An artificial intelligence edge computing terminal for industrial equipment fault diagnosis

Abstract

This invention discloses an artificial intelligence edge computing terminal for fault diagnosis of industrial equipment, belonging to the field of intelligent operation and maintenance technology for industrial equipment. In this invention, the multi-source heterogeneous data acquisition and synchronization module uses hardware-level timestamp synchronization to ensure strict alignment of multi-source data; the heterogeneous computing core module supports dynamic partitioning to achieve parallel inference and training; the hierarchical adaptive inference module dynamically schedules models of different complexities based on equipment health; the multimodal feature fusion module uses a cross-modal attention mechanism for weighted feature fusion; and the online model evolution module introduces an elastic weight consolidation algorithm to prevent catastrophic forgetting and combines it with knowledge distillation to achieve continuous model evolution. This invention can achieve low-latency, high-precision, and adaptive fault diagnosis at the edge, possessing extremely strong licensing stability and resilience against invalidation.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for industrial equipment, and in particular to an artificial intelligence edge computing terminal for fault diagnosis of industrial equipment. Background Technology

[0002] With the increasing complexity of industrial equipment and its growing automation, the economic losses and safety risks caused by equipment failures are also increasing. Traditional equipment maintenance models mainly include reactive maintenance and periodic preventive maintenance, both of which have their limitations. In recent years, predictive maintenance technology based on condition monitoring has received widespread attention and application. By continuously monitoring the operating status parameters of equipment, data analysis and machine learning methods are used to identify the health status and fault symptoms of the equipment, thereby providing early warnings and making maintenance decisions before failures occur.

[0003] Currently, industrial equipment fault diagnosis systems primarily employ a centralized cloud architecture. This involves deploying sensors on-site to collect data, uploading the data to a cloud server for storage and analysis, and then distributing the diagnostic results back to the field. This architecture suffers from problems such as high data transmission latency, significant network bandwidth pressure, high data security risks, and high cloud computing resource costs. Edge computing technology offers a new technical approach to address these issues by pushing computing and storage capabilities to the network edge, significantly reducing data transmission latency, minimizing bandwidth consumption, and improving data security.

[0004] However, deploying fault diagnosis algorithms to the edge still faces many challenges: edge devices have limited computing resources, making it difficult to run complex deep learning models; fault samples in industrial sites are scarce, making model training difficult; equipment operating conditions are highly variable, requiring continuous model updates to adapt to new fault modes; the fusion and processing of multi-source heterogeneous sensor data is complex, affecting diagnostic accuracy; in addition, there is a catastrophic forgetting problem during edge model updates, causing the model to forget old faults when learning new faults, seriously affecting the reliability of the diagnostic system. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence edge computing terminal for fault diagnosis of industrial equipment, which can achieve low-latency and high-precision fault diagnosis at the edge, while having adaptive computing resource scheduling, intelligent fusion of multimodal data, and online model update capabilities, thus solving the problem of catastrophic forgetting.

[0006] To achieve the above objectives, the present invention provides an artificial intelligence edge computing terminal for fault diagnosis of industrial equipment, comprising: The multi-source heterogeneous data acquisition and synchronization module is configured to acquire multi-source raw monitoring data of industrial equipment in a hardware synchronization manner. The raw monitoring data includes at least vibration signals, temperature data, sound signals and current waveform data. The heterogeneous computing core module includes an ARM processor core and a reconfigurable neural network computing unit. The reconfigurable neural network computing unit supports dynamic partitioning and can perform forward inference and backward propagation computations simultaneously. The hierarchical adaptive inference module runs on top of the heterogeneous computing core module. It is configured to dynamically select diagnostic models of different complexities for fault diagnosis based on the device's operating status. The hierarchical adaptive inference module includes a lightweight diagnostic submodule, a standard diagnostic submodule, and a deep diagnostic submodule with progressively increasing computational complexity. The multimodal feature fusion module is configured to map preprocessed multi-source monitoring data to a unified feature space and use a cross-modal attention mechanism to perform weighted feature fusion to generate a fused feature vector. An online model evolution module based on incremental learning is configured to update the diagnostic model locally at the edge. The online update includes an elastic weight consolidation algorithm to prevent catastrophic forgetting and a knowledge distillation mechanism to transfer knowledge from complex cloud models to lightweight edge models. The diagnostic results output module is configured to output the fault type, fault location, fault severity, and maintenance suggestions.

