Fault diagnosis method and device for air conditioner outdoor unit module and air conditioner

Abstract

The application relates to the technical field of fault diagnosis, and discloses a fault diagnosis method and device for an outdoor unit module of an air conditioner and the air conditioner. The method comprises the following steps: collecting air conditioner operation data and after-sales work order data, and constructing a multi-modal training set, wherein the after-sales work order data comprises a maintenance time stamp, a fault phenomenon description and actual replacement component information; inputting the multi-modal training set into a lightweight Transformer model for model training to obtain a fault diagnosis model; in the case that an outdoor unit module fault code of the air conditioner is triggered, inputting operation data before the fault into the fault diagnosis model to output a component fault probability; and recommending a maintenance personnel to carry corresponding spare parts for on-site maintenance according to the component fault probability. According to the scheme, the maintenance personnel can more accurately predict the fault component, reduce repeated maintenance caused by the absence of spare parts, significantly improve the one-time maintenance success rate, reduce the after-sales cost, and improve the user experience.

Description

Technical Field

[0001] This application relates to the field of air conditioning fault diagnosis technology, for example to a fault diagnosis method, device and air conditioner for an outdoor unit module of an air conditioner. Background Technology

[0002] With the development of air conditioning technology, the diagnostic methods for outdoor unit modules have evolved from simple to complex. Early methods relied primarily on manual repair, with technicians using experience and simple testing tools to systematically check each module and pinpoint the fault. This approach was adequate in the early days when air conditioning technology was relatively simple and fault types were limited. However, with continuous advancements in air conditioning technology, the structure and operating principles of outdoor unit modules have become increasingly complex, leading to a greater diversity of fault types. Traditional manual repair methods have gradually revealed numerous problems and are no longer sufficient to meet the demands of modern air conditioning fault diagnosis.

[0003] To address the challenges in the development of air conditioner outdoor unit module fault diagnosis, this technology offers a fault monitoring method based on real-time monitoring data. This method involves installing multiple monitoring modules within the air conditioner to collect various monitoring data during operation, such as key parameters like current, voltage, and temperature. When anomalies are detected, the system can promptly determine whether these anomalies will cause an air conditioner malfunction and, based on the severity of the fault, control whether to shut down the air conditioner. This approach advances fault monitoring to before a fault occurs, improving the timeliness and accuracy of fault monitoring compared to traditional manual repair methods, and reducing the risk of equipment damage due to undetected faults.

[0004] In the process of implementing the embodiments of this disclosure, at least the following problems were found in the related art:

[0005] While related technologies have addressed some issues related to early fault detection, significant shortcomings remain. Their data dimensions are relatively limited, relying primarily on real-time monitoring parameters during air conditioner operation, and the diagnostic logic remains relatively simple, resulting in insufficient generalization ability of the algorithms. When faced with faults in outdoor unit modules of air conditioners from different brands, models, and operating environments, they struggle to accurately adapt and diagnose, leading to a still low rate of accurate spare parts carrying by maintenance personnel.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0008] This disclosure provides a fault diagnosis method, device, and air conditioner for an outdoor unit module of an air conditioner, which can effectively improve the accuracy of spare parts carried by maintenance personnel.

[0009] In some embodiments, the fault diagnosis method for an air conditioner outdoor unit module includes: collecting air conditioner operation data and after-sales work order data to construct a multimodal training set, wherein the after-sales work order data includes a repair timestamp, a description of the fault phenomenon, and information on the actual replaced parts; inputting the multimodal training set into a lightweight Transformer model for model training to obtain a fault diagnosis model; when the air conditioner triggers a fault code in the outdoor unit module, inputting the operation data before the fault into the fault diagnosis model to output the component fault probability; and recommending repair personnel to bring the corresponding spare parts for on-site repair based on the component fault probability.

[0010] In some embodiments, the fault diagnosis device for an air conditioner outdoor unit module includes: a construction module configured to collect air conditioner operation data and after-sales work order data to construct a multimodal training set, wherein the after-sales work order data includes a repair timestamp, a description of the fault phenomenon, and information on the actual replaced parts; a training module configured to input the multimodal training set into a lightweight Transformer model for model training to obtain a fault diagnosis model; a diagnosis module configured to input the pre-fault operation data into the fault diagnosis model when the air conditioner triggers a fault code in the outdoor unit module to output the component fault probability; and a recommendation module configured to recommend repair personnel to bring corresponding spare parts for on-site repair based on the component fault probability.

[0011] In some embodiments, the fault diagnosis device for an air conditioner outdoor unit module includes a processor and a memory storing program instructions, wherein the processor is configured to execute the aforementioned fault diagnosis method for an air conditioner outdoor unit module when running the program instructions.

[0012] In some embodiments, the air conditioner includes: an outdoor unit; and the aforementioned fault diagnosis device for the outdoor unit module, which is installed on the outdoor unit.

[0013] The fault diagnosis method, apparatus, and air conditioner for air conditioner outdoor unit modules provided in this disclosure can achieve the following technical effects:

[0014] This solution constructs a multimodal training set by integrating air conditioner operation data and after-sales work order data. A lightweight Transformer model is employed for fault diagnosis, effectively capturing deep-seated features and complex dependencies within the data. Compared to traditional manual repair methods, this approach not only improves the efficiency and accuracy of fault diagnosis but also significantly enhances the model's generalization ability through multi-source data fusion and deep learning techniques. This allows repair personnel to more accurately predict faulty components, reducing rework due to missing spare parts, significantly increasing the first-time repair success rate, lowering after-sales costs, and improving user experience.

[0015] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0016] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein:

[0017] Figure 1 This is a schematic diagram of a fault diagnosis method for an air conditioner outdoor unit module provided in an embodiment of this disclosure;

[0018] Figure 2 This is a schematic diagram of a method for obtaining a fault diagnosis model provided in an embodiment of this disclosure;

[0019] Figure 3 This is a schematic diagram of a method for obtaining the failure probability of each component provided in an embodiment of this disclosure;

[0020] Figure 4 This is a schematic diagram of another fault diagnosis method for an air conditioner outdoor unit module provided in this embodiment of the disclosure;

[0021] Figure 5 This is a schematic diagram of a fault diagnosis device for an air conditioner outdoor unit module provided in an embodiment of this disclosure;

[0022] Figure 6 This is a schematic diagram of another fault diagnosis device for an air conditioner outdoor unit module provided in an embodiment of this disclosure.

