Test data processing method and device based on edge calculation, medium and product

Abstract

The invention provides a test data processing method and device based on edge computing, a medium and a product. The method comprises the following steps that a cloud end trains and issues an update model to an edge end based on fault information and features uploaded by the edge of the last period; performing preliminary diagnosis on the edge end, and performing selective feature extraction on the data of the current period based on a preliminary diagnosis result to construct a feature vector; if the suspected fault exists, the edge end performs deep diagnosis by using an updating model to obtain the fault type and grade; and uploading the fault information and features of the current period to a cloud for subsequent model updating. According to the technical scheme, selective feature extraction and condition triggering deep diagnosis are guided through edge preliminary diagnosis, the edge calculation load and the network bandwidth are reduced, meanwhile, continuous optimization of the cloud model is ensured, and the data processing efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to a test data processing method, device, medium and product based on edge computing. Background Technology

[0002] Condition monitoring of industrial equipment (such as motors) is crucial for ensuring production. Utilizing sensors to collect operational data for testing and analysis to achieve fault early warning is a common requirement. With increasing demands for data dimensionality and real-time performance, efficient data processing methods are needed.

[0003] In existing technologies, edge-cloud collaborative test data processing methods typically involve edge nodes performing analysis models on all sensor data and systematically uploading the processed features to the cloud for model updates. However, in practical applications, this mechanism of uniformly processing all data and non-selectively uploading features leads to continuous high loads on edge computing resources and network bandwidth being occupied by a large amount of invalid feature data when faced with a large amount of redundant information (such as normal operation data) or large-scale deployments. This results in low processing efficiency and hinders the effective iteration of cloud models. Summary of the Invention

[0004] In view of this, this application provides a test data processing method, device, medium, and product based on edge computing to solve the above problems.

[0005] Firstly, an edge computing-based test data processing method is provided, applied to a data processing system including a cloud and multiple edge terminals. The method includes: The cloud acquires the first fault type, first fault level, and first multi-dimensional fusion feature vector uploaded from each of the edge terminals in the previous test cycle or preset by the cloud. Through the cloud, based on the first fault type, the first fault level and the first multi-dimensional fusion feature vector, a preset first machine learning model is trained to obtain a second machine learning model, which includes a convolutional neural network and a recurrent neural network. The second machine learning model is distributed to each of the edge devices via the cloud. The current operating conditions of the motor in the current test cycle are obtained through each of the edge terminals. Based on the current operating conditions, the acquisition frequencies of various preset sensors are selected, and the multi-source data acquired based on the acquisition frequencies are preliminarily processed. The multi-source data after preliminary processing is preliminarily diagnosed by each of the aforementioned edge terminals to determine whether a suspected fault is output and to generate a preliminary diagnostic result. Based on the preliminary diagnostic results, feature extraction and fusion are performed on the pre-processed multi-source data through each of the aforementioned edge terminals to construct the second multi-dimensional fusion feature vector for the current testing cycle; When the suspected fault is determined and output, the second multi-dimensional fusion feature vector corresponding to the suspected fault is input into the second machine learning model through each of the edge terminals to obtain the second fault type and the second fault level of the current test cycle. The second fault type, the second fault level, and the second multi-dimensional fusion feature vector are uploaded to the cloud via each of the edge terminals.

[0006] The above technical solution introduces a hierarchical processing and feature selection mechanism based on feature validity assessment at the edge node side. A lightweight analysis model performs preliminary screening of raw sensor data, triggering deep processing modules only on data showing detected state changes or abnormal fluctuations, thus significantly reducing redundant computation on normal data. An adaptive upload strategy based on feature confidence is established between the edge and the cloud, uploading only filtered and compressed key feature information, thereby reducing network bandwidth consumption. Since the uploaded dataset focuses on highly correlated fault features, the cloud model can iteratively optimize on higher-quality data samples, achieving efficient model training convergence and improved diagnostic accuracy. Through this chain-like improvement of "data hierarchical processing—feature selection and uploading—high-quality model update," the solution systematically addresses the problems of high edge computing resource consumption, network transmission redundancy, and inefficient cloud model updates in existing technologies, comprehensively improving the data processing efficiency and model evolution performance of industrial equipment condition monitoring.

[0007] Optionally, the step of training a preset first machine learning model through the cloud based on the first fault type, the first fault level, and the first multi-dimensional fusion feature vector to obtain a second machine learning model specifically includes: The cloud platform aggregates the first fault type, the first fault level, and the first multi-dimensional fusion feature vector uploaded from each of the edge terminals. The data quality of the first multi-dimensional fusion feature vectors collected is evaluated, and target multi-dimensional fusion feature vectors that meet the preset quality standards are selected. A global training dataset is constructed using the target multi-dimensional fusion feature vector and the first fault type and first fault level corresponding to the target multi-dimensional fusion feature vector. The first machine learning model is trained based on the global training dataset to obtain the second machine learning model.

[0008] The above technical solution assesses and filters the quality of data uploaded from each edge device, using only data that meets quality standards to build the training set, thus avoiding the impact of noisy data or incorrect labels on model performance. By aggregating and constructing a global training dataset, the cloud can obtain comprehensive training samples covering different working conditions and equipment differences, thereby improving the generalization ability of the machine learning model and ensuring that the model deployed to the edge device maintains stable recognition accuracy under different motors or environments.

[0009] Optionally, training the first machine learning model based on the global training dataset to obtain the second machine learning model specifically includes: The global training dataset is divided into multiple data subsets according to the source of each edge, and the multiple data subsets correspond to each edge. Based on each of the multiple data subsets, the parameters of the first machine learning model are loaded as initial parameters, and the initial parameters are iteratively trained on the cloud using each of the multiple data subsets to obtain multiple candidate update models. The performance of each candidate update model is tested using the global validation dataset pre-stored in the cloud, and a performance score for each candidate update model is obtained. The weighting coefficients of each candidate update model are determined based on the performance scores of each candidate update model. The weighting coefficient of each candidate update model is the proportion of the performance score of each candidate update model to the total performance score of all candidate update models. The second machine learning model is generated by weighting the candidate update models based on the weighting coefficients.

[0010] The above technical solution divides the global training dataset into multiple data subsets based on their edge source, trains candidate update models separately in the cloud, and then weights and fuses them based on the performance scores of each model on a unified validation set. This avoids model imbalance caused by biases in the distribution of data from a single edge. The second machine learning model after weighted averaging has a more balanced performance, taking into account the differences in working conditions at various edge points, making the diagnostic effect of the model more consistent across all field nodes.

[0011] Optionally, the step of performing preliminary diagnosis on the pre-processed multi-source data through each of the edge terminals to determine whether to output a suspected fault and generate a preliminary diagnosis result specifically includes: A fault rule base is configured at each of the aforementioned edge terminals. The fault rule base is preset with multiple sets of fault feature thresholds corresponding to the preliminary features extracted from the multi-source data after preliminary processing. The multiple sets of fault feature thresholds correspond to different operating conditions. Based on the current operating condition, select a set of fault feature thresholds corresponding to the current operating condition from the multiple sets of fault feature thresholds; Extract the preliminary features from the pre-processed multi-source data; The preliminary features are matched with the selected set of fault feature thresholds; When any one of the preliminary features exceeds the corresponding fault feature threshold in the set of fault feature thresholds, the suspected fault is determined and output, and a preliminary diagnostic result indicating the corresponding abnormality type is generated.

