Equipment state prediction method and device, equipment and readable storage medium

Abstract

The invention discloses an equipment state prediction method and device, equipment and a readable storage medium, and is applied to the technical field of internet equipment management, and the method comprises the steps: determining a feature vector corresponding to each type of equipment data based on a data modal characteristic and business demand coupling principle, and carrying out the alignment of each feature vector, and obtaining an equipment state feature vector; uploading the equipment state feature vector to a server, so that the server updates an equipment state prediction model by utilizing cross-equipment group knowledge on the premise of not sharing equipment data; predicting by utilizing a latest equipment state prediction model based on the equipment state feature vector, and determining an equipment abnormal state; the latest equipment state prediction model is a model for global updating and local incremental updating of the edge node based on the server. According to the method, the information of each mode is deeply associated, global updating and local incremental updating are carried out on the model based on the server, the latest model is used for prediction, and the accuracy of state prediction is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of Internet device management technology, and in particular to a device status prediction method, apparatus, device, and readable storage medium. Background Technology

[0002] Currently, when predicting equipment status, lightweight convolutional neural networks are deployed on the device to process single-type sensor data (such as vibration signals), extract preliminary feature vectors, and upload them to the cloud. However, there is a problem of cross-modal spatiotemporal correlation breaks, which makes the accuracy of equipment status prediction low.

[0003] Therefore, how to improve the accuracy of equipment status prediction is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method, apparatus, device and readable storage medium for predicting equipment status, which solves the problem of low accuracy in equipment status prediction in the prior art.

[0005] To address the aforementioned technical problems, this invention provides a method for predicting equipment status, comprising:

[0006] Obtain device data of a preset type for the target device, determine the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, and align each feature vector to obtain the device status feature vector;

[0007] Based on the federated learning framework, the device state feature vector is uploaded to the server so that the server can update the device state prediction model using cross-device group knowledge without sharing device data.

[0008] Based on the device status feature vector, the latest device status prediction model is used to predict and determine the abnormal state of the device; wherein, the latest device status prediction model is a model based on global updates of the server and local incremental updates of edge nodes.

[0009] Optionally, acquire device data of a preset type for the target device, determine the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, and align each feature vector to obtain a device status feature vector, including:

[0010] Extract the frequency domain feature vectors from vibration signals and current waveform data;

[0011] Extract the trend feature vector from temperature time series data;

[0012] The operational logic relationships in the device control log data are modeled using a graph neural network to obtain graph embedding feature vectors; wherein, the preset types of device data include the vibration signal, the current waveform data, the temperature time series data, and the device control log data;

[0013] The frequency domain feature vector, the trend feature vector, and the graph embedding feature vector are compressed and fused to obtain the device state feature vector.

[0014] Optionally, the device state feature vector is uploaded to the server based on a federated learning framework, so that the server can update the device state prediction model using cross-device group knowledge without sharing device data, including:

[0015] Random noise is added to the device state feature vector, and the noise-added feature vector is uploaded to the server through an encrypted channel. The server dynamically selects the device state feature vectors corresponding to the edge nodes participating in the aggregation based on the similarity of device operating conditions and data quality. After aggregation, the device state prediction model is updated. During the update process, the knowledge graph constructed from historical failure cases is embedded into the model as a knowledge enhancement mechanism as a regularization term.

[0016] Optionally, after determining the abnormal state of the equipment by predicting using the latest equipment state prediction model based on the equipment state feature vector, the method further includes:

[0017] Determine whether the probability of the abnormal state of the device exceeds a preset abnormal threshold;

[0018] When the anomaly probability exceeds the preset anomaly probability threshold, the fault type, fault cause, and maintenance strategy are output based on the multi-level fault tracing mechanism.

[0019] Optionally, when the anomaly probability exceeds the preset anomaly probability threshold, the fault type, fault cause, and maintenance strategy are output based on a multi-level fault tracing mechanism, including:

[0020] The fault type is determined using a fault mode matching library based on the abnormal state of the equipment.

[0021] The cause of the failure is determined by analyzing the dependencies between abnormal features related to the failure type using a causal reasoning engine.

[0022] The digital twin simulation module is invoked to verify the cause of the fault and to generate the maintenance strategy containing operating procedures.

[0023] Optionally, the latest equipment status prediction model includes a temporal convolutional network and a spatiotemporal graph neural network. Based on the equipment status feature vector, the latest equipment status prediction model is used to predict and determine the abnormal equipment status, including:

[0024] For short-term prediction, the temporal convolutional network is used to analyze the evolution law of individual device states to obtain short-term abnormal device states; wherein, the evolution law of individual device states is the inherent pattern and trend of the device's own health status parameters changing over time; the temporal convolutional network adopts a set number of dilated convolutional layers and a sequence of dilation factors;

[0025] The spatiotemporal neural network is used to model the synergistic effect between equipment clusters for setting long-term predictions, thereby obtaining long-term abnormal equipment states. The synergistic effect between equipment clusters refers to the mutual influence between multiple devices in a production system or production line due to physical connections, process flows, or control logic. The time period for setting short-term predictions is shorter than the time period for setting long-term predictions.