[0023] Figure label:

[0024] 30: Processor; 31: Memory; 32: Communication interface; 33: Bus; 200: Fault diagnosis device for air conditioner outdoor unit module; 300: Fault diagnosis device for air conditioner outdoor unit module; 51: Construction module; 52: Training module; 53: Diagnostic module; 54: Recommendation module. Detailed Implementation

[0025] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0026] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0027] Unless otherwise stated, the term "multiple" means two or more.

[0028] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0029] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0030] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0031] Combination Figure 1 As shown, optionally, this disclosure provides a fault diagnosis method for an air conditioner outdoor unit module, including:

[0032] S11, the air conditioner collects air conditioner operation data and after-sales work order data to build a multimodal training set. The after-sales work order data includes repair timestamps, fault descriptions, and information on actual replaced parts.

[0033] S12, the air conditioner inputs the multimodal training set into the lightweight Transformer model for model training to obtain the fault diagnosis model.

[0034] S13, when the air conditioner triggers the outdoor unit module fault code, the air conditioner inputs the air conditioner operation data before the fault into the fault diagnosis model to output the component fault probability.

[0035] S14. Based on the probability of component failure, it is recommended that repair personnel bring the corresponding spare parts for on-site repair of the air conditioner.

[0036] In this solution, the air conditioner can collect operating data and after-sales work order data. The operating data includes key parameters such as indoor and outdoor ambient temperature, compressor frequency, coil temperature, electronic expansion valve opening, fan speed, and current and voltage. In essence, the operating data comprehensively reflects the status of the outdoor unit module during operation. Specifically, the operating data can be collected at a certain sampling frequency f to capture real-time changes in the air conditioner's operating conditions. In one example, the sampling frequency f can be determined based on the rate of change of the operating data and the real-time requirements of fault diagnosis, ensuring that key operating status changes are captured while avoiding data redundancy and wasted computing resources.

[0037] In this solution, after-sales work order data can be obtained from the after-sales service module of the ERP (Enterprise Resource Planning) system. This data includes the work order number, repair timestamp, a description of the fault symptoms filled out by the repair engineer on-site (e.g., "compressor not starting," "abnormal noise from the outdoor fan," etc.), and information on the actual replaced parts. This information may include the part code, model, and specifications. The work order data may also include the equipment status verification results after repair. Thus, after collecting air conditioning operation data and after-sales work order data, a multimodal training set can be constructed based on these data. This solution, by constructing a multimodal training set, allows the model to simultaneously learn from the air conditioning operation status and historical repair experience, thereby gaining a more comprehensive understanding of fault modes and improving the accuracy and generalization ability of fault diagnosis.

[0038] Furthermore, the air conditioner first inputs the multimodal training set into a lightweight Transformer model. This model, with its unique architecture, can handle complex time-series data and output the probability of faults in each component. These fault probabilities are the model's preliminary diagnostic results on the input data, providing a foundation for subsequent model optimization. Further, the air conditioner can calculate the model's loss value based on the difference between the fault probabilities output by the lightweight Transformer model and the actual fault labels. Here, the loss value is an important indicator of the model's prediction accuracy and is calculated using a loss function. As an example, the loss function can be the cross-entropy loss function. Thus, the air conditioner can update the model's parameters using the backpropagation algorithm based on the calculated loss value. This process allows the model to gradually improve its prediction accuracy by continuously adjusting the parameters. The iterative training process continues until preset conditions are met, such as the loss value reaching a preset convergence threshold or the number of iterations reaching a set number. When these conditions are met, the training process terminates, and the resulting model is the trained fault diagnosis model, which can be used for actual fault diagnosis tasks.

[0039] In another approach, air conditioners can also use different methods to input the multimodal training set into a lightweight Transformer model for training. For example, a learning rate decay mechanism can be introduced during training, gradually reducing the learning rate as training progresses, allowing the model to finely adjust parameters in the later stages of training, thus improving the model's stability and accuracy. Furthermore, an early stopping mechanism can be employed, monitoring the model's performance on the validation set. If the model's performance no longer improves after a certain number of iterations, training is terminated early to prevent overfitting and ensure the model has good generalization ability. Using this approach, by inputting the multimodal training set into the lightweight Transformer model for training, the model achieves a significant improvement in fault diagnosis accuracy.

[0040] Furthermore, when the air conditioning unit triggers the outdoor unit module fault code (F1), the air conditioner automatically initiates a fault diagnosis process. Specifically, the air conditioner automatically extracts operating data from the t minutes prior to the fault occurrence. This data includes key parameters such as indoor and outdoor ambient temperature, compressor frequency, coil temperature, electronic expansion valve opening, fan speed, and current and voltage. This operating data comprehensively reflects the operating status of the outdoor unit module before the fault occurred. The extracted data is then input into a pre-trained fault diagnosis model. This model, based on a lightweight Transformer architecture, can efficiently process time-series data and output the probability of faults for each component. The model's output is a list containing the fault probabilities of each component, which helps maintenance personnel quickly locate potentially faulty components.

[0041] In one optimized solution, to save internal storage space in the air conditioner and improve diagnostic efficiency, the trained fault diagnosis model can be deployed to a cloud server. In this scenario, when the device triggers a fault code in the outdoor unit module, the air conditioner automatically extracts operating data from the minutes preceding the fault and sends it to the cloud server. The inference engine on the cloud server receives this data and performs real-time diagnosis using the cloud-deployed fault diagnosis model. The diagnostic results not only include the three components with the highest probability of failure but also the probability confidence level of each component's failure. After manual review, this information is pushed to the after-sales service system, allowing repair personnel to bring the appropriate spare parts for on-site repair based on this accurate diagnostic information. This solution, by deploying the fault diagnosis model to a cloud server, not only saves internal storage space in the air conditioner but also improves diagnostic accuracy and efficiency. Compared to existing technologies, this cloud deployment method can process large amounts of data more quickly, providing more accurate fault diagnosis results, thereby significantly improving repair efficiency, reducing repair time and costs, and enhancing the user experience.