[0012] The above technical solution configures a fault rule base at the edge and selects an appropriate threshold set for feature matching based on the current operating conditions, directly filtering out abnormal signals from multi-source data. Through this threshold judgment mechanism executed locally at the edge, the system can detect potential anomalies in real time at the data acquisition site, reducing the latency of data upload and waiting for cloud analysis. This allows for rapid location of suspected faults during testing, supporting real-time alarms or subsequent feature extraction processing.

[0013] Optionally, the step of extracting and fusing features from the pre-processed multi-source data based on the preliminary diagnostic results at each of the edge endpoints to construct a second multi-dimensional fusion feature vector for the current testing period specifically includes: When the preliminary diagnostic result indicates a specific anomaly type, a feature subset related to the specific anomaly type is preferentially extracted from the pre-processed multi-source data, and the feature subset is combined with a set of preset basic features used to characterize the normal operating state of the motor to construct the second multi-dimensional fusion feature vector; When the preliminary diagnostic results do not indicate a specific abnormality type, a preset standard feature set is extracted from the pre-processed multi-source data to construct the second multi-dimensional fusion feature vector. The standard feature set includes electrical features, vibration features, dynamic features, and thermal features.

[0014] The above technical solution dynamically adjusts the sensor sampling frequency according to the operating conditions and employs timestamp synchronization, wavelet denoising, and outlier detection algorithms to unify and reduce noise in multi-source data, thereby improving the accuracy of subsequent feature analysis. When a specific anomaly type is detected, a multi-dimensional fusion feature vector is constructed by prioritizing the extraction of relevant feature subsets and combining them with basic features. This retains key feature information while reducing interference from irrelevant data, making the generated feature vector more reflective of the current fault characteristics or equipment operating status.

[0015] Optionally, when the suspected fault is determined and output, the second multi-dimensional fusion feature vector corresponding to the suspected fault is input into the second machine learning model through each of the edge terminals to obtain the second fault type and the second fault level of the current test period, specifically including: The high-dimensional feature map is extracted from the second multi-dimensional fusion feature vector by the convolutional neural network. The high-dimensional feature map is a combination of features extracted by the convolutional neural network that can characterize the motor's operating state. Construct a time-series feature sequence based on the high-dimensional feature map; The time-series feature sequence is input into the recurrent neural network to obtain the dynamic change trend of the time-series feature sequence over time. Based on the dynamic change trend, the second fault type and the second fault level are output.

[0016] The above technical solution uses a model composed of convolutional neural networks and recurrent neural networks at the edge to perform deep identification of suspected fault data. The convolutional network extracts high-dimensional spatial features representing the motor state from the fused feature vector, while the recurrent network further captures the dynamic trends of these features over time. By combining spatiotemporal features, more accurate fault types and fault levels can be directly output at the edge, enabling graded judgment of complex motor faults.

[0017] Optionally, the step of extracting the high-dimensional feature map from the second multi-dimensional fused feature vector through the convolutional neural network specifically includes: The second multi-dimensional fused feature vector is input into the convolutional layer of the convolutional neural network, and the convolution operation is performed on the second multi-dimensional fused feature vector by the convolutional kernel in the convolutional layer to generate a feature map. The feature map is downsampled by the pooling layer of the convolutional neural network to generate the high-dimensional feature map.

[0018] The above technical solution extracts local features by performing convolution operations in convolutional layers and generates high-dimensional feature maps by downsampling in pooling layers, which can reduce the amount of computation while preserving key feature patterns. This structure can effectively compress the data dimensionality of the original multidimensional feature vectors, remove redundant information, and enable subsequent recurrent neural networks to process time series changes more efficiently, thereby improving the stability and speed of fault feature identification.

[0019] Secondly, embodiments of this application provide an electronic device, including a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, and both the user interface and the network interface are used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method described in any of the above-mentioned methods.

[0020] Thirdly, embodiments of this application provide a computer-readable storage medium storing instructions that, when executed, perform the method described in any of the above-mentioned embodiments.

[0021] Fourthly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a server, cause the server to perform the method described in the first aspect and any possible implementation thereof.

[0022] It is understood that the electronic device provided in the second aspect, the computer-readable storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the methods provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0023] In summary, implementing one or more technical solutions provided in this application has at least the following technical effects or advantages: By constructing a dynamic collaboration mechanism between the edge and the cloud, hierarchical autonomous processing and global intelligent optimization of test data are achieved. The edge dynamically adjusts sampling and processing depth based on operating conditions, reducing redundant data and computational load; the cloud performs quality assessment and distributed balanced training based on high-confidence features, avoiding bandwidth consumption caused by full data transmission. Through closed-loop optimization of "data filtering—feature uploading—model feedback—local application," rapid model iteration and synergistic improvement of system performance are achieved. Attached Figure Description

[0024] Figure 1 This is an exemplary system architecture diagram of a test data processing method based on edge computing according to this application; Figure 2 This is a flowchart illustrating a test data processing method based on edge computing in an embodiment of this application. Figure 3 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application.

[0025] Explanation of reference numerals in the attached figures: 100, System architecture; 101, First terminal device; 102, Second terminal device; 103, Third terminal device; 104, Network; 105, Server; 301, Processor; 302, Communication bus; 303, User interface; 304, Network interface; 305, Memory. Detailed Implementation

[0026] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0027] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0028] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0029] Figure 1 An exemplary system architecture diagram is shown, illustrating an embodiment of a test data processing method based on edge computing that can be applied according to this application.

[0030] like Figure 1 As shown, the system architecture 100 may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used as a medium to provide communication links between the terminal devices 101, 102, 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0031] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as model training applications, video recognition applications, web browser applications, social platform software, etc.

[0032] Terminal devices 101, 102, and 103 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices with displays, including but not limited to smartphones, tablets, e-book readers, MP3 (Moving Picture Experts Group Audio Layer III) players, MP4 (Moving Picture Experts Group Audio Layer IV) players, laptops, and desktop computers, etc. When terminal devices 101, 102, and 103 are software, they can be installed in the aforementioned electronic devices. They can be implemented as multiple software programs or software modules (e.g., multiple software programs or software modules used to provide distributed services) or as a single software program or software module. No specific limitations are imposed here.

[0033] When terminals 101, 102, and 103 are hardware devices, video capture devices can also be installed on them. These video capture devices can be various devices capable of capturing video, such as cameras, sensors, etc. Users can use the video capture devices on terminals 101, 102, and 103 to capture video.

[0034] Server 105 can be a server that provides various services, such as a backend server for processing data displayed on terminal devices 101, 102, and 103. The backend server can analyze and process the received data and can feed back the processing results (such as recognition results) to the terminal devices.

