Urban rail equipment defect adaptive identification method and system for networkless environment

Abstract

The application discloses a kind of urban rail equipment defect adaptive identification methods and systems for networkless environment, belong to offline equipment defect self-identification technical field.It includes: presetting offline SQLite database containing equipment operation history;Extract five characteristic parameters;Deploy lightweight YOLOv8 model to carry out end-side AI inference, obtain calibration probability by temperature scaling and calculate confidence uncertainty;Adaptive threshold is calculated by fusing cross-product term and dynamic threshold formula of Sigmoid mapping;Compare maximum probability with adaptive threshold to determine defects, generate event package offline storage;Upload event package after network recovery, server generates work order and feedbacks incremental update package, terminal pulls to complete closed loop.The application uses equipment operation history to realize differentiated dynamic threshold, covers five types of defects, significantly reduces false negative rate and false positive rate, fills the technical gap of urban rail equipment adaptive identification under networkless environment.

Description

Technical Field

[0001] This invention relates to the field of offline equipment defect self-identification technology, and in particular to an adaptive identification method and system for urban rail transit equipment defects in network-free environments. Background Technology

[0002] With the continuous expansion of urban rail transit networks, the number of devices in stations and tunnels has increased dramatically. Traditional inspection methods, primarily relying on manual visual inspection and paper-based record-keeping, are no longer sufficient to meet the efficiency and standardization requirements of operation and management. In recent years, the maturity of edge computing and lightweight deep learning technologies has made it possible to deploy target detection models on mobile terminals for automatic equipment defect identification. Detection algorithms, represented by the YOLO series, have been widely used in industrial defect identification. Combined with local inference capabilities on the terminal, they can achieve real-time on-site identification in scenarios such as subway stations and tunnels without network coverage, significantly improving inspection efficiency and the standardization of data collection. Integrating historical equipment maintenance information with edge AI recognition to form an offline intelligent identification—online closed-loop synchronous maintenance system has become an important development direction for urban rail intelligent maintenance technology.

[0003] However, existing technologies still have the following shortcomings: First, defect judgment generally uses a globally fixed confidence threshold, which makes it difficult to simultaneously consider the false negative rate and false positive rate in scenarios with extremely imbalanced defect samples. Second, existing methods use the same identification threshold for all devices, ignoring individual differences in criticality level, years of operation, historical defect frequency, maintenance status, and remaining lifespan, and failing to utilize the device's own maintenance history to optimize the identification strategy. Third, the confidence level output by deep learning models often has an overconfidence problem, and existing solutions lack confidence level calibration and uncertainty quantification, resulting in insufficient reliability of threshold judgment.

[0004] Therefore, there is an urgent need for a defect identification method that can integrate equipment operation and maintenance history and achieve dynamic threshold adaptation on the end side. Summary of the Invention

[0005] Therefore, this invention provides an adaptive identification method and system for urban rail equipment defects in network-free environments to overcome the problems in the prior art.

[0006] To achieve the above objectives, the present invention provides an adaptive defect identification method for urban rail transit equipment in network-free environments, comprising: S1 is a pre-configured offline SQLite database containing the operation and maintenance history of urban rail equipment; S2, based on an offline SQLite database, automatically extracts five characteristic parameters of the current device to be identified: device critical level coefficient, device aging coefficient, historical defect frequency coefficient, maintenance effect coefficient, and predicted risk score. S3 deploys a lightweight YOLOv8 model for edge AI inference, obtains the calibration probability by temperature scaling of the original log odds output by the model, and calculates the confidence uncertainty based on the calibration probability. S4, input the five feature parameters and confidence uncertainty into the dynamic threshold calculation module, and calculate the adaptive threshold corresponding to the device according to the dynamic threshold formula that integrates the cross-product term and the Sigmoid nonlinear mapping; S5. Compare the maximum probability in the calibration probability with the adaptive threshold. If the maximum probability is not less than the adaptive threshold, a defect is determined to exist and an event package is generated for offline storage; otherwise, no defect is determined. S6. After the network is restored, the event packet is uploaded to the server, the server generates a work order and sends back an incremental update packet, and the terminal pulls the incremental update packet to complete the closed loop.

[0007] Furthermore, the unique device code in the offline SQLite database consists of the line number, station number, device type, and serial number; the device maintenance history includes static device attributes, maintenance work order records, inspection records, defect reporting records, and the remaining life prediction value output by predictive maintenance.

[0008] Furthermore, in S2, the equipment criticality level coefficient K crit Obtained from the equipment criticality level mapping table; the higher the criticality level, the larger the coefficient. The equipment aging coefficient ,in, The duration of operation For the designed service life, This is an aging amplification factor. The historical defect frequency coefficient ,in This represents the average number of defects per month for the current equipment over the past 12 months. This represents the average monthly defect rate for similar equipment. Maintenance effect coefficient ,in This is the number of days since the last maintenance was completed. To account for the extent of the impact of the repairs This is the recovery time constant; Predictive Risk Score ,in This is the predicted remaining lifespan. This is a lifetime-scale parameter.

