An industrial time series data-oriented cloud-edge collaborative adaptive fault diagnosis method

Abstract

The application discloses a cloud-edge collaborative adaptive fault diagnosis method for industrial time series data, and belongs to the technical field of industrial internet of things and intelligent manufacturing. In view of the problems of strong noise of industrial time series data, easy drift of working conditions, limited edge computing power, and interrupted service of model upgrading, an adaptive optimization mechanism of three-layer cooperation of data-model-deployment is proposed: the edge end adaptively denoises and purifies the original vibration and temperature time series data, and completes low-latency fault reasoning through a lightweight model; the cloud end drives the adaptive iteration of the model through distributed drift detection, and performs secondary lightweight according to the edge computing power; finally, the optimized model is downloaded to the edge in a hot replacement manner, forming an adaptive diagnosis closed loop without interrupting service, so that the fault diagnosis system dynamically adapts to the data distribution, working condition change and edge computing power, and the robustness, real-time performance and engineering landing ability of the industrial equipment diagnosis are improved.

Description

Technical Field

[0001] This application relates to the fields of industrial Internet of Things and intelligent manufacturing technology, specifically to a cloud-edge collaborative adaptive fault diagnosis method for scenarios with strong noise, variable operating conditions, and limited edge computing power. It is particularly suitable for real-time processing, fault diagnosis, and dynamic model optimization of vibration and temperature time series data of industrial equipment. Background Technology

[0002] With the deepening of Industry 4.0 and intelligent manufacturing, the stable operation of core processing equipment such as machine tools, milling machines, and spindles directly determines the efficiency of the production line and the quality of products. Online fault diagnosis based on time-series sensor data such as vibration and temperature has become a key technical means for intelligent operation and maintenance and predictive maintenance of equipment. However, time-series data collected in industrial fields generally have characteristics such as high noise, non-stationarity, high redundancy, and strong time-series correlation. Weak fault characteristics are easily drowned out by strong noise, which places high demands on the real-time performance and robustness of fault diagnosis.

[0003] To address the aforementioned issues, existing technologies have proposed corresponding solutions at different levels. At the signal preprocessing level, wavelet transform, empirical mode decomposition and its variants, and variational mode decomposition are commonly used for denoising and feature extraction of time-series data. At the fault identification level, deep learning models such as convolutional neural networks, recurrent neural networks, and temporal convolutional networks are used to complete end-to-end fault classification. At the deployment architecture level, a cloud-edge collaboration scheme has been proposed, with the edge handling low-latency inference and the cloud handling model training and centralized management. For data distribution drift, model update methods based on incremental learning or transfer learning have emerged. To address the limited computing power at the edge, model compression techniques such as channel pruning and weight quantization are used to reduce model size.

[0004] However, the existing technologies still have significant shortcomings: First, single filtering or decomposition methods struggle to effectively remove strong noise while retaining subtle fault characteristics, resulting in high feature redundancy and insufficient discriminative power. Second, uploading all raw data to the cloud leads to high bandwidth consumption and transmission latency, failing to meet real-time diagnostic requirements. Third, existing models are mostly based on static assumptions, making it difficult to adaptively update with changing operating conditions, leading to missed or false diagnoses, while full retraining in the cloud is inefficient and costly. Fourth, although pruned and quantized models can be deployed at the edge, model upgrades often require pausing diagnostic services, impacting production continuity. Fifth, existing cloud-edge collaboration solutions are mostly general-purpose architectures, failing to form a complete adaptive closed loop of data-model-deployment tailored to the characteristics of industrial time-series data, making it difficult to meet actual production needs. Therefore, there is an urgent need for an industrial time-series data fault diagnosis method that can dynamically adapt to data distribution, changes in operating conditions, and edge computing power, while ensuring uninterrupted service during model upgrades. Summary of the Invention

[0005] In view of this, this application provides a cloud-edge collaborative adaptive fault diagnosis method for industrial time-series data, which helps to solve the problems of strong noise, high redundancy, easy distribution drift of operating conditions, limited edge computing power, difficulty in adapting static models, and easy production interruption during upgrades in the existing technology.