[0035] It should be noted that a server can be either hardware or software. When the server is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the server is software, it can be implemented as multiple software programs or software modules (e.g., multiple software programs or software modules used to provide distributed services), or as a single software program or software module. No specific limitations are made here.

[0036] It should be understood that Figure 1The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included. In particular, if the target data does not need to be obtained remotely, the above system architecture may exclude the network and include only terminal devices or servers. Figure 2 This is a flowchart illustrating a test data processing method based on edge computing, as described in this application. This method can be implemented using a computer program or a microcontroller. The computer program can be integrated into an application or run as a standalone utility application. The specific steps of this edge computing-based test data processing method are described in detail below.

[0037] S201: Obtain the first fault type, first fault level, and first multi-dimensional fusion feature vector uploaded from each of the edge terminals in the previous test cycle through the cloud.

[0038] In this embodiment of the application, the first multi-dimensional fusion feature vector refers to the set of feature data that is constructed and uploaded to the cloud after the edge device processes the multi-source data collected in the previous test cycle (e.g., after preliminary processing, feature extraction and fusion) and is used to input the machine learning model. This vector contains key information representing the motor's operating state at that time and will be used as the input feature for cloud model training in the current cycle.

[0039] Specifically, at the start of a new test cycle (e.g., cycle N, where N is a positive integer greater than 1), historical data for cloud model training in the current cycle (cycle N) needs to be acquired. This data represents the processing results of the previous test cycle (cycle N-1). The previous test cycle refers to the test execution interval immediately preceding the current test cycle. Specifically, data uploaded by each pre-configured edge device (a computing node deployed near the motor or equipment site with certain computing and storage capabilities) at the end of cycle N-1 is retrieved from the pre-configured cloud (a remote server or platform providing centralized data storage and computing capabilities). This data includes: a first fault type (the specific fault category diagnosed by the edge machine learning model in cycle N-1, such as bearing failure, winding short circuit, etc.), a first fault level (the level indicating the severity of the fault, such as minor, moderate, severe, etc., accompanying the first fault type output), and a first multi-dimensional fusion feature vector (the set of feature data constructed by the edge device after processing the collected multi-source data in cycle N-1, used as input to the machine learning model). This step aims to ensure that all necessary real-time status information (current operating conditions) and historical data from the previous cycle (first fault type, level, and feature vector) for model training in the current cycle are ready before entering the model training (subsequent step S202) and edge processing phases.

[0040] It should be noted that claim 1 and Figure 2 The first step of the illustrated process includes the qualifier "or preset," which is intended to cover the first test cycle (N=1) when the system is first deployed or launched. In the first test cycle, since there is no data uploaded from the "previous test cycle," the cloud training step (as in subsequent step S202) cannot be based on historical feedback. In this initialization scenario, it can be launched as follows: the cloud can directly distribute a pre-trained, general initial machine learning model (as the "second machine learning model" at this time) to each edge device (performing the operation as in subsequent step S203). Each edge device will then use this initial model to execute the processing flow of the current (first) test cycle (i.e., subsequent steps S204 to S207 and subsequent output). At the end of the first test cycle, the edge device will upload the second fault type, second fault level, and second multi-dimensional fusion feature vector generated in that cycle to the cloud, as specified in subsequent step S208. Thus, starting from the second test cycle, when the cloud executes step S201, it can obtain the data actually uploaded in the first cycle, thereby entering an iterative optimization closed loop based on real feedback data. Therefore, the method described in this application can effectively start from the initial state and enter a working mode of continuous adaptive optimization. S202: Through the cloud, based on the first fault type, the first fault level, and the first multi-dimensional fusion feature vector, a preset first machine learning model is trained to obtain a second machine learning model, the second machine learning model including a convolutional neural network and a recurrent neural network.

[0041] For example, this step describes the core training update process performed by the cloud at the beginning of each new test cycle to achieve continuous adaptive optimization of the model. This process utilizes the actual diagnostic results (i.e., the first fault type and the first fault level) collected from various edge devices in the previous test cycle, along with their associated input features (i.e., the first multi-dimensional fusion feature vector). This real-world data is used as new training samples to iteratively train the pre-set machine learning model in the current system (i.e., the first machine learning model from the previous test cycle, which serves as the starting point for this training). Through this training process, a new model version (i.e., the second machine learning model) is generated, incorporating the latest operational data experience and designed to improve performance. This new model version is then deployed to the edge devices for real-time data processing and diagnostic tasks in the current test cycle.

[0042] In one possible implementation, the step of training a preset first machine learning model through the cloud based on the first fault type, the first fault level, and the first multi-dimensional fusion feature vector to obtain a second machine learning model specifically includes: collecting the first fault type, the first fault level, and the first multi-dimensional fusion feature vector uploaded from each of the edge terminals through the cloud; performing data quality assessment on the collected first multi-dimensional fusion feature vectors to select target multi-dimensional fusion feature vectors that meet preset quality standards; constructing a global training dataset using the target multi-dimensional fusion feature vectors and the first fault type and the first fault level corresponding to the target multi-dimensional fusion feature vectors; and training the first machine learning model based on the global training dataset to obtain the second machine learning model.

[0043] In this embodiment, data quality assessment refers to the process of checking and evaluating the collected multi-dimensional fused feature vectors to determine whether these feature vectors meet preset requirements or standards, so as to select high-quality data for model training. For example, the assessment can be based on indicators such as the completeness, consistency, signal-to-noise ratio, outlier detection results, or the confidence score of the model when generating the vector.

[0044] Specifically, the cloud platform will aggregate (i.e., collect and integrate) all data uploaded from various edge devices in the previous testing cycle. This data includes the first fault type, the first fault level, and the corresponding first multi-dimensional fusion feature vector. After aggregation, to ensure the quality of the training data, a data quality assessment will be performed on all aggregated first multi-dimensional fusion feature vectors. This assessment process may include checking whether the feature vector dimensions are complete, whether there are missing values, whether the values ​​are within the expected reasonable range (e.g., determining upper and lower limits through statistical analysis), or calculating the signal-to-noise ratio (SNR). These assessment metrics will be compared with preset quality standards, which are defined, for example, as: the proportion of missing values ​​in the feature vector is less than 5%, all values ​​are within 3 standard deviations of the historical statistical range, and the SNR is higher than 20 dB. Only feature vectors that meet all assessment metrics of the preset quality standards will be selected. These feature vectors that pass the quality check are called target multi-dimensional fusion feature vectors. Using these selected target multi-dimensional fusion feature vectors, and their corresponding (unselected but must be paired with their qualified feature vectors) first fault types and first fault levels, a global training dataset is constructed. The global training dataset is the high-quality, cleaned dataset used for this model training. Based on this constructed global training dataset, incremental training is performed on the pre-defined first machine learning model (i.e., the model version trained and deployed to the edge in the previous test cycle, serving as the basis for this update). This training process aims to iteratively adjust the existing parameters of the first machine learning model using new data from the global training dataset to update the model's internal parameters (such as weights and biases). After training, the optimized model obtained is the second machine learning model.