[0042] Furthermore, after the air conditioner obtains the component failure probabilities, it can recommend that repair personnel bring the corresponding spare parts for on-site repair based on these probabilities. Specifically, after the fault diagnosis model outputs the failure probabilities of each component, the air conditioner will sort the faulty components according to these probabilities. In one example, the fault diagnosis model will output a list containing the failure probabilities of each component; for example, the probability of compressor failure is 70%, the probability of outdoor fan failure is 20%, and the probability of circuit board failure is 10%. Based on these probabilities, the air conditioner will generate a detailed repair suggestion list, recommending that repair personnel prioritize bringing the spare parts of the components with the highest failure probabilities for on-site repair. After receiving a repair task, the repair personnel will prepare the corresponding spare parts according to the repair suggestion list provided by the air conditioner. For example, if the compressor failure probability is the highest, the repair personnel will prioritize bringing compressor spare parts; if the outdoor fan failure probability is the second highest, the repair personnel will also bring outdoor fan spare parts, and so on. This spare parts recommendation mechanism based on failure probabilities ensures that repair personnel have brought the most likely replacement parts when they arrive at the repair site, thereby reducing duplicate repairs due to insufficient spare parts and significantly reducing after-sales costs. Meanwhile, precise spare parts recommendations enable repair personnel to complete repairs faster, increasing the first-time repair success rate and improving user experience. Furthermore, because they can quickly prepare the necessary spare parts based on a clear repair suggestion list, preparation time for repair personnel is reduced, further improving the overall efficiency of repair work.

[0043] In an optimized approach, the fault diagnosis model outputs both the probability and confidence score of each component's failure. This allows the air conditioner to combine the failure probability and confidence score to more accurately recommend that repair personnel bring the appropriate spare parts for on-site repair. In one example, the air conditioner sorts the faulty components based on the model's output failure probabilities and evaluates the failure prediction for each component based on the confidence score. For instance, the fault diagnosis model might output the following results: a 70% probability of compressor failure with a confidence score of 0.9; a 20% probability of outdoor fan failure with a confidence score of 0.8; and a 10% probability of circuit board failure with a confidence score of 0.7. In this case, the air conditioner will comprehensively consider both the failure probability and confidence score to generate a repair recommendation list. If a component has a high failure probability and a high confidence score, the air conditioner will prioritize recommending that repair personnel bring spare parts for that component. For example, if the compressor failure probability is 70% with a confidence score of 0.9, this indicates that the fault diagnosis model's prediction of compressor failure is highly reliable, therefore the air conditioner will recommend that repair personnel bring compressor spare parts.

[0044] In an optimized approach, the air conditioner can also set a confidence threshold, such as 0.8. Only when the confidence level of the fault diagnosis model's prediction of a component's failure exceeds this threshold will the air conditioner recommend that maintenance personnel carry a spare part for that component. This further improves the reliability of maintenance recommendations and avoids incorrect spare part carrying due to inaccurate model predictions. For example, if the probability of an outdoor fan failure is 20%, but the confidence level is 0.8, exceeding the set threshold, the air conditioner will also recommend that maintenance personnel carry a spare outdoor fan. Conversely, if the probability of a circuit board failure is 10%, but the confidence level is only 0.7, below the set threshold, the air conditioner will advise maintenance personnel not to carry a spare circuit board for the time being, but to make a decision based on the on-site situation. This approach, by combining the component failure probability and confidence level output by the fault diagnosis model for maintenance recommendations, not only improves the accuracy and reliability of maintenance recommendations but also further optimizes the allocation of maintenance resources. Compared to solutions that rely solely on failure probabilities, this method combining confidence levels can more accurately predict faulty components and reduce incorrect spare part carrying caused by the uncertainty of the fault diagnosis model's predictions. This not only improves repair efficiency and reduces after-sales costs, but also enhances the work efficiency of repair personnel and the user experience. At the same time, by setting confidence thresholds, the air conditioner can better balance prediction accuracy and the use of repair resources, ensuring that repair personnel can carry the most likely replacement parts when resources are limited, further improving the first-time repair success rate.

[0045] The fault diagnosis method for air conditioner outdoor unit modules provided in this disclosure constructs a multimodal training set by fusing air conditioner operation data and after-sales work order data. Employing a lightweight Transformer model for fault diagnosis effectively captures deep-seated features and complex dependencies in the data. Compared to traditional manual repair methods, this approach not only improves the efficiency and accuracy of fault diagnosis but also significantly enhances the model's generalization ability through multi-source data fusion and deep learning techniques. This allows repair personnel to more accurately predict faulty components, reduces rework due to missing spare parts, significantly improves the first-time repair success rate, reduces after-sales costs, and enhances user experience.

[0046] Combination Figure 2 As shown, optionally, in S12, the air conditioner inputs the multimodal training set into the lightweight Transformer model for model training to obtain a fault diagnosis model, including:

[0047] S21, the air conditioner inputs the multimodal training set into the lightweight Transformer model to obtain the failure probability of each component output by the lightweight Transformer model.

[0048] S22, the air conditioner calculates the model loss value based on the failure probability of each component output.

[0049] S23, the air conditioner updates the model parameters based on the model loss value until the preset conditions are met and the iteration terminates.

[0050] S24, when the iteration terminates, the air conditioner obtains the trained fault diagnosis model.

[0051] The preset conditions include the model loss value meeting the preset convergence condition or the current iteration number reaching the set iteration number.

[0052] In this approach, the air conditioner can input the collected multimodal training set into a lightweight Transformer model. Through its unique architecture, it can process complex time-series data and output the probability of failure for each component. These failure probabilities represent the model's preliminary diagnostic results from the input data, providing a foundation for subsequent model optimization.