[0006] In a first aspect, embodiments of this application provide a cloud-edge collaborative adaptive fault diagnosis method for industrial time-series data, comprising the following steps: The edge device performs adaptive denoising and feature extraction on the raw time-series data of vibration and temperature of industrial equipment to obtain a fault-sensitive compact feature vector. At the edge, a pre-deployed lightweight time-series diagnostic model is used. The compact feature vector is input to complete real-time fault reasoning, and the reasoning deviation information corresponding to this reasoning is recorded. The edge device uploads the compact feature vector and the corresponding inference bias information to the cloud via a lightweight transmission protocol; The cloud receives data uploaded from the edge terminal, performs data distribution drift detection based on the compact feature vector and the inference bias information, and performs adaptive iterative optimization of the model according to the drift type and drift degree obtained from the drift detection. The cloud performs secondary lightweighting of the optimized model based on computing power matching and distributes it to the edge. The edge completes the model upgrade in a hot-swap manner without interrupting the service.

[0007] In one possible implementation, the edge device performs adaptive denoising and feature extraction on the collected raw time-series data of industrial equipment vibration and temperature to obtain a fault-sensitive compact feature vector, including: The original time series data is decomposed into three levels and multiple scales using the db4 wavelet. The decomposition yields a low-frequency approximate component as the low-frequency effective signal, and at the same time, a high-frequency component composed of multiple high-frequency detail components is obtained. The high-frequency component is mixed with fault impact effective information and high-frequency noise. Variational mode decomposition is performed on the high-frequency components. The number of modes is adaptively determined according to the signal complexity. The correlation coefficient of each mode is calculated and the noise-dominant mode is removed to complete the fusion and denoising. Multi-domain features in the time domain, frequency domain, and time-frequency domain are extracted from the denoised time-series data, and significance is tested by analysis of variance to remove redundant features, thereby obtaining the fault-sensitive compact feature vector.

[0008] In one possible implementation, the edge device employs a pre-deployed lightweight temporal diagnostic model, inputs the compact feature vector to complete real-time fault inference, and records the inference deviation information corresponding to this inference, including: A basic network for constructing a lightweight time-series diagnostic model is selected, and a combined strategy of channel pruning and weight quantization is applied to the basic network to obtain a lightweight time-series diagnostic model adapted to edge hardware. The lightweight time-series diagnostic model is deployed at the edge to perform real-time fault reasoning on the compact feature vector and record reasoning deviation information.

[0009] In one possible implementation, the edge device uploads the compact feature vector and the corresponding inference bias information to the cloud via a lightweight transmission protocol, including: The operating condition is determined based on the fault reasoning output of the lightweight timing diagnostic model; If the inference output is a normal operating condition category, the operating condition is determined to be normal; if the inference output is a fault category, the operating condition is determined to be abnormal. Different upload frequencies are set according to whether the operating conditions are normal or abnormal; The edge device uploads the compact feature vector and the inference bias information to the cloud via a lightweight transmission protocol, according to the differentiated upload frequency, and verifies the integrity of the data uploaded to the cloud.

[0010] In one possible implementation, the cloud receives data uploaded from the edge, performs data distribution drift detection based on the compact feature vector and inference bias information, and performs adaptive iterative optimization of the model according to the drift type and degree obtained from the drift detection, including: The cloud-based system uses two metrics, KL divergence and maximum mean difference, to construct a joint drift score, DriftScore. Based on the DriftScore, the degree of drift is divided into three categories: no drift, slow drift, and abrupt drift. When the drift level is zero, the current model remains unchanged. When the drift level is slow drift or abrupt drift, the corresponding incremental learning or transfer learning update strategy is triggered.

