Fault detection and diagnosis method for train purging robot

Abstract

The invention discloses a fault detection and diagnosis method for a train purging robot, and relates to the technical field of rail transit maintenance. The method comprises the following steps: collecting multi-source operation data of the purging robot in real time, wherein the multi-source operation data comprises equipment operation parameters, vibration data, temperature data and high-definition image data reflecting the state of key components of the equipment; performing feature extraction on the multi-source operation data; inputting the feature data set into a convolutional neural network model which is pre-trained based on a multi-fault type sample and is adapted through a transfer learning technology for real-time reasoning, and outputting a fault type, a fault position and a fault matching degree; making a decision according to the matching degree and triggering a corresponding alarm; and closed-loop feedback is formed after fault processing so as to optimize the model. According to the invention, real-time and accurate diagnosis and early warning of faults of the train purging robot are realized, and the operation and maintenance efficiency and the equipment reliability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of rail transit vehicle maintenance technology, specifically to a fault detection and diagnosis method for a sweeping robot used in subway, railway and other rail vehicles. Background Technology

[0002] During operation, subway and railway vehicles accumulate large amounts of dust, oil, and metal debris on their undercarriages and sides. Traditionally, cleaning these areas has relied on manual high-pressure airflow or fixed / pass-through cleaning equipment. These methods suffer from numerous drawbacks, including low efficiency, uneven cleaning results, high energy consumption, and safety hazards.

[0003] With the development of intelligent technology, intelligent cleaning robots have begun to be applied in this field. However, current intelligent cleaning equipment mainly focuses on cleaning path planning and dust removal execution, and its ability to monitor its own operating status in real time and provide fault early warning is generally insufficient. Cleaning robots typically operate in complex environments that are humid and dusty, and their key components (such as fans, robotic arms, and pipelines) are prone to wear, blockage, or damage during long-term operation. Once a fault occurs, it often requires shutdown for troubleshooting, affecting normal vehicle maintenance plans, and may even cause greater losses due to the escalation of the fault. Existing technology lacks an effective method for real-time and accurate online fault detection and diagnosis of the cleaning robot itself, making it difficult to achieve the transformation from "reactive maintenance" to "predictive maintenance."

[0004] Therefore, there is an urgent need for a method that can detect and diagnose operational faults of train cleaning robots in real time and with high accuracy, so as to ensure stable operation of the equipment and improve maintenance efficiency and intelligence. Summary of the Invention

[0005] (a) Purpose of the invention The purpose of this invention is to overcome the shortcomings of the prior art and provide a fault detection and diagnosis method for train cleaning robots. This method aims to achieve real-time online monitoring of the operating status of the cleaning robot, early and accurate identification and warning of faults, thereby reducing the rate of sudden equipment failure, reducing unplanned downtime, and improving operation and maintenance efficiency and safety.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides a fault detection and diagnosis method for a train cleaning robot, comprising the following steps: S1. Real-time acquisition of multi-source operating data of the purging robot, including equipment operating parameters, vibration data, temperature data, and high-definition image data reflecting the status of key components of the equipment; S2. Perform multimodal feature extraction on the multi-source operational data to obtain a structured feature dataset; wherein: Extract time-domain statistical features and frequency-domain spectral features from the operating parameters of the equipment; The vibration data is subjected to wavelet packet transform to extract the energy features of the fault characteristic frequency band associated with the preset fault type; For the high-definition image data, the key component regions of the device in the image are first identified and located based on the semantic segmentation model, and then texture and shape features are extracted in the region. S3. Input the structured feature dataset into the pre-trained fault identification model for real-time inference; the fault identification model is a multi-fault parallel diagnostic model built on a convolutional neural network, which is pre-trained using a labeled sample set containing multiple fault types, and the last fully connected layer of the model is fine-tuned using transfer learning technology to adapt to the target device; the model output includes fault type, fault location, and feature matching degree. S4. Make a hierarchical decision based on the feature matching degree: if the matching degree is greater than or equal to the first threshold, it is determined to be a confirmed fault and a level one alarm is triggered; if the matching degree is between the second threshold and the first threshold, it is determined to be a suspected fault and a level two alarm is triggered, where the first threshold is greater than the second threshold. S5. After triggering the alarm, generate and push alarm information containing the fault location, type and handling suggestions; after the fault is handled, store the characteristic data, diagnosis results and handling results of this fault in association to form a closed-loop feedback sample, which is used to incrementally learn and optimize the fault identification model.

