A test diagnosis system and method for home appliance repair

Abstract

The application discloses a kind of test diagnosis system and method of household appliance maintenance, belong to the technical field of intelligent fault diagnosis of household appliance. Specifically include the following steps: S1, historical data construction and pre-processing: the historical running data of multiple categories, multiple brands of household appliances, fault labeling data, maintenance case data are collected, and the original fault dataset is constructed;The original fault dataset is pre-processed in multiple levels, and the standardized training dataset is obtained;S2, multi-level feature extraction: a multi-level feature extraction network is constructed, and the training dataset is progressively deep feature extraction. Through three levels of cascade, progressive deep feature extraction from bottom signal level→component level→system level is realized in the vertical feature extraction architecture of step-by-step dependence, fully conforms to the evolution law that household appliance fault evolves from bottom parameter anomaly to component performance degradation, and then to system fault outbreak, fully excavates the hierarchical correlation features and causal relationship of fault.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology for household appliances, and in particular to a testing and diagnostic system and method for repairing household appliances. Background Technology

[0002] As home appliances become increasingly intelligent and integrated, their functions are becoming more and more diverse, and their internal circuits and mechanical structures are becoming more and more complex, posing a great challenge to after-sales maintenance and fault diagnosis.

[0003] Currently, fault diagnosis in the home appliance repair industry mainly relies on the manual experience of repair personnel, using methods such as reading fault codes, visual inspection, multimeter measurement, and component replacement. This approach has the following core drawbacks: First, it suffers from low diagnostic efficiency and poor fault location accuracy, making it difficult to identify hidden faults and faults coupled across components. Second, it is highly dependent on the professional experience of repair personnel, making it difficult for novice repair personnel to learn and resulting in high repair costs. Third, for new smart home appliances, manual diagnosis cannot fully analyze the hidden fault information in the main control board's operating logs and timing data, leading to a high rate of misdiagnosis.

[0004] Most existing AI-based home appliance fault diagnosis technologies employ traditional machine learning algorithms or simple deep learning models. Their core deficiency lies in the inadequacy of the feature extraction stage. Existing solutions often use single-level or multi-branch parallel feature extraction architectures, failing to fully exploit the hierarchical and progressive correlation features of home appliance faults, from underlying signal anomalies to component performance degradation and finally to system-level failures. For complex scenarios involving the same fault manifesting differently, or the same phenomenon presenting with different faults, as well as cross-component coupled faults and early-stage latent faults, existing solutions suffer from low accuracy and poor generalization ability, making them unsuitable for home appliances from multiple brands, categories, and age groups. Furthermore, most existing systems only support offline data analysis, failing to achieve a complete closed-loop process of "on-site testing - real-time diagnosis - repair guidance," resulting in insufficient practicality and applicability.

[0005] In view of the many shortcomings of existing technologies, there is an urgent need to propose a home appliance repair, testing and diagnosis solution that can fully explore the characteristics of fault hierarchy, has high diagnostic accuracy, strong generalization ability and can realize a closed loop throughout the entire process. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to provide a testing and diagnostic system and method for home appliance repair; it can solve the problems of existing technologies that fail to fully explore the characteristics of fault hierarchy, have low diagnostic accuracy, poor generalization ability, and cannot achieve a closed-loop process.

[0007] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a testing and diagnostic method for repairing household appliances, specifically including the following steps:

[0008] S1. Historical Data Construction and Preprocessing: Collect historical operating data, fault labeling data, and repair case data of multiple categories and brands of household appliances to construct the original fault dataset; perform multi-level preprocessing on the original fault dataset to obtain a standardized training dataset;

[0009] S2. Multi-level Feature Extraction: Construct a multi-level feature extraction network to perform progressive deep feature extraction on the training dataset, where:

[0010] First-level signal feature matrix: Using preprocessed standardized data as input, extract time-domain, frequency-domain, and statistical basic signal features of household appliance operation data;

[0011] Second-level signal feature matrix: The input is the first-level signal feature matrix. The features are grouped according to the core functional components of the household appliance to be diagnosed, and the time-series dependence, operating condition matching, and parameter correlation features within each component are extracted.

[0012] The third-level signal feature matrix: The input is the second-level signal feature matrix, which is used to extract cross-component coupling correlation features, fault propagation path features, and fault mode mapping features;

[0013] S3. Model Training and Optimization: Using the third-level signal feature matrix as input and the corresponding fault labeling data, fault location information, and maintenance plan data as labels, a multi-task learning fault diagnosis model is constructed. The model training and hyperparameter optimization are completed through the training set and validation set to obtain the target fault diagnosis model.

[0014] S4. On-site testing and diagnosis: Real-time operating test data of the household appliance to be diagnosed is collected through a portable test acquisition terminal. After preprocessing, the data is input into the target fault diagnosis model. The model outputs fault diagnosis results, accurate fault location information, and corresponding standardized repair guidance schemes.

[0015] S5. Model Iteration Update: Supplement the original fault dataset with the actual maintenance results of on-site diagnosis and newly added fault case data, and repeat steps S1-S3 to incrementally update and iteratively optimize the fault diagnosis model.

[0016] Furthermore, the historical operating data includes electrical operating parameters of household appliances, sensor timing data, main control board operating logs, fault code data, and operating mode data; the fault labeling data includes fault type, fault location, fault cause, and fault severity; the repair case data includes repair steps, replacement part models, and acceptance standards; and the multi-level preprocessing sequentially includes data cleaning, data alignment, data encoding, data normalization, data augmentation, and dataset partitioning.