[0011] In one possible implementation, triggering a corresponding incremental learning or transfer learning update strategy when the drift degree is a slow drift or abrupt drift includes: When the drift is slow, the model is locally updated using adaptive learning rate incremental learning, without full retraining. When the drift is a sudden drift, if there are enough samples, incremental learning with an increased learning rate is used to update the model parameters; if there are few samples and it is a new working condition, transfer learning is used to transfer general fault diagnosis knowledge to the new scenario, and only the high-level feature layer and classification layer are fine-tuned to achieve rapid adaptation.

[0012] In one possible implementation, the cloud performs a secondary lightweighting of the optimized model based on computing power matching and distributes it to the edge. The edge then completes the model upgrade via hot replacement without service interruption, including: Based on the computing power constraints of the target edge, the cloud performs a computing power matching secondary lightweighting process on the iteratively optimized model to obtain an updated model adapted to the edge. The cloud distributes the updated model to the corresponding edge device; the edge device uses a hot-swap mechanism to load the updated model, completing the model version upgrade without interrupting the fault diagnosis service, thus realizing cloud-edge collaborative closed-loop fault diagnosis.

[0013] In one possible implementation, the cloud distributes the updated model to the corresponding edge device; the edge device uses a hot-swap mechanism to load the updated model, completing the model version upgrade without interrupting the fault diagnosis service, thus achieving cloud-edge collaborative closed-loop fault diagnosis, including: The edge device divides the memory into a running area and a backup cache area. The running area loads the current online diagnostic model, and the backup cache area receives and loads the updated model sent from the cloud. By atomically switching the model inference entry pointer, the inference call object is switched from the old model in the running area to the new model in the backup cache area, completing the seamless model upgrade and forming a complete adaptive closed loop.

[0014] In one possible implementation, the edge is an ARM-based low-computing-power edge gateway, and the cloud is a high-performance computing cluster.

[0015] Secondly, embodiments of this application provide an electronic device, including: processor; Memory; And a computer program, wherein the computer program is stored in the memory, the computer program including instructions that, when executed by the processor, cause the electronic device to perform the method described in any one of the first aspects.

[0016] Thirdly, embodiments of this application provide a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the method described in any one of the first aspects.

[0017] In this embodiment, to address the problems of strong noise in industrial time-series data, easy drift in operating conditions, limited edge computing power, and easy service interruption during model upgrades, an adaptive optimization mechanism with three-layer collaboration of data, model, and deployment is proposed: the edge adaptively denoises and purifies the original vibration and temperature time-series data, and completes low-latency fault inference through a lightweight model; the cloud drives the model to adaptively iterate through distributed drift detection and performs secondary lightweighting according to the edge computing power; finally, the optimized model is distributed to the edge in a hot-swap manner, forming an adaptive diagnostic closed loop without service interruption. This enables the fault diagnosis system to dynamically adapt to data distribution, changes in operating conditions, and edge computing power, improving the robustness, real-time performance, and engineering implementation capability of industrial equipment diagnosis. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an overall flowchart of the cloud-edge collaborative adaptive fault diagnosis method provided in the embodiments of this application; Figure 2 This is a flowchart of edge-end adaptive denoising and multi-domain feature extraction provided in the embodiments of this application; Figure 3 A flowchart of lightweight TCN model construction and real-time inference provided for embodiments of this application; Figure 4 This application provides a precise interaction strategy diagram for cloud-edge time-series data in its embodiments. Figure 5 A flowchart of cloud data distribution drift detection and model adaptive iteration provided in the embodiments of this application; Figure 6 A flowchart illustrating the cloud-based model secondary lightweighting and edge-end hot-swap upgrade process provided in this application embodiment; Figure 7 This is a flowchart illustrating the business architecture of the cloud-edge collaborative fault diagnosis system provided in this application embodiment. Figure 8 This is a schematic diagram of the overall architecture and data flow of the cloud-edge collaborative intelligent production line provided in the embodiments of this application; Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] To better understand the technical solution of this application, the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0021] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0022] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0023] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0024] In the embodiments of this application, Figure 1 For an overall flowchart of the cloud-edge collaborative adaptive fault diagnosis method provided in the embodiments of this application, please refer to [link / reference]. Figure 1 The cloud-edge collaborative adaptive fault diagnosis method in this embodiment includes: S101: The edge device performs adaptive denoising and feature extraction on the raw time-series data of industrial equipment vibration and temperature to obtain a fault-sensitive compact feature vector.