[0045] In one possible implementation, training the first machine learning model based on the global training dataset to obtain the second machine learning model specifically includes: dividing the global training dataset into multiple data subsets according to the source of each edge endpoint, wherein the multiple data subsets correspond to each edge endpoint; loading the parameters of the first machine learning model as initial parameters based on each of the multiple data subsets, and iteratively training the initial parameters using each of the data subsets in the cloud to obtain multiple candidate update models; performing performance tests on each of the candidate update models using a global validation dataset pre-stored in the cloud to obtain a performance score for each candidate update model; determining the weighting coefficient of each candidate update model based on each performance score, wherein the weighting coefficient of each candidate update model is the proportion of the performance score of each candidate update model to the total performance score of all candidate update models; and performing a weighted average of the candidate update models based on the weighting coefficients to generate the second machine learning model.

[0046] In this embodiment, the performance score refers to a numerical metric used to quantify the performance of each candidate update model on a specific task. This metric is obtained by performing performance tests on the candidate update models on a pre-stored global validation dataset. For example, the performance score can be the accuracy, precision, recall, or other predefined evaluation criteria achieved by the model on the validation dataset.

[0047] Specifically, this step details an implementation method for training a second machine learning model based on a global training dataset. Its core lies in generating the final updated model through a weighted fusion strategy based on performance evaluation. The constructed global training dataset is divided according to the edge from which the data originally originated; that is, all data is grouped according to their corresponding edge identifiers, forming multiple data subsets. Each data subset contains all training data from a specific edge. For each data subset, an independent training process is performed: the existing parameters of the pre-defined first machine learning model are loaded as initial parameters for training, and then these initial parameters are iteratively trained using this data subset in the cloud. The iterative training refers to the process of optimizing the model through multiple rounds of data processing and parameter updates. Since each training process is based on a different data subset, this results in multiple different localized model versions representing optimizations based on data from each edge; these models are called candidate updated models. To evaluate the actual performance of these candidate updated models, a global validation dataset, different from the training data, is pre-stored in the cloud. The performance of each candidate update model is tested using the global validation dataset. For example, predictions are made by the model and compared with the true labels in the validation data to calculate the performance score of each candidate update model. After obtaining the performance scores of all candidate update models, their respective contribution weights in the final fusion model are determined based on these scores; that is, the weighting coefficients of each candidate update model are determined. Specifically, the calculation method is as follows: the weighting coefficient of each candidate update model is equal to its performance score divided by the sum of the performance scores of all candidate update models, i.e., the proportion of that model's performance score to the total performance score. After obtaining all weighting coefficients, a weighted average operation is performed: each candidate update model is multiplied by its corresponding weighting coefficient, and then all products are summed to obtain the final fusion model. This model generated by the weighted average is the second machine learning model.

[0048] Furthermore, to ensure the efficiency and robustness of cloud-based iterative training, the following more specific implementation methods can be adopted: The iterative training process can use the Adam (Adaptive Moment Estimation) optimizer, as it can adaptively adjust the learning rate for different parameters and is suitable for processing complex industrial data; the initial learning rate can be set to a small value, such as 0.001, to allow for stable fine-tuning based on the parameters of the first machine learning model; simultaneously, to prevent oscillations or overfitting in the later stages of training, a learning rate decay strategy can be used, such as exponential decay, which multiplies the learning rate by a decay factor less than 1 (e.g., 0.95) every certain number of training steps (or rounds). The choice of loss function should be closely related to the specific output task: for fault type prediction (usually a multi-class problem), the cross-entropy loss function can be used, which can effectively measure the difference between the predicted probability distribution and the true label; for fault level prediction, if it is regarded as a continuous value regression problem, mean squared error loss or mean absolute error loss can be used; if it is regarded as an ordered multi-class problem (e.g., "mild", "moderate", "severe"), ranking loss or a classification loss function with order information can be used. The termination condition for training can be set to completing a preset maximum number of epochs (e.g., 50 epochs, where 1 epoch represents the entire training dataset being processed by the model once) to ensure sufficient learning; or, more preferably, an early stopping strategy can be adopted: during training, the performance of the candidate updated model is periodically evaluated on the global validation dataset (e.g., after each epoch). If it is found that the performance no longer shows significant improvement (e.g., the loss no longer decreases or the key performance score no longer increases) over several consecutive evaluation epochs (e.g., the "patience value" is set to 5 or 10 epochs), the training of the model is terminated early to prevent overfitting and save computational resources. The construction of the global validation dataset is crucial. It should be independent of the training dataset and contain balanced samples from different operating conditions and different edge cases to ensure the unbiasedness and comprehensiveness of the model performance evaluation. Balance means that the data of certain common operating conditions or fault types should not be too high, causing the model evaluation to be biased towards these majority classes. Its size is recommended to be no less than 10% of the total training data to ensure the statistical stability of the evaluation results. When calculating the performance scores of each candidate updated model, if there is class imbalance in the fault types (some faults are rare), the macro-average F1 score is preferred. The macro-average F1 score calculates the F1 score for each class separately and then averages them, treating each class equally and better reflecting the model's recognition performance on rare fault categories, which is particularly important for industrial fault diagnosis. When generating the final second machine learning model by weighted averaging based on the performance scores, the calculated weighting coefficients (based on the proportion of performance scores) can be smoothed to increase the robustness of the fusion process.For example, a minimum weight (such as 0.01 or dynamically adjusted according to the number of models) can be set for each candidate update model. This means that even if a candidate model performs slightly worse on the validation set, the information it has learned (which may be valuable for specific edge cases or operating conditions) will not be completely ignored, which helps to improve the generalization ability and adaptability of the final fusion model to unseen scenarios. S203: The second machine learning model is distributed to each of the edge terminals via the cloud.

[0049] In this application embodiment, "deployment" refers to the process by which the cloud transmits and deploys the trained or updated machine learning model file or its parameters to a designated edge node via the network. The purpose is to enable the edge node to use the latest version of the model for local data processing and inference tasks.

[0050] Specifically, in step S202, after the cloud successfully trains an updated model—the second machine learning model—based on the data uploaded in the previous test cycle, it needs to deploy this new model to each edge computing node that actually performs test data processing. The cloud initiates the model distribution process, transmitting the newly trained second machine learning model (which includes a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN), and may include model structure files, parameter files, etc.) to each pre-configured edge terminal via a network (e.g., a wireless or wired network). After receiving this new model file, the edge terminal can, for example, first load and verify the new model in the background. After successful verification, during the next processing interval or according to a preset strategy (e.g., during low-load periods), switch the processing task to the new model and unload the old model, thereby completing the online update and replacement of the model. This step ensures that all edge terminals can promptly obtain and apply the model optimized with the latest data in the cloud, preparing for the subsequent execution of deep diagnostics (triggered by preliminary diagnostics) in the current test cycle (the Nth cycle). S204: Obtain the current operating condition of the motor in the current test cycle through each of the edge terminals, select the preset acquisition frequency of multiple sensors based on the current operating condition, and perform preliminary processing on the multi-source data acquired based on the acquisition frequency.