[0017] Furthermore, the data cleaning process employs linear interpolation to fill missing values ​​and a combination of the 3σ criterion and the isolated forest algorithm to remove outliers; the data alignment process resamples multi-source time-series data, unifying the sampling frequency and time dimension; the data encoding process performs one-hot encoding on categorical data and sliding window encoding on time-series data; the data normalization process uses Min-Max normalization to map numerical data to the [0,1] interval; the data augmentation process expands the sample size using sliding window sampling, Gaussian noise addition, and working condition translation enhancement; and the dataset partitioning process divides the samples into training, validation, and test sets according to a preset ratio.

[0018] Furthermore, step S2 specifically includes:

[0019] S21. First-level signal feature matrix extraction: The first-level signal feature extraction layer adopts a structure of a 3-layer one-dimensional convolutional neural network (1D-CNN) cascaded with 2 fully connected layers. The convolutional kernel sizes are 7, 5, and 3 respectively, and the stride is 2. Basic signal features are extracted through convolution operations, and the first-level signal feature matrix is ​​output.

[0020] S22. Second-level signal feature matrix extraction: The second-level component-level correlation feature extraction layer sets up parallel feature extraction units that match the number of core functional components of the household appliance. Each unit adopts a structure of a bidirectional long short-term memory network Bi-LSTM cascaded attention mechanism. Each unit only inputs the correlation signal features of the corresponding functional component in the first-level signal feature matrix. The temporal dependency features inside the component are extracted by Bi-LSTM, and the fault-sensitive features are highlighted by weighting through the attention mechanism. The component feature vectors output by each unit are concatenated to obtain the second-level component feature matrix.

[0021] S23. Third-level signal feature matrix extraction: The third-level system-level fault deep feature extraction layer adopts the structure of a graph attention network (GAT) cascaded cross-attention mechanism. The feature vectors of each functional component in the second-level component feature matrix are used as graph nodes, and the electrical connections and control logic associations of the household appliance system are used as edges to construct a system topology graph. The coupling associations and fault propagation path features across components are extracted through GAT, and the accurate mapping between features and fault modes is achieved through the cross-attention mechanism, outputting the third-level signal feature matrix.

[0022] Furthermore, the multi-task learning fault diagnosis model includes a fault classification branch, a fault location regression branch, and a maintenance plan matching branch. The fault classification branch uses the Softmax activation function to identify fault types and normal operating conditions. The fault location regression branch outputs accurate location information and fault severity for faulty components and parts. The maintenance plan matching branch uses a cosine similarity matching algorithm to output standardized maintenance guidance plans that match the fault. The model loss function is a weighted sum of cross-entropy loss and mean squared error loss, and the optimizer uses AdamW, with a cosine annealing learning rate decay strategy and an early stopping mechanism.

[0023] Furthermore, the portable test and acquisition terminal includes an electrical parameter acquisition unit, a sensor acquisition unit, an appliance communication unit, a data processing unit, and a display unit. The electrical parameter acquisition unit is used to acquire the voltage, current, resistance, and frequency electrical parameters of the appliance to be diagnosed. The sensor acquisition unit is used to acquire temperature, vibration, and noise operating condition data. The appliance communication unit communicates with the main control board of the appliance to be diagnosed via infrared, Bluetooth, WiFi, or CAN bus to read operating logs and fault codes. The data processing unit performs real-time data preprocessing and model inference.

[0024] According to another aspect of the present invention, a testing and diagnostic system for home appliance repair is provided, characterized in that: the system is used in the above-described testing and diagnostic method for home appliance repair, comprising: a data acquisition module, a data preprocessing module, a three-level vertical feature extraction module, a model training module, a diagnostic execution module, and a human-computer interaction module;

[0025] Data acquisition module: used to collect historical fault data of household appliances and real-time test data of the appliance to be diagnosed;

[0026] Data preprocessing module: used to perform multi-level preprocessing on the collected raw data and output standardized data;

[0027] The three-level vertical feature extraction module includes a first-level signal-level feature extraction unit, a second-level signal-level feature extraction unit, and a third-level signal-level feature extraction unit. The only input of the second-level signal-level feature extraction unit is the output of the first-level signal-level feature extraction unit, and the only input of the third-level signal-level feature extraction unit is the output of the second-level signal-level feature extraction unit.

[0028] Model training module: Used to train, optimize and incrementally update the fault diagnosis model based on the extracted deep fault feature set, and output the target fault diagnosis model;

[0029] Diagnostic execution module: Used to input standardized real-time data of the home appliance to be diagnosed into the target fault diagnosis model, perform inference, and output diagnostic results and repair guidance plan;

[0030] Human-computer interaction module: used to display diagnostic results, receive user operation commands, and output visual maintenance guidance;

[0031] The data acquisition module includes a historical data acquisition submodule and a field test acquisition submodule. The historical data acquisition submodule connects to the after-sales database of home appliance manufacturers and the database of third-party repair platforms to acquire historical fault and repair data in batches. The field test acquisition submodule is a portable test acquisition terminal used to collect real-time operating data and electrical test data of the home appliance to be diagnosed on-site.