[0025] like Figure 2 As shown, the edge end uses db4 wavelet to perform three-level multi-scale decomposition on the collected vibration and temperature raw time series data. The decomposition yields low-frequency approximate components as low-frequency effective signals, and at the same time, high-frequency components composed of multiple high-frequency detail components are obtained. The high-frequency components are mixed with fault impact effective information and high-frequency noise.

[0026] The high-frequency components obtained from the decomposition are subjected to VMD (Variational Mode Decomposition). The number of modes is adaptively determined based on the signal complexity, the correlation coefficient of each mode is calculated, and the noise-dominant mode is removed to complete the fusion and denoising process. The formula for VMD is: In the formula, Let be the value of the k-th modal component at time t. The modal components obtained from the decomposition, The center frequency corresponds to each mode. The original time-series signal to be decomposed. Indicates time Find the partial derivative. The term is a complex exponential term used in the construction of analytical signals. This represents minimizing the objective function. "Subject to" is an abbreviation for the constraint condition, indicating that the superposition of all modal components should completely reconstruct the original input signal, i.e., satisfy... .

[0027] To determine the total number of mode decompositions, an adaptive solution is performed based on the permutation entropy of high-frequency components, which characterizes the signal complexity: a larger permutation entropy value indicates a more complex signal, thus requiring a larger number of VMD modes. Conversely, a smaller permutation entropy indicates a simpler signal, thus requiring a smaller number of modes. This adaptive number of modes K avoids the under-decomposition problem caused by manually setting a fixed K.

[0028] After denoising, multi-domain features in the time domain, frequency domain, and time-frequency domain are extracted periodically within a fixed sliding time window. ANOVA (Analysis of Variance) is used for significance testing to remove redundant features, resulting in a compact and highly discriminative fault-sensitive feature vector. The time-frequency domain is a joint analysis domain that simultaneously represents the signal's time and frequency dimensions, constructed from the intrinsic mode components obtained through VMD decomposition.

[0029] ANOVA The formula for the statistic is: In the formula, For the sum of squares between groups, For the sum of squares within the group, The number of feature category groups, This represents the total number of samples. This is an ANOVA statistic used to assess the significance of differences in feature values ​​between different fault categories; The larger the value, the more significant the difference between different fault categories, and the stronger its fault identification ability. The smaller the value, the weaker the difference in the feature across various fault samples, indicating a redundant feature. In this embodiment, the value corresponding to each candidate feature is calculated. The statistic is set and a significance threshold is set. Features with statistics below a threshold are identified as redundant and removed, while features with statistics above a threshold are retained, ultimately forming a compact and highly discriminative fault-sensitive feature vector.

[0030] S102: The edge terminal adopts a pre-deployed lightweight time-series diagnostic model, inputs the compact feature vector to complete real-time fault reasoning, and records the reasoning deviation information corresponding to this reasoning.

[0031] like Figure 3As shown, a diagnostic model adapted to long-term temporal correlation characteristics is constructed based on TCN (Temporal Convolutional Network). The model has a depth of 4 layers, a kernel size of 32, and an expansion rate that increases layer by layer in increments of 1 / 2 / 4 / 8. A joint strategy of channel pruning and weight quantization is adopted to quantize the 32-bit floating-point parameters into 8-bit integers, forming a lightweight diagnostic model adapted to edge hardware.