[0009] Furthermore, in S3, the lightweight YOLOv8 model identifies five types of defects: cracks, water leakage, deformation, abnormal sounds, and overheating. Among them, cracks, water leakage, and deformation are detected and identified by the YOLOv8 model through visible light images captured by the mobile phone camera; abnormal sounds are detected by the YOLOv8 model through the YOLOv8 model after the sound of the device being operated is collected by the mobile phone microphone, and the spectrum is generated by short-time Fourier transform; overheating is detected by the YOLOv8 model through infrared thermal imaging photos collected by the external USB-C infrared thermal imager module of the mobile phone.

[0010] Furthermore, the mapping relationship between different equipment defect types is executed according to the preset equipment defect type mapping table; when judging defects, if the defect type identified by the YOLOv8 model is not within the mapping range of the current equipment, the confidence weight of the detection box is automatically reduced.

[0011] Furthermore, in S3, the temperature scaling is probabilistically calibrated using the following formula: ; in For the first The primitive logarithmic odds of a class The optimal temperature coefficient is determined by minimizing the negative log-likelihood on the validation set. The optimal temperature coefficient is used as the temperature coefficient; the confidence uncertainty is defined as: in This represents the total number of defect categories. For the first time after calibration The probability of a class.

[0012] Furthermore, in S4, the dynamic threshold formula is: ; Among them, T adapt For adaptive threshold, To set a preset baseline threshold, The comprehensive adjustment factor is defined as follows: ; in For the five characteristic parameters, These are the cross-term weight coefficients; ultimately Limit to a preset lower limit and upper limit between.

[0013] Furthermore, in S5, the event package adopts JSON format and includes at least the event unique identifier, device number, defect type code, calibration probability, confidence uncertainty, adaptive threshold, timestamp, and detection image path; the event package is written using transactional atomic write and supports a first-in-first-out cleanup strategy for local storage space, only cleaning up event packages that have been successfully uploaded and confirmed.

[0014] Furthermore, it includes a penalty mechanism for unrepaired defects, specifically by counting the number of unrepaired defects in the current equipment. Define the penalty factor ,in The adaptive threshold is multiplied by the penalty factor as a penalty coefficient to suppress repeated reporting of existing unprocessed defects.

[0015] This invention also provides an adaptive identification system for urban rail transit equipment defects in network-free environments. The system is used to implement any of the adaptive identification methods for urban rail transit equipment defects in network-free environments described in this invention. The system includes: The offline database pre-configuration module is used to pre-configure an offline SQLite database containing the operation and maintenance history of urban rail equipment; The feature parameter extraction module is used to automatically extract five feature parameters of the current device to be identified, based on an offline SQLite database: device critical level coefficient, device aging coefficient, historical defect frequency coefficient, maintenance effect coefficient, and predicted risk score. The confidence calibration module is used to deploy a lightweight YOLOv8 model for edge AI inference. It obtains the calibration probability by temperature scaling the original log odds output by the model and calculates the confidence uncertainty based on the calibration probability. An adaptive threshold calculation module is used to input five feature parameters and confidence uncertainty into a dynamic threshold calculation module, and calculate the adaptive threshold corresponding to the device according to a dynamic threshold formula that integrates cross-product terms and Sigmoid nonlinear mapping. The defect determination module is used to compare the maximum probability in the calibration probability with the adaptive threshold. If the maximum probability is not less than the adaptive threshold, a defect is determined to exist and an event package is generated for offline storage; otherwise, no defect is determined. The network synchronization module is used to upload the event packet to the server after the network is restored. The server generates a work order and sends back an incremental update packet. The terminal pulls the incremental update packet to complete the closed loop.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: Firstly, this invention constructs a dynamic adaptive threshold mechanism that integrates multi-dimensional operation and maintenance characteristics of equipment, enabling differentiated defect identification tailored to specific equipment in network-free environments. By automatically extracting five feature parameters from an offline SQLite database—equipment critical level coefficient, equipment aging coefficient, historical defect frequency coefficient, maintenance effect coefficient, and predicted risk score—and designing a dynamic threshold formula that integrates cross-product terms and a Sigmoid nonlinear mapping, each piece of equipment can obtain an independent discrimination threshold based on its own critical level, years of operation, historical defect frequency, post-maintenance recovery status, and remaining lifespan. This effectively overcomes the contradiction between false alarms and false negatives caused by globally fixed thresholds in existing technologies, significantly improving the targeting and accuracy of defect identification.