[0032] Beneficial effects: This invention achieves progressive deep feature extraction from the bottom signal level to the component level to the system level through a three-level cascaded, progressively dependent vertical feature extraction architecture. It perfectly matches the evolution of household appliance faults from abnormal bottom parameters to component performance degradation, and then to system-level fault outbreak. It fully explores the hierarchical correlation features and causal relationships of faults, and solves the core pain points of insufficient feature extraction and low recognition rate of latent and coupled faults in existing technologies. It significantly improves the accuracy and generalization ability of fault diagnosis and can be adapted to household appliances of multiple brands, categories and different service years.

[0033] This invention constructs a closed-loop solution covering the entire process of "data acquisition - preprocessing - feature extraction - model training - on-site diagnosis - repair guidance - model iteration". It supports both offline model training and data analysis, as well as synchronous acquisition, preprocessing and online diagnosis of real-time test data on-site. With the accompanying portable test acquisition terminal, it can be directly implemented on-site in home appliance repair, making it highly practical and significantly reducing the reliance on the experience of repair personnel. Novice repair personnel can quickly complete accurate diagnosis and repair.

[0034] This invention employs a multi-task learning framework, which can simultaneously achieve fault type identification, accurate fault location, fault severity assessment, and repair plan matching. It can not only identify major fault categories but also accurately locate specific faulty components, distinguish the causal relationship of faults, avoid misdiagnosis, and output standardized repair steps, part numbers, and acceptance standards. It is highly instructive for repairs, effectively improving the first-time pass rate and reducing repair costs.

[0035] The three-level vertical feature extraction architecture of this invention has strong scalability. It can flexibly adjust component grouping, network parameters and topology according to the structural characteristics of different types of household appliances. It can be seamlessly adapted to fault diagnosis of all types of household appliances such as air conditioners, washing machines, refrigerators, water heaters, televisions, and kitchen appliances, and has a very wide range of applications. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the method. Detailed Implementation

[0037] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] Example 1

[0039] This embodiment targets household inverter wall-mounted air conditioners, adapting to mainstream brands such as Midea, Gree, and Haier with a service life of 1-10 years. It enables on-site testing, intelligent diagnosis, and standardized repair guidance for air conditioner malfunctions, fully implementing the testing and diagnosis methods and systems described in claims 1-8 of this invention. The specific implementation process is as follows:

[0040] I. Historical Data Construction and Preprocessing

[0041] In this embodiment, the historical data acquisition submodule of the data acquisition module connects to the after-sales databases of three leading air conditioner manufacturers and the databases of two nationwide home appliance repair platforms to collect full historical data of multiple brands and models of inverter air conditioners, constructing an original fault dataset. Specific data collection includes:

[0042] Historical operating data: electrical operating parameters of the air conditioner (compressor input voltage / current / frequency, indoor and outdoor fan speed, electronic expansion valve opening, total power), sensor time-series data (indoor / outdoor ambient temperature, evaporator / condenser coil temperature, compressor discharge temperature, return gas pressure sensor data), main control board operation log, fault code data, cooling / heating / dehumidification / standby mode data, with a sampling frequency of 1Hz and a duration of 60s for each time-series data entry;

[0043] Fault labeling data: Labels the fault type (compressor phase loss, electronic expansion valve sticking, fan bearing wear, refrigerant leakage, main control board IPM module damage, etc., 28 common air conditioning faults), fault location (outdoor unit compressor, electronic expansion valve, indoor unit cross-flow fan, main control board IPM module, etc.), fault cause, and fault severity (minor warning, moderate fault, severe shutdown).

[0044] Repair case data: Standardized repair steps for corresponding faults, replacement part models, and post-repair acceptance standards (acceptance indicators such as operating pressure, inlet and outlet air temperature difference, and operating noise).

[0045] The original fault dataset, after data collection, contains 128,000 samples, including 32,000 samples from normal operating conditions and 96,000 samples from various fault types. The original dataset undergoes multi-level preprocessing via a data preprocessing module, as follows:

[0046] Data cleaning: Linear interpolation was used to fill missing values ​​in the time series data, and samples with a missing value rate of more than 30% were directly removed; outliers were removed by combining the 3σ criterion and the isolated forest algorithm. First, the 3σ criterion was used to remove numerical outliers that exceeded 3 times the standard deviation, and then the isolated forest algorithm was used to identify discrete outliers and pseudo-fault samples. In the end, more than 4,200 invalid samples were removed.

[0047] Data alignment: Resample multi-source time series data with a uniform sampling frequency of 1Hz to align the time dimension and ensure that the time series length of all data in a single sample is consistent at 60 time steps;

[0048] Data encoding: One-hot encoding is performed on classified data such as operating mode and fault code, and sliding window encoding is used for 60-second time series data. The sliding window size is 5 and the step size is 1 to generate time series feature sequences.

[0049] Data normalization: Min-Max normalization is used to map all numerical data to the [0,1] interval to eliminate the influence of data with different units;

[0050] Data augmentation: The sample size was expanded by sliding window sampling, Gaussian noise addition, and working condition shifting. Sliding window sampling resampled 60-second time series data with a step size of 2. Gaussian noise addition was set with a signal-to-noise ratio of 20dB. Working condition shifting was performed by shifting features by ±2℃ for different ambient temperature working conditions. Finally, the sample size was expanded to 220,000.

[0051] Dataset partitioning: The samples are divided into a training set (176,000 records), a validation set (22,000 records), and a test set (22,000 records) according to a preset ratio of 8:1:1, resulting in a standardized training dataset.