[0026] Optionally, after determining the abnormal state of the equipment by predicting using the latest equipment state prediction model based on the equipment state feature vector, the method further includes:

[0027] The system monitors changes in the operating conditions of target equipment in real time. When a distribution drift is detected, it uses the latest equipment data to locally retrain the key layers of the model and dynamically compresses the size of the equipment condition prediction model to obtain an incrementally optimized equipment condition prediction model.

[0028] This invention also provides a device for predicting device status, comprising:

[0029] The device status feature vector determination module is used to obtain device data of a preset type of target device, determine the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, and align each feature vector to obtain the device status feature vector.

[0030] The global update module is used to upload the device state feature vector to the server based on the federated learning framework, so that the server can update the device state prediction model using cross-device group knowledge without sharing device data.

[0031] The device abnormal state determination module is used to predict and determine the device abnormal state based on the device state feature vector using the latest device state prediction model; wherein, the latest device state prediction model is a model based on global updates of the server and local incremental updates of edge nodes.

[0032] This invention also provides a device state prediction device, comprising:

[0033] Memory, used to store computer programs;

[0034] A processor for executing the computer program to implement the steps of the device state prediction method described above.

[0035] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the device state prediction method described above.

[0036] The present invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the above-described device state prediction method.

[0037] As can be seen, this invention acquires device data of preset types from the target device, determines the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, aligns each feature vector to obtain a device state feature vector, uploads the device state feature vector to the server based on a federated learning framework, so that the server can update the device state prediction model using cross-device group knowledge without sharing device data, and makes predictions based on the latest device state prediction model using the device state feature vector to determine the abnormal state of the device. The latest device state prediction model is a model based on global updates from the server and incremental updates from local edge nodes. This invention extracts feature vectors from device data based on the coupling principle of data modality characteristics and business requirements, thereby deeply integrating the spatiotemporal information of various modalities, eliminating feature omissions at the source, and simultaneously performing global updates to the model based on the server and incremental updates based on local nodes, using the latest model for prediction, significantly improving the accuracy of state prediction.

[0038] In addition, the present invention also provides a device state prediction apparatus, device, and readable storage medium, which also have the above-mentioned beneficial effects. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0040] Figure 1 A flowchart of a device status prediction method provided in an embodiment of the present invention;

[0041] Figure 2 A flowchart of a device status prediction and fault diagnosis method provided in an embodiment of the present invention;

[0042] Figure 3This is a schematic diagram illustrating feature extraction based on a lightweight feature fusion model, provided as an embodiment of the present invention.

[0043] Figure 4 A schematic diagram of a federated learning framework provided in an embodiment of the present invention;

[0044] Figure 5 A schematic diagram illustrating a maintenance strategy tracing mechanism provided in an embodiment of the present invention;

[0045] Figure 6 This is a schematic diagram of the structure of the device status prediction device provided in an embodiment of the present invention;

[0046] Figure 7 This is a schematic diagram of the structure of a device state prediction device provided in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Some terms that appear in the description of the embodiments of this application are subject to the following interpretation:

[0049] Multi-source heterogeneous data: Industrial equipment simultaneously generates vibration waveforms, temperature curves, current signals, and control logs, which have different physical dimensions, sampling frequencies, and protocol formats.

[0050] Edge multimodal fusion: A technique for achieving spatiotemporal alignment and feature extraction of multi-source data at edge computing nodes through parallel neural networks (CNN (convolutional neural network) for vibration / current processing + Attention-LSTM (long short-term memory network with attention mechanism) for temperature analysis + GNN (graph neural network) for log modeling).

[0051] Federated knowledge enhancement: When aggregating feature vectors of edge devices in the cloud, a knowledge-sharing mechanism is established by embedding historical fault knowledge graphs as regularization terms into model training, under the premise of privacy protection (using differential privacy technology with added noise).

[0052] Adaptive edge optimization: Based on KL divergence (which is the difference in information (i.e., relative entropy) generated when approximating the true distribution Q with a distribution P), it detects working condition drift and achieves continuous adaptation of the edge model through incremental fine-tuning (local retraining of key layers) and model pruning (dynamically compressing parameters).

[0053] Please refer to Figure 1 , Figure 1 This invention provides a method for predicting device status. The method may include:

[0054] S101, Obtain device data of preset types of target devices, determine the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, and align each feature vector to obtain the device status feature vector.

[0055] In this embodiment, the execution entity for each step is the edge node, which provides services to physical devices and sensors. The target device in this embodiment is the device whose device status prediction is required by the current edge node. It can be understood that multi-source heterogeneous sensor data, consisting of preset types of device data (e.g., vibration signals, temperature time-series data, current waveforms, and device control logs), is collected in real-time by edge nodes deployed in the industrial field. The principle of coupling data modal characteristics and business requirements in this embodiment to determine the feature vector corresponding to each type of device data means that the feature extraction method corresponding to each type of device data can be determined based on the coupling principle of data modal characteristics and business requirements. The collection of preset types of device data is implemented through an industrial protocol adaptive parser. This parser supports plug-and-play conversion between OPC UA protocol (an architecture protocol for process control), Modbus protocol (the oldest, simplest, and most commonly used field-level communication protocol in the industrial field), and MQTT protocol (message queue telemetry transport protocol), and automatically handles clock synchronization for data with different sampling frequencies.