[0007] Furthermore, in step S1, the equipment operating parameters include the fan motor current, fan speed, and the motor current and speed of each joint of the robotic arm; the acquisition frequency of the high-definition image data is not less than 15fps, and the resolution is not less than 1920×1080.

[0008] Furthermore, in step S2, the semantic segmentation model is a U-Net++ structure that introduces an attention mechanism. Its encoder part includes a channel attention module and a spatial attention module, which are used to enhance the feature weights of key component regions in complex backgrounds.

[0009] Furthermore, in step S3, the convolutional neural network includes an input layer, at least three convolutional-pooling layer groups, a global average pooling layer, and a fully connected output layer connected in sequence; wherein, the transfer learning technique specifically includes: based on the pre-trained model, freezing all weight parameters except for the last fully connected layer, and fine-tuning the fully connected layer using normal operating data of the target device.

[0010] Furthermore, in step S4, the first threshold is 95% and the second threshold is 80%.

[0011] Furthermore, in step S4, the triggering methods for the first-level alarm include activating the local audible and visual alarm, popping up a full-screen warning window on the control terminal, and pushing alarm information to the mobile terminal of the maintenance personnel; the triggering methods for the second-level alarm include controlling the local yellow warning light to flash and providing a text prompt on the control terminal. If the alarm is not deactivated within a preset 5 minutes or the matching degree continues to rise, it will automatically escalate to a first-level alarm.

[0012] Furthermore, in step S5, the incremental learning optimization specifically includes: periodically adding closed-loop feedback samples to the training dataset, retraining the weights of the fully connected layer of the fault identification model, and using knowledge distillation technology to retain the original fault category identification ability.

[0013] (III) Beneficial Effects Compared with the prior art, the present invention has the following beneficial effects: 1. Real-time performance and high accuracy: By fusing multi-source operational data and using an efficient CNN transfer learning model for real-time inference, fault feature matching can be completed within milliseconds, reducing fault response time from several hours in traditional manual inspection to minutes, while the fault identification accuracy can reach over 98%.

[0014] 2. Predictive and proactive: The method can issue early warnings (level two alarms) in the early or potential stages of a fault through in-depth analysis of features such as vibration spectrum, temperature trend, and image texture, transforming "post-fault maintenance" into "predictive maintenance" and effectively avoiding the escalation of faults and unplanned downtime.

[0015] 3. Adaptive and Continuous Optimization: The transfer learning technique employed enables the fault identification model to quickly adapt to different models or working environments of blow-drying robots, reducing model deployment costs. The closed-loop feedback mechanism continuously accumulates fault cases for incremental model learning, giving the diagnostic system self-evolution capabilities, and its diagnostic capabilities continuously improve over time.

[0016] 4. Intelligent and Easy-to-Use: The entire detection, diagnosis, and alarm process is completed automatically without manual intervention. Diagnostic results are directly linked to handling suggestions and pushed through multiple channels (sound, light, pop-ups, and mobile devices), greatly reducing the technical threshold and workload for maintenance personnel and improving maintenance efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1This is an overall flowchart of the fault detection and diagnosis method for the train cleaning robot provided by the present invention.

[0018] Figure 2 This is a flowchart of the fault detection and diagnosis method for the train cleaning robot provided in Embodiment 1 of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the key terms used in this invention will be explained below: Key components: These refer to the core functional components of the purging robot that are prone to failure and require status monitoring, including but not limited to the joints of the robotic arm, the blades and bearings of the blower system, and the high-pressure air pipelines and their interfaces.

[0021] Preset fault characteristic frequency bands: These refer to specific frequency ranges in the vibration signal spectrum where different types of faults exhibit significantly concentrated energy, determined through analysis of historical fault data. For example, wear faults in robotic arm joints may correspond to the 500-1000Hz frequency band.