[0052] II. Multi-level Feature Extraction

[0053] A three-level cascaded vertical feature extraction network is constructed using a three-level vertical feature extraction module to perform progressive deep feature extraction on the training dataset. The specific steps are as follows:

[0054] First-level signal feature matrix extraction: The first-level signal-level primary feature extraction layer adopts a structure of 3-layer one-dimensional convolutional neural network (1D-CNN) cascaded with 2 fully connected layers. The convolution kernel sizes are 7, 5, and 3 respectively, with a stride of 2 for each layer. The activation function is ReLU. Through convolution operations, the basic signal features of the air conditioner operation data in the time domain, frequency domain, and statistical categories are extracted, including the time domain fluctuation features of voltage / current, frequency domain harmonic features, and statistical features such as mean / variance / peak value. The final output is a first-level signal feature matrix with a dimension of 256×1.

[0055] Second-level signal feature matrix extraction: The second-level component-level correlation feature extraction layer sets up five parallel feature extraction units for the five core functional components of the variable frequency air conditioner (compressor drive unit, indoor fan unit, outdoor fan unit, throttling control unit, and environmental sensing unit). Each unit adopts a structure of a bidirectional long short-term memory network (Bi-LSTM) cascaded attention mechanism, with a Bi-LSTM hidden layer dimension of 128 and an attention mechanism weight dimension of 64. Each unit only inputs the correlation signal features of the corresponding functional component in the first-level signal feature matrix. For example, the compressor drive unit only inputs correlation features such as compressor voltage, current, frequency, and exhaust temperature. The Bi-LSTM extracts the temporal dependency features within the component, capturing the performance degradation law of the component over time. The attention mechanism weights and highlights fault-sensitive features and suppresses interference from irrelevant features. The component feature vectors output by the five units are concatenated to obtain a second-level component feature matrix with a dimension of 640×1.

[0056] The third-level signal feature matrix extraction layer: The third-level system-level fault deep feature extraction layer adopts a structure of cascaded graph attention network (GAT) with a cross-attention mechanism. Using the feature vectors of the five core functional components in the second-level component feature matrix as graph nodes, and the electrical connections and refrigerant circulation control logic of the air conditioning system as edges, a system topology graph is constructed. The GAT sets two attention heads, each with eight attention heads, and uses LeakyReLU as the activation function. The GAT extracts cross-component coupling and fault propagation path features, such as the cascading impact of compressor fault-induced exhaust temperature anomalies on the throttling unit and heat exchange unit. Then, the cross-attention mechanism achieves accurate mapping between deep features and fault modes, ultimately outputting a third-level signal feature matrix with a dimension of 512×1.

[0057] III. Model Training and Optimization

[0058] Through the model training module, a multi-task learning fault diagnosis model is constructed using the third-level signal feature matrix as input and the corresponding fault labeling data, fault location information, and maintenance plan data as labels. The specific architecture and training settings are as follows:

[0059] The multi-task learning-based fault diagnosis model includes three parallel branches:

[0060] Fault classification branch: The input is the third-level signal feature matrix. It adopts 2 fully connected layers + Softmax activation function and outputs 29 classification results (28 faults + 1 normal condition) to identify fault types and normal conditions.

[0061] Fault location regression branch: It adopts a 3-layer fully connected layer to output the accurate location information of the faulty component and the faulty part, and at the same time outputs the fault severity value in the range of 0-1. The closer the value is to 1, the more severe the fault.

[0062] Repair Solution Matching Branch: A standardized repair solution feature library is pre-stored. The cosine similarity matching algorithm is used to match the fault features output by the model with the solution feature library, and output the top 3 standardized repair guidance solutions with the highest similarity, including repair steps, replacement parts models and acceptance standards.

[0063] The core parameters for model training are set as follows: The loss function is a weighted sum of cross-entropy loss (fault classification branch) and mean squared error loss (fault location regression branch), with weights of 0.6 and 0.4, respectively; the optimizer is AdamW, with an initial learning rate of 1e-4, a cosine annealing learning rate decay strategy, T_max of 20, and an early stopping mechanism. If the validation set loss does not decrease for 10 consecutive epochs, training is terminated early to prevent overfitting.

[0064] The model was trained using the training set, and hyperparameters were optimized using the validation set. Finally, the model was tested on the test set. The model achieved a fault classification accuracy of 98.72% and a fault location accuracy (down to the component level) of 96.35%, thus obtaining the target fault diagnosis model, which was then deployed to a portable test acquisition terminal and a cloud backend.

[0065] IV. On-site Testing and Diagnostic Implementation

[0066] Through the diagnostic execution module and the human-machine interaction module, in conjunction with a portable test and acquisition terminal, on-site testing and diagnosis are completed. The portable test and acquisition terminal used in this embodiment strictly follows the architecture of claim 6, integrating an electrical parameter acquisition unit, a sensor acquisition unit, a home appliance communication unit, a data processing unit, and a touch display unit, wherein:

[0067] Electrical parameter acquisition unit: Employs a high-precision multimeter module and oscilloscope module to acquire electrical parameters such as voltage, current, resistance, and frequency of the air conditioner under diagnosis. Measurement accuracy: voltage ±0.5%FS, current ±0.8%FS;

[0068] Sensor acquisition unit: integrates temperature sensor, vibration sensor, and noise sensor to collect temperature, vibration, and noise operating data of the air conditioner;

[0069] Home appliance communication unit: Supports infrared, Bluetooth, WiFi and CAN bus communication, can communicate with the main control board of mainstream brand air conditioners, and read operation logs and fault codes;

[0070] Data processing unit: It adopts an ARM Cortex-M7 core processor, with built-in preprocessing algorithms and a lightweight target fault diagnosis model to complete real-time data preprocessing and local model inference;

[0071] Touch display unit: Uses a 7-inch touch screen for human-computer interaction, display of diagnostic results, and visualization of maintenance guidance.