[0007] Preferably, the multi-source heterogeneous data acquisition and synchronization module includes a hardware-level timestamp synchronization unit; The hardware-level timestamp synchronization unit includes a phase-locked loop and a high-speed counter inside the FPGA. It is configured to send sampling trigger pulses to all data acquisition channels at the same time, so that the phase difference between multi-channel data with different sampling rates is less than 1 microsecond, and generates a strictly time-aligned multimodal fusion data frame.

[0008] Preferably, the reconfigurable neural network computing unit is a neural network accelerator based on a coarse-grained reconfigurable array; The reconfigurable neural network computing unit has a dynamic resource partitioning function, which can be configured to divide an independent computing island for performing backpropagation computation for incremental learning while performing regular inference tasks, thereby realizing a parallel working mode of inference and training.

[0009] Preferably, the hierarchical adaptive inference module further includes a state evaluation unit and a model scheduling unit; The status assessment unit is configured to calculate the device health index H based on the fused feature vector, and the calculation formula is as follows: ; in, This is a health indicator for the equipment, with a value range of [0, 1]. For the first Weight coefficients for each feature dimension; For the first The degree of deviation of each feature dimension from the normal benchmark. ; For the current eigenvalue, This is the normal baseline value. The standard deviation of this feature dimension; The model scheduling unit is configured to schedule based on device health indicators. Select the diagnostic submodule: When When the value exceeds the first preset threshold, the lightweight diagnostic submodule is scheduled; when When the threshold value is between the first and second preset thresholds, the standard diagnostic submodule is scheduled; when... When the value is less than the second preset threshold, the deep diagnostic submodule is scheduled.

[0010] Preferably, the multimodal feature fusion module includes: The modal feature extraction unit is configured to use corresponding feature extraction networks for different types of monitoring data. Specifically, a one-dimensional convolutional neural network is used to extract time-frequency features for vibration signals, a long short-term memory network is used to extract temporal evolution features for temperature data, Mel spectrum transform combined with a two-dimensional convolutional neural network is used to extract acoustic features for sound signals, and wavelet packet decomposition is used to extract energy features of each frequency band for current waveform data. The feature mapping unit is configured to map the features extracted from each modality to a unified feature space of the same dimension through a fully connected layer. The cross-modal attention unit is configured to calculate the correlation matrix between features of each modality and generate attention weights for each modality based on the correlation matrix. The formula for calculating the attention weights is as follows: ; in, For the first Attention weight vectors for each modality; For the first A query vector of modalities; The key vector matrix for all modes; The dimension of the key vector; and the feature fusion unit, configured to perform a weighted summation of the features of each modality based on the attention weights to generate a fused feature vector.

[0011] Preferably, the online model evolution module based on incremental learning includes: The concept drift detector is configured to monitor the KL divergence between the feature distribution of new input data and the historical baseline. When the KL divergence exceeds a preset drift threshold, the feature caching process is triggered. The feature buffer is configured to cache sample features marked as low confidence by the inference engine, as well as sample features triggered by the concept drift detector; The knowledge distillation unit is configured to retrieve the soft-label output of the teacher model from the cloud server and use the soft-label output as a supervision signal to guide the training of the student model at the edge. The loss function for knowledge distillation is: in, This is the total loss function; Cross-entropy loss based on real labels; The knowledge distillation loss is based on the soft labeling of the teacher model; This is the balance coefficient; The model update unit is configured to use the elastic weight consolidation algorithm to update model parameters. The loss function of the elastic weight consolidation algorithm is: ; in, The loss function for the current task; The diagonal element of the Fisher information matrix represents the first element. The importance of each parameter to the old task; For the current parameter; These are the optimal parameters learned from the old task; The hyperparameter controls the penalty strength for the importance of old tasks; and the version rollback controller is configured to retain the previous version model as a rollback node and monitor the average loss value of the new version model on multiple consecutive batches of data; if the average loss value increases by more than a safety threshold compared to the previous version model, it will automatically roll back to the previous version model.