[0051] In this embodiment, preliminary processing refers to basic preprocessing operations performed on the newly collected raw multi-source data at the edge. The purpose is to clean, format, or preliminarily straighten the data to prepare it for subsequent preliminary diagnosis or more complex feature extraction steps. For example, preliminary processing may include filtering out obvious noise interference, data unit conversion, or basic timestamp alignment.

[0052] Specifically, after the updated second machine learning model is distributed from the cloud to each edge device, the edge device needs to start processing the data for the current test cycle. This involves adjusting the data acquisition strategy based on the real-time operating status of the motor. Using the acquired current operating condition of the motor, the acquisition frequency of various preset sensors (e.g., vibration sensors, temperature sensors, current sensors, etc.) is selected. The acquisition frequency refers to the rate at which the sensor collects data points per unit time; the system will preset at least two different acquisition frequencies. Based on the specific circumstances of the current operating condition (e.g., whether it is dynamic or stable), the corresponding preset acquisition frequency is selected. For example, the edge device continuously monitors key parameters in the current operating condition, such as motor speed and load. The rate of change or standard deviation of the key parameters within a preset time window (e.g., 1 second) is calculated and compared with a preset stability threshold. When the rate of change or standard deviation is lower than its respective preset stability threshold, the current operating condition is determined to be a stable operating condition. Conversely, when the rate of change or standard deviation of any parameter exceeds its preset stability threshold, or when the edge device receives a speed adjustment or load change command from the motor control system, the current operating condition is determined to be a dynamic operating condition. When the current operating condition is the stable operating condition, the edge device uses the second sampling frequency (e.g., 1kHz for vibration sensors and 5kHz for current sensors) to collect data; when the current operating condition is the dynamic operating condition, it immediately switches to the first sampling frequency (e.g., 20kHz for vibration sensors and 50kHz for current sensors) to collect data. This setting ensures that data is collected at a lower frequency when the operating condition is stable, saving edge device processing resources; while when the operating condition changes drastically (dynamic operating condition), data can be collected at a higher frequency (the first sampling frequency is greater than the second sampling frequency), ensuring that transient high-frequency signals that may indicate faults are captured. Through each edge device, multiple sensors connected to them are controlled to begin acquiring real-time multi-source data according to the selected sampling frequency. The multi-source data refers to various types of data streams from different types of sensors that reflect the motor's operating status. After acquiring this multi-source data based on the sampling frequency, each edge device immediately performs preliminary processing on this raw data. This initial processing aims to provide data with a consistent format and reduced interference factors for subsequent steps through timestamp synchronization, noise reduction, and outlier handling. S205: Perform a preliminary diagnosis on the pre-processed multi-source data through each of the edge terminals, determine whether to output a suspected fault, and generate a preliminary diagnosis result.

[0053] For example, this step describes a state assessment performed on pre-processed data at an edge computing node. Its purpose is to use a diagnostic mechanism with preset rules and thresholds to perform a preliminary analysis of the current device operating data to determine if there are any signs of potential anomalies or malfunctions (i.e., suspected faults). This preliminary diagnostic process is intended to act as a screening checkpoint, and the resulting preliminary diagnostic results will be used to determine whether more complex processing procedures need to be initiated (e.g., calling a deep learning model for fine-tuning or performing specific feature engineering), thereby optimizing the allocation of computing resources at the edge.

[0054] In one possible implementation, the step of performing preliminary diagnosis on the pre-processed multi-source data through each of the edge terminals, determining whether to output a suspected fault, and generating a preliminary diagnosis result specifically includes: configuring a fault rule base at each of the edge terminals, wherein the fault rule base has preset multiple sets of fault feature thresholds corresponding to the preliminary features extracted from the pre-processed multi-source data, and the multiple sets of fault feature thresholds correspond to different operating conditions; selecting a set of fault feature thresholds corresponding to the current operating condition from the multiple sets of fault feature thresholds according to the current operating condition; extracting the preliminary features from the pre-processed multi-source data; matching the preliminary features with the selected set of fault feature thresholds; and when any feature value in the preliminary features exceeds the corresponding fault feature threshold in the set of fault feature thresholds, determining and outputting the suspected fault, and generating the preliminary diagnosis result indicating the corresponding abnormality type.

[0055] In this application embodiment, the fault rule base refers to a data structure or database that is pre-configured and stored at each edge end. It defines rules for initially judging whether the device status is abnormal. These rules are mainly reflected in thresholds associated with specific feature dimensions.

[0056] Specifically, a fault rule base needs to be pre-configured at each edge. The core of this fault rule base is a pre-defined set of multiple fault feature thresholds. These multiple fault feature thresholds refer to pre-defined threshold standards for different equipment operating states. For example, one set of thresholds is set for high-speed operation, and another set is set for low-speed operation, ensuring the applicability of the thresholds. Each set of fault feature thresholds contains specific threshold values ​​corresponding to various dimensions of the preliminary features to be extracted (e.g., current amplitude, peak vibration frequency, specific point temperature, etc.). During diagnosis, based on the currently acquired motor operating condition (i.e., the current operating condition of the current test cycle), the set of fault feature thresholds that precisely matches the current operating condition needs to be selected (or called) from the pre-defined multiple sets of fault feature thresholds. Furthermore, features for preliminary judgment need to be extracted from the pre-processed multi-source data obtained in step S204. These features are called preliminary features, such as time-domain statistics or frequency-domain main frequency amplitude. After extracting the preliminary features, a matching operation is performed: the actual values ​​of each extracted preliminary feature are compared with the corresponding thresholds in the set of fault feature thresholds selected based on the operating conditions. The judgment logic is as follows: if the value of any of these preliminary features exceeds (or falls below, depending on whether the threshold is defined as an upper or lower limit) the specific fault feature threshold in that set of thresholds, then the current state is determined to be abnormal, i.e., a suspected fault is output. Simultaneously, based on which feature(s) exceed their corresponding thresholds, a preliminary diagnostic result is generated. This result not only identifies the suspected fault but also indicates the corresponding abnormality type related to the exceeding feature (e.g., "abnormal vibration," "excessive temperature," etc.).

[0057] In a preferred embodiment, the extraction of the preliminary features may include: calculating the root mean square (RMS) value of the current signal within a sliding time window as the current amplitude feature; performing a Fast Fourier Transform (FFT) on the vibration signal and using the frequency corresponding to the maximum amplitude within a preset frequency band as the vibration frequency peak feature; and reading the instantaneous value of the temperature sensor and performing a moving average filter to obtain the temperature feature at a specific point. The fault rule base can be implemented as a key-value storage structure, where the first-level key is the operating condition identifier (e.g., a string defined by a combination of speed range and load range), the second-level key is the name of the preliminary feature (e.g., "CurrentRMS"), and the corresponding value is the threshold value of that feature under that operating condition. The division of different operating conditions can be based on motor nameplate parameters and historical operating data. For example, a speed below 30% of the rated speed can be defined as a low-speed condition, 30%-70% as a medium-speed condition, and above 70% as a high-speed condition. During matching, each extracted preliminary feature value is compared with the threshold corresponding to its operating condition. For example, if the current RMS value is greater than the upper limit threshold for current under high-speed operating conditions, a suspected fault is identified, and "abnormal current" is indicated in the preliminary diagnosis result. For features that require judgment intervals, the rule base can store both upper and lower limit thresholds simultaneously. S206: Based on the preliminary diagnosis results, feature extraction and fusion are performed on the pre-processed multi-source data through each of the aforementioned edge terminals to construct the second multi-dimensional fusion feature vector for the current testing cycle.