[0017] Secondly, this invention introduces a temperature scaling calibration and confidence uncertainty quantification mechanism, effectively improving the reliability of edge AI discrimination results. By applying temperature scaling to the raw log probability output by the YOLOv8 model, the overconfidence problem commonly found in deep learning models is alleviated, making the calibration probability closer to the true accuracy. Simultaneously, the confidence uncertainty is calculated based on the calibration probability and incorporated as a threshold adjustment factor into the dynamic threshold formula. This allows the model to automatically raise the judgment threshold when the prediction confidence is insufficient, thereby further suppressing false reporting of low-quality detection results and enhancing the robustness of defect discrimination.

[0018] Thirdly, this invention achieves end-side adaptive identification and online closed-loop synchronization of defects in urban rail transit equipment in a network-free environment, filling a technological gap in this scenario. Through an offline SQLite database, a lightweight YOLOv8 model, dynamic threshold calculation, offline storage of event packets, and automatic uploading and incremental updates after network recovery, a complete closed-loop system of "offline intelligent identification—online work order generation—incremental policy feedback" is constructed. This enables the inspection terminal to efficiently and reliably complete defect identification even when completely disconnected from the network, and automatically completes data synchronization and policy updates after reconnection, significantly improving the intelligence level and operational efficiency of urban rail transit equipment inspection. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 A flowchart of an adaptive defect identification method for urban rail transit equipment in a network-free environment provided in an embodiment of the present invention; Figure 2This is a structural block diagram of an adaptive identification system for urban rail equipment defects in network-free environments, provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0022] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0024] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0025] Example 1 like Figure 1 As shown, this invention provides an adaptive defect identification method for urban rail transit equipment in network-free environments, comprising: S1 is a pre-configured offline SQLite database containing the operation and maintenance history of urban rail equipment; The unique device code in the offline SQLite database consists of the line number, station number, device type, and serial number; the device operation and maintenance history includes static device attributes, maintenance work order records, inspection records, defect reporting records, and the remaining life prediction value output by predictive maintenance.

[0026] In one possible implementation, taking the Phase I project of Metro Line 8 in a certain city as an example, the entire line has 16 stations, covering key equipment in tunnel sections and station equipment rooms. This embodiment selects one standard station as the data deployment node. This station contains four main categories of equipment: ventilation and air conditioning system, water supply and drainage system, power and lighting system, and platform screen door system, totaling approximately 320 pieces of equipment. An offline SQLite database is pre-installed in the local storage of the inspection terminal. This database is pre-packaged and synchronously distributed to each inspection terminal by the central server before the line is opened, ensuring that the terminal can still completely read the equipment operation and maintenance history even in the event of a network interruption.

[0027] The unique code for equipment in the database consists of four segments: line number (e.g., L08 for Line 8), station number, equipment type (e.g., CH for chiller unit), and sequence number (a four-digit number, such as 0032). This results in a globally unique code like L08-S12-CH-0032, which serves as the primary key for all data tables and is used for cross-table joins.

[0028] The equipment operation and maintenance history includes at least the following five types of structured data: (1) Equipment static attribute table, including equipment name, model, design service life, commissioning date, and criticality level (A / B / C / D levels); (2) Maintenance work order record table, including maintenance completion date, maintenance content, and replacement parts list; (3) Inspection record table, including inspection date, inspection personnel, and qualitative description of equipment appearance and operating status; (4) Defect reporting record table, including defect type, reporting time, and processing status, including unrepaired, repaired, and pending observation; (5) Predictive maintenance output table, including the most recent remaining life prediction value and prediction date. In this embodiment, the database adopts WAL (Write-Ahead Logging) mode to improve concurrent read and write performance, and uses AES-256 encryption algorithm to fully encrypt the database file to prevent data leakage after equipment loss.

[0029] S2, based on an offline SQLite database, automatically extracts five characteristic parameters of the current device to be identified: device critical level coefficient, device aging coefficient, historical defect frequency coefficient, maintenance effect coefficient, and predicted risk score. In S2, the equipment criticality level coefficient K crit Obtained from the equipment criticality level mapping table; the higher the criticality level, the larger the coefficient. The equipment aging coefficient ,in, The duration of operation For the designed service life, This is an aging amplification factor. The historical defect frequency coefficient ,in This represents the average number of defects per month for the current equipment over the past 12 months. This represents the average monthly defect rate for similar equipment. Maintenance effect coefficient ,in This is the number of days since the last maintenance was completed. To account for the extent of the impact of the repairs This is the recovery time constant; Predictive Risk Score ,in This is the predicted remaining lifespan. This is a lifetime-scale parameter.

[0030] In one possible implementation, this embodiment takes the tunnel ventilation fan of the fifth station of a certain subway line 1 as an example, with the unique equipment code L1-S05-FAN-101. This fan was put into operation on January 15, 2018, with a designed service life of E... life =10 years, equipment criticality level is 2, i.e., highly critical equipment. The current inspection date is June 26, 2026. The parameters involved in this embodiment are: aging amplification factor α=0.4, maintenance impact magnitude β=0.5, recovery time constant τ=45 days, and life scale parameter η=0.2×Elife=2 years. It can be understood that the above parameters are all empirical values ​​preset in the terminal configuration file. In actual applications, corresponding adjustments can be made according to different actual situations. The data in this embodiment are all exemplary data, only used to intuitively demonstrate the specific calculation process, and do not limit the invention in any way.