[0012] Preferably, it also includes an anomaly detection module; The anomaly detection module is configured as follows: A feature distribution model under normal operating conditions is constructed, and the feature distribution model is represented by a Gaussian mixture model. Calculate the likelihood probability of the current monitoring data relative to the feature distribution model; When the likelihood probability is lower than the preset anomaly threshold, it is determined to be an abnormal state and the deep diagnosis submodule is triggered to perform fault diagnosis. At the same time, the abnormal sample and its feature vector are stored in the abnormal sample library for use by the online model evolution module.

[0013] Preferably, the diagnostic method includes the following steps: S1: Collects multi-source raw monitoring data of industrial equipment through a hardware synchronization method using a multi-source heterogeneous data acquisition and synchronization module; S2: The raw monitoring data is preprocessed by the data preprocessing module to obtain the preprocessed monitoring data; S3: The preprocessed monitoring data is subjected to feature extraction and cross-modal attention fusion through the multimodal feature fusion module to generate a fused feature vector; S4: The hierarchical adaptive inference module selects a diagnostic model of appropriate complexity based on the equipment's operating status, performs fault diagnosis on the fused feature vector, and obtains the diagnostic results; S5: An online model evolution module based on incremental learning updates the diagnostic model online using newly collected fault samples. This online update includes employing an elastic weight consolidation algorithm to prevent catastrophic forgetting and a knowledge distillation mechanism to transfer knowledge from complex cloud models to lightweight edge models; and S6: Outputs the fault type, fault location, fault severity, and maintenance suggestions through the diagnostic result output module.

[0014] Preferably, in step S5, updating the diagnostic model online further includes: When the confidence level of the diagnosis result is lower than the preset confidence threshold, the current sample is marked as a sample to be confirmed. When the concept drift detector detects that the KL divergence between the feature distribution of new input data and the historical baseline exceeds a preset drift threshold, feature caching is triggered. Incremental learning tasks are performed on independent computing islands partitioned by reconfigurable neural network computing units, based on samples accumulated in the feature buffer. After the model is updated, the version rollback controller monitors the average loss value of the new version model. If the loss value increases by more than a safety threshold compared to the previous version model, the rollback operation is automatically executed.

[0015] Preferably, the terminal is used in an industrial equipment fault diagnosis system, comprising: At least one AI edge computing terminal is deployed in an industrial site and configured to perform real-time fault diagnosis of industrial equipment; a cloud server is connected to the AI ​​edge computing terminal and configured to store historical diagnostic data, train complex diagnostic models, generate soft tags, and send model update parameters to the AI ​​edge computing terminal; and a monitoring terminal is connected to the AI ​​edge computing terminal and the cloud server and configured to display diagnostic results, receive alarm information, and remotely configure diagnostic parameters.

[0016] Therefore, the artificial intelligence edge computing terminal for fault diagnosis of industrial equipment using the above structure of the present invention has the following beneficial effects: (1) This invention achieves time alignment of multi-source data at the microsecond level through a hardware-level timestamp synchronization unit, laying a solid foundation for subsequent multimodal fusion and accurate fault diagnosis, and avoiding feature extraction errors caused by data asynchrony.

[0017] (2) The present invention adopts a heterogeneous architecture that combines ARM processor cores with reconfigurable neural network computing units, especially a coarse-grained reconfigurable array that supports dynamic partitioning, thereby realizing parallel execution of inference and training, significantly improving the utilization of computing resources and reducing power consumption.

[0018] (3) This invention is based on equipment health indicators Its hierarchical adaptive reasoning mechanism can dynamically select a diagnostic model of appropriate complexity based on the actual operating status of the device, thereby maximizing the conservation of computing resources and extending the terminal's lifespan while ensuring diagnostic accuracy.

[0019] (4) Through a cross-modal attention mechanism, the model can automatically learn the importance weights of each sensor data under different fault types, realize adaptive weighted fusion of features, and significantly improve diagnostic robustness and accuracy. Furthermore, the online model evolution module introduces an elastic weight consolidation algorithm, which effectively prevents catastrophic forgetting by constraining the update magnitude of important parameters during incremental learning; at the same time, combined with knowledge distillation, the knowledge of the large cloud model is transferred to the lightweight edge model, enabling the terminal to continuously evolve without losing historical experience.