[0058] For example, this step describes how the edge device constructs the final input features for subsequent deep analysis based on the preliminary diagnostic results after obtaining preliminary diagnostic information. This process uses the preliminary diagnostic results generated in the previous step as guidance to perform feature extraction and fusion operations on the previously pre-processed multi-source data. Crucially, the specific methods of feature extraction and fusion are not fixed but dynamically adjusted according to the state indicated by the preliminary diagnostic results (e.g., whether it is a specific type of suspected fault). The aim is to generate a second, multi-dimensional fused feature vector containing information related to the preliminary diagnostic results for subsequent machine learning model judgment.

[0059] In one possible implementation, the step of extracting and fusing features from the pre-processed multi-source data based on the preliminary diagnostic results through each of the edge terminals to construct a second multi-dimensional fusion feature vector for the current test cycle specifically includes: when the preliminary diagnostic results indicate a specific anomaly type, preferentially extracting a feature subset related to the specific anomaly type from the pre-processed multi-source data, and combining the feature subset with a set of preset basic features used to characterize the normal operating state of the motor to construct the second multi-dimensional fusion feature vector; when the preliminary diagnostic results do not indicate a specific anomaly type, extracting a preset standard feature set from the pre-processed multi-source data to construct the second multi-dimensional fusion feature vector, wherein the standard feature set includes electrical features, vibration features, dynamic features, and thermal features.

[0060] In the embodiments of this application, a feature subset refers to a set of features selected from all possible features that are most relevant to a specific anomaly type indicated by the preliminary diagnostic results. This subset is used for more targeted analysis when a suspected fault occurs. For example, when the preliminary diagnosis indicates a vibration anomaly, the feature subset may preferentially include refined vibration features such as the spectral peak value, kurtosis factor, and specific frequency band energy of the vibration signal.

[0061] Specifically, after the preliminary processing in step S204, the feature extraction and fusion method will be determined based on the preliminary diagnostic results generated in step S205: If the preliminary diagnostic results clearly indicate a specific type of anomaly (e.g., abnormal vibration or overheating), then features will be preferentially extracted from the pre-processed multi-source data. For example, a specific advanced signal processing algorithm (such as envelope spectrum analysis) can be called to calculate the feature subset. Then, this specifically extracted feature subset will be combined with a set of other preset basic features. "Other basic features" refer to standard features other than the preferentially extracted feature subset used to characterize the normal operating state of the motor. Through this combination method, the second multi-dimensional fusion feature vector for the current test cycle is constructed. On the other hand, if the preliminary diagnostic results do not indicate any specific type of anomaly (e.g., only judged as "suspected fault" but cannot be classified, or no anomaly detected), then a preset standard feature set will be extracted from the pre-processed multi-source data according to a preset standard procedure. The standard feature set is a relatively comprehensive set of features, including, for example, the electrical characteristics of the motor (such as current and voltage harmonics), vibration characteristics (such as total vibration value and basic spectrum), dynamic characteristics (such as speed fluctuations), and thermal characteristics (such as temperature of key components). In this case, this extracted standard feature set is directly used as the second multi-dimensional fusion feature vector for the current test cycle.

[0062] In a preferred embodiment, when the preliminary diagnostic results indicate vibration anomaly, the feature subset may include detailed spectral information obtained through short-time Fourier transform, envelope spectrum analysis results, and time-domain statistical features such as kurtosis and margin factor. When indicating "overheating," the feature subset may include temperature gradients and heating rates of multiple key components (such as windings and bearings). The "priority extraction" can be manifested as calling a specific, computationally intensive, high-level feature extraction algorithm to calculate the features of the subset based on the indicated anomaly type, while "other basic features" are extracted using a preset basic feature extraction algorithm (e.g., a standard feature set extraction algorithm). "Other basic features" may partially overlap with the "standard feature set." For example, if fine vibration features are extracted first, other basic features may include electrical, dynamic, and thermal features from the standard feature set, as well as simplified vibration features (such as total vibration value RMS). The extraction of the standard feature set may include: calculating the amplitude of each harmonic of the current signal using FFT as electrical features; calculating the time-domain RMS, peak-to-peak value, and frequency-domain dominant frequency amplitude of the vibration signal as vibration features; calculating the average value and standard deviation of the rotational speed signal as dynamic features; and reading the average value of each temperature sensor as thermal features. Feature fusion typically involves concatenating all extracted features (whether subset + base or standard set) into a fixed-length vector according to a predefined dimensional order (e.g., electrical features first, then vibration features, then dynamic features, and finally thermal features, or sorted according to feature importance), which serves as the second multi-dimensional fusion feature vector. This adaptive feature extraction and fusion method based on preliminary diagnostic results makes the feature vectors input to the second machine learning model more targeted, helping to improve the accuracy of deep diagnosis. It may also reduce the computational overhead of the feature engineering stage by avoiding complex calculations for all features. S207: When the suspected fault is determined and output, the second multi-dimensional fusion feature vector corresponding to the suspected fault is input into the second machine learning model through each edge terminal to obtain the second fault type and the second fault level of the current test cycle.

[0063] For example, this step describes a deep diagnostic process performed at the edge for precise fault identification and severity rating. This process is not continuous but is activated only when a previous preliminary diagnostic step determines that the current data indicates a suspected fault. Once triggered, this step uses a previously constructed second multi-dimensional fusion feature vector, which may have been adaptively adjusted based on the preliminary diagnostic results, as input. It then calls a second machine learning model, updated and deployed from the cloud, for analysis, aiming to output a final judgment on the specific type and severity level of the device fault within the current testing period.

[0064] In one possible implementation, when the suspected fault is determined and output, the second multi-dimensional fusion feature vector corresponding to the suspected fault is input into the second machine learning model through each of the edge terminals to obtain the second fault type and the second fault level of the current test cycle. Specifically, this includes: extracting a high-dimensional feature map from the second multi-dimensional fusion feature vector through the convolutional neural network, wherein the high-dimensional feature map is a combination of features extracted by the convolutional neural network that can characterize the motor operating state; constructing a time-series feature sequence based on the high-dimensional feature map; inputting the time-series feature sequence into the recurrent neural network to obtain the dynamic change trend of the time-series feature sequence over time; and outputting the second fault type and the second fault level based on the dynamic change trend.

[0065] In this embodiment of the application, a high-dimensional feature map refers to an intermediate data representation extracted by a CNN from the second multi-dimensional fused feature vector of the input. It is not the original sensor reading, but a set of combined features generated after convolution and pooling operations that can characterize the motor's operating state. It usually contains multiple feature channels and is used to capture local patterns or abstract relationships in the input data.