[0031] Based on the above data, the terminal calculates five characteristic parameters respectively: (1) Equipment criticality level coefficient K crit The device criticality level is read as 2 from the device information master table. According to the criticality level mapping table in the terminal configuration file, criticality level 2 corresponds to the coefficient K. crit =0.9.

[0032] (2) Equipment aging coefficient Aage The current date is June 26, 2026, and the commissioning date is January 15, 2018. Calculate the commissioning time T. age =8.4 years. Design service life E life =10 years, aging amplification factor α=0.4. Substitute into the formula. The calculation yields: A age =1+0.4×8.4 / 10=1.336.

[0033] (3) Historical defect frequency coefficient F hist A search of the defect history table for the current equipment over the past 12 months revealed only one defect record: a repair request submitted on December 20, 2025, and repair completed on December 22. Therefore, the average monthly defect count for this equipment is λ = 1 / 12 ≈ 0.0833 times / month. The average monthly defect rate for the entire line of similar equipment (wind turbines) is 0.02 times / month. This value is derived from the historical defect records of all wind turbine equipment on the entire line and is pre-set in the terminal configuration file. Substituting into the formula... The calculation yields: F hist =max(1.0,0.0833 / 0.02)=max(1.0, 4.1667) = 4.1667.

[0034] (4) Maintenance effect coefficient R repair The most recent maintenance completion date for this equipment, as found in the defect history table, was December 22, 2025. The number of days remaining until the current date, June 26, 2026, is d = 186 days. The maintenance impact magnitude is β = 0.5, and the recovery time constant is τ = 45 days. Substituting these values ​​into the formula... The calculation yields: R repair =1+0.5×exp(-186 / 45)≈1+0.5×0.0162≈1.0081.

[0035] (5) Predicted risk score S risk Retirement life (RUL) prediction value is read from the equipment information master table as 1.2 years, and lifespan parameter η = 0.2 × Elife = 2 years. Substitute these values ​​into the formula. The calculation yields: S risk =exp(-1.2 / 2)=exp(-0.6)≈0.5488.

[0036] In summary, the five characteristic parameters of this equipment during this inspection are as follows: K crit =0.9, A age =1.336, F hist =4.1667, R repair =1.0081, S risk =0.5488. After the above parameters are extracted, they are temporarily stored in the terminal memory for use in dynamic threshold calculation.

[0037] This invention solves the problem of inaccurate defect identification caused by ignoring individual differences of equipment and using the same fixed threshold for all equipment in the prior art by automatically extracting five feature parameters from an offline SQLite database: equipment critical level coefficient, equipment aging coefficient, historical defect frequency coefficient, maintenance effect coefficient, and predicted risk score. This improves the pertinence and accuracy of defect identification and reduces the false alarm rate and false negative rate.

[0038] S3 deploys a lightweight YOLOv8 model for edge AI inference, obtains the calibration probability by temperature scaling of the original log odds output by the model, and calculates the confidence uncertainty based on the calibration probability. In S3, the lightweight YOLOv8 model identifies five types of defects: cracks, water leakage, deformation, abnormal sounds, and overheating. Cracks, water leakage, and deformation are detected and identified by the YOLOv8 model using visible light images captured by the phone's camera. Abnormal sounds are detected by the YOLOv8 model using the phone's microphone to capture the sound of the device in operation, which is then converted into a spectrum by a short-time Fourier transform. Overheating is detected by the YOLOv8 model using an external USB-C infrared thermal imager module connected to the phone to capture infrared thermal images.

[0039] The mapping relationship between different equipment defect types is executed according to the preset equipment defect type mapping table; when judging defects, if the defect type identified by the YOLOv8 model is not within the mapping range of the current equipment, the confidence weight of the detection box is automatically reduced.

[0040] In S3, the temperature scaling is probabilistically calibrated using the following formula: ; in For the first The primitive logarithmic odds of a class The optimal temperature coefficient is determined by minimizing the negative log-likelihood on the validation set. The optimal temperature coefficient is used as the temperature coefficient; the confidence uncertainty is defined as: in This represents the total number of defect categories. For the first time after calibration The probability of a class.

[0041] In one possible implementation, a pre-defined device-defect type mapping table is used, supporting five types of defects: cracks, water leakage, deformation, abnormal sounds, and overheating. In this embodiment, the lightweight YOLOv8 model uses the YOLOv8 nano version. After structured channel pruning and INT8 quantization, the model file size is 8.6 megabytes. It is deployed on a handheld terminal in TensorFlow Lite format and loaded and run through the NNAPI inference framework. The model input tensor size is (1, 640, 640, 3), and the output is the bounding box coordinates and the classification log odds.