[0056] Different data use different parsing protocols. The core function of an industrial protocol adaptive parser is to handle this complexity of "different data using different protocols." By unifying all data to the same time and format standard, it provides clean, standardized, and usable input data for subsequent feature fusion models.

[0057] It should be further explained that, based on any of the above embodiments, the process of obtaining device data of a preset type from the target device, determining the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, and aligning each feature vector to obtain a device state feature vector may include:

[0058] S1011, extract the frequency domain feature vector of vibration signal and current waveform data;

[0059] S1012, extract the trend feature vector of temperature time series data;

[0060] S1013, The operation logic relationship in the device control log data is modeled by a graph neural network to obtain a graph embedding feature vector; wherein, the preset types of device data include vibration signal, current waveform data, temperature time series data and device control log data;

[0061] S1014 compresses and fuses the frequency domain feature vector, trend feature vector, and graph embedding feature vector to obtain the device status feature vector.

[0062] In this embodiment, a lightweight feature fusion model can be used at edge nodes to perform spatiotemporal alignment and feature extraction on preset types of device data to generate a unified-dimensional device state feature vector. When extracting temperature time-series data, a one-dimensional convolutional neural network can be used to process vibration signals and current waveform data. A temporal attention mechanism can be used to extract long-term dependent features from the temperature time-series data. A graph neural network is used to model the operational logic relationships in the device control logs, and a feature compression layer is used to map the extracted features to a low-dimensional vector space to form the device state feature vector. It can be understood that after processing by the lightweight feature fusion model, the preset types of device data ultimately yield a 128-dimensional device state feature vector. This vector integrates deep features from multiple sources such as vibration, current, temperature, and control logs, forming a unified, machine-understandable "device health fingerprint." It is not limited to using one-dimensional convolutional neural networks (1D-CNN), but 1D-CNN is particularly suitable for processing waveform sequence data such as vibration and current, as it can effectively extract local frequency domain features and waveform patterns. Using other networks (such as 2D-CNN or RNN) may increase computational complexity or lose local details, which does not meet the lightweight requirements of edge devices. The operational logic relationships in the device control log are not feature vectors themselves, but graph embedding vectors extracted after modeling the instruction sequences in the control log using graph neural networks (GNN), and then concatenated and compressed with other features to form the final vector. Considering the constraints of edge computing resources, all models adopt a lightweight design (such as small convolutional kernels and low-dimensional embedding) to ensure that feature extraction is both sufficient and efficient.

[0063] S102, based on the federated learning framework, uploads the device state feature vector to the server so that the server can update the device state prediction model using cross-device group knowledge without sharing device data.