[0032] Channel-level pruning includes: calculating the L1 norm of the convolutional kernel weights for each output channel of the TCN convolutional layer; identifying channels with an L1 norm below a preset pruning threshold as redundant channels and removing the corresponding convolutional kernels to reduce the number of network channels; and then fine-tuning the network using an industrial fault dataset to compensate for the accuracy loss caused by pruning and to compress the model structure.

[0033] The formula for weight quantization is: In the formula, round() This is the rounding function. For the original 32-bit floating-point parameters, For quantification Bit integer parameter, Scaling factor This is the zero-point offset. To quantize the number of bits, in this embodiment =8, corresponding to 8-bit integer quantization, zero-point offset. =0; , These represent the maximum and minimum values ​​of the parameter, respectively.

[0034] The lightweight model is deployed on an ARM architecture edge gateway to perform real-time inference on fault types such as spindle wear, bearing damage, and temperature overload, and to record inference deviation information such as diagnostic accuracy, inference output deviation, missed diagnosis rate, and false diagnosis rate.

[0035] S103: The edge device uploads the compact feature vector and the corresponding inference bias information to the cloud via a lightweight transmission protocol.

[0036] like Figure 4As shown, the edge end adopts MQTT (Message Queuing Telemetry Transport) protocol to upload compact feature vectors and inference deviation information to the cloud end, and judges the working condition state according to the real-time fault inference output of the edge-end lightweight time-series diagnosis model: if the model inference output is a normal working condition category, the working condition is determined to be normal; if the model inference output is various fault categories, the working condition is determined to be abnormal.

[0037] Set differentiated upload frequencies according to the working condition state: low-frequency upload for normal working conditions, and automatically increase the upload frequency for abnormal working conditions to ensure rapid response to anomalies. In this embodiment, the upload frequency for normal working conditions is once every 1 minute, and the upload frequency for abnormal working conditions is once every 10 seconds.

[0038] During the transmission process, the CRC32 data check mechanism is adopted to ensure transmission integrity, and after receiving the data, the cloud end classifies and stores the data by device number, working condition type and timestamp.

[0039] S104: The cloud end receives the data uploaded by the edge end, performs data distribution drift detection based on the compact feature vectors and the inference deviation information, and executes adaptive iterative optimization of the model.

[0040] As shown in Figure 5 , the cloud end adopts KL divergence and MMD (Maximum Mean Discrepancy) to jointly detect data distribution drift, and divides the drift degree into no drift, slow drift and sudden drift according to the drift degree.

[0041] The present application integrates the two indicators of KL divergence and MMD to construct a joint drift score DriftScore, and uses T1 and T2 as demarcation thresholds to divide the drift degree into three categories: no drift, slow drift, and sudden drift. The division rules are as follows: (1) No drift: DriftScore<T1, the distribution of the current working condition is basically consistent with the historical reference data, no model iteration is triggered, and the original model is maintained.

[0042] (2) Slow drift: T1 ≤ DriftScore<T2, the working condition has a gradual small offset, corresponding to the slow aging and wear of the equipment, and incremental learning with a learning rate of 0.001-0.005 is used for local model update.

[0043] (3) Sudden drift: DriftScore ≥ T2, the working condition has a sharp jump; when samples are sufficient, incremental learning with a learning rate of 0.005-0.01 is adopted, and transfer learning fine-tuning is adopted under the new working condition with few samples.

[0044] T1 and T2 are preset thresholds. In this example, T1=0.2 and T2=0.6 are used, which can be adaptively adjusted according to the device type.

[0045] The KL divergence formula is: In the formula, The sample feature values ​​representing the industrial production line are random variables in the industrial data sample space collected at the edge, and all possible values ​​of all sample features are traversed. This represents the probability distribution of the current edge data. This represents the historical data distribution of the cloud-based benchmark model. Used to measure the difference between two distributions, the larger the value, the more severe the distribution drift.

[0046] The MMD formula is: In the formula, It is taken from the distribution Sample characteristics (of the current data distribution at the edge). It is taken from the distribution Sample characteristics of (historical baseline data distribution in the cloud); and For kernel function mapping, , Distributions and Expectations For the regenerating nucleus Hilbert space; It is used to measure the global difference between two distributions, complements KL divergence, and improves the robustness of drift detection.