[0042] The terminal camera captured visible light images of the wind turbine surface. The original image resolution was 1920×1080 pixels. The terminal scaled the image to 640×640 pixels before inputting it into the model. The model output 8400 candidate boxes, and the top 100 candidate boxes were retained based on their highest original scores. The candidate box with the highest score had an original log-odds vector of z=[2.8, -0.5, -1.2, -0.8, -1.5], corresponding to five types of defects: cracks, water leakage, deformation, abnormal noise, and overheating. The defect type identified by this candidate box was a crack, which is a valid defect type supported by the wind turbine equipment. Therefore, this detection box was retained normally, and the confidence weight reduction process was not triggered.

[0043] Using the optimal temperature coefficient T*=1.5 obtained beforehand by minimizing the negative log-likelihood of the validation set, and substituting it into the temperature scaling formula: Where C=5 and T=1.5, the calculated calibration probabilities for each defect category are: p1 (crack) = 0.72, p2 (water leakage) = 0.08, p3 (deformation) = 0.07, p4 (abnormal sound) = 0.06, p5 (overheating) = 0.07. The maximum probability after calibration is max(p) = 0.72, corresponding to the defect category of crack.

[0044] Based on the calibration probability vector p = [0.72, 0.08, 0.07, 0.06, 0.07], the Shannon entropy is calculated to be 0.38. Substituting this into the confidence uncertainty formula: The calculated confidence uncertainty is 0.164. After non-maximum suppression, this candidate box is the only valid detection box. Since the maximum calibration probability of 0.72 is greater than the early stopping threshold of 0.1, this detection result proceeds to the subsequent dynamic threshold calculation and defect determination steps.

[0045] This invention addresses the problems of unreliable confidence levels caused by overconfidence in deep learning models and false alarms caused by mismatch between defect types and devices in existing technologies by deploying a lightweight YOLOv8 model for edge AI inference and introducing temperature scaling calibration and confidence uncertainty quantification mechanisms. This improves the reliability and robustness of defect identification and reduces the false alarm rate caused by model overconfidence and defect type mismatch.

[0046] S4, input the five feature parameters and confidence uncertainty into the dynamic threshold calculation module, and calculate the adaptive threshold corresponding to the device according to the dynamic threshold formula that integrates the cross-product term and the Sigmoid nonlinear mapping; In S4, the dynamic threshold formula is: ; Among them, T adapt For adaptive threshold, To set a preset baseline threshold, The comprehensive adjustment factor is defined as follows: ; in For the five characteristic parameters, These are the cross-term weight coefficients; ultimately Limit to a preset lower limit and upper limit between.

[0047] In one possible implementation, the equipment criticality level coefficient K crit =0.9, Equipment aging factor A age =1.336, historical defect frequency coefficient F hist =4.1667, maintenance effect coefficient R repair =1.0081, Predicted Risk Score risk =0.5488; Confidence level uncertainty U conf =0.164. Preset baseline threshold T0=0.55. Cross-term weight coefficient λ cross =0.1, lower threshold T min =0.2, upper limit of threshold T max =0.9, and all of the above parameters are preset in the terminal configuration file.

[0048] The eigenvectors x = {0.9, 1.336, 4.1667, 1.0081, 0.5488} are denoted as x1 to x5, respectively. Calculate all pairwise cross products: x1x2= 0.9 × 1.336 = 1.2024; x1x3= 0.9 × 4.1667 = 3.7500; x1x4= 0.9 × 1.0081 = 0.9073; x1x5= 0.9 × 0.5488 = 0.4939; x2x3= 1.336 × 4.1667 = 5.5671; x2x4= 1.336 × 1.0081 = 1.3468; x2x5= 1.336 × 0.5488 = 0.7332; x3x4= 4.1667 × 1.0081 = 4.2004; x3x5= 4.1667 × 0.5488 = 2.2867; x4x5= 1.0081 × 0.5488 = 0.5532; Summing each cross product term by term yields a result of 21.0410. Denominator = K crit × A age ×F his t × R repair × (1 + S risk ) ≈ 7.8216, the denominator reflects the independent combined influence of the five characteristic parameters, and the larger the value, the higher the overall risk of the equipment.

[0049] According to the formula: The calculated value is θ = 1.0476. θ is then mapped to the (0,1) interval using the Sigmoid function, resulting in Sigmoid(θ) = 1 / (1 + exp(-θ)) ≈ 0.7403. This mapping compresses the comprehensive adjustment factor to a reasonable range, allowing the threshold to fluctuate smoothly around the baseline threshold and avoiding extreme jumps.

