstructed based on the priority weight w, and used for decision output in the next stage.

7. The IoT log analysis system for elevator fault diagnosis according to claim 6, characterized in that: The method for dynamically correcting alarm priorities by the alarm priority iteration control module also includes: Set the initial weight vector W initial ; The gradient descent method is used to iteratively optimize the value of W: Where λ is the learning rate, L(W) is the loss function designed based on the false alarm rate and actual feedback, and ∇L(W) is the gradient of the loss function L(W) with respect to the weight vector W. new With W old These are the weight vectors before and after iterative optimization, respectively; After each optimization, the current training feature set is normalized.

8. The IoT log analysis system for elevator fault diagnosis according to claim 7, characterized in that, It also includes an interference source shielding module, configured as follows: Interference sources during data transmission are identified and marked, and the frequency bands of the interference sources are analyzed and targeted shielding is implemented. The interference source shielding module is communicatively connected to the log collection and feature extraction module, and receives the data transmission quality indicators reported by the module as an auxiliary judgment for identifying interference.

9. The IoT log analysis system for elevator fault diagnosis according to claim 1, characterized in that, The specific configuration of the diagnostic result visualization and interpretation module is as follows: The failure probability distribution map is overlaid on the corresponding components of the elevator's three-dimensional structural model in the form of a heat map; It generates a natural language explanation based on the contribution of key features, which includes the top K abnormal features that lead to an increased probability of failure and their contribution weights.

10. The IoT log analysis system for elevator fault diagnosis according to claim 1, characterized in that: In the data quality assessment and repair module, the data repair procedure includes methods based on time series interpolation and prediction based on historical data patterns; The data quality assessment and repair module also marks the repaired data as estimated values ​​and inputs them into the subsequent feature extraction and diagnosis module.

Images