[0066] Specifically, this step is only triggered when the suspected fault is output in the preliminary diagnosis. During execution, the second multi-dimensional fused feature vector corresponding to the suspected fault event is used as input data and provided to the second machine learning model deployed at each of the edge devices. This second machine learning model is explicitly defined as a composite structure containing both CNN and RNN components. The second multi-dimensional fused feature vector is fed into the CNN part of the model. The CNN performs feature extraction, generating the high-dimensional feature map from the input vector. This map represents the combined features extracted by the CNN that characterize the motor's operating state. A temporal feature sequence is constructed based on the high-dimensional feature map, for example, by performing dimensionality transformation (e.g., flattening) on ​​the high-dimensional feature map and constructing it into a time-series format (e.g., arranging the feature map vectors extracted from continuous time windows by time steps), for processing by the RNN. RNNs excel at processing sequential data; they analyze the dynamic trends of the temporal feature sequence over time, i.e., capturing the patterns of how these combined features evolve over time. Based on the analysis results of this dynamic trend by RNN, the model outputs the final diagnostic conclusion, that is, to determine the specific second fault type (e.g., "bearing outer ring damage") and the corresponding second fault level (e.g., "severe") of the current test cycle.

[0067] In one possible implementation, the step of extracting the high-dimensional feature map from the second multi-dimensional fused feature vector through the convolutional neural network specifically includes: inputting the second multi-dimensional fused feature vector into the convolutional layer of the convolutional neural network, performing convolution operations on the second multi-dimensional fused feature vector through the convolutional kernel in the convolutional layer to generate a feature map; and performing downsampling processing on the feature map through the pooling layer of the convolutional neural network to generate the high-dimensional feature map.

[0068] In the embodiments of this application, a convolutional layer refers to the core layer in a CNN responsible for performing the main feature extraction function. It performs sliding window convolution operations on the input data by using one or more learnable convolutional kernels (also called filters) to detect local patterns or features in the input.

[0069] Specifically, this step details the implementation process of "extracting high-dimensional feature maps from the second multi-dimensional fused feature vector using the convolutional neural network" mentioned in the previous step. The second multi-dimensional fused feature vector is input into the convolutional layer of the CNN. Inside the convolutional layer, there are a preset number and size of convolutional kernels. These kernels perform convolution operations on the input second multi-dimensional fused feature vector, that is, by using a sliding window and weighted summation, to detect specific local patterns or features (such as specific edges or combinations of frequency components in a signal waveform) in the data. The result of the convolution operation is the generation of one or more feature maps, which are the response representations obtained after filtering the input data through specific convolutional kernels. To reduce data dimensionality, decrease computational load, and improve feature robustness, these generated feature maps are fed into the pooling layer of the CNN. The pooling layer performs downsampling processing on the input feature maps, such as using max pooling or average pooling, to reduce the spatial size of the feature maps while preserving the main feature information. After downsampling by the pooling layer, the resulting dimensionality-reduced feature map is defined as the high-dimensional feature map and serves as the final output of this feature extraction stage for use by subsequent network layers (such as RNNs).

[0070] In a preferred embodiment, the CNN can be constructed by alternating stacks of two convolutional layers and two max-pooling layers. For example, the first convolutional layer uses 32 3x3 convolutional kernels with a stride of 1 and the ReLU activation function; followed by a 2x2 max-pooling layer. The second convolutional layer uses 64 3x3 convolutional kernels, also using ReLU activation and 2x2 max-pooling. The RNN can be a bidirectional long short-term memory (Bi-LSTM) network layer containing 128 hidden units. Before inputting the high-dimensional feature map output by the CNN into the RNN, it may be necessary to flatten it and perform dimensionality adjustment through a fully connected layer, and then construct the input sequence required by the RNN from the continuously output feature vectors within the time window. The final hidden state or output sequence of the RNN layer can be connected to two output heads through a fully connected layer: one outputs the probability distribution of each fault type using the Softmax activation function, and the other can directly output the predicted value of the fault level (if considered as a regression task) or output the probability of different levels through another Softmax layer (if considered as a classification task).

[0071] In a preferred embodiment, after obtaining the second fault type and second fault level for the current test cycle through a second machine learning model, the diagnostic results need to be output for users to view, record, or trigger subsequent maintenance processes. For example, the fault type and level can be displayed on a monitoring interface, stored in a test log database, or notified to relevant personnel through an alarm system. S208: Upload the second fault type, the second fault level, and the second multi-dimensional fusion feature vector to the cloud via each of the aforementioned edge terminals.

[0072] For example, after completing the deep fault diagnosis and outputting the test results for the current test cycle at the edge, these diagnostic results and feature data need to be fed back to the cloud to support the continuous learning and model iteration of the entire system. Each edge device packages the second fault type, the second fault level, and the second multi-dimensional fusion feature vector used to generate the diagnostic results, which it finally determined in the current test cycle (the Nth cycle), according to a predefined data format (e.g., JSON or Protobuf), and sends (i.e., uploads) it to the pre-configured cloud via a secure network connection (e.g., HTTPS (Hypertext Transfer Protocol Secure) or encrypted MQTT (Message Queuing Telemetry Transport)). Simultaneously, to optimize network bandwidth resources, the edge device will, through configuration changes or command control, stop transmitting unprocessed raw data streams from various sensors to the cloud, and only upload processed and filtered high-value information. This upload operation ensures that the cloud can collect the latest labeled data related to actual (or suspected) faults from all edge devices (where the second fault type and second fault level can be used as training labels, and the second multi-dimensional fusion feature vector can be used as training input features). This uploaded data will be used by the cloud to train and update the machine learning model at the beginning of the next test cycle (the N+1th cycle) (i.e., it becomes the data basis for the "first fault type", "first fault level", and "first multi-dimensional fusion feature vector" required when executing step S202 in the next cycle), thus forming a closed loop for model optimization.

[0073] Accordingly, if it is determined in step S205 that no suspected fault is output, then the in-depth diagnosis in step S207 and the data upload in step S208 are not performed. In this way, only high-value data containing suspected or confirmed faults are uploaded to the cloud, thereby reducing the network transmission of redundant data under normal operation and achieving the technical effect of reducing network bandwidth consumption. This embodiment also discloses an electronic device, as shown in the reference. Figure 3 The electronic device may include: at least one processor 301, at least one communication bus 302, user interface 303, network interface 304, and at least one memory 305.

[0074] The communication bus 302 is used to enable communication between these components.

[0075] The user interface 303 may include a display screen and a camera. Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0076] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0077] The processor 301 may include one or more processing cores. The processor 301 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 305, and by calling data stored in memory 305. Optionally, the processor 301 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 301 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 301 and may be implemented as a separate chip.

[0078] The memory 305 may include random access memory (RAM) or read-only memory. Optionally, the memory 305 may include a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 305 may also be at least one storage device located remotely from the aforementioned processor 301. Figure 3 As shown, the memory 305, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a test data processing method based on edge computing.

[0079] exist Figure 3 In the electronic device shown, the user interface 303 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 301 can be used to call an application program stored in the memory 305 that is a test data processing method based on edge computing. When executed by one or more processors 301, the electronic device executes one or more methods as described in the above embodiments.