[0022] Structured feature dataset: refers to a set of unified format feature vectors extracted from raw multi-source operational data (operational parameters, vibration, temperature, images) after specific processing, which can characterize the health status of equipment and can be used for model inference.

[0023] The fault detection and diagnosis method of this invention logically comprises the following collaborative modules: a multi-source data acquisition module, responsible for real-time acquisition of operating parameters, vibration, temperature, and image data; a feature extraction and preprocessing module, responsible for filtering and transforming the raw data and extracting time-domain, frequency-domain, and image features; an intelligent fault reasoning module, with a built-in fault recognition model based on CNN and transfer learning, responsible for real-time diagnosis; a decision and alarm output module, implementing tiered alarms based on the diagnostic results; and a closed-loop learning module, used to collect feedback and optimize the model. The method flow described below illustrates the collaboration of the above modules.

[0024] Please see Figure 1 As shown, the present invention provides a fault detection and diagnosis method for a train cleaning robot, comprising the following steps: S1. Real-time acquisition of multi-source operating data of the purging robot, including equipment operating parameters, vibration data, temperature data, and high-definition image data reflecting the status of key components of the equipment; S2. Perform multimodal feature extraction on the multi-source operational data to obtain a structured feature dataset; wherein: Extract time-domain statistical features and frequency-domain spectral features from the operating parameters of the equipment; The vibration data is subjected to wavelet packet transform to extract the energy features of the fault characteristic frequency band associated with the preset fault type; For the high-definition image data, the key component regions of the device in the image are first identified and located based on the semantic segmentation model, and then texture and shape features are extracted in the region. S3. Input the structured feature dataset into the pre-trained fault identification model for real-time inference; the fault identification model is a multi-fault parallel diagnostic model built on a convolutional neural network, which is pre-trained using a labeled sample set containing multiple fault types, and the last fully connected layer of the model is fine-tuned using transfer learning technology to adapt to the target device; the model output includes fault type, fault location, and feature matching degree. S4. Make a hierarchical decision based on the feature matching degree: if the matching degree is greater than or equal to the first threshold, it is determined to be a confirmed fault and a level one alarm is triggered; if the matching degree is between the second threshold and the first threshold, it is determined to be a suspected fault and a level two alarm is triggered, where the first threshold is greater than the second threshold. S5. After triggering the alarm, generate and push alarm information containing the fault location, type and handling suggestions; after the fault is handled, store the characteristic data, diagnosis results and handling results of this fault in association to form a closed-loop feedback sample, which is used to incrementally learn and optimize the fault identification model.

[0025] In step S1, the equipment operating parameters include the fan motor current, fan speed, and the motor current and speed of each joint of the robotic arm; the acquisition frequency of the high-definition image data is not less than 15fps, and the resolution is not less than 1920×1080.

[0026] In step S2, the semantic segmentation model is a U-Net++ structure that introduces an attention mechanism. Its encoder part includes a channel attention module and a spatial attention module, which are used to enhance the feature weights of key component regions in complex backgrounds.

[0027] In step S3, the convolutional neural network includes an input layer, at least three convolutional-pooling layer groups, a global average pooling layer, and a fully connected output layer connected in sequence; wherein, the transfer learning technique specifically includes: based on the pre-trained model, freezing all weight parameters except the last fully connected layer, and fine-tuning the fully connected layer using normal operating data of the target device.

[0028] The feature matching degree is a value between 0% and 100%, which physically represents the confidence score of the fault identification model in classifying the currently input structured feature dataset as a specific fault type. This value is directly derived from the output probability value of the fault identification model's output layer (Softmax activation function) corresponding to the fault category. For example, if the model's output probability for the "wind turbine bearing wear" category is 0.965, then the feature matching degree reported by the system is 96.5%. This value serves as the core basis for subsequent classification decisions, and its calculation process is entirely embedded in the model's forward inference process, requiring no additional computation.

[0029] In step S4, the first threshold is 95% and the second threshold is 80%.

[0030] The triggering methods for the Level 1 alarm include activating the local audible and visual alarm, popping up a full-screen warning window on the control terminal, and pushing alarm information to the mobile terminals of maintenance personnel; the triggering methods for the Level 2 alarm include controlling the local yellow warning light to flash and providing a text prompt on the control terminal. If the alarm is not deactivated within a preset 5 minutes or the matching degree continues to rise, it will automatically escalate to a Level 1 alarm.