[0020] (5) In this invention, the version rollback controller monitors the performance of the new model in real time, and automatically rolls back to the stable version once a performance degradation is detected; the concept drift detector actively senses changes in operating conditions and triggers model updates in advance to avoid diagnostic failures; the anomaly detection module can identify unknown fault types and provide data support for model evolution. The edge terminal is responsible for real-time diagnosis and rapid response, the cloud terminal is responsible for complex model training and knowledge generation, and the monitoring terminal provides visual interaction and remote management, forming a three-in-one intelligent operation and maintenance closed loop.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 Structural block diagram of an artificial intelligence edge computing terminal; Figure 2 Workflow diagram of the hierarchical adaptive inference module; Figure 3 A schematic diagram of the structure of the multimodal feature fusion module; Figure 4 A diagram illustrating the working principle of an online model evolution module based on incremental learning. Figure 5 Architecture diagram of industrial equipment fault diagnosis system. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0025] Example 1 like Figure 1 As shown, the AI ​​edge computing terminal provided in this embodiment includes a multi-source heterogeneous data acquisition and synchronization module, a data preprocessing module, a heterogeneous computing core module, a hierarchical adaptive inference module, a multimodal feature fusion module, an online model evolution module based on incremental learning, a diagnostic result output module, and an anomaly detection module. The modules communicate with each other via an internal bus or high-speed interface to collaboratively complete the fault diagnosis task of industrial equipment.

[0026] The multi-source heterogeneous data acquisition and synchronization module is configured to acquire multi-source raw monitoring data from industrial equipment in a hardware synchronous manner. This module integrates interfaces for various sensor types, including an IEPE interface for connecting to an accelerometer to acquire vibration signals, a PT100 / thermocouple interface for acquiring temperature data, an ultrasonic sensor interface for acquiring high-frequency acoustic emission signals, and a Hall current sensor interface for acquiring three-phase current waveforms from a motor.

[0027] Furthermore, the core of this module lies in its hardware-level timestamp synchronization unit, which includes a phase-locked loop (PLL) and a high-speed counter within the FPGA. The PLL locks the system clock, and the high-speed counter generates a synchronization trigger pulse. This pulse simultaneously reaches the analog-to-digital converters of all data acquisition channels via hardware circuitry, ensuring that the phase difference between multi-channel data at different sampling rates is less than 1 microsecond, generating strictly time-aligned multimodal fusion data frames. In addition, this module integrates a sensor health self-check unit, which periodically sends test excitation signals to the connected sensors and receives echoes to determine whether the sensors have failed or their accuracy has decreased.

[0028] The data preprocessing module is connected to the multi-source heterogeneous data acquisition and synchronization module and is configured to preprocess the raw monitoring data, including data cleaning, normalization, and time alignment. The data cleaning step removes baseline deviations caused by sensor drift and uses moving average filtering to remove high-frequency noise; the normalization process converts various types of data to a uniform numerical range; since the hardware synchronization module has already achieved strict start time alignment, the time alignment step mainly performs linear interpolation or spline interpolation to match the uniform sampling rate required for subsequent feature extraction.

[0029] The heterogeneous computing core module adopts a heterogeneous architecture combining an ARM processor core and a reconfigurable neural network computing unit. The ARM processor core is responsible for running a lightweight embedded operating system, handling industrial protocol conversion, and executing non-computationally intensive control logic. The reconfigurable neural network computing unit is a neural network accelerator based on a coarse-grained reconfigurable array, featuring dynamic resource partitioning. While performing regular inference tasks, it can allocate independent computing islands for incremental learning backpropagation computation, achieving parallel operation of inference and training. The entire module employs a fan-out fanless cooling structure, suitable for high-temperature industrial environments.