[0064] The federated learning framework in this embodiment includes: a differentiated client selection strategy that dynamically selects edge nodes participating in aggregation based on equipment operating condition similarity and data quality; a privacy protection mechanism that adds homomorphic encryption noise before feature vector uploading to achieve gradient obfuscation; and a knowledge enhancement mechanism that embeds a knowledge graph constructed from historical fault cases as a regularization term into the global model training. The core difference between the federated learning framework of this invention and existing traditional frameworks lies in its upgrade from simple aggregation to intelligent enhanced aggregation. Traditional frameworks typically select clients randomly or based on system resources (such as power consumption, network), and directly perform average aggregation on model parameters or gradients, lacking deep integration of industrial data characteristics and knowledge. This framework introduces a triple enhancement mechanism: dynamic edge node selection based on business semantics (rather than random selection), privacy protection by adding homomorphic noise (at the feature level rather than the raw data level), and embedding a historical fault knowledge graph as a regularization term into the training (achieving a combination of data-driven and knowledge-driven approaches), making it more suitable for multi-source heterogeneous, privacy-sensitive, and deeply diagnostic scenarios in the Industrial Internet. "Client" refers to each edge node participating in federated learning (i.e., data acquisition and processing nodes in the industrial field). "Selecting clients" refers to selecting a subset of devices from all edge devices during each round of federated learning aggregation, uploading their locally trained feature vectors (or model updates) to the cloud to participate in the global model aggregation. "Similarity" refers to the similarity of operating conditions between devices, such as: device type (e.g., all CNC machine tools); processing materials (e.g., all aluminum alloys); operating load, environmental conditions, etc. By calculating the cosine similarity between the feature vectors of devices in the client and comparing the cosine similarity with the corresponding similarity threshold (e.g., a similarity threshold of 0.7), it is determined whether the devices are in similar operating conditions. The reason for selecting devices with similar operating conditions is to avoid interference from distributional differences: industrial equipment varies greatly; if devices are randomly selected for aggregation, data from devices with large distributional differences will interfere with each other, leading to a decrease in the generalization ability of the global model. Selecting devices with similar operating conditions and high data quality (completeness > 95%) ensures that the feature vectors used for aggregation are within the same distribution range, improving the model's convergence speed and accuracy. This allows for the training of more accurate models for specific types / operating condition device clusters, avoiding the problem of insufficient generalization from a "one-size-fits-all" approach. The selected client (edge ​​node) uploads its encrypted feature vector to the cloud (server). The cloud then aggregates these feature vectors (e.g., using a weighted average) to update the global fault diagnosis model. In other words, data from similar devices participates in the training and knowledge sharing of the global model through feature vectors. The reason for dynamically selecting clients based on operating condition similarity and data quality is that industrial equipment varies greatly (e.g., new and old machine tools, equipment processing different materials), and their data distribution can be drastically different.If devices (clients) are randomly selected for aggregation, data from devices with vastly different distributions may interfere with each other, rendering the aggregated global model unsuitable and unable to focus on learning accurate fault modes under specific operating conditions. By selecting devices with similar operating conditions and high data quality, it is ensured that the feature vectors used for aggregation are within the same distribution range, thereby improving the effectiveness of aggregation and the stability of model convergence. It should be further noted that, based on any of the above embodiments, uploading device state feature vectors to the server using a federated learning framework, allowing the server to update the device state prediction model using cross-device group knowledge without sharing device data, can include: adding random noise to the device state feature vectors, uploading the noisy feature vectors to the server via an encrypted channel, and allowing the server to dynamically select the device state feature vectors corresponding to the edge nodes participating in aggregation based on device operating condition similarity and data quality, then updating the device state prediction model. During the update process, a knowledge graph constructed from historical fault cases is embedded as a regularization term in the model's knowledge enhancement mechanism. This embodiment adds homomorphic encryption noise before feature vector uploading to achieve a gradient obfuscation privacy protection mechanism. In this embodiment, device condition similarity refers to the similarity of device condition data corresponding to edge nodes. Data from similar devices participates in the training and knowledge sharing of the global model through feature vectors. This embodiment deeply integrates knowledge-driven and data-driven methods through a knowledge enhancement mechanism. The advantages of this approach are a significant improvement in the efficiency and effectiveness of federated learning. First, it improves the convergence speed and accuracy of the global model because aggregation is based on more relevant and higher-quality data sources, reducing bias caused by noise and distribution differences. Second, it enhances the model's specialization capability, enabling the training of more accurate models for device clusters of specific types or conditions, avoiding the problem of insufficient generalization from a one-size-fits-all approach. Finally, this strategy also optimizes the utilization of communication and computing resources, avoiding the waste of bandwidth and computing power caused by invalid or low-quality data participating in aggregation, making the entire federated learning process more efficient and economical.

[0065] S103, based on the device status feature vector, use the latest device status prediction model to predict and determine the abnormal status of the device; wherein, the latest device status prediction model is a model based on global updates from the server and local incremental updates from edge nodes.

[0066] The latest device status prediction model in this embodiment is based on a global update from the server and an incremental update from the edge nodes. The probability of device anomalies can be determined in this embodiment. This embodiment can also assess reliability by calculating the confidence interval of the prediction results (device anomaly status) and the Monte Carlo Dropout uncertainty index. When the uncertainty index exceeds a dynamic threshold, it automatically requests collaborative inference from the cloud model. Data transmission between the edge nodes and the server uses a breakpoint-resume compression protocol. This protocol dynamically adjusts the compression rate between 30% and 70% based on network bandwidth conditions and employs differential coding technology to reduce redundant transmission of consecutive data packets.

[0067] It should be further explained that, based on any of the above embodiments, after predicting and determining the abnormal state of the equipment using the latest equipment state prediction model based on the equipment state feature vector, the method may further include: determining whether the abnormal probability of the abnormal state of the equipment exceeds a preset abnormal threshold; when the abnormal probability exceeds the preset abnormal probability threshold, outputting the fault type, fault cause, and maintenance strategy based on a multi-level fault tracing mechanism. This embodiment does not limit the specific preset abnormal threshold; for example, the preset abnormal threshold in this embodiment can be 80%; or the preset abnormal threshold can also be 85%, etc. The multi-level fault tracing mechanism in this embodiment is a three-level fault tracing mechanism, with the first level used to determine the fault type, the second level used to determine the fault cause, and the third level used to determine the maintenance strategy. This embodiment does not limit the specific method for determining the fault type. For example, this embodiment can be based on a multi-classification model, where the obtained abnormal state of the equipment is directly output as a probability distribution of a fault type through a classifier such as a neural network, support vector machine, or gradient boosting tree. Alternatively, it can be implemented based on multi-model fusion and ensemble learning, where multiple dedicated classifiers can be trained and their results integrated through a meta-model or voting mechanism to improve accuracy and robustness. Alternatively, similarity matching can be used to calculate the similarity between the obtained abnormal state and the states in the historical fault case library (such as cosine similarity or Euclidean distance). The K most similar historical cases are found, and the fault type is the most likely fault type at present. This embodiment does not limit the specific method for determining the cause of the fault. For example, this embodiment can determine the cause based on knowledge graph reasoning; or this embodiment can analyze the sequence patterns of sensor data and event logs in the period before the fault occurred, and determine the cause of the fault based on the sequence model. This embodiment does not limit the specific method for determining the maintenance strategy. For example, this embodiment can determine the maintenance strategy based on the reasoning of cases; or this embodiment can determine the strategy based on a rule engine. This embodiment can upgrade from superficial diagnosis to root cause eradication through three-level tracing, outputting a comprehensive strategy such as "replacing the bearing + upgrading the voltage regulator", reducing the fault recurrence rate, reducing unplanned downtime, and improving the overall efficiency of the equipment.