[0031] In step S5, the incremental learning optimization specifically includes: periodically adding closed-loop feedback samples to the training dataset, retraining the weights of the fully connected layer of the fault identification model, and using knowledge distillation technology to retain the original fault category identification ability.

[0032] The following detailed explanation of each step is provided with reference to the embodiments, and the core fault identification model construction and training method is explained.

[0033] Methods for constructing and training fault identification models To achieve high-precision, parallel, real-time diagnosis of multiple faults, this invention constructs a fault identification model based on a convolutional neural network (CNN). The structure of this model mainly includes the following layers: Input layer: Receives the structured feature dataset generated in step S2. This dataset consists of normalized multidimensional feature vectors.

[0034] Feature fusion and abstraction layer: This layer consists of three consecutive convolutional-pooling block (Conv-Block). Each convolutional-pooling block comprises two convolutional layers (3×3 kernel size, ReLU activation function) and one max-pooling layer (2×2 pooling window, stride 2). This part of the network automatically learns deep correlations and non-linear abstract representations between features from different sensors.

[0035] Global Feature Layer: A global average pooling layer is applied after the last pooling layer to transform the feature map into a fixed-length feature vector. This design replaces the traditional fully connected layer, effectively reducing the number of model parameters, suppressing overfitting, and enhancing the model's robustness to changes in the input feature space.

[0036] Classification output layer: This is a fully connected layer with the same number of neurons as the system's preset number of fault types. The activation function is Softmax, and the output is a probability vector, where each element represents the confidence that the current input data belongs to a specific fault type, i.e., the feature matching degree.

[0037] The training of the model is divided into two stages: Pre-training phase: In a laboratory or historical operation and maintenance environment, a large-scale sample library covering common fault types of purging robots (such as wear of robotic arm joints, cracked fan blades, bearing overheating, pipe blockage / leakage, etc.) is constructed. This sample library contains multi-source operational data (operational parameters, vibration, temperature, images) synchronously collected under various fault conditions, as well as accurate fault labels verified by humans. The CNN model described above is trained end-to-end using this sample library. During training, the Adam optimization algorithm is used, with an initial learning rate of 1e-3, a categorical cross-entropy loss function, and Dropout technology (dropout rate of 0.5) is used to enhance the model's generalization ability.

[0038] Transfer learning adaptation phase: When deploying the diagnostic system on a specific model, batch, or working environment of a blower robot, transfer learning technology is used for rapid adaptation to overcome model performance degradation caused by individual differences and changes in working conditions. The specific operation process is as follows: Load the base model weights obtained during the pre-training phase.

[0039] The weight parameters of the first two convolutional-pooling layers of the model are frozen. These underlying networks have learned generalized fault feature patterns (such as edges, textures, and frequency domain basic features), exhibiting good generalization ability, and do not require modification.

[0040] Unfreeze the last convolutional-pooling layer group and subsequent global average pooling layers and classification output layers (fully connected layers). These high-level networks are responsible for learning specific feature combinations that are more relevant to the specific device model and operating conditions.

[0041] Use a continuous operating data segment (typically 24-48 hours, approximately 100,000 to 200,000 data records) collected from the target device under normal, fault-free conditions as the fine-tuning dataset. This dataset does not require fault labels.

[0042] Fine-tuning is performed on the thawed layers using a low learning rate (e.g., 1e-4 to 1e-5), with the training objective being to make the model's output feature distribution for these normally functioning data more concentrated and stable. Good results are typically achieved after 20-50 iterations.

[0043] This transfer learning strategy greatly reduces the dependence on the number of fault samples in the target equipment, enables rapid and low-cost transfer of diagnostic capabilities, and significantly improves the engineering practicality and universality of the method.

[0044] The fault detection and diagnosis method of the train cleaning robot provided by the present invention will be described in detail below with specific embodiments.

[0045] Example 1 like Figure 1 As shown, this embodiment provides a fault detection and diagnosis method for a train cleaning robot, specifically including the following steps: Step S101: Real-time acquisition of multi-source operation data.