[0030] The hierarchical adaptive inference module runs on top of the heterogeneous computing core module and is configured to dynamically select diagnostic models of varying complexity for fault diagnosis based on the device's operating status. This module includes a lightweight diagnostic submodule, a standard diagnostic submodule, a deep diagnostic submodule, a status assessment unit, and a model scheduling unit. The lightweight submodule uses a lightweight neural network with fewer than 1M parameters and can complete a diagnostic inference within 10ms; the standard submodule uses a neural network with between 1M and 10M parameters and can complete a diagnosis within 50ms; the deep submodule uses a deep neural network with more than 10M parameters and enables a multi-model integrated diagnostic mechanism, enabling a diagnosis within 200ms.

[0031] The status assessment unit calculates the equipment health index based on the fused feature vector. Its calculation formula is ,in This represents the degree of deviation of the feature dimension from the normal baseline. The model scheduling unit selects the diagnostic submodule based on the H value: when When the value exceeds a first preset threshold (e.g., 0.8), the lightweight submodule is scheduled; when... When the value is between 0.5 and 0.8, the standard submodule is scheduled; when... When the value is less than 0.5, the scheduling depth submodule is used. This adaptive scheduling mechanism can rationally allocate computing resources based on the device's operating status.

[0032] The multimodal feature fusion module is configured to map preprocessed multi-source monitoring data to a unified feature space and employ a cross-modal attention mechanism for weighted feature fusion. This module includes a modal feature extraction unit, a feature mapping unit, a cross-modal attention unit, and a feature fusion unit. The modal feature extraction unit uses appropriate networks for different data types: vibration signals use a one-dimensional convolutional neural network to extract time-frequency features; temperature data uses a long short-term memory network to extract temporal evolution features; sound signals use Mel-spectrum transform combined with a two-dimensional convolutional neural network to extract acoustic features; and current waveform data uses wavelet packet decomposition to extract energy features for each frequency band. The feature mapping unit maps each modal feature to a unified feature space of the same dimension through a fully connected layer. The cross-modal attention unit calculates the correlation matrix between each modal feature and generates attention weights. The feature fusion unit performs a weighted summation of the features from each modality based on attention weights to generate a fused feature vector. .

[0033] One of the core innovations is the online model evolution module based on incremental learning, configured to update the diagnostic model locally at the edge while overcoming the catastrophic forgetting problem. This module includes a concept drift detector, a feature buffer, a knowledge distillation unit, a model update unit, and a version rollback controller. The concept drift detector monitors the KL divergence between the feature distribution of new input data and the historical baseline, triggering feature caching when the divergence exceeds a preset drift threshold. The feature buffer caches low-confidence samples and samples triggered by concept drift. The knowledge distillation unit obtains the soft-label output of the teacher model from the cloud server and uses the soft labels as a supervision signal to guide the training of the student model at the edge, with the loss function being: ; The model update unit uses the elastic weight consolidation algorithm for parameter updates, and the loss function is: ; in, These are the diagonal elements of the Fisher information matrix, representing the importance of parameters to the old task. The version rollback controller retains the previous version model as a rollback node, monitors the average loss value of the new version model on multiple consecutive batches of data, and automatically rolls back if the loss value exceeds a safety threshold.

[0034] The diagnostic results output module is configured to output fault type, fault location, fault severity, and maintenance recommendations. Fault types include bearing faults, gear faults, motor faults, etc.; fault location is determined by a localization algorithm; fault severity is divided into three levels: minor, moderate, and severe; maintenance recommendations are generated based on the fault type and severity.

[0035] The anomaly detection module is configured to identify unknown types of abnormal states. This module first constructs a feature distribution model under normal operating conditions, represented by a Gaussian mixture model; calculates the likelihood probability of the current monitoring data relative to this model; when the likelihood probability is lower than a preset anomaly threshold, it is determined to be an abnormal state and triggers the deep diagnosis submodule to perform fault diagnosis. At the same time, the abnormal samples and their feature vectors are stored in the abnormal sample library for use by the online model evolution module.