[0072] Specific procedures for on-site diagnosis:

[0073] The maintenance personnel connected the portable test and acquisition terminal to the electrical interface and main control board communication interface of the inverter air conditioner to be diagnosed, started the equipment, and selected the brand, model, and service life of the corresponding air conditioner on the human-machine interface.

[0074] The terminal synchronously collects 60 seconds of real-time operation test data under the air conditioner's cooling / heating conditions through the electrical parameter acquisition unit, sensor acquisition unit, and home appliance communication unit. This data includes electrical parameters, sensor timing data, main control board operation logs, and fault codes.

[0075] The terminal's data processing unit performs the same preprocessing steps as described above on the real-time collected data, including data cleaning, alignment, encoding, and normalization, to generate standardized real-time data.

[0076] Standardized real-time data is input into the lightweight target fault diagnosis model. After the model completes inference, it outputs fault diagnosis results, accurate fault location information, and corresponding standardized maintenance guidance.

[0077] The diagnostic results are displayed visually via a touch screen unit. Example output: Fault type: compressor IPM module damaged; Fault location: outdoor unit main control board IPM power module; Fault severity: severe shutdown. Standardized repair plan is output simultaneously: ① Power off and disassemble; ② Remove the main control board protective cover; ③ Check the continuity of IPM module pins; ④ Replace with the same model IPM module; ⑤ Power on and test; ⑥ Complete acceptance according to the standard of inlet and outlet air temperature difference ≥8℃ and operating current stable within the rated range, while indicating the model of compatible parts.

[0078] V. Model Iteration and Update

[0079] The actual repair results of this on-site diagnosis and the newly added fault case data (including post-repair verification data) are uploaded to the cloud backend through the terminal to supplement the original fault dataset. The entire process of data preprocessing, feature extraction and model training mentioned above is repeated monthly to incrementally update and iteratively optimize the fault diagnosis model and continuously improve the model's ability to identify new air conditioner models and new fault modes.

[0080] Example 2

[0081] This embodiment targets household front-loading washing machines, adapting to mainstream brands of 8kg-12kg capacity front-loading washing machines, covering household models with a service life of 0.5-12 years. It addresses the pain points of low identification rate of hidden motor faults and water circuit coupling faults in washing machine repair by manual diagnosis, and the difficulty for novice repair personnel to learn how to handle them. It fully implements the testing and diagnostic method and system described in claims 1-8 of this invention. The specific implementation process is as follows:

[0082] I. Historical Data Construction and Preprocessing

[0083] In this embodiment, the historical data acquisition submodule of the data acquisition module connects to the after-sales databases of four leading washing machine manufacturers and the database of a nationwide home appliance repair service platform to collect full historical data of multiple brands and models of drum washing machines, constructing an original fault dataset. Specific data collection includes:

[0084] Historical operating data: electrical operating parameters of the washing machine (motor input voltage / current / speed, water inlet valve / drain pump / door lock on / off parameters, heating element power, main control board power supply parameters), sensor timing data (data from tub water temperature sensor, water level sensor, motor vibration sensor, and door lock status sensor), main control board operation log, fault code data, and data on water inlet / washing / rinsing / spin-drying operating modes. The sampling frequency is uniformly 2Hz, and the duration of a single timing data is 30s.

[0085] Fault labeling data: Labels the fault type (22 common washing machine faults such as motor carbon brush wear, drain pump blockage, inlet valve failure, door lock failure, heating element open circuit, main control board communication failure, etc.), fault location (inverter motor, drain pump, inlet solenoid valve, door lock, heating element, main control board MCU unit, etc.), fault cause, and fault severity (minor warning, moderate fault, severe inability to operate);

[0086] Repair case data: standardized repair procedures for corresponding faults, replacement part models, and no-load / load test acceptance standards after repair.

[0087] The original fault dataset, after data collection, contains 96,000 samples, including 24,000 samples from normal operating conditions and 72,000 samples from various fault types. The original dataset undergoes multi-level preprocessing via a data preprocessing module, as follows:

[0088] Data cleaning: Linear interpolation was used to fill missing values ​​in time series data, and samples with a missing value rate of more than 25% were directly removed; outliers were removed by combining the 3σ criterion and the isolated forest algorithm. First, numerical outliers were removed by the 3σ criterion, and then invalid samples such as false alarms and abnormal operating conditions were identified by the isolated forest algorithm. In the end, more than 2,800 invalid samples were removed.

[0089] Data alignment: Resample multi-source time series data with a uniform sampling frequency of 2Hz to align the time dimension and ensure that the time series length of all data in a single sample is consistent at 60 time steps;

[0090] Data encoding: One-hot encoding is performed on classified data such as operating mode, fault code, and door lock status. Sliding window encoding is used for 30-second time series data. The sliding window size is 4 and the step size is 1 to generate a time series feature sequence.

[0091] Data normalization: Min-Max normalization is used to map all numerical data to the [0,1] interval to eliminate dimensional differences;

[0092] Data augmentation: The sample size was expanded by sliding window sampling, Gaussian noise addition, and working condition translation. The sliding window sampling step size was 2, the signal-to-noise ratio of Gaussian noise addition was set to 25dB, and the working condition translation was performed on water temperature and load weight working conditions. Finally, the sample size was expanded to 168,000.