[0047] The formula for the joint drift score (DriftScore) is: In the formula, These are weighting coefficients used to balance the contributions of KL divergence and MMD. The joint scoring comprehensively determines the type and degree of data distribution drift. Weighting coefficients The value of is determined through historical simulation tests and cross-validation of industrial production lines: when the noise interference of the production line is strong and more emphasis is placed on the differences in local probability distribution, a larger value is selected. (0.7-0.8), increase the weight of KL divergence; when the production line samples have high dispersion and focus more on global distribution differences, take a smaller value. (0.5-0.65), increasing the weight of the MMD indicator. This embodiment preferably uses this value. =0.65.

[0048] For slow drift and mutation drift with sufficient samples, the model's adaptive learning rate is: In the formula, The initial learning rate for the model is set under two different scenarios: for the slow drift scenario... =0.001, mutation drift scenario is taken =0.005. This is the learning rate amplification factor, which is taken in this embodiment. =0.8. and All of these can be adaptively adjusted based on device and dataset characteristics. DriftScore is a joint drift score. The higher the drift level, the larger the learning rate, enabling adaptive updates of model parameters.

[0049] The formula for incremental learning parameter update is: In the formula, For the first Model parameters for the next iteration For adaptive learning rate, For the new sample set In the current model parameters The gradient of the loss function is used; only parameters related to drift are updated to avoid catastrophic forgetting.

[0050] For drift scenarios with few samples under new operating conditions, transfer learning is used to fine-tune the model. By leveraging the basic feature extraction capabilities of the model pre-trained under historical operating conditions, only the high-level drift-related parameters of the model are fine-tuned, reducing the model's dependence on the number of labeled samples under new operating conditions.

[0051] S105: The cloud performs secondary lightweighting of the optimized model by matching computing power and distributes it to the edge. The edge completes the model upgrade in a hot-swap manner without interrupting the service.

[0052] like Figure 6 As shown, based on the memory and computing power constraints of the edge gateway, the cloud performs secondary channel pruning and parameter sparsification on the iterative model to adapt the model size to edge deployment requirements. The lightweight model is then encrypted and distributed using AES. The edge terminal seamlessly upgrades the model using a hot-swap method without interrupting diagnostic services, forming a complete adaptive closed loop.

[0053] The hot-swap upgrade described in this application refers to the process of switching model versions at the edge without stopping the ongoing fault diagnosis inference business process. The specific implementation logic consists of four steps: 1. Dual Model Cache Partition: The edge memory is divided into two independent model storage areas. One is the running area, which stores the online diagnostic model that is currently executing inference; the other is the spare cache area, which is used to receive, decrypt, and temporarily store new lightweight models that are matched with computing power and sent from the cloud. 2. New model verification: The edge device receives the encrypted model from the cloud, completes decryption and integrity verification. After the verification is successful, the new model is loaded into the backup cache. During this stage, the original running model continues to provide fault diagnosis and inference services to the outside world, and the business is completely uninterrupted. 3. Atomic pointer switching: After verification, the process handle pointer is atomically switched to change the inference task's calling object from the old model in the runtime area to the new model in the standby cache area; the switch is an atomic operation, completed in milliseconds, and will not cause inference task interruption or frame loss; 4. Resource reclamation: After the pointer switch is completed and the new model is confirmed to be working normally for inference, the memory resources occupied by the old model are released; the spare cache is reset and awaits the next round of cloud model delivery.

[0054] Figure 7 This is a business architecture flowchart of the cloud-edge collaborative fault diagnosis system of this application, used to show the overall business data flow of the system; after inference is performed at the edge, it is divided into two business flows: one is to complete real-time early warning and health management of intelligent equipment operation and maintenance at the edge; the other is to upload data to the cloud through the MQTT protocol, and the cloud completes storage, drift detection, model iteration, and secondary lightweighting before sending it to the edge to complete hot replacement and upgrade of the model.