[0036] Example 2 like Figures 2 to 4 As shown, the industrial equipment fault diagnosis method provided in this embodiment is implemented using the aforementioned artificial intelligence edge computing terminal, and includes the following steps: Step S1: Collect multi-source raw monitoring data of industrial equipment through a hardware synchronization method using a multi-source heterogeneous data acquisition and synchronization module. Taking the spindle of a certain type of CNC machine tool as the diagnostic object, collect vibration acceleration signal (sampling frequency 25.6kHz), spindle bearing temperature data (sampling frequency 1Hz), spindle motor current waveform (sampling frequency 10kHz), and spindle running sound signal (sampling frequency 44.1kHz).

[0037] Step S2: The raw monitoring data is preprocessed by the data preprocessing module, including detrending, filtering, normalization and time series alignment.

[0038] Step S3: The preprocessed monitoring data is subjected to feature extraction and cross-modal attention fusion through a multimodal feature fusion module to generate a 256-dimensional fused feature vector. Specifically, frequency domain features are extracted from vibration signals, time-series statistical features are extracted from temperature data, Mel-frequency cepstral coefficients are extracted from sound signals, wavelet packet decomposition is performed on current waveforms, and then a multi-head attention mechanism is used for deep fusion.

[0039] Step S4: The hierarchical adaptive inference module selects a diagnostic model of appropriate complexity based on the equipment's operating status for fault diagnosis. Assuming the current health index H is 0.65, which is between 0.5 and 0.8, the standard diagnostic submodule is scheduled, and the output diagnostic results are: Fault type - early wear of the bearing outer ring, Fault location - rear bearing of the spindle, Fault severity - minor, Diagnostic confidence - 0.82.

[0040] Step S5: The diagnostic model is updated online based on newly collected fault samples using an online model evolution module based on incremental learning. Since the diagnostic confidence is high (0.82 > 0.7), the sample is not considered a pending confirmation sample, but the concept drift detector detects that the KL divergence is close to the threshold, so the sample's features are cached. When the cumulative number of newly added samples reaches the preset batch size, the model update unit triggers the model update process, performing an incremental learning task on the independent computing island of the reconfigurable neural network computing unit, using knowledge distillation and elastic weight consolidation algorithms to update the model parameters. After the update, the version rollback controller monitors the average loss value of the new version model on the subsequent three batches of data. If it does not exceed the safety threshold, the new model is retained.

[0041] Step S6: Output the fault type, fault location, fault severity, and maintenance suggestions through the diagnostic result output module.

[0042] Example 3 like Figure 5 As shown, the industrial equipment fault diagnosis system provided in this embodiment includes: multiple artificial intelligence edge computing terminals deployed next to different equipment in the industrial site; a cloud server that communicates with each artificial intelligence edge computing terminal via an industrial Ethernet or 5G network; and a monitoring terminal that communicates with both the artificial intelligence edge computing terminals and the cloud server.

[0043] The AI ​​edge computing terminal is responsible for real-time fault diagnosis of industrial equipment, achieving millisecond-level diagnostic responses, and periodically uploading diagnostic results and key monitoring data to the cloud server. The cloud server is responsible for storing historical diagnostic data, training complex diagnostic models, generating soft tags, and distributing model update parameters to the AI ​​edge computing terminal. The monitoring terminal can be an industrial computer, tablet, or smartphone, providing a graphical human-machine interface that displays the real-time health status of each device, historical trend curves, fault statistics reports, and other information, and allows for remote configuration of diagnostic parameters.

[0044] The system architecture in this embodiment implements an edge-cloud collaborative fault diagnosis mode: the edge is responsible for real-time diagnosis and rapid response, the cloud is responsible for data storage and model training, and the monitoring end is responsible for result display and remote management.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An artificial intelligence edge computing terminal for fault diagnosis of industrial equipment, characterized in that, include: The multi-source heterogeneous data acquisition and synchronization module is configured to acquire multi-source raw monitoring data of industrial equipment in a hardware synchronization manner. The raw monitoring data includes at least vibration signals, temperature data, sound signals and current waveform data. The heterogeneous computing core module includes an ARM processor core and a reconfigurable neural network computing unit. The reconfigurable neural network computing unit supports dynamic partitioning and can perform forward inference and backward propagation computations simultaneously. The hierarchical adaptive inference module runs on top of the heterogeneous computing core module. It is configured to dynamically select diagnostic models of different complexities for fault diagnosis based on the device's operating status. The hierarchical adaptive inference module includes a lightweight diagnostic submodule, a standard diagnostic submodule, and a deep diagnostic submodule with progressively increasing computational complexity. The multimodal feature fusion module is configured to map preprocessed multi-source monitoring data to a unified feature space and use a cross-modal attention mechanism to perform weighted feature fusion to generate a fused feature vector. The online model evolution module based on incremental learning is configured to update the diagnostic model locally at the edge. The online update includes preventing catastrophic forgetting based on the elastic weight consolidation algorithm and using a knowledge distillation mechanism to transfer the knowledge of the complex model in the cloud to the lightweight model at the edge. as well as The diagnostic results output module is configured to output the fault type, fault location, fault severity, and maintenance suggestions.