[0093] Dataset partitioning: The samples are divided into a training set (126,000 records), a validation set (25,200 records), and a test set (16,800 records) according to a preset ratio of 7.5:1.5:1, resulting in a standardized training dataset.

[0094] II. Multi-level Feature Extraction

[0095] A three-level cascaded vertical feature extraction network is constructed using a three-level vertical feature extraction module to perform progressive deep feature extraction on the training dataset. The specific steps are as follows:

[0096] First-level signal feature matrix extraction: The first-level signal-level primary feature extraction layer adopts a structure of 3-layer one-dimensional convolutional neural network (1D-CNN) cascaded with 2 fully connected layers. The convolution kernel sizes are 7, 5, and 3 respectively, with a stride of 2 for each layer. The activation function is ReLU. Through convolution operations, the time-domain, frequency-domain, and statistical basic signal features of the washing machine operation data are extracted, including the time-domain fluctuation features of motor current, the frequency-domain vibration features under the spin-drying condition, and the statistical features of voltage / power. The final output is a first-level signal feature matrix with a dimension of 192×1.

[0097] Second-level signal feature matrix extraction: The second-level component-level correlation feature extraction layer sets up six parallel feature extraction units for the six core functional components of the drum washing machine (motor drive unit, water inlet control unit, drainage control unit, door safety unit, heating and drying unit, and main control unit). Each unit adopts a structure of a bidirectional long short-term memory network Bi-LSTM cascaded attention mechanism, with a Bi-LSTM hidden layer dimension of 96 and an attention mechanism weight dimension of 48. Each unit only inputs the correlation signal features of the corresponding functional component in the first-level signal feature matrix. For example, the motor drive unit only inputs correlation features such as motor voltage, current, speed, and vibration. The Bi-LSTM extracts the temporal dependency features inside the component, capturing the temporal degradation law of early latent faults such as motor carbon brush wear. The attention mechanism weights and highlights fault-sensitive features and suppresses irrelevant features such as vibration interference under the dehydration condition. The component feature vectors output by the six units are concatenated to obtain the second-level component feature matrix with a dimension of 576×1.

[0098] The third-level signal feature matrix extraction: The third-level system-level fault depth feature extraction layer adopts a structure of cascaded graph attention network (GAT) with a cross-attention mechanism. Using the feature vectors of the six core functional components in the second-level component feature matrix as graph nodes, and the electrical connections and program control logic associations of the washing machine system as edges, a system topology graph is constructed. The GAT sets two attention heads, each with six attention heads, and uses LeakyReLU as the activation function. The GAT extracts cross-component coupling and fault propagation path features, such as the main control unit program interruption caused by door lock failure and the water level sensor abnormality cascading fault features caused by inlet valve failure. Then, the cross-attention mechanism achieves accurate mapping between depth features and fault modes, ultimately outputting a third-level signal feature matrix with a dimension of 384×1.

[0099] III. Model Training and Optimization

[0100] Through the model training module, a multi-task learning fault diagnosis model is constructed using the third-level signal feature matrix as input and the corresponding fault labeling data, fault location information, and maintenance plan data as labels. The specific architecture and training settings are as follows:

[0101] The multi-task learning-based fault diagnosis model includes three parallel branches:

[0102] Fault classification branch: The input is the third-level signal feature matrix. It adopts 2 fully connected layers + Softmax activation function and outputs 23 classification results (22 types of faults + 1 type of normal operating condition), which are used to identify fault types and normal operating conditions. The focus is on optimizing the classification and identification ability of early hidden faults of motors.

[0103] Fault location regression branch: It adopts a 3-layer fully connected layer to output the accurate location information of faulty parts and components, and outputs the fault severity value in the range of 0-1, realizing three-level positioning from whole machine → parts → components.

[0104] Repair Solution Matching Branch: A pre-stored feature library of standardized repair solutions for drum washing machines is used. The cosine similarity matching algorithm is employed to match the fault features output by the model with the feature library of solutions, and output the standardized repair guidance solution with the highest similarity, which includes safe operating procedures, repair steps, parts models, and test and acceptance standards.

[0105] The core parameters for model training are set as follows: The loss function is a weighted sum of cross-entropy loss (fault classification branch) and mean squared error loss (fault location regression branch), with weights of 0.55 and 0.45, respectively; the optimizer is AdamW, with an initial learning rate of 8e-5, a cosine annealing learning rate decay strategy, T_max of 15, and an early stopping mechanism. If the validation set loss does not decrease for 8 consecutive epochs, training is terminated early to prevent overfitting.

[0106] The model was trained using the training set, and hyperparameters were optimized using the validation set. Finally, the model was tested on the test set. The model achieved a fault classification accuracy of 99.15%, with an early latent fault identification rate of 95.28% for motors and a fault location accuracy (down to the component level) of 97.62%. The target fault diagnosis model was obtained and deployed to the cloud backend and portable test acquisition terminal, respectively.