[0046] Various sensors are deployed in key parts of the cleaning robot to build a comprehensive sensor network.

[0047] High-precision current sensors and photoelectric encoders are installed at the drive motors of the fan motor and each joint of the robotic arm to collect equipment operating parameters in real time, including fan current and power, and motor current and speed of each robotic arm joint. The data acquisition frequency is set to 10Hz. These electrical parameters can directly reflect the motor load status, efficiency changes, and potential electrical or mechanical jamming.

[0048] Triaxial vibration acceleration sensors and platinum resistance temperature sensors (PT100) are installed on key mechanical parts such as the fan bearing housing and the joint housings of the robotic arm to collect vibration and temperature data. The acquisition frequency is set to 50Hz. The vibration signal contains rich mechanical condition information (such as imbalance, misalignment, wear, and looseness), while the temperature is a direct indicator of overload, poor lubrication, or abnormal friction.

[0049] The robot is equipped with at least two high-definition industrial cameras (1920×1080 resolution), featuring wide-angle lenses and LED supplementary lighting. These cameras are positioned so that the lenses face the robotic arm's working area and the fan / pipeline area respectively, capturing high-definition images of the surface condition of key components at a frequency of 15fps. These images allow for the direct monitoring of visually identifiable faults such as cracks, rust, oil leaks, and loose or missing connectors on the component surfaces.

[0050] All sensor data is transmitted in real time via CAN bus or industrial Ethernet to the edge computing unit located in the robot body or a nearby electrical cabinet.

[0051] To accurately associate specific frequency bands in vibration signals with specific fault types, this method requires the establishment of a "fault-frequency band" mapping knowledge base before deployment. The construction process of this knowledge base is as follows: Historical fault data collection and analysis: Collect raw vibration acceleration data of key components of the purging robot (such as fan bearings and robotic arm joints) over a long period of time under known fault conditions (such as wear, cracks, and imbalance).

[0052] Signal processing and feature extraction: Wavelet packet transform (WPT) is performed on the vibration data samples for each type of fault to decompose them into multiple frequency bands. Statistical characteristics such as energy, kurtosis, and crescendo factor are calculated for each frequency band.

[0053] Sensitive Frequency Band Identification: By comparing the characteristic distribution differences between faulty samples and normal samples in each frequency band (e.g., using t-tests or calculating relative energy growth ratios), 1-3 key frequency bands that are most sensitive to this type of fault and exhibit the most significant energy changes are identified and designated as the "preset fault characteristic frequency bands" for this type of fault. For example, analysis may show that the characteristic energy of "robotic arm joint wear" faults is mainly concentrated in the 480-520Hz and 950-1050Hz frequency bands.

[0054] Knowledge base storage: The mapping relationship between "fault type - component location - preset characteristic frequency band (center frequency and bandwidth)" is stored in the system configuration file. During real-time feature extraction in step S102, the system calls the corresponding preset frequency band to calculate energy characteristics based on the currently monitored component.

[0055] It should be further explained that combining domain expert knowledge with data-driven analysis gives the feature extraction process a clear physical meaning, which greatly improves the sensitivity of vibration features to early faults and the interpretability of diagnosis, and avoids the computational redundancy and feature dilution caused by blind full-band analysis.

[0056] Step S102: Multimodal feature extraction of multi-source running data.

[0057] The edge computing unit processes the raw data collected in step S101 in real time, extracts multi-dimensional features that can characterize the health status of the device, and constructs a structured feature dataset.

[0058] Operating parameter feature extraction: For the collected current and speed sequences, calculate their time domain statistical features (including mean, standard deviation, peak-to-peak value, skewness, and kurtosis) and frequency domain spectral features (obtain the spectrum through Fast Fourier Transform (FFT) and extract the fundamental frequency amplitude, main harmonic amplitude and their ratio to the fundamental frequency).

[0059] Vibration signal feature extraction: The vibration acceleration signal undergoes a four-level wavelet packet transform (WPT) to decompose the signal into different frequency bands. Based on historical fault analysis, sensitive frequency bands for specific faults are preset (e.g., the frequency band corresponding to a fault in the outer ring of a wind turbine bearing is 800-1200Hz). The energy proportion and RMS value of the signal within these preset fault characteristic frequency bands are calculated as features.