2. The artificial intelligence edge computing terminal for fault diagnosis of industrial equipment according to claim 1, characterized in that, The multi-source heterogeneous data acquisition and synchronization module includes a hardware-level timestamp synchronization unit; The hardware-level timestamp synchronization unit includes a phase-locked loop and a high-speed counter inside the FPGA. It is configured to send sampling trigger pulses to all data acquisition channels at the same time, so that the phase difference between multi-channel data with different sampling rates is less than 1 microsecond, and generates a strictly time-aligned multimodal fusion data frame.

3. The artificial intelligence edge computing terminal for fault diagnosis of industrial equipment according to claim 1, characterized in that, The reconfigurable neural network computing unit is a neural network accelerator based on a coarse-grained reconfigurable array; The reconfigurable neural network computing unit has a dynamic resource partitioning function, which can be configured to divide an independent computing island for performing backpropagation computation for incremental learning while performing regular inference tasks, thereby realizing a parallel working mode of inference and training.

4. The artificial intelligence edge computing terminal for fault diagnosis of industrial equipment according to claim 1, characterized in that, The hierarchical adaptive inference module also includes a state evaluation unit and a model scheduling unit; The status assessment unit is configured to calculate the device health index H based on the fused feature vector, and the calculation formula is as follows: ； in, This is a health indicator for the equipment, with a value range of [0, 1]. For the first Weight coefficients for each feature dimension; For the first The degree of deviation of each feature dimension from the normal benchmark. ； For the current eigenvalue, This is the normal baseline value. The standard deviation of this feature dimension; The model scheduling unit is configured to schedule based on device health indicators. Select the diagnostic submodule: When When the value exceeds the first preset threshold, the lightweight diagnostic submodule is scheduled; when When the threshold value is between the first and second preset thresholds, the standard diagnostic submodule is scheduled; when... When the value is less than the second preset threshold, the deep diagnostic submodule is scheduled.

5. An artificial intelligence edge computing terminal for fault diagnosis of industrial equipment according to claim 1, characterized in that, The multimodal feature fusion module includes: The modal feature extraction unit is configured to use corresponding feature extraction networks for different types of monitoring data. Specifically, a one-dimensional convolutional neural network is used to extract time-frequency features for vibration signals, a long short-term memory network is used to extract temporal evolution features for temperature data, Mel spectrum transform combined with a two-dimensional convolutional neural network is used to extract acoustic features for sound signals, and wavelet packet decomposition is used to extract energy features of each frequency band for current waveform data. The feature mapping unit is configured to map the features extracted from each modality to a unified feature space of the same dimension through a fully connected layer. The cross-modal attention unit is configured to calculate the correlation matrix between features of each modality and generate attention weights for each modality based on the correlation matrix. The formula for calculating the attention weights is as follows: ； in, For the first Attention weight vectors for each modality; For the first A query vector of modalities; The key vector matrix for all modes; The dimension of the key vector; and the feature fusion unit, configured to perform a weighted summation of the features of each modality based on the attention weights to generate a fused feature vector.