[0055] Figure 8 This diagram illustrates the overall architecture and data flow of a cloud-edge collaborative intelligent production line, showcasing a three-layer deployment architecture consisting of the edge device layer, the edge terminal, and the cloud data center.

[0056] It should be noted that the specific content involved in the embodiments of this application can be found in the description of the above method embodiments, and will not be repeated here for the sake of brevity.

[0057] Corresponding to the above embodiments, this application also provides an electronic device.

[0058] See Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 9 As shown, the electronic device 30 may include a processor 301, a memory 302, and a communication unit 303. These components communicate via one or more buses. Those skilled in the art will understand that the electronic device structure shown in the figures does not constitute a limitation on the embodiments of this application. It may be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0059] The communication unit 303 is used to establish a communication channel, thereby enabling the electronic device to communicate with other devices.

[0060] The processor 301 serves as the control center of the electronic device, connecting various parts of the device via various interfaces and lines. It executes software programs and / or modules stored in the memory 302, and calls data stored in the memory to perform various functions and / or process data. The processor can be composed of integrated circuits (ICs), such as a single packaged IC or multiple packaged ICs with the same or different functions connected together. For example, the processor 301 may consist only of a central processing unit (CPU). In this embodiment, the CPU may have a single processing core or include multiple processing cores.

[0061] Memory 302 is used to store the execution instructions of processor 301. Memory 302 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0062] When the execution instructions in memory 302 are executed by processor 301, the electronic device 30 is able to perform some or all of the steps in the above method embodiments.

[0063] Corresponding to the above embodiments, this application also provides a computer-readable storage medium, wherein the computer-readable storage medium may store a program, wherein when the program runs, it can control the device where the computer-readable storage medium is located to execute some or all of the steps in the above method embodiments. In specific implementation, the computer-readable storage medium may be a magnetic disk, an optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0064] Corresponding to the above embodiments, this application also provides a computer program product containing executable instructions that, when executed on a computer, cause the computer to perform some or all of the steps in the above method embodiments.

[0065] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0066] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. 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 this application.

[0067] Those skilled in the art will 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.

[0068] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium 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 described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0069] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A cloud-edge collaborative adaptive fault diagnosis method for industrial time-series data, characterized in that, Includes the following steps: The edge device performs adaptive denoising and feature extraction on the raw time-series data of vibration and temperature of industrial equipment to obtain a fault-sensitive compact feature vector. At the edge, a pre-deployed lightweight time-series diagnostic model is used. The compact feature vector is input to complete real-time fault reasoning, and the reasoning deviation information corresponding to this reasoning is recorded. The edge device uploads the compact feature vector and the corresponding inference bias information to the cloud via a lightweight transmission protocol; The cloud receives data uploaded from the edge terminal, performs data distribution drift detection based on the compact feature vector and the inference bias information, and performs adaptive iterative optimization of the model according to the drift type and drift degree obtained from the drift detection. The cloud performs secondary lightweighting of the optimized model based on computing power matching and distributes it to the edge. The edge completes the model upgrade in a hot-swap manner without interrupting the service.

2. The method according to claim 1, characterized in that, The edge device performs adaptive denoising and feature extraction on the collected raw time-series data of industrial equipment vibration and temperature to obtain a fault-sensitive compact feature vector, including: The original time series data is decomposed into three levels and multiple scales using the db4 wavelet. The decomposition yields a low-frequency approximate component as the low-frequency effective signal, and at the same time, a high-frequency component composed of multiple high-frequency detail components is obtained. The high-frequency component is mixed with fault impact effective information and high-frequency noise. Variational mode decomposition is performed on the high-frequency components. The number of modes is adaptively determined according to the signal complexity. The correlation coefficient of each mode is calculated and the noise-dominant mode is removed to complete the fusion and denoising. Multi-domain features in the time domain, frequency domain, and time-frequency domain are extracted from the denoised time-series data, and significance is tested by analysis of variance to remove redundant features, thereby obtaining the fault-sensitive compact feature vector.