[0107] IV. On-site Testing and Diagnostic Implementation

[0108] Through the diagnostic execution module and the human-machine interaction module, in conjunction with a portable test and acquisition terminal, on-site testing and diagnosis are completed. The portable test and acquisition terminal used in this embodiment strictly follows the architecture of claim 6, is adapted to washing machine repair scenarios, and integrates an electrical parameter acquisition unit, a sensor acquisition unit, an appliance communication unit, a data processing unit, and a touch display unit, wherein:

[0109] Electrical parameter acquisition unit: Adapted to the parameter acquisition of low-voltage DC and 220V AC circuits of washing machines, it can accurately acquire the voltage, current, resistance, and on / off status of motor, water inlet valve, drain pump, and heating element. Measurement accuracy: resistance ±0.2%FS, voltage ±0.5%FS.

[0110] Sensor acquisition unit: integrates a triaxial vibration sensor, a noise sensor, and a contact temperature sensor to collect vibration data, operating noise, and water temperature data during the washing machine's spin-drying process;

[0111] Home appliance communication unit: Supports infrared, Bluetooth, WiFi and RS485 bus communication, can communicate with the main control board of mainstream brand washing machines, and read operation logs, fault codes and program running status;

[0112] Data processing unit: It adopts an ARM Cortex-M7 core processor, with built-in preprocessing algorithms and a lightweight target fault diagnosis model, supports local offline inference, and can complete on-site diagnosis without network connection;

[0113] Touch display unit: It adopts a 5-inch touch screen, which is suitable for handheld operation by maintenance personnel and supports the graphic and textual visualization of maintenance steps.

[0114] Specific procedures for on-site diagnosis:

[0115] After the novice repairman arrives at the site, he connects the portable test acquisition terminal to the communication interface and test terminals of the main control board of the washing machine to be diagnosed. After turning on the machine, he selects the brand, model, years of use and fault symptoms of the washing machine on the human-machine interaction touch screen.

[0116] The terminal controls the washing machine to sequentially perform empty water intake, washing, and spin-drying operations, and simultaneously collects 30 seconds of real-time operation test data, including electrical parameters, sensor timing data, main control board operation log, and fault codes. At the same time, the static resistance and continuity tests of key components are completed through the electrical parameter acquisition unit.

[0117] The terminal's data processing unit performs a preprocessing procedure on the real-time collected data that is consistent with the aforementioned steps, generating standardized real-time data;

[0118] Standardized real-time data is input into the lightweight target fault diagnosis model. After the model completes local inference, it outputs fault diagnosis results, accurate fault location information, and corresponding standardized maintenance guidance.

[0119] The touchscreen displays the diagnostic results visually. Example output: Fault type is excessive wear of motor carbon brushes (early latent fault), fault location is carbon brush assembly of roller inverter motor, and fault severity is moderate. Simultaneously, a standardized repair plan is output: first, perform a power-off safety operation, then disassemble the motor end cover step by step, replace the carbon brushes with the same model, reinstall the motor, perform no-load test, and perform load test. The acceptance criteria are clearly stated as the dehydration speed reaching the rated value, the operating vibration amplitude ≤0.5mm, and no abnormal noise. The model of the compatible carbon brush accessories is also marked.

[0120] V. Model Iteration and Update

[0121] The actual repair results from on-site diagnosis, user feedback, and data on newly added fault cases are uploaded to the cloud backend via portable terminals to supplement the original fault dataset. Small-batch incremental training is performed every two weeks, and full-scale iterative optimization is performed every quarter. The entire process of data preprocessing, feature extraction, and model training is repeated to continuously improve the model's adaptability and recognition accuracy for new washing machine models, niche brand models, and newly added fault modes.

[0122] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A testing and diagnostic method for repairing household appliances, characterized in that, Specifically, the following steps are included: S1. Historical Data Construction and Preprocessing: Collect historical operating data, fault labeling data, and repair case data of multiple categories and brands of household appliances to construct the original fault dataset; perform multi-level preprocessing on the original fault dataset to obtain a standardized training dataset; S2. Multi-level Feature Extraction: Construct a multi-level feature extraction network to perform progressive deep feature extraction on the training dataset, where: First-level signal feature matrix: Using preprocessed standardized data as input, extract time-domain, frequency-domain, and statistical basic signal features of household appliance operation data; Second-level signal feature matrix: The input is the first-level signal feature matrix. The features are grouped according to the core functional components of the household appliance to be diagnosed, and the time-series dependence, operating condition matching, and parameter correlation features within each component are extracted. The third-level signal feature matrix: The input is the second-level signal feature matrix, which is used to extract cross-component coupling correlation features, fault propagation path features, and fault mode mapping features; S3. Model Training and Optimization: Using the third-level signal feature matrix as input and the corresponding fault labeling data, fault location information, and maintenance plan data as labels, a multi-task learning fault diagnosis model is constructed. The model training and hyperparameter optimization are completed through the training set and validation set to obtain the target fault diagnosis model. S4. On-site testing and diagnosis: Real-time operating test data of the household appliance to be diagnosed is collected through a portable test acquisition terminal. After preprocessing, the data is input into the target fault diagnosis model. The model outputs fault diagnosis results, accurate fault location information, and corresponding standardized repair guidance schemes. S5. Model Iteration Update: Supplement the original fault dataset with the actual maintenance results of on-site diagnosis and newly added fault case data, and repeat steps S1-S3 to incrementally update and iteratively optimize the fault diagnosis model.

2. The testing and diagnostic method for repairing household appliances according to claim 1, characterized in that: The historical operating data includes electrical operating parameters of household appliances, sensor timing data, main control board operating logs, fault code data, and operating mode data; The fault labeling data includes fault type, fault location, fault cause, and fault severity; the repair case data includes repair steps, replacement part models, and acceptance criteria; the multi-level preprocessing includes data cleaning, data alignment, data encoding, data normalization, data augmentation, and dataset partitioning.