[0060] Image feature extraction: This process consists of two sub-steps: Component Region Localization: Each high-resolution image frame acquired is input into a pre-trained semantic segmentation model. This model employs a U-Net++ architecture with an attention mechanism. During the encoder stage, channel attention and spatial attention modules enhance the feature weights related to key components such as "robotic arm links," "wind turbine blades," and "pipe joints." The model output is a pixel-level component region mask, accurately eliminating irrelevant regions such as the background, tracks, and vehicles.

[0061] Targeted Feature Calculation: Feature calculation is performed within the segmented component region mask. Texture Feature Extraction: The Gray-Level Co-occurrence Matrix (GLCM) is calculated to obtain indices such as contrast, correlation, energy, and homogeneity, which are used to identify surface scratches, corrosion, or oil adhesion. Shape Feature Extraction: The component contour is analyzed, and its area, perimeter, roundness, and rectangularity are calculated to identify component deformation, breakage, or missing bolts.

[0062] Finally, all extracted time-domain, frequency-domain, vibration energy, image texture, and shape features are combined and normalized in a predetermined order to form a unified multidimensional feature vector, i.e., a structured feature dataset.

[0063] Step S103: Fault reasoning based on CNN transfer learning model.

[0064] The structured feature dataset obtained in step S102 is input into the fault identification model that has completed transfer learning adaptation. The model completes forward propagation inference within an average of 30ms.

[0065] Model Output: The model outputs a diagnostic result that includes the fault type, fault location, and feature matching degree. For example, the output might be: {Fault Type: “Wind turbine bearing wear”, Fault Location: “Main wind turbine non-drive end bearing”, Feature Matching Degree: 96.5%}. The feature matching degree is the probability value of the corresponding category output by the model's Softmax layer, ranging from 0-100%.

[0066] Step S104: Hierarchical decision-making and alarm triggering.

[0067] The system makes automated hierarchical decisions based on the feature matching degree output by the model. In this embodiment, the first threshold is set to 95%, and the second threshold is set to 80%.

[0068] If the matching degree is ≥ 95%, it is judged as "confirmed fault" and a level 1 alarm is triggered immediately.

[0069] If 80% ≤ matching degree < 95%, it is judged as "suspected fault" and a level 2 alarm is triggered.

[0070] If the matching degree is less than 80%, it is judged as "no fault" and the data is archived for long-term trend analysis.

[0071] Actions to be performed after an alarm is triggered: Level 1 Alarm: Simultaneously activate the robot's local audible and visual alarm (high-frequency buzzer sounds continuously, red warning light stays on); a red full-screen emergency alarm window pops up on the monitoring screen of the remote centralized control center, accompanied by a prompt sound; and pushes structured alarm information, including fault location, detailed type, possible cause, and preliminary handling suggestions, to the mobile terminals of all members of the on-duty maintenance team in real time via WeChat, SMS, or a dedicated maintenance APP.

[0072] Level 2 Alarm: Triggers a local yellow warning light on the robot to flash at a frequency of 1Hz; a yellow text message is displayed in the sidebar or designated area of ​​the control center monitoring interface. The system starts a 5-minute observation timer. During this period, if the matching degree of subsequent consecutive inferences remains above 95% or shows an upward trend, the system automatically escalates the alarm to Level 1; if the matching degree decreases and stabilizes below 80%, the alarm is automatically cleared. This mechanism effectively filters out false alarms caused by transient electromagnetic interference, occasional sensor noise, etc.

[0073] Step S105: Closed-loop feedback and model optimization.

[0074] After the maintenance personnel have handled the fault on-site, they must enter the finally confirmed cause of the fault, the handling measures taken (such as "replacing the fan drive end bearing SKF 6308", "cleaning and tightening the No. 3 robotic arm air pipe joint") and the handling results into the maintenance log system of the control terminal.

[0075] The system automatically stores the complete data chain of this fault (from raw sensor data, extracted features, model inference results to the final report confirmed by humans) as a closed-loop feedback sample in the "Fault Diagnosis Case Library".