6. An artificial intelligence edge computing terminal for fault diagnosis of industrial equipment according to claim 1, characterized in that, The online model evolution module based on incremental learning includes: The concept drift detector is configured to monitor the KL divergence between the feature distribution of new input data and the historical baseline. When the KL divergence exceeds a preset drift threshold, the feature caching process is triggered. The feature buffer is configured to cache sample features marked as low confidence by the inference engine, as well as sample features triggered by the concept drift detector; The knowledge distillation unit is configured to retrieve the soft-label output of the teacher model from the cloud server and use the soft-label output as a supervision signal to guide the training of the student model at the edge. The loss function for knowledge distillation is: ； in, This is the total loss function; Cross-entropy loss based on real labels; The knowledge distillation loss is based on the soft labeling of the teacher model; This is the balance coefficient; The model update unit is configured to use the elastic weight consolidation algorithm to update model parameters. The loss function of the elastic weight consolidation algorithm is: ； in, The loss function for the current task; The diagonal element of the Fisher information matrix represents the first element. The importance of each parameter to the old task; For the current parameter; These are the optimal parameters learned from the old task; The hyperparameter controls the penalty strength for the importance of old tasks; and the version rollback controller is configured to retain the previous version model as a rollback node and monitor the average loss value of the new version model on multiple consecutive batches of data; if the average loss value increases by more than a safety threshold compared to the previous version model, it will automatically roll back to the previous version model.

7. An artificial intelligence edge computing terminal for fault diagnosis of industrial equipment according to claim 1, characterized in that, It also includes an anomaly detection module; The anomaly detection module is configured as follows: A feature distribution model under normal operating conditions is constructed, and the feature distribution model is represented by a Gaussian mixture model. Calculate the likelihood probability of the current monitoring data relative to the feature distribution model; When the likelihood probability is lower than the preset anomaly threshold, it is determined to be an abnormal state and the deep diagnosis submodule is triggered to perform fault diagnosis. At the same time, the abnormal sample and its feature vector are stored in the abnormal sample library for use by the online model evolution module.

8. An artificial intelligence edge computing terminal for fault diagnosis of industrial equipment according to claim 1, characterized in that, Its diagnostic method includes the following steps: S1: Collects multi-source raw monitoring data of industrial equipment through a hardware synchronization method using a multi-source heterogeneous data acquisition and synchronization module; S2: The raw monitoring data is preprocessed by the data preprocessing module to obtain the preprocessed monitoring data; S3: The preprocessed monitoring data is subjected to feature extraction and cross-modal attention fusion through the multimodal feature fusion module to generate a fused feature vector; S4: The hierarchical adaptive inference module selects a diagnostic model of appropriate complexity based on the equipment's operating status, performs fault diagnosis on the fused feature vector, and obtains the diagnostic results; S5: An online model evolution module based on incremental learning updates the diagnostic model online using newly collected fault samples. This online update includes employing an elastic weight consolidation algorithm to prevent catastrophic forgetting and a knowledge distillation mechanism to transfer knowledge from complex cloud models to lightweight edge models; and S6: Outputs the fault type, fault location, fault severity, and maintenance suggestions through the diagnostic result output module.

9. An artificial intelligence edge computing terminal for fault diagnosis of industrial equipment according to claim 8, characterized in that, In step S5, updating the diagnostic model online also includes: When the confidence level of the diagnosis result is lower than the preset confidence threshold, the current sample is marked as a sample to be confirmed. When the concept drift detector detects that the KL divergence between the feature distribution of new input data and the historical baseline exceeds a preset drift threshold, feature caching is triggered. Incremental learning tasks are performed on independent computing islands partitioned by reconfigurable neural network computing units, based on samples accumulated in the feature buffer. After the model is updated, the version rollback controller monitors the average loss value of the new version model. If the loss value increases by more than a safety threshold compared to the previous version model, the rollback operation is automatically executed.

10. An artificial intelligence edge computing terminal for fault diagnosis of industrial equipment according to claim 1, characterized in that, This terminal is used in an industrial equipment fault diagnosis system, including: At least one AI edge computing terminal is deployed in an industrial site and configured to perform real-time fault diagnosis of industrial equipment; a cloud server is connected to the AI ​​edge computing terminal and configured to store historical diagnostic data, train complex diagnostic models, generate soft tags, and send model update parameters to the AI ​​edge computing terminal; and a monitoring terminal is connected to the AI ​​edge computing terminal and the cloud server and configured to display diagnostic results, receive alarm information, and remotely configure diagnostic parameters.

Images