[0053] Furthermore, the air conditioner can calculate the model's loss value based on the difference between the failure probabilities of each component output by the model and the actual failure labels. Here, the failure labels refer to the manually labeled real failure states used for supervised learning, serving as the benchmark for calculating the loss value during model training. The loss value can be calculated using a loss function. Based on the calculated model loss value, the air conditioner can update the model's parameters using the backpropagation algorithm. By continuously adjusting the parameters, the fault diagnosis model can gradually improve its predictive accuracy. The iterative training process continues until preset conditions are met, such as the loss value reaching a preset convergence threshold or the number of iterations reaching a set number. When these conditions are met, the training process terminates, and the resulting model is the trained fault diagnosis model, ready for use in practical fault diagnosis tasks.

[0054] This approach, by inputting a multimodal training set into a lightweight Transformer model for training, fully leverages the rich information in air conditioning operation data and after-sales work order data to improve the model's understanding and identification of fault modes. Compared to traditional methods, it can not only handle complex time-series data but also integrate multi-source data, providing more comprehensive fault diagnosis results. By accurately calculating the loss value and updating the model parameters accordingly, the model's accuracy and generalization ability are significantly improved, enabling more precise prediction of faulty components and reducing false positives and false negatives. This not only improves the efficiency and accuracy of fault diagnosis but also reduces repair time and costs, enhancing the user experience.

[0055] Understandably, for rare fault samples occurring less than 5% of the time, air conditioners can employ a meta-learning optimization strategy for special handling: First, a small sample task set containing K historical cases of the fault type is constructed. Then, by calculating the meta-gradient of this task set and the base model on the current batch of data, the model parameters are adjusted specifically, with the adjustment magnitude controlled by the meta-learning rate. This process significantly improves the ability to capture sparse fault patterns while maintaining the main structure of the base model. In this way, through the meta-learning mechanism, a high recognition accuracy can still be maintained for rare fault types with insufficient data.

[0056] Combination Figure 3 As shown, optionally, in S21, the air conditioner inputs the multimodal training set into the lightweight Transformer model to obtain the failure probabilities of each component output by the lightweight Transformer model, including:

[0057] S31, the air conditioner inputs the multimodal training set into the embedding layer of the lightweight Transformer model for feature mapping and position encoding to obtain a multidimensional temporal feature vector.

[0058] S32, the air conditioner inputs the multi-dimensional temporal feature vector into the encoder of the lightweight Transformer model for feature extraction, so as to obtain the feature representation output by the encoder of the lightweight Transformer model.

[0059] S33, the air conditioner performs global averaging compression on the feature representation output of the encoder of the lightweight Transformer model to obtain a fixed-length feature vector.

[0060] S34, the air conditioner inputs a fixed-length feature vector into the three-dimensional classification head for classification, so as to obtain the failure probability of each component output by the three-dimensional classification head.

[0061] In this approach, the air conditioner first preprocesses the collected multimodal training set. Preprocessing steps include, but are not limited to, data cleaning, normalization, and feature extraction. These steps remove noise and outliers from the data, transform data of different dimensions to the same scale, and extract key features to ensure data quality and consistency. After preprocessing, the air conditioner can input the processed data into the embedding layer of a lightweight Transformer model. Understandably, the main task of the embedding layer of the lightweight Transformer model is to map the input data to a high-dimensional vector space through linear transformation, obtaining embedding vectors. This process allows the data to better reveal its inherent characteristics and patterns in this new space, providing richer information for subsequent feature extraction and model training. Furthermore, to preserve the temporal order information of the data, the air conditioner adds positional encoding to the embedding vector for each time step. It should be noted that positional encoding is a special type of vector used to represent the positional information of a time step. Thus, by adding positional encoding to the embedding vector, the model can capture the temporal order information in the time series data. Further, the combination of embedding vectors and positional encoding generates multidimensional temporal feature vectors containing temporal information, which will serve as input for subsequent model processing. This approach effectively transforms multimodal data into multidimensional time-series feature vectors suitable for model processing. This transformation not only preserves the key features of the original data but also enhances the model's ability to understand time-series data, thereby improving the accuracy and efficiency of fault diagnosis.

[0062] Furthermore, the air conditioner inputs the pre-processed and lightweight Transformer model's embedding layer's multi-dimensional temporal feature vector into the encoder of the lightweight Transformer model for feature extraction. Here, the encoder of the lightweight Transformer model has a multi-layer structure. Each encoder layer contains a multi-head attention mechanism and a feedforward network. These two components work together to extract complex features and dependencies in the data. Specifically, the multi-head attention mechanism can simultaneously focus on different parts of the input sequence, projecting the input data into multiple low-dimensional subspaces by calculating linear transformations of the query, key, and value, and independently calculating attention scores in each subspace. These scores are then weighted and summed to obtain the multi-head attention output, thereby capturing long-distance dependencies in the data. This process allows the model to process information in parallel in different subspaces, improving the model's ability to capture complex patterns. Further, the obtained attention output is then passed to the feedforward network. In this embodiment, the feedforward network is a two-layer fully connected neural network containing the ReLU activation function. The feedforward network performs a non-linear transformation on the attention output to further extract features. This process enhances the model's ability to represent input data, enabling it to learn more complex feature representations. To maintain model stability and training efficiency, the output of the feedforward network is residually connected to the input, meaning the output and input of the feedforward network are added together and then normalized. Residual connections help solve the vanishing gradient problem in deep networks, while layer normalization ensures consistency in data distribution across different feature dimensions, thereby improving the model's training stability and convergence speed. After completing the above operations in each encoder layer, the output becomes the input to the next encoder layer. Through this layer-by-layer transmission and processing, the model can progressively extract more complex feature representations. As the number of layers increases, the model's understanding of the data deepens, ultimately yielding the feature representations output by the encoder. These feature representations capture deep-seated features and complex dependencies in the data, providing rich information for subsequent fault diagnosis.

[0063] This approach, by streamlining the feature extraction process of the Transformer model's encoder, effectively handles complex time-series data and progressively extracts deeper feature representations. Compared to traditional feature extraction methods, this Transformer-based architecture not only captures long-range dependencies in the data but also enhances the model's ability to learn complex patterns through a combination of multi-head attention mechanisms and feedforward networks. Furthermore, residual connections and layer normalization further improve the model's training stability and convergence speed, enabling it to learn useful feature representations more efficiently. Ultimately, these feature representations provide a solid foundation for fault diagnosis, significantly improving its accuracy and efficiency.