[0068] It should be further explained that, based on any of the above embodiments, the output of fault type, fault cause, and maintenance strategy based on the multi-level fault tracing mechanism may include: determining the fault type based on the abnormal equipment state using a fault mode matching library; determining the fault cause by analyzing the dependency relationship between abnormal features related to the fault type through a causal reasoning engine; verifying the fault cause by calling a digital twin simulation module; and generating the maintenance strategy containing operating procedures. The multi-level fault tracing mechanism in this embodiment may include a three-level diagnostic process: the first level quickly identifies known fault types based on the fault mode matching library; the second level locates the root cause by analyzing the dependency relationship between abnormal features through a causal reasoning engine; and the third level calls a digital twin simulation module to verify the diagnostic results and generate a maintenance strategy containing spare parts replacement paths and operating procedures.

[0069] It should be further noted that, based on any of the above embodiments, the latest equipment state prediction model includes temporal convolutional networks and spatiotemporal graph neural networks. Based on the equipment state feature vector, the latest equipment state prediction model is used to predict and determine the abnormal state of the equipment, which may include:

[0070] S1031, for short-term prediction, a temporal convolutional network is used to analyze the evolution law of the state of individual equipment to obtain the short-term abnormal state of the equipment; wherein, the evolution law of the state of individual equipment is the inherent pattern and trend of the health status parameters of the equipment itself changing over time; the temporal convolutional network adopts a set number of dilated convolutional layers and a sequence of dilation factors.

[0071] S1032, for setting long-term prediction, the spatiotemporal graph neural network is used to model the synergistic effect between equipment clusters to obtain the long-term abnormal state of equipment; wherein, the synergistic effect between equipment clusters is the mutual influence between multiple equipment in a production system or production line due to physical connection, process flow or control logic association; wherein, the time period of setting short-term prediction is shorter than the time period of setting long-term prediction.

[0072] In this embodiment, a hybrid prediction model architecture is adopted for state prediction: for short-term predictions (e.g., within one hour), a temporal convolutional network is used to analyze the state evolution of individual devices, while for long-term predictions (e.g., more than 24 hours), a spatiotemporal graph neural network is used to model the synergistic effects between device clusters. The state evolution of an individual device refers to the inherent patterns and trends in the changes of its own health state parameters (such as vibration, temperature, and current) over time. These changes are mainly driven by the device's own operating process and internal physicochemical changes (such as wear, fatigue, and aging). The synergistic effect between device clusters refers to the mutual influence between multiple devices in a production system or production line due to physical connections, technological processes, or control logic. A change in the state of one device can act as an "input" or "disturbance," affecting the state of other connected devices. In this embodiment, the temporal convolutional network (TCN) uses a specific number of dilated convolutional layers (4 layers) and a dilation factor sequence ([1,2,4,8]) to capture specific periodic patterns of device state changes; the spatiotemporal graph neural network (STGNN) is specifically used to construct and parse the physical connection topology between device clusters to model the synergistic effects between machines. This strategy, which leverages existing robust architecture and employs refined adaptation and integration for specific industrial scenarios, ensures the model's sophistication and effectiveness.

[0073] It should be further explained that, based on any of the above embodiments, after predicting and determining the abnormal state of the equipment using the latest equipment state prediction model based on the equipment state feature vector, it may also include: real-time monitoring of changes in the operating condition distribution of the target equipment; when distribution drift is detected, using the latest equipment data to locally retrain the key layers of the model; and dynamically compressing the size of the equipment state prediction model to obtain an incrementally optimized equipment state prediction model. The method for determining distribution shift in this embodiment may include KL divergence, JS divergence (a metric derived from KL divergence that measures the similarity between two probability distributions. KL divergence is asymmetric, while JS divergence solves the asymmetry problem by calculating the average value of KL divergence, making the result symmetric and smooth, and its value is always between 0 and 1 (when using base-2 logarithms)), or Wasserstein distance (measuring the minimum amount of work required to "move" one probability distribution into another), or methods based on statistical hypothesis testing (such as the KS test (a nonparametric statistical test method)) to monitor changes in data distribution. For ease of understanding, the edge nodes in this embodiment can adaptively learn online. Examples include a drift detection unit that monitors real-time changes in equipment operating condition distribution based on KL divergence; an incremental fine-tuning unit that uses the latest data to locally retrain key layers of the model when distribution drift is detected; and a model pruning unit that dynamically compresses the edge model size to adapt to computational resource constraints. KL divergence, due to its high computational efficiency and sensitivity to differences in probability distribution, is more suitable for the real-time computational constraints of edge devices. Distribution drift typically refers to a significant deviation of the data distribution during equipment operation from the baseline distribution during training. This is specifically manifested as a continuous change in statistical characteristics (such as mean, variance, and distribution pattern). For example, an increase in ambient temperature may cause a 10°C shift in temperature sensor readings, or component wear may cause a continuous increase in the energy of the vibration spectrum in a specific frequency band. When this shift exceeds a threshold, it is determined to be a drift. The key layers mainly refer to the output layers and adjacent hidden layers in the model that are sensitive to changes in operating conditions, such as the fully connected classification layer and the feature compression layer before it. These layers are directly related to the semantic mapping of the equipment state. The bottom layers of the model (such as convolutional kernels) usually extract general features and are not easily affected by changes in operating conditions. Therefore, during incremental fine-tuning, only the last 20% of the layers (such as the last two fully connected layers) are retrained. This can quickly adapt to the new distribution and avoid destroying the existing knowledge structure of the model.