3. The method according to claim 1, characterized in that, The edge device employs a pre-deployed lightweight temporal diagnostic model, inputs the compact feature vector to complete real-time fault inference, and records the inference deviation information corresponding to this inference, including: A basic network for constructing a lightweight time-series diagnostic model is selected, and a combined strategy of channel pruning and weight quantization is applied to the basic network to obtain a lightweight time-series diagnostic model adapted to edge hardware. The lightweight time-series diagnostic model is deployed at the edge to perform real-time fault reasoning on the compact feature vector and record reasoning deviation information.

4. The method according to claim 1, characterized in that, The edge device uploads the compact feature vector and corresponding inference bias information to the cloud via a lightweight transmission protocol, including: The operating condition is determined based on the fault reasoning output of the lightweight timing diagnostic model; If the inference output is a normal operating condition category, the operating condition is determined to be normal; if the inference output is a fault category, the operating condition is determined to be abnormal. Different upload frequencies are set according to whether the operating conditions are normal or abnormal; The edge device uploads the compact feature vector and the inference bias information to the cloud via a lightweight transmission protocol, according to the differentiated upload frequency, and verifies the integrity of the data uploaded to the cloud.

5. The method according to claim 1, characterized in that, The cloud receives data uploaded from the edge, performs data distribution drift detection based on the compact feature vector and inference bias information, and performs adaptive iterative optimization of the model according to the drift type and degree obtained from the drift detection, including: The cloud-based system uses two metrics, KL divergence and maximum mean difference, to construct a joint drift score, DriftScore. Based on the DriftScore, the degree of drift is divided into three categories: no drift, slow drift, and abrupt drift. When the drift level is zero, the current model remains unchanged. When the drift level is slow drift or abrupt drift, the corresponding incremental learning or transfer learning update strategy is triggered.

6. The method according to claim 5, characterized in that, When the drift is a slow drift or a sudden drift, a corresponding incremental learning or transfer learning update strategy is triggered, including: When the drift is slow, the model is locally updated using adaptive learning rate incremental learning, without full retraining. When the drift is a sudden drift, if there are enough samples, incremental learning with an increased learning rate is used to update the model parameters; if there are few samples and it is a new working condition, transfer learning is used to transfer general fault diagnosis knowledge to the new scenario, and only the high-level feature layer and classification layer are fine-tuned to achieve rapid adaptation.

7. The method according to claim 1, characterized in that, The cloud performs secondary lightweighting of the optimized model based on computing power matching and distributes it to the edge. The edge performs model upgrades via hot replacement without service interruption, including: Based on the computing power constraints of the target edge, the cloud performs a computing power matching secondary lightweighting process on the iteratively optimized model to obtain an updated model adapted to the edge. The cloud distributes the updated model to the corresponding edge device; the edge device uses a hot-swap mechanism to load the updated model, completing the model version upgrade without interrupting the fault diagnosis service, thus realizing cloud-edge collaborative closed-loop fault diagnosis.

8. The method according to claim 7, characterized in that, The cloud distributes the updated model to the corresponding edge device; the edge device uses a hot-swap mechanism to load the updated model, completing the model version upgrade without interrupting the fault diagnosis service, thus achieving cloud-edge collaborative closed-loop fault diagnosis, including: The edge device divides the memory into a running area and a backup cache area. The running area loads the current online diagnostic model, and the backup cache area receives and loads the updated model sent from the cloud. By atomically switching the model inference entry pointer, the inference call object is switched from the old model in the running area to the new model in the backup cache area, completing the seamless model upgrade and forming a complete adaptive closed loop.

9. The method according to any one of claims 1-8, characterized in that: The edge is a low-computing-power edge gateway based on ARM architecture, while the cloud is a high-performance computing cluster.

Images