[0064] In this embodiment, the feature representation processed by the encoder of the lightweight Transformer model is a high-dimensional temporal feature matrix. Each row of this matrix represents a feature vector at a time step, and each column represents a feature dimension. To convert this high-dimensional temporal feature matrix into a fixed-length feature vector, a global average pooling operation is performed on the feature representation output by the encoder. Specifically, the global average pooling operation averages all time steps across each feature dimension to obtain a fixed-length feature vector. This fixed-length feature vector retains the average information of the original temporal data across each feature dimension, effectively capturing the overall features of the data while reducing the feature dimensionality, making subsequent processing more efficient. This approach compresses the high-dimensional temporal feature matrix into a fixed-length feature vector. This process not only preserves the key features of the data but also reduces the feature dimensionality, improving the computational efficiency of the model. Compared to directly using the high-dimensional temporal feature matrix, the global average pooling operation can more effectively capture the overall features of the data while reducing feature redundancy, resulting in better model performance in subsequent classification or other tasks. In addition, global average pooling has a certain degree of noise resistance, which can improve the robustness of the model.

[0065] As alternatives to global average pooling, other methods can be considered to obtain fixed-length feature vectors. For example, max pooling can be used. Specifically, max pooling selects the maximum value in each feature dimension, thus preserving salient features in the data. Another option is adaptive pooling, which adaptively adjusts the pooling window size based on a preset target dimension to obtain a fixed-length feature vector. These alternatives can be selected based on the specific application scenario and data characteristics to further optimize the performance of the fault diagnosis model.

[0066] Furthermore, the air conditioner can input a fixed-length feature vector obtained through global average pooling into a 3D classification head for classification. Here, the 3D classification head is a simple fully connected layer whose main function is to map the fixed-length feature vector to a fault probability space. Specifically, the fully connected layer transforms the input feature vector into a vector with the same number of fault categories through a linear transformation, where each element represents the raw probability of a component failure. These raw probabilities are then processed by the Softmax activation function, which converts each raw probability into a value between 0 and 1, ensuring that the sum of all component failure probabilities is 1. Finally, the model outputs a probability distribution representing the probability of each component failure; for example, the probability of compressor failure is 70%, outdoor fan failure is 20%, and circuit board failure is 10%. This approach effectively transforms the feature vector into a probability distribution of component failures. The use of the Softmax activation function ensures that the output probability distribution has good mathematical properties, making the model output more intuitive and easier to interpret. Compared to directly outputting the raw probabilities, the probability distribution after Softmax processing more accurately reflects the likelihood of each component failure, thereby improving the accuracy and reliability of fault diagnosis. In addition, this classification method has good generalization ability and can adapt to the fault diagnosis needs of air conditioner outdoor unit modules of different brands, models and usage environments.

[0067] In one optimized approach, the air conditioner also includes a causal graph inference module. Specifically, this module can perform deep analysis on the feature representation output by the encoder of the lightweight Transformer model. In one example, a causal directed acyclic graph (DAG) between variables is first learned from historical fault data, for example, establishing a causal chain of "capacitor failure → abnormal current → compressor shutdown". Then, the calculated causal weight matrix (e.g., the causal strength of capacitor failure on abnormal current is 0.92) is weighted and fused with the original attention score of the Transformer. A gating mechanism is used to adjust the ratio between the two; when the causal confidence exceeds a threshold (e.g., >0.8), the corresponding attention score is increased by 30% to 50%, thereby strengthening the weight of the root cause feature. Thus, in the 3D classification head, a fixed-length feature vector corrected for causality can be concatenated with the causal weight vector. An additional causal constraint loss function is added when inputting into the fully connected layer to ensure that the classification result is consistent with the root cause derived from the causal graph. Compared to existing technologies that rely solely on statistical correlation for diagnosis, this solution, through explicit modeling of causal relationships, can effectively distinguish between fundamental faults (such as capacitor damage) and derivative phenomena (such as current fluctuations). It effectively reduces the false positive rate in complex concurrent fault scenarios, while the output diagnostic results are interpretable (such as showing the causal chain of "capacitor damage leading to abnormal current"), enabling maintenance personnel to quickly locate the root cause of the problem rather than the superficial symptoms.

[0068] Optionally, in step S31, the air conditioner inputs the multimodal training set into the embedding layer of the lightweight Transformer model for feature mapping and position encoding to obtain a multidimensional temporal feature vector, including:

[0069] The air conditioning system preprocesses the air conditioning operation data and after-sales work order data from the multimodal training set.

[0070] The air conditioner inputs the preprocessed data into the embedding layer of the lightweight Transformer model, and maps the input data to a high-dimensional vector space through linear transformation to obtain the embedding vector.

[0071] The air conditioner adds positional encoding to the embedding vector at each time step to obtain a multidimensional temporal feature vector containing time information.

[0072] In this solution, the air conditioner preprocesses the air conditioner operation data and after-sales work order data in the multimodal training set, including: data cleaning to remove noise and outliers, ensuring data accuracy and completeness; data normalization to convert data of different dimensions to the same scale for easier model processing; and feature extraction to extract key features from the raw data to enhance the model's understanding of the data. In a specific embodiment, data cleaning includes filling missing values ​​in the air conditioner operation data through interpolation or using default values. For text descriptions in the after-sales work order data, irrelevant characters and formatted text are removed. Data normalization includes normalizing temperature data to the range of 0 to 1 and current and voltage data to the same dimension. Feature extraction includes extracting the rate of change of compressor frequency, peak module temperature, etc., to enhance the model's understanding of the data. This solution significantly improves data quality and consistency, providing a solid foundation for subsequent model training. Data cleaning ensures data accuracy and completeness, reducing the interference of noise and outliers on model training. Data normalization enables data of different dimensions to be compared and processed on the same scale, improving the training efficiency and stability of the model. Feature extraction further enhances the model's ability to understand the data, allowing it to learn key features and patterns more effectively. These preprocessing steps work together to significantly improve the model's training performance and the accuracy of fault diagnosis, providing an important guarantee for achieving efficient and accurate fault diagnosis.