[0050] According to the formula: The calculated adaptive threshold was 0.7940. This final calculated value of 0.7940 falls between the preset lower limit of 0.2 and the upper limit of 0.9; therefore, the final adaptive threshold was set to 0.7940. It is understandable that this threshold is significantly higher than the preset baseline threshold of 0.55, mainly because of the historical defect frequency coefficient F of the device. hist The value of 4.1667 indicates that its defect frequency over the past 12 months is significantly higher than the average level of similar equipment. The cross-product term further amplifies this effect, resulting in a comprehensive adjustment factor greater than 1, which, after Sigmoid mapping, is greater than 0.5. This, combined with the confidence uncertainty U... confThe amplification effect ultimately raises the judgment threshold from 0.55 to 0.7940. This means that for this device, it is only judged as defective when the calibration probability output by the model reaches 0.7940 or higher, thus effectively suppressing false alarms caused by the fixed threshold being too low for this high-risk device. At the same time, if the device condition further deteriorates (such as the feature parameters continuing to increase), the threshold will also be dynamically adjusted accordingly to achieve adaptive discrimination.

[0051] S5. Compare the maximum probability in the calibration probability with the adaptive threshold. If the maximum probability is not less than the adaptive threshold, a defect is determined to exist and an event package is generated for offline storage; otherwise, no defect is determined. In S5, the event packet adopts JSON format and includes at least the event unique identifier, device number, defect type code, calibration probability, confidence uncertainty, adaptive threshold, timestamp, and detection image path; the event packet is written using transactional atomic write and supports a first-in-first-out cleanup strategy for local storage space, only cleaning up event packets that have been successfully uploaded and confirmed.

[0052] In one possible implementation, the terminal compares the maximum calibration probability of 0.72 with the adaptive threshold of 0.7940. Since 0.72 is less than 0.7940, the condition that the maximum calibration probability is not less than the adaptive threshold is not met, therefore the detection result is determined not to be a defect. The inspection result for this instance is defect-free, no event packet is generated, and no local storage space is occupied.

[0053] In contrast, if a preset fixed baseline threshold T0=0.55 is used, then 0.72 is greater than 0.55, triggering a false alarm. This embodiment avoids this false alarm by using a dynamic threshold. After the event packet is generated, the terminal saves it to the local SQLite database or file system using a transactional atomic write method. Specifically, the event packet content is first written to a temporary file, and after the write is completed, the data integrity is verified. Then, the temporary file is replaced with the official file through an atomic renaming operation, ensuring that the event packet is not corrupted due to power outages or abnormalities during the write process. The terminal's local storage space is set with an upper limit for event packet storage, for example, the total size does not exceed 500 megabytes. When the storage space is close to the upper limit, a first-in-first-out cleanup strategy is triggered, deleting only those event packets that have been successfully uploaded to the server and have received confirmation receipts; event packets that have not been uploaded or have been uploaded but not confirmed are retained to ensure that no data is lost.

[0054] This invention compares the maximum probability in the calibration probability with an adaptive threshold, generates a JSON-formatted event package when the conditions are met, and performs a transactional atomic write to offline storage. At the same time, it adopts a first-in-first-out cleanup strategy to clean only the confirmed uploaded event packages. This solves the problems of high false alarm rate caused by fixed threshold judgment in the prior art, as well as the easy loss or damage of event data in the absence of network environment. It improves the reliability of defect judgment results, the integrity and traceability of event data, and reduces the false alarm rate and the risk of data loss.

[0055] S6. After the network is restored, the event packet is uploaded to the server, the server generates a work order and sends back an incremental update packet, and the terminal pulls the incremental update packet to complete the closed loop.

[0056] It also includes a penalty mechanism for unrepaired defects, specifically by counting the number of unrepaired defects in the current equipment. Define the penalty factor ,in The adaptive threshold is multiplied by the penalty factor as a penalty coefficient to suppress repeated reporting of existing unprocessed defects.

[0057] In one possible implementation, after the inspection is completed, the network is restored. The terminal detects that the network is available and idle, and automatically triggers the event packet upload process, uploading all locally stored but unuploaded event packets to the server one by one. Upon receiving the event packets, the server performs the following processing: Data validation and deduplication: The server verifies the uniqueness of the event identifier in the event packet to confirm that the event has not been submitted repeatedly; at the same time, it checks whether the device number exists in the device information master table. After successful verification, proceed to the next step. The server automatically generates standardized maintenance work orders based on information such as the device number, defect type code, and calibration probability in the event packet, following a preset work order template. The work order includes device location information, defect type, confidence level, and a thumbnail of the detected image, and is automatically assigned to the corresponding specialized maintenance team. After generation, the work order is stored in the server's work order database and marked as "pending dispatch." The database is updated, the server adds a new defect record to the defect history table, updates the corresponding field of the device in the device information master table, and triggers the predictive maintenance model to recalculate the remaining life prediction value. An incremental update package is generated. The server compiles statistics on the detection of defects in all types of equipment across the entire line after this report, dynamically adjusts the baseline thresholds and temperature coefficients for each equipment type, and packages the changed data into an incremental update package using SQLite incremental backup format or JSON Patch format, containing only the records that have been changed. The server database version number is updated accordingly, and the incremental update package is published to the update service. When idle, the terminal automatically queries the server for the latest version number. If the server version number is higher than the local version number, it automatically downloads the incremental update package. After downloading, it verifies the integrity of the update. If the verification is successful, it merges the update into the local SQLite database. Before merging, the terminal automatically backs up the current local database. If an error occurs during the merging process, it automatically rolls back to the backup version. After the update is complete, the terminal reloads the new baseline thresholds and temperature coefficients for subsequent offline inspections.