[0074] This invention provides a device status prediction method, which may include: S101, acquiring device data of a preset type for the target device, determining the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, and aligning each feature vector to obtain a device status feature vector; S102, uploading the device status feature vector to a server based on a federated learning framework, so that the server can update the device status prediction model using cross-device group knowledge without sharing device data; S103, making a prediction based on the device status feature vector using the latest device status prediction model to determine the abnormal state of the device; wherein, the latest device status prediction model is a model based on global updates from the server and incremental updates from local edge nodes. This invention deeply associates the spatiotemporal information of various modalities, eliminates feature omissions from the root, and simultaneously performs global updates to the model based on the server and incremental updates based on local nodes, using the latest model for prediction, significantly improving the accuracy of status prediction.

[0075] For a clearer understanding of this invention, please refer to the following details. Figure 2 , Figure 2 The flowchart of an equipment status prediction and fault diagnosis method provided in this embodiment of the invention specifically includes: real-time acquisition of multi-source heterogeneous sensor data, consisting of vibration signals, temperature time-series data, current waveforms, and equipment control logs of the target equipment (equipment sensors), through edge computing nodes deployed in industrial sites; spatiotemporal alignment and feature extraction of the multi-source heterogeneous data using a lightweight feature fusion model at the edge computing nodes to generate a unified-dimensional equipment status feature vector; uploading the feature vector to a cloud aggregation server based on a federated learning framework, and updating the global fault diagnosis model using cross-equipment group knowledge without sharing the original data. It should be noted that the prediction is continuous, using the latest local model, and model updates are periodic. The updated global model parameters are distributed to the edge nodes, and the local prediction model is incrementally optimized using an adaptive online learning module; real-time status prediction is performed based on the optimized edge model, and when the predicted anomaly probability exceeds a preset threshold, a multi-level fault tracing mechanism is triggered to output diagnostic results and maintenance strategies.

[0076] In this process, edge computing nodes (such as those using ARM Cortex-A72 processors, the core of the system's main control chip) are deployed in the industrial field. These nodes connect to the device sensor network via multi-protocol adapters to collect vibration accelerometer signals (10kHz sampling rate), infrared temperature sensor time-series data (1Hz), current transformer waveforms (5kHz), and PLC control logs (OPC UA protocol) in real time. The edge nodes (edge ​​computing nodes) run a lightweight feature fusion model, aligning the four types of heterogeneous data with timestamps and normalizing their spatial coordinates to generate a 128-dimensional device status feature vector. This feature vector is uploaded to a cloud-based federated learning aggregation server (AWS IoT Core architecture) via the MQTT protocol. The server integrates feature vectors from over 200 similar devices to update the global fault diagnosis model (ResNet-18 architecture) and distributes the updated model parameters to the edge nodes. The edge nodes' adaptive online learning module optimizes the local model based on the new parameters. When the output anomaly probability > 0.85, a three-level fault tracing process is triggered, and a PDF-format maintenance strategy report is generated.

[0077] Among them, such as Figure 3 As shown, Figure 3This is a schematic diagram illustrating feature extraction based on a lightweight feature fusion model, provided by an embodiment of the present invention. It includes: processing vibration signal and current waveform data using a one-dimensional convolutional neural network; extracting long-term dependency features from temperature time-series data using a temporal attention mechanism; modeling the operational logic relationships in the device control log using a graph neural network; and using a feature compression layer to map the extracted features to a low-dimensional vector space to form a device state feature vector. The lightweight feature fusion model comprises four parallel processing branches: vibration / current data is input into a one-dimensional convolutional neural network (kernel size 64, stride 8, ReLU activation) to extract frequency domain features; temperature data is input into a BiLSTM layer (hidden units 32) and combined with an attention mechanism (scaled dot product attention) to capture long-term trends; and the control log is parsed into an operation instruction sequence graph (node ​​= operation instruction, edge = temporal dependency), input into a graph convolutional network (GCN layer number 2, aggregation function = mean) to learn the logical relationships. The three feature outputs are concatenated in the fusion layer and then compressed through a fully connected compression layer (128 dimensions, Dropout=0.2) to output the device state feature vector. The total number of model parameters is controlled within 1.2MB to adapt to the memory limitations of edge devices. Data acquisition is achieved through an industrial protocol adaptive parser. This parser supports plug-and-play conversion between OPC UA, Modbus, and MQTT protocols and automatically handles clock synchronization for data with different sampling frequencies. The industrial protocol adaptive parser is deployed on the edge node Linux system and includes a protocol conversion layer (converting Modbus RTU messages to JSON format and MQTT messages to Protobuf) and a clock synchronization engine. Clock synchronization adopts the IEEE 1588 PTP protocol, interpolating and resampling data at different sampling rates to a unified time axis (reference clock accuracy ±100ns). Abnormal data is cleaned by median filtering (window size 5) and Z-score normalization (threshold ±3).