[0073] Furthermore, the air conditioner can input preprocessed data into the embedding layer of the lightweight Transformer model. The main task of the embedding layer of the lightweight Transformer model is to map the raw input data into a high-dimensional vector space to extract feature representations of the data. Specifically, the embedding layer of the lightweight Transformer model transforms the m-dimensional input features into n-dimensional embedding vectors through linear transformation. This process not only transforms the data into a new space but also better reveals the inherent characteristics and patterns of the data in this space. For example, through linear transformation, features that are difficult to distinguish in low-dimensional space may become clearer in high-dimensional space, thus providing richer information for subsequent feature extraction and model training. In this way, the use of the embedding layer of the lightweight Transformer model significantly enhances the model's ability to understand the input data. By mapping the data to a high-dimensional vector space, the model can more effectively capture complex patterns and feature relationships in the data. This high-dimensional representation makes the model more adept at handling complex fault diagnosis tasks, enabling it to more accurately identify fault modes, thereby improving the accuracy and efficiency of fault diagnosis. Furthermore, the linear transformation of the embedding layer in the lightweight Transformer model provides a flexible way to adjust the representation of features, enabling the model to better adapt to different data characteristics and task requirements.

[0074] Furthermore, after processing the embedding layer of the lightweight Transformer model, the air conditioner adds positional encoding to the embedding vector at each time step to obtain a multi-dimensional temporal feature vector containing time information. The positional encoding is calculated as follows:

[0075] ;and,

[0076] ;

[0077] Where pos represents the position of the time step, i is the dimension index of the embedding vector, and d model This refers to the dimension of the embedding vector. In this embodiment, the dimension of the embedding vector is 64. This allows the model to capture the temporal sequence information of the data, enabling the model to better understand feature changes over time when processing time-series data.

[0078] Optionally, in step S32, the air conditioner inputs a multi-dimensional temporal feature vector into the encoder of a lightweight Transformer model for feature extraction to obtain the feature representation output by the encoder, including:

[0079] The air conditioner inputs multi-dimensional temporal feature vectors into the encoder of the lightweight Transformer model. The encoder of the lightweight Transformer model consists of multiple layers, and each layer includes a multi-head attention mechanism and a feedforward network.

[0080] In each encoder layer, the air conditioner uses a multi-head attention mechanism to calculate the attention output and passes the attention output to the feedforward network to extract features through nonlinear transformation.

[0081] The air conditioner performs residual connection and layer normalization on the output of the feedforward network to obtain the output of the encoder of that layer.

[0082] The air conditioner uses the output of the previous encoder as the input of the next encoder. Through layer-by-layer transmission and processing, it obtains the characteristic representation of the encoder output.

[0083] In this scheme, the air conditioner inputs the multi-dimensional temporal feature vector, processed by the embedding layer of the lightweight Transformer model, into the encoder of the lightweight Transformer model for feature extraction. Here, the encoder of the lightweight Transformer model consists of multiple layers, each containing a multi-head attention mechanism and a feedforward network. These two components work together to extract complex features and dependencies from the data. Specifically, the multi-head attention mechanism allows the model to focus on different parts of the input sequence in different subspaces, thereby capturing dependencies and feature information in the data more comprehensively. It projects the input query, key, and value into multiple low-dimensional subspaces through linear transformations, calculates attention scores in each subspace, and finally sums these scores with weights to obtain the output. This process can be represented as:

[0084]

[0085] Where, d k is the dimension of the key vector.

[0086] Alternatively, the Softmax algorithm formula is as follows:

[0087]

[0088] Here, K represents the dimension of the input variable, and each element in the input variable is calculated according to the formula.

[0089] Furthermore, the output of the multi-head attention layer is passed to the feedforward network. In this embodiment, the feedforward network is a simple two-layer fully connected neural network containing a ReLU activation function, which can be expressed by the following formula:

[0090]

[0091] Furthermore, the output of the feedforward network is residually connected to the input, essentially adding the feedforward network's output and input together. The result of the residual connection undergoes layer normalization to ensure the data has a similar distribution across different feature dimensions. After layer normalization, the output of that layer's encoder is obtained. Thus, the air conditioner uses the output of the previous encoder as the input to the next, passing and processing it layer by layer to obtain the feature representation of the encoder output. This approach, combining multi-head attention and a feedforward network, enhances the model's ability to learn complex patterns. Moreover, residual connections and layer normalization further improve the model's training stability and convergence speed, enabling the model to learn useful feature representations more efficiently.

[0092] Optionally, the 3D classification head includes a fully connected layer and a Softmax activation function; S34, the air conditioner inputs a fixed-length feature vector into the 3D classification head for classification to obtain the failure probabilities of each component output by the 3D classification head, including:

[0093] Air conditioners use fully connected layers to map fixed-length feature vectors to a fault probability space, obtaining the original probability of each component's failure.

[0094] The air conditioner uses the Softmax activation function to convert the original probability into a normalized fault probability distribution, so as to output the fault probability of each component.

[0095] In this scheme, the air conditioner can input a fixed-length feature vector obtained through global average pooling into a 3D classification head for classification. Here, the 3D classification head is a simple fully connected layer whose main function is to map the fixed-length feature vector to a fault probability space. Specifically, the fully connected layer transforms the input feature vector into a vector with the same number of fault categories through a linear transformation, where each element represents the raw probability of a component failure. These raw probabilities are then processed by the Softmax activation function, which converts each raw probability into a value between 0 and 1, ensuring that the sum of all component failure probabilities is 1. Finally, the model outputs a probability distribution representing the probability of each component failure; for example, the probability of compressor failure is 70%, outdoor fan failure is 20%, and circuit board failure is 10%. This scheme effectively transforms the feature vector into a probability distribution of component failures. The use of the Softmax activation function ensures that the output probability distribution has good mathematical properties, making the model output more intuitive and easier to interpret. Compared to directly outputting the raw probabilities, the probability distribution after Softmax processing more accurately reflects the likelihood of each component failure, thereby improving the accuracy and reliability of fault diagnosis. In addition, this classification method has good generalization ability and can adapt to the fault diagnosis needs of air conditioner outdoor unit modules of different brands, models and usage environments.