[0058] The unrepaired defect penalty mechanism is applied after dynamic threshold calculation and before defect determination. Its function is to further increase the determination threshold for equipment with reported but unrepaired defects when such defects exist, thus suppressing repeated reporting of the same type of defect. Specifically, before defect determination, the terminal queries the defect history table in the offline SQLite database to count the number of unrepaired defects currently existing in the device, i.e., the number of defect records where the repair completion date field is empty, denoted as nopen. A penalty coefficient γ is then used to define the penalty factor. The adaptive threshold calculated by S4 is multiplied by the penalty factor to obtain the final threshold used for judgment. If the equipment has no unrepaired defects, i.e., nopen=0, the penalty factor is 1, and the threshold remains unchanged. Through this mechanism, equipment with unresolved defects requires a higher confidence level in subsequent inspections to trigger defect reporting again, thereby effectively reducing redundant work orders and avoiding repeated alarms for the same defect.

[0059] This invention completes the closed loop by automatically uploading event packets to the server after network recovery, generating maintenance work orders and feeding back incremental update packets from the server, and then having the terminal pull the incremental update packets. At the same time, it introduces an unrepaired defect penalty mechanism to adjust the adaptive threshold for devices with unprocessed defects. This solves the problems of event data not being synchronized in a timely manner in the absence of network, identification strategies not being updated dynamically, and the repeated reporting of the same defect. It improves the timeliness of the operation and maintenance closed loop and the accuracy of defect identification, and reduces the risk of data retention and redundant work orders.

[0060] Example 2 like Figure 2 As shown, the present invention also provides an adaptive identification system for urban rail transit equipment defects in network-free environments. The system is used to implement the adaptive identification method for urban rail transit equipment defects in network-free environments described in any one of Embodiment 1. The system includes: The offline database pre-configuration module is used to pre-configure an offline SQLite database containing the operation and maintenance history of urban rail equipment; The feature parameter extraction module is used to automatically extract five feature parameters of the current device to be identified, based on an offline SQLite database: device critical level coefficient, device aging coefficient, historical defect frequency coefficient, maintenance effect coefficient, and predicted risk score. The confidence calibration module is used to deploy a lightweight YOLOv8 model for edge AI inference. It obtains the calibration probability by temperature scaling the original log odds output by the model and calculates the confidence uncertainty based on the calibration probability. An adaptive threshold calculation module is used to input five feature parameters and confidence uncertainty into a dynamic threshold calculation module, and calculate the adaptive threshold corresponding to the device according to a dynamic threshold formula that integrates cross-product terms and Sigmoid nonlinear mapping. The defect determination module is used to compare the maximum probability in the calibration probability with the adaptive threshold. If the maximum probability is not less than the adaptive threshold, a defect is determined to exist and an event package is generated for offline storage; otherwise, no defect is determined. The network synchronization module is used to upload the event packet to the server after the network is restored. The server generates a work order and sends back an incremental update packet. The terminal pulls the incremental update packet to complete the closed loop.

[0061] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An adaptive defect identification method for urban rail transit equipment in network-free environments, characterized in that, include: S1 is a pre-configured offline SQLite database containing the operation and maintenance history of urban rail equipment; S2, based on an offline SQLite database, automatically extracts five characteristic parameters of the current device to be identified: device critical level coefficient, device aging coefficient, historical defect frequency coefficient, maintenance effect coefficient, and predicted risk score. S3 deploys a lightweight YOLOv8 model for edge AI inference, obtains the calibration probability by temperature scaling of the original log odds output by the model, and calculates the confidence uncertainty based on the calibration probability. S4, input the five feature parameters and confidence uncertainty into the dynamic threshold calculation module, and calculate the adaptive threshold corresponding to the device according to the dynamic threshold formula that integrates the cross-product term and the Sigmoid nonlinear mapping; S5. Compare the maximum probability in the calibration probability with the adaptive threshold. If the maximum probability is not less than the adaptive threshold, a defect is determined to exist and an event package is generated for offline storage; otherwise, no defect is determined. S6. After the network is restored, the event packet is uploaded to the server, the server generates a work order and sends back an incremental update packet, and the terminal pulls the incremental update packet to complete the closed loop.

2. The adaptive defect identification method for urban rail transit equipment in network-free environments according to claim 1, characterized in that, The unique device code in the offline SQLite database consists of the line number, station number, device type, and serial number; the device operation and maintenance history includes static device attributes, maintenance work order records, inspection records, defect reporting records, and the remaining life prediction value output by predictive maintenance.