[0076] Every time a certain number of new samples are accumulated (e.g., 50), or periodically (e.g., quarterly), the system initiates incremental learning optimization of the model. The optimization process employs a strategy combining fine-tuning and knowledge distillation. The classification output layer of the fault identification model is fine-tuned using newly added closed-loop feedback samples, enabling it to learn new fault modes or optimize the decision boundary of existing modes.

[0077] To prevent the forgetting of old knowledge when adapting to new data (catastrophic forgetting), knowledge distillation is introduced. The model before optimization is used as the "teacher model," and the model during optimization is used as the "student model." A distillation loss term is added to the training loss function so that the student model can learn from new data while trying to mimic the output distribution of the teacher model for the original fault categories.

[0078] Through this closed loop, the diagnostic system achieves continuous evolution from "initial intelligence" to "experience intelligence," with diagnostic accuracy and reliability constantly improving over time.

[0079] It should be further explained that in the incremental optimization of the model in step S105, in order to balance "learning new faults" and "retaining old knowledge", this invention adopts knowledge distillation technology, which is implemented as follows: Model role definition: The fault identification model before optimization is fixed as the "Teacher Model", and the copy of the model to be optimized is used as the "Student Model".

[0080] Distillation loss design: The total loss function when training student models. Ltotal It consists of two parts: Ltotal = α x Ltask +(1- α )× Ldistill ; in: Ltask The standard cross-entropy loss is used to calculate the difference between the student model output and the true label of the new batch of closed-loop feedback samples.

[0081] Ldistill ; represents the distillation loss, used to measure the similarity between the output probability distributions of the student and teacher models. Specifically, the inputs to the Softmax layers of both the teacher and student models are first "softened" by dividing them by a temperature parameter T (T=2 in this example) to obtain a smoother probability distribution P. T and P S Then, the KL divergence between the two is calculated as... Ldistill ; .

[0082] α is a balancing weight used to adjust the degree of learning new tasks and retention of old knowledge; in this embodiment, it is set to 0.7.

[0083] Optimization process: Using a mixed dataset containing both new and old samples, with the total loss function L... total The student model is trained. After optimization, the student model can maintain the recognition accuracy of the original fault database to the greatest extent while mastering new fault modes.

[0084] The beneficial effects of this design are as follows: through temperature parameters and smooth probability distribution, knowledge distillation technology effectively transmits the "hidden knowledge" such as inter-category relationships implicit in the teacher model, making the generalization ability of the student model stronger, systematically alleviating the catastrophic forgetting problem in incremental learning, and ensuring the stability of the long-term evolution of the diagnostic system.

[0085] Example 2: Model Transfer Learning Adaptation and Effect Validation This embodiment demonstrates the actual adaptation process and effect of migrating this diagnostic system from a model A purging robot to a model B robot.

[0086] Adaptation Operation: Model B robots differ from Model A in their robotic arm structure and fan power. We collected 48 consecutive hours (approximately 173,000 data points) of normal operating data from Model B robots. Following the aforementioned transfer learning method, we froze the first four convolutional layers of the pre-trained model (trained based on Model A data), and then fine-tuned the remaining network layers using normal data from Model B, iterating for a total of 40 rounds with a learning rate of 1e-4.

[0087] Effect verification: After adaptation, the B-type robot's own fault test set (containing 150 samples, covering 8 types of faults) was used for evaluation.

[0088] Accuracy: The overall fault identification accuracy reached 98.2%, which is a significant improvement compared to directly using the A-type pre-trained model (accuracy 89.5%).

[0089] Response time: The average time for single fault reasoning is 32ms, which meets the real-time requirements.

[0090] Energy consumption impact: Through successful early warning of wear and tear failures, three instances of abnormal high-power operation of the equipment due to the expansion of failures were avoided. It is estimated that each robot can save approximately 18% of its annual power consumption.

[0091] Example 3: Optimization of Image Segmentation Model This embodiment further optimizes the semantic segmentation model used in the "part region localization" sub-step of image feature extraction in step S102, based on embodiment 1. The model adopts an improved U-Net++ structure that integrates a dual attention mechanism (CBAM) and dense connection paths.