[0096] Combination Figure 4 As shown, optionally, after recommending that a repairman bring the appropriate spare parts for on-site repair, the method also includes:

[0097] S41, upon receiving maintenance feedback information, the air conditioner continuously updates the multimodal training set based on the maintenance feedback information.

[0098] S42, the air conditioner re-inputs the updated multimodal training set into the lightweight Transformer model for model training, in order to improve the model's accuracy in fault diagnosis.

[0099] In this solution, to further improve the model's accuracy and adaptability, the air conditioner continuously optimizes the model based on repair feedback after recommending repair personnel to bring the appropriate spare parts for on-site repair. Specifically, after the repair personnel complete the repair and provide feedback, the air conditioner receives feedback information including the actual faulty component, repair measures, and repair results. This information is used to update the multimodal training set, that is, adding new repair cases to the existing training data. Furthermore, the air conditioner re-inputs the updated multimodal training set into the lightweight Transformer model for model training. In this way, the model can learn the latest fault modes and repair experience, thereby continuously improving the accuracy of fault diagnosis. With this solution, by continuously updating the training dataset and model parameters, the model can gradually improve its ability to identify complex fault modes, reduce false positives and false negatives, thereby improving repair efficiency, reducing after-sales costs, and enhancing user experience. In addition, this method can reduce reliance on large amounts of historical data, allowing the model to adapt to new equipment and environmental conditions more quickly and maintain the reliability of its diagnostic capabilities.

[0100] Combination Figure 5 As shown, optionally, this disclosure provides a fault diagnosis device 200 for an air conditioner outdoor unit module, including a construction module 51, a training module 52, a diagnosis module 53, and a recommendation module 54. The construction module 51 is configured to collect air conditioner operation data and after-sales service order data to construct a multimodal training set, wherein the after-sales service order data includes a repair timestamp, a description of the fault phenomenon, and information on the actual replaced parts. The training module 52 is configured to input the multimodal training set into a lightweight Transformer model for model training to obtain a fault diagnosis model. The diagnosis module 53 is configured to input the pre-fault operation data into the fault diagnosis model when the air conditioner triggers a fault code in the outdoor unit module, to output the component fault probability. The recommendation module 54 is configured to recommend repair personnel to bring corresponding spare parts for on-site repair based on the component fault probability.

[0101] The fault diagnosis device 200 for air conditioner outdoor unit modules provided in this disclosure not only improves the efficiency and accuracy of fault diagnosis, but also significantly enhances the generalization ability of the model through multi-source data fusion and deep learning technology. This allows maintenance personnel to more accurately predict faulty components, reduce repeated repairs due to missing spare parts, significantly improve the first-time repair success rate, reduce after-sales costs, and enhance the user experience.

[0102] Combination Figure 6As shown, this disclosure provides an apparatus 300 for an air conditioner outdoor unit module, including a processor 30 and a memory 31. Optionally, the apparatus 300 may further include a communication interface 32 and a bus 33. The processor 30, communication interface 32, and memory 31 can communicate with each other via the bus 33. The communication interface 32 can be used for information transmission. The processor 30 can call logical instructions in the memory 31 to execute the fault diagnosis method for the air conditioner outdoor unit module described in the above embodiment.

[0103] Furthermore, the logical instructions in the aforementioned memory 31 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0104] The memory 31, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor 30 executes functional applications and data processing by running the program instructions / modules stored in the memory 31, thereby implementing the fault diagnosis method for the air conditioner outdoor unit module in the above embodiments.

[0105] The memory 31 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 31 may include high-speed random access memory and may also include non-volatile memory.

[0106] This disclosure provides an air conditioner, including an outdoor unit and the aforementioned fault diagnosis device 200 (300) for the outdoor unit module. The fault diagnosis device 200 (300) for the outdoor unit module is installed in the outdoor unit. The installation relationship described herein is not limited to placement inside the outdoor unit, but also includes installation connections with other components of the air conditioner, including but not limited to physical connections, electrical connections, or signal transmission connections. Those skilled in the art will understand that the fault diagnosis device 200 (300) for the outdoor unit module can be adapted to any feasible outdoor unit, thereby enabling other feasible embodiments.

[0107] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to perform the above-described fault diagnosis method for an air conditioner outdoor unit module.

[0108] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc., and other media capable of storing program code.

[0109] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0110] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0111] The methods and products disclosed in the embodiments herein (including but not limited to devices and equipment) can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0112] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A fault diagnosis method for an air conditioner outdoor unit module, characterized in that, include: Collect air conditioner operation data and after-sales work order data to construct a multimodal training set. The air conditioner operation data includes indoor and outdoor ambient temperature, compressor frequency, coil temperature, electronic expansion valve opening, and fan speed. The after-sales work order data includes maintenance timestamp, fault description, and actual replaced parts information. The multimodal training set is input into a lightweight Transformer model for training to obtain a fault diagnosis model. The lightweight Transformer model includes a multi-layer encoder, each layer comprising a multi-head attention mechanism and two fully connected feedforward neural networks. The feedforward neural network includes a ReLU activation function, and the lightweight Transformer model injects temporal information into the input data through sinusoidal position encoding. The process of inputting the multimodal training set into the lightweight Transformer model for training to obtain a fault diagnosis model includes: inputting the multimodal training set into the embedding layer of the lightweight Transformer model for feature mapping and position encoding to obtain a multi-dimensional temporal feature vector; and inputting the multi-dimensional temporal feature vector into the encoder of the lightweight Transformer model for feature extraction to obtain the lightweight Transformer model. The feature representation output by the encoder of the lightweight Transformer model is analyzed in depth by a causal graph inference module. The causal graph inference module performs the following operations: learns a causal directed acyclic graph (DAG) between variables from historical fault data to establish causal chains; calculates a causal weight matrix; weights and fuses the calculated causal weight matrix with the original attention score of the Transformer model, using a gating mechanism to adjust the ratio between the two. When the causal confidence exceeds a threshold, the corresponding attention score is increased by 30% to 50% to strengthen the weight of the root cause feature; the threshold is 0.8; the feature representation strengthened by the causal graph inference module is compressed using global average pooling to obtain a fixed-length feature vector; the fixed-length feature vector is input into a 3D classification head for classification to obtain the failure probability of each component output by the 3D classification head. When the air conditioner triggers a fault code in the outdoor unit module, the air conditioner's operating data before the fault is input into the fault diagnosis model to output the component failure probability. Based on the probability of component failure, it is recommended that repair personnel bring the corresponding spare parts to the site for repair, so that repair personnel can complete the repair task more quickly and improve the success rate of repair. The method also includes: When the fault diagnosis model outputs the failure probability of each component, it also outputs the corresponding confidence level. When the confidence level of a certain component exceeds the confidence level threshold, the air conditioner recommends that the repair personnel carry the spare part of that component.