[0078] Among them, such as Figure 4 As shown, Figure 4This is a schematic diagram of a federated learning framework provided in an embodiment of the present invention. It may include: a differentiated client selection strategy that dynamically selects edge nodes participating in aggregation based on device operating condition similarity and data quality; a privacy protection mechanism that adds homomorphic encryption noise before feature vector uploading to achieve gradient obfuscation; and a knowledge enhancement mechanism that embeds a knowledge graph constructed from historical fault cases as a regularization term into the global model training. The federated learning framework implementation includes three core mechanisms: client selection uses dynamic filtering based on operating condition similarity (devices with cosine similarity > 0.7 and data integrity > 95% participate in aggregation); privacy protection is achieved by adding Gaussian noise (μ=0, σ=0.01) to the feature vectors to achieve gradient obfuscation (a differential privacy protection mechanism); the knowledge enhancement module converts the fault knowledge graph (5000+ historical cases stored in the Neo4j database) into graph embedding vectors, which are added as a regularization term (L2 penalty coefficient 0.01) to the global model loss function; the aggregation process uses the FedAvg algorithm (federated averaging algorithm), performing a model update every 24 hours, reducing communication bandwidth usage to 1 / 20 of the original data.

[0079] The adaptive model update process at edge nodes includes: a drift detection unit that monitors changes in device operating conditions in real time based on KL divergence; an incremental fine-tuning unit that locally retrains key layers of the model using the latest data when distribution drift is detected; and a model pruning unit that dynamically compresses the size of the edge model to adapt to computational resource constraints. The adaptive online learning module consists of three units working collaboratively: the drift detection unit calculates the KL divergence between the current data distribution and the historical baseline every 5 minutes (sliding window = 1000 samples), triggering an alarm when the divergence value > 1.5; the incremental fine-tuning unit freezes the weights of the first 80% of the model layers, training only the last two fully connected layers online (learning rate 0.001, batch size 16); and the model pruning unit uses iterative magnitude pruning, removing 20% ​​of the lowest weight channels when the available memory on the edge device is < 100MB, dynamically compressing the model size by 30%-50%.

[0080] like Figure 5 As shown, Figure 5This is a schematic diagram of a maintenance strategy tracing mechanism provided in an embodiment of the present invention. It can include a three-level diagnostic process: the first level quickly identifies known fault types based on a fault mode matching library; the second level analyzes the dependencies between abnormal features using a causal reasoning engine to locate the root cause; and the third level calls a digital twin simulation module to verify the diagnostic results and generate a maintenance strategy including spare parts replacement paths and operating procedures. The multi-level fault tracing mechanism execution process is as follows: the first level matches a pre-set library of 200 fault modes (a SQLite database stores feature thresholds, such as bearing fault = vibration spectrum 6kHz component > 0.5g); the second level activates a causal reasoning engine (based on a fault propagation model, structural equation model, causal graph, and causal reasoning constructed using a Bayesian network). It analyzes dependency chains such as "sudden current increase → temperature rise → increased vibration" to locate the root cause; the third level calls a digital twin of the equipment (Unity3D simulation environment) to inject fault parameters, simulates the maintenance operation effect, and outputs the optimal maintenance path (such as prioritizing replacement of spare parts number #B-2032). The maintenance strategy includes torque standards (e.g., bolt tightening force 85±5Nm) and links to safe operation videos.