3. The testing and diagnostic method for repairing household appliances according to claim 2, characterized in that: The data cleaning process uses linear interpolation to fill missing values ​​and a combination of the 3σ criterion and the isolated forest algorithm to remove outliers. The data alignment process resamples multi-source time-series data, unifying the sampling frequency and time dimension. The data encoding process performs one-hot encoding on categorical data and sliding window encoding on time-series data. The data normalization process uses Min-Max normalization to map numerical data to the [0,1] interval. The data augmentation process expands the sample size using sliding window sampling, Gaussian noise addition, and working condition translation enhancement. The dataset partitioning process divides the samples into training, validation, and test sets according to a preset ratio.

4. The testing and diagnostic method for repairing household appliances according to claim 1, characterized in that: Step S2 specifically includes: S21. First-level signal feature matrix extraction: The first-level signal feature extraction layer adopts a structure of a 3-layer one-dimensional convolutional neural network (1D-CNN) cascaded with 2 fully connected layers. The convolutional kernel sizes are 7, 5, and 3 respectively, and the stride is 2. Basic signal features are extracted through convolution operations, and the first-level signal feature matrix is ​​output. S22. Second-level signal feature matrix extraction: The second-level component-level correlation feature extraction layer sets up parallel feature extraction units that match the number of core functional components of the household appliance. Each unit adopts a structure of a bidirectional long short-term memory network Bi-LSTM cascaded attention mechanism. Each unit only inputs the correlation signal features of the corresponding functional component in the first-level signal feature matrix. The temporal dependency features inside the component are extracted by Bi-LSTM, and the fault-sensitive features are highlighted by weighting through the attention mechanism. The component feature vectors output by each unit are concatenated to obtain the second-level component feature matrix. S23. Third-level signal feature matrix extraction: The third-level system-level fault deep feature extraction layer adopts the structure of a graph attention network (GAT) cascaded cross-attention mechanism. The feature vectors of each functional component in the second-level component feature matrix are used as graph nodes, and the electrical connections and control logic associations of the household appliance system are used as edges to construct a system topology graph. The coupling associations and fault propagation path features across components are extracted through GAT, and the accurate mapping between features and fault modes is achieved through the cross-attention mechanism, outputting the third-level signal feature matrix.

5. The testing and diagnostic method for repairing household appliances according to claim 1, characterized in that: The multi-task learning fault diagnosis model includes a fault classification branch, a fault location regression branch, and a maintenance plan matching branch. The fault classification branch uses the Softmax activation function to identify fault types and normal operating conditions. The fault location regression branch is used to output the precise location information and severity of the faulty component; the maintenance plan matching branch uses a cosine similarity matching algorithm to output a standardized maintenance guidance plan that matches the fault; the model loss function is a weighted sum of cross-entropy loss and mean squared error loss, the optimizer is AdamW, and a cosine annealing learning rate decay strategy and an early stopping mechanism are set.

6. The testing and diagnostic method for repairing household appliances according to claim 1, characterized in that: The portable test and acquisition terminal includes an electrical parameter acquisition unit, a sensor acquisition unit, an appliance communication unit, a data processing unit, and a display unit. The electrical parameter acquisition unit is used to acquire the voltage, current, resistance, and frequency electrical parameters of the appliance to be diagnosed. The sensor acquisition unit is used to acquire temperature, vibration, and noise operating condition data. The appliance communication unit communicates with the main control board of the appliance to be diagnosed via infrared, Bluetooth, WiFi, or CAN bus to read the operation log and fault codes. The data processing unit performs real-time data preprocessing and model inference.

7. A testing and diagnostic system for repairing household appliances, characterized in that: This system is used to implement a testing and diagnostic method for home appliance repair as described in any one of claims 1-6, comprising: a data acquisition module, a data preprocessing module, a three-level vertical feature extraction module, a model training module, a diagnostic execution module, and a human-computer interaction module; Data acquisition module: used to collect historical fault data of household appliances and real-time test data of the appliance to be diagnosed; Data preprocessing module: used to perform multi-level preprocessing on the collected raw data and output standardized data; The three-level vertical feature extraction module includes a first-level signal-level feature extraction unit, a second-level signal-level feature extraction unit, and a third-level signal-level feature extraction unit. The only input of the second-level signal-level feature extraction unit is the output of the first-level signal-level feature extraction unit, and the only input of the third-level signal-level feature extraction unit is the output of the second-level signal-level feature extraction unit. Model training module: Used to train, optimize and incrementally update the fault diagnosis model based on the extracted deep fault feature set, and output the target fault diagnosis model; Diagnostic execution module: Used to input standardized real-time data of the home appliance to be diagnosed into the target fault diagnosis model, perform inference, and output diagnostic results and repair guidance plan; Human-computer interaction module: used to display diagnostic results, receive user operation commands, and output visual maintenance guidance.

8. The testing and diagnostic system for home appliance repair according to claim 7, characterized in that: The data acquisition module includes a historical data acquisition submodule and an on-site test acquisition submodule; the historical data acquisition submodule connects to the after-sales database of home appliance manufacturers and the database of third-party repair platforms to acquire historical fault and repair data in batches. The on-site test acquisition submodule is a portable test acquisition terminal used to collect real-time operating data and electrical test data of the home appliance to be diagnosed on-site.

Images