[0080] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0081] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0082] In some embodiments of this application, a computer-readable storage medium is provided, including instructions that, when executed on the electronic device, cause the electronic device to perform a test data processing method based on edge computing as described in this application.

[0083] In some embodiments of this application, a computer program product is also provided, which, when run on an electronic device, causes the electronic device to execute a test data processing method based on edge computing as described in this application.

[0084] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only 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. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0085] 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 achieve the purpose of this embodiment according to actual needs.

[0086] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0087] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 305 and includes several 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 methods of the various embodiments of this application. The aforementioned memory 305 includes various media capable of storing program code, such as a USB flash drive, external hard drive, magnetic disk, or optical disk.

[0088] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the disclosure in this specification. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are considered exemplary only, and the scope of this application is defined by the claims.

Claims

1. A test data processing method based on edge computing, characterized in that, Applied to a data processing system, which includes a cloud and multiple edge endpoints, the method includes: The cloud acquires the first fault type, first fault level, and first multi-dimensional fusion feature vector uploaded from each of the edge terminals in the previous test cycle or preset by the cloud. Through the cloud, based on the first fault type, the first fault level and the first multi-dimensional fusion feature vector, a preset first machine learning model is trained to obtain a second machine learning model, which includes a convolutional neural network and a recurrent neural network. The second machine learning model is distributed to each of the edge devices via the cloud. The current operating conditions of the motor in the current test cycle are obtained through each of the edge terminals. Based on the current operating conditions, the acquisition frequencies of various preset sensors are selected, and the multi-source data acquired based on the acquisition frequencies are preliminarily processed. The multi-source data after preliminary processing is preliminarily diagnosed by each of the aforementioned edge terminals to determine whether a suspected fault is output and to generate a preliminary diagnostic result. Based on the preliminary diagnostic results, feature extraction and fusion are performed on the pre-processed multi-source data through each of the aforementioned edge terminals to construct the second multi-dimensional fusion feature vector for the current testing cycle; When the suspected fault is determined and output, the second multi-dimensional fusion feature vector corresponding to the suspected fault is input into the second machine learning model through each of the edge terminals to obtain the second fault type and the second fault level of the current test cycle. The second fault type, the second fault level, and the second multi-dimensional fusion feature vector are uploaded to the cloud via each of the aforementioned edge terminals.

2. The method according to claim 1, characterized in that, The step of training a preset first machine learning model through the cloud, based on the first fault type, the first fault level, and the first multi-dimensional fusion feature vector, to obtain a second machine learning model, specifically includes: The cloud platform aggregates the first fault type, the first fault level, and the first multi-dimensional fusion feature vector uploaded from each of the edge terminals. The data quality of the first multi-dimensional fusion feature vectors collected is evaluated, and target multi-dimensional fusion feature vectors that meet the preset quality standards are selected. A global training dataset is constructed using the target multi-dimensional fusion feature vector and the first fault type and first fault level corresponding to the target multi-dimensional fusion feature vector. The first machine learning model is trained based on the global training dataset to obtain the second machine learning model.

3. The method according to claim 2, characterized in that, The step of training the first machine learning model based on the global training dataset to obtain the second machine learning model specifically includes: The global training dataset is divided into multiple data subsets according to the source of each edge, and the multiple data subsets correspond to each edge. Based on each of the multiple data subsets, the parameters of the first machine learning model are loaded as initial parameters, and the initial parameters are iteratively trained on the cloud using each of the multiple data subsets to obtain multiple candidate update models. The performance of each candidate update model is tested using the global validation dataset pre-stored in the cloud, and a performance score for each candidate update model is obtained. The weighting coefficients of each candidate update model are determined based on the performance scores of each candidate update model. The weighting coefficient of each candidate update model is the proportion of the performance score of each candidate update model to the total performance score of all candidate update models. The second machine learning model is generated by weighting the candidate update models based on the weighting coefficients.

4. The method according to claim 1, characterized in that, The preliminary diagnosis of the multi-source data after initial processing through each of the aforementioned edge terminals, determining whether to output a suspected fault, and generating a preliminary diagnosis result specifically includes: A fault rule base is configured at each of the aforementioned edge terminals. The fault rule base is preset with multiple sets of fault feature thresholds corresponding to the preliminary features extracted from the multi-source data after preliminary processing. The multiple sets of fault feature thresholds correspond to different operating conditions. Based on the current operating condition, select a set of fault feature thresholds corresponding to the current operating condition from the multiple sets of fault feature thresholds; Extract the preliminary features from the pre-processed multi-source data; The preliminary features are matched with the selected set of fault feature thresholds; When any one of the preliminary features exceeds the corresponding fault feature threshold in the set of fault feature thresholds, the suspected fault is determined and output, and a preliminary diagnostic result indicating the corresponding abnormality type is generated.

5. The method according to claim 1, characterized in that, The step involves extracting and fusing features from the pre-processed multi-source data based on the preliminary diagnostic results at each of the edge endpoints to construct a second multi-dimensional fusion feature vector for the current testing period. Specifically, this includes: When the preliminary diagnostic result indicates a specific anomaly type, a feature subset related to the specific anomaly type is preferentially extracted from the pre-processed multi-source data, and the feature subset is combined with a set of preset basic features used to characterize the normal operating state of the motor to construct the second multi-dimensional fusion feature vector; When the preliminary diagnostic results do not indicate a specific abnormality type, a preset standard feature set is extracted from the pre-processed multi-source data to construct the second multi-dimensional fusion feature vector. The standard feature set includes electrical features, vibration features, dynamic features, and thermal features.

6. The method according to claim 1, characterized in that, When the suspected fault is determined and output, the second multi-dimensional fusion feature vector corresponding to the suspected fault is input into the second machine learning model through each of the edge terminals to obtain the second fault type and the second fault level of the current test period, specifically including: The high-dimensional feature map is extracted from the second multi-dimensional fusion feature vector by the convolutional neural network. The high-dimensional feature map is a combination of features extracted by the convolutional neural network that can characterize the motor's operating state. Construct a time-series feature sequence based on the high-dimensional feature map; The time-series feature sequence is input into the recurrent neural network to obtain the dynamic change trend of the time-series feature sequence over time. Based on the dynamic change trend, the second fault type and the second fault level are output.

7. The method according to claim 6, characterized in that, The step of extracting the high-dimensional feature map from the second multi-dimensional fused feature vector through the convolutional neural network specifically includes: The second multi-dimensional fused feature vector is input into the convolutional layer of the convolutional neural network, and the convolution operation is performed on the second multi-dimensional fused feature vector by the convolutional kernel in the convolutional layer to generate a feature map. The feature map is downsampled by the pooling layer of the convolutional neural network to generate the high-dimensional feature map.

8. An electronic device, characterized in that, Including processor and memory; The memory is used to store computer program code, the computer program code including computer instructions, and the processor invokes the computer instructions to cause the electronic device to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed on an electronic device, the electronic device causes the electronic device to perform the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on an electronic device, it causes the electronic device to perform the method as described in any one of claims 1-7.

Images