[0092] A Convolutional Block Attention (CBAM) module is inserted after each downsampling stage of the encoder. This module performs channel attention calculation and spatial attention calculation in sequence, enabling the model to adaptively focus on feature channels and spatial locations in the image that are more important for part segmentation.

[0093] Dense connections are introduced between the encoder and decoder, as well as between different layers of the decoder, to facilitate the reuse and fusion of features at different scales.

[0094] This optimization improved the mIoU (mean intersection-union ratio) of component region segmentation by approximately 4.7% to 96.8% under complex conditions such as drastic changes in lighting, partial occlusion of components, or oil contamination, providing more accurate and robust target regions for subsequent feature extraction.

[0095] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for fault detection and diagnosis of a train cleaning robot, characterized in that, Includes the following steps: S1. Real-time acquisition of multi-source operating data of the purging robot, including equipment operating parameters, vibration data, temperature data, and high-definition image data reflecting the status of key components of the equipment; S2. Perform multimodal feature extraction on the multi-source operational data to obtain a structured feature dataset; wherein: Extract time-domain statistical features and frequency-domain spectral features from the operating parameters of the equipment; The vibration data is subjected to wavelet packet transform to extract the energy features of the fault characteristic frequency band associated with the preset fault type; For the high-definition image data, the key component regions of the device in the image are first identified and located based on the semantic segmentation model, and then texture and shape features are extracted in the region. S3. Input the structured feature dataset into the pre-trained fault identification model for real-time inference; the fault identification model is a multi-fault parallel diagnostic model built on a convolutional neural network, which is pre-trained using a labeled sample set containing multiple fault types, and the last fully connected layer of the model is fine-tuned using transfer learning technology to adapt to the target device; the model output includes fault type, fault location, and feature matching degree. S4. Make a hierarchical decision based on the feature matching degree: if the matching degree is greater than or equal to the first threshold, it is determined to be a confirmed fault and a level one alarm is triggered; if the matching degree is between the second threshold and the first threshold, it is determined to be a suspected fault and a level two alarm is triggered, where the first threshold is greater than the second threshold. S5. After triggering the alarm, generate and push alarm information containing the fault location, type and handling suggestions; after the fault is handled, store the characteristic data, diagnosis results and handling results of this fault in association to form a closed-loop feedback sample, which is used to incrementally learn and optimize the fault identification model.

2. The method according to claim 1, characterized in that, In step S1, the equipment operating parameters include the fan motor current, fan speed, and the motor current and speed of each joint of the robotic arm; the acquisition frequency of the high-definition image data is not less than 15fps, and the resolution is not less than 1920×1080.

3. The method according to claim 1, characterized in that, In step S2, the semantic segmentation model is a U-Net++ structure that introduces an attention mechanism. Its encoder part includes a channel attention module and a spatial attention module, which are used to enhance the feature weights of key component regions in complex backgrounds.

4. The method according to claim 1, characterized in that, In step S3, the convolutional neural network includes an input layer, at least three convolutional-pooling layer groups, a global average pooling layer, and a fully connected output layer connected in sequence; wherein, the transfer learning technique specifically includes: based on the pre-trained model, freezing all weight parameters except the last fully connected layer, and fine-tuning the fully connected layer using normal operating data of the target device.

5. The method according to claim 1, characterized in that, In step S4, the first threshold is 95% and the second threshold is 80%.

6. The method according to claim 1 or 5, characterized in that, In step S4, the triggering methods for the first-level alarm include activating the local audible and visual alarm, popping up a full-screen warning window on the control terminal, and pushing alarm information to the mobile terminal of the maintenance personnel; the triggering methods for the second-level alarm include controlling the local yellow warning light to flash and providing a text prompt on the control terminal. If the alarm is not deactivated within a preset 5 minutes or the matching degree continues to rise, it will automatically be upgraded to a first-level alarm.

7. The method according to claim 1, characterized in that, In step S5, the incremental learning optimization specifically includes: periodically adding closed-loop feedback samples to the training dataset, retraining the weights of the fully connected layer of the fault identification model, and using knowledge distillation technology to retain the original fault category identification ability.

Images