3. The adaptive defect identification method for urban rail transit equipment in network-free environments according to claim 1, characterized in that, In S2, the equipment criticality level coefficient K crit Obtained from the equipment criticality level mapping table; the higher the criticality level, the larger the coefficient. The equipment aging coefficient ,in, The length of time the equipment has been in operation. For the designed service life, This is an aging amplification factor. The historical defect frequency coefficient ,in This represents the average number of defects per month for the current equipment over the past 12 months. This represents the average monthly defect rate for similar equipment. Maintenance effect coefficient ,in This is the number of days since the last maintenance was completed. To account for the extent of the impact of the repairs This is the recovery time constant; Predictive Risk Score ,in This is the predicted remaining lifespan. This is a lifetime-scale parameter.

4. The adaptive defect identification method for urban rail transit equipment in network-free environments according to claim 1, characterized in that, In S3, the lightweight YOLOv8 model identifies five types of defects: cracks, water leakage, deformation, abnormal sounds, and overheating. Cracks, water leakage, and deformation are detected and identified by the YOLOv8 model using visible light images captured by the phone's camera. Abnormal sounds are detected by the YOLOv8 model using the phone's microphone to capture the sound of the device in operation, which is then converted into a spectrum by a short-time Fourier transform. Overheating is detected by the YOLOv8 model using an external USB-C infrared thermal imager module connected to the phone to capture infrared thermal images.

5. The adaptive defect identification method for urban rail transit equipment in network-free environments according to claim 1, characterized in that, The mapping relationship between different equipment defect types is executed according to the preset equipment defect type mapping table; When determining defects, if the defect type identified by the YOLOv8 model is not within the mapping range of the current device, the confidence weight of the detection box will be automatically reduced.

6. The adaptive defect identification method for urban rail transit equipment in network-free environments according to claim 3, characterized in that, In S3, the temperature scaling is probabilistically calibrated using the following formula: ； in For the first The primitive logarithmic odds of a class The optimal temperature coefficient is determined by minimizing the negative log-likelihood on the validation set. The optimal temperature coefficient is used as the temperature coefficient; the confidence uncertainty is defined as: in This represents the total number of defect categories. For the first time after calibration The probability of a class.

7. The adaptive defect identification method for urban rail transit equipment in a network-free environment according to claim 6, characterized in that, In S4, the formula for calculating the adaptive threshold is: ； Among them, T adapt For adaptive threshold, To set a preset baseline threshold, The comprehensive adjustment factor is defined as follows: ； in For the five characteristic parameters, These are the cross-term weight coefficients; ultimately Limit to a preset lower limit and upper limit between.

8. The adaptive defect identification method for urban rail transit equipment in network-free environments according to claim 1, characterized in that, In S5, the event packet adopts JSON format and includes at least the event unique identifier, device number, defect type code, calibration probability, confidence uncertainty, adaptive threshold, timestamp, and detection image path; the event packet is written using transactional atomic write and supports a first-in-first-out cleanup strategy for local storage space, only cleaning up event packets that have been successfully uploaded and confirmed.

9. The adaptive defect identification method for urban rail transit equipment in a network-free environment according to claim 1, characterized in that, It also includes a penalty mechanism for unrepaired defects, specifically by counting the number of unrepaired defects in the current equipment. Define the penalty factor ,in The adaptive threshold is multiplied by the penalty factor as a penalty coefficient to suppress repeated reporting of existing unprocessed defects.

10. An adaptive defect identification system for urban rail transit equipment in network-free environments, characterized in that: The system is used to implement the adaptive identification method for urban rail transit equipment defects in network-free environments as described in any one of claims 1 to 9, and the system comprises: The offline database pre-configuration module is used to pre-configure an offline SQLite database containing the operation and maintenance history of urban rail equipment; The feature parameter extraction module is used to automatically extract five feature parameters of the current device to be identified, based on an offline SQLite database: device critical level coefficient, device aging coefficient, historical defect frequency coefficient, maintenance effect coefficient, and predicted risk score. The confidence calibration module is used to deploy a lightweight YOLOv8 model for edge AI inference. It obtains the calibration probability by temperature scaling the original log odds output by the model and calculates the confidence uncertainty based on the calibration probability. An adaptive threshold calculation module is used to input five feature parameters and confidence uncertainty into a dynamic threshold calculation module, and calculate the adaptive threshold corresponding to the device according to a dynamic threshold formula that integrates cross-product terms and Sigmoid nonlinear mapping. The defect determination module is used to compare the maximum probability in the calibration probability with the adaptive threshold. If the maximum probability is not less than the adaptive threshold, a defect is determined to exist and an event package is generated for offline storage; otherwise, no defect is determined. The network synchronization module is used to upload the event packet to the server after the network is restored. The server generates a work order and sends back an incremental update packet. The terminal pulls the incremental update packet to complete the closed loop.

Images