2. The method according to claim 1, characterized in that, The step of inputting the multimodal training set into a lightweight Transformer model for model training to obtain a fault diagnosis model further includes: The model loss value is calculated based on the failure probability of each component in the output. The model parameters are updated based on the model loss value until the iteration terminates when a preset condition is met. When the iteration terminates, the trained fault diagnosis model is obtained; The preset conditions include the model loss value satisfying a preset convergence condition or the current iteration number reaching a set number of iterations.

3. The method according to claim 1, characterized in that, The step of inputting the multimodal training set into the embedding layer of a lightweight Transformer model for feature mapping and position encoding to obtain a multidimensional temporal feature vector includes: The air conditioner operation data and the after-sales work order data in the multimodal training set are preprocessed; The preprocessed data is input into the embedding layer of the lightweight Transformer model, and the input data is mapped to a high-dimensional vector space through linear transformation to obtain the embedding vector. Position encoding is added to the embedding vector at each time step to obtain a multidimensional temporal feature vector containing time information.

4. The method according to claim 1, characterized in that, The step of inputting the multidimensional temporal feature vector into the encoder of the lightweight Transformer model for feature extraction, so as to obtain the feature representation output by the encoder of the lightweight Transformer model, includes: The multidimensional temporal feature vector is input into the encoder of the lightweight Transformer model, which includes multiple layers, each of which includes a multi-head attention mechanism and a feedforward network. In each encoder layer, the attention output is calculated using the multi-head attention mechanism and then passed to the feedforward network to extract features through nonlinear transformation. The output of the feedforward network is subjected to residual connection and layer normalization to obtain the output of the encoder of that layer; The output of the previous encoder layer is used as the input of the next encoder layer, so that the feature representation of the encoder output of the lightweight Transformer model can be obtained by passing and processing it layer by layer.

5. The method according to claim 1, characterized in that, The 3D classification head includes a fully connected layer and a Softmax activation function; the step of inputting the fixed-length feature vector into the 3D classification head for classification to obtain the failure probability of each component output by the 3D classification head includes: The fixed-length feature vector is mapped to the fault probability space using the fully connected layer to obtain the original probability of each component failure. The original probabilities are converted into a normalized failure probability distribution using the Softmax activation function to output the failure probability of each component.

6. The method according to claim 1, characterized in that, After the recommended repair personnel bring the corresponding spare parts to the site for repair, the method also includes: Upon receiving maintenance feedback information, the multimodal training set is continuously updated based on the maintenance feedback information; The updated multimodal training set is re-input into the lightweight Transformer model for model training to improve the model's accuracy in fault diagnosis.

7. A fault diagnosis device for an air conditioner outdoor unit module, characterized in that, include: The construction module is configured to collect air conditioner operation data and after-sales work order data to build a multimodal training set. The after-sales work order data includes repair timestamps, fault descriptions, and actual replaced parts information. The air conditioner operation data includes indoor and outdoor ambient temperatures, compressor frequency, coil temperature, electronic expansion valve opening, and fan speed. The training module is configured to input the multimodal training set into a lightweight Transformer model for model training to obtain a fault diagnosis model. The lightweight Transformer model includes a multi-layer encoder, each encoder layer including a multi-head attention mechanism and two fully connected feedforward neural networks. The feedforward neural network contains a ReLU activation function, and the lightweight Transformer model injects temporal information into the input data through sinusoidal position encoding. The process of inputting the multimodal training set into the lightweight Transformer model for model training to obtain the fault diagnosis model includes: inputting the multimodal training set into the embedding layer of the lightweight Transformer model for feature mapping and position encoding to obtain a multi-dimensional temporal feature vector; inputting the multi-dimensional temporal feature vector into the encoder of the lightweight Transformer model for feature extraction to obtain the feature representation output by the encoder of the lightweight Transformer model; and processing the model through a causal graph inference module. The feature representation output by the encoder of the lightweight Transformer model is subjected to deep analysis. The causal graph inference module performs the following operations: learns a causal directed acyclic graph between variables from historical fault data to establish causal chains; calculates a causal weight matrix; weights and fuses the calculated causal weight matrix with the original attention score of the Transformer, using a gating mechanism to adjust the ratio between the two. When the causal confidence exceeds a threshold, the corresponding attention score is increased by 30% to 50% to strengthen the weight of the root cause feature; the threshold is 0.8; the feature representation strengthened by the causal graph inference module is compressed using global average pooling to obtain a fixed-length feature vector; the fixed-length feature vector is input into a 3D classification head for classification to obtain the failure probability of each component output by the 3D classification head. The diagnostic module is configured to input the pre-fault operating data into the fault diagnosis model when the air conditioner triggers a fault code in the outdoor unit module, so as to output the component failure probability. The recommendation module is configured to recommend maintenance personnel to bring the corresponding spare parts for on-site repair based on the failure probability of the component, so that maintenance personnel can complete the repair task more quickly and improve the success rate of repair. Also includes: When the fault diagnosis model outputs the failure probability of each component, it also outputs the corresponding confidence level. When the confidence level of a certain component exceeds the confidence level threshold, the air conditioner recommends that the repair personnel carry the spare part of that component.

8. A fault diagnosis device for an air conditioner outdoor unit module, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute, when running the program instructions, the fault diagnosis method for an air conditioner outdoor unit module as described in any one of claims 1 to 6.

9. An air conditioner, characterized in that, include: Air conditioner outdoor unit; The fault diagnosis device for an air conditioner outdoor unit module as described in claim 7 or 8 is installed on the air conditioner outdoor unit.

Images