[0081] The state prediction employs a hybrid prediction model architecture: for short-term predictions within one hour, a temporal convolutional network is used to analyze the evolution of individual device states; for long-term predictions exceeding 24 hours, a spatiotemporal graph neural network is used to model the synergistic effects between device clusters. The hybrid prediction model architecture is divided as follows: short-term predictions use a temporal convolutional network (TCN, 4 dilated convolutional layers, dilation factor [1,2,4,8]), inputting 60 minutes of historical data to output the state for the next 60 minutes; long-term predictions use a spatiotemporal graph neural network (STGNN, graph attention layer + Gated-RNN) to construct a device cluster topology graph (nodes = devices, edges = physical connections), inputting 24 hours of cluster data to predict the degradation trend for the next 72 hours. The prediction results are displayed on the HMI interface as a line graph overlaid with confidence intervals. Because the objectives of short-term and long-term predictions differ, their prediction results (outputs) are also completely different. These forecasts are ultimately displayed visually on the HMI (Human-Machine Interface) (e.g., line charts, dashboards) to provide decision support for operations and maintenance personnel. The core objective of short-term forecasting is "real-time early warning," therefore its output consists of high-frequency, specific physical parameter predictions, such as the parameters of each key sensor on a single device, and the predicted sensor readings and their confidence intervals at each point in time (e.g., the next minute, five minutes later, and one hour later) within the next hour (60 minutes), along with vibration, temperature, current, pressure, and flow rates. The core objective of long-term forecasting is "planning and prevention," therefore its output consists of macroscopic, comprehensive health indicators and trends, such as the remaining useful life health index or degradation score and failure probability.

[0082] The system can also deploy a model reliability assessment module. This module evaluates reliability by calculating the confidence interval of the prediction results and using the Monte Carlo Dropout uncertainty index. When the uncertainty index exceeds a dynamic threshold, it automatically requests collaborative inference from the cloud model. The model reliability assessment module is implemented as follows: the confidence interval is calculated using quantile regression (outputting 10%-90% quantiles), and the uncertainty index is obtained through Monte Carlo Dropout (randomly dropping 10 times during inference and calculating the standard deviation). When the uncertainty > 0.3, the edge node uploads its current feature vector to the cloud. The cloud model (with 10 times the number of parameters as the edge model) performs inference and returns the results, with the collaborative latency controlled within 300ms.

[0083] The data transmission between edge computing nodes and cloud servers employs a breakpoint-resume compression protocol. This protocol dynamically adjusts the compression rate between 30% and 70% based on network bandwidth conditions and uses differential coding to reduce redundant transmission of consecutive data packets. The breakpoint-resume compression protocol workflow is as follows: before data transmission, edge nodes use the LZ77 algorithm for compression (the compression rate is dynamically adjusted based on network latency: 30% for latency < 50ms, 70% for latency > 200ms). The compressed data is divided into blocks (256KB each) and a CRC32 checksum is added. When transmission is interrupted, the breakpoint location is recorded. Upon reconnection, only the differential block (differentially encoded Δ data packets) is sent, reducing communication bandwidth usage to less than 5% of the original data.

[0084] This invention comprehensively addresses four technical problems that collectively constitute the core pain points of intelligent operation and maintenance in the industrial internet. The existence of any one of these problems can lead to system failure.

[0085] 1. Solving the dilemma of multi-source heterogeneous data fusion: This invention addresses this problem through a lightweight multi-path fusion model at the edge: 1D-CNN is used to extract vibration / current frequency domain features, Attention-LSTM (Attention Mechanism Long Short-Term Memory Network) is used to capture long-term temperature trends, and GNN is used to model log operation logic, ultimately fusing and compressing the data into a unified 128-dimensional feature vector. This approach deeply correlates cross-modal spatiotemporal information, eliminating feature omissions at the source, significantly reducing the misdiagnosis rate, and laying the foundation for accurate diagnosis.

[0086] 2. Resolving the conflict between real-time response and bandwidth: This invention constructs a dual-mode architecture of "real-time edge early warning - asynchronous cloud aggregation". Feature extraction is completed at the edge, data is compressed to 1 / 20 of its original size, and adaptive transmission is achieved through a dynamic compression ratio breakpoint resume protocol. This avoids the bottleneck of original data transmission, shortening the critical early warning response time from >500ms to within 100ms, providing a millisecond-level emergency stop window for high-speed devices, and effectively preventing fault escalation.

[0087] 3. Addressing the disconnect between static modeling and operational condition drift: This invention introduces an adaptive online learning module: based on KL divergence (>1.5 threshold), it detects data distribution drift in real time, triggering incremental fine-tuning only on the last 20% of key layers of the model, and initiating dynamic volume compression through model pruning when resources are scarce. This enables the model to continuously track equipment aging and environmental changes, maintaining high accuracy at all times, significantly reducing false alarms and the frequency of manual intervention, and achieving adaptive and intelligent operation and maintenance.

[0088] 4. Addressing Superficial Fault Diagnosis: This invention establishes a three-tiered fault tracing mechanism: the first tier rapidly matches a fault feature database; the second tier activates a Bayesian network causal engine to analyze the root causes of fault propagation; and the third tier invokes digital twin simulation to verify the maintenance plan. This mechanism upgrades from superficial diagnosis to root cause eradication, outputting comprehensive strategies such as "replacing the bearing + upgrading the voltage regulator," reducing fault recurrence rates, minimizing unplanned downtime, and improving overall equipment efficiency.

[0089] In a multi-source data joint modeling scenario, the process of the above-mentioned equipment status prediction method may include:

[0090] Step 1: Deploy sensors and connect edge devices to capture physical signals and control status of device operation in real time, forming the basis for analysis.

[0091] Install physical sensors on key parts of the industrial equipment: for example, install a triaxial vibration sensor on the motor bearing housing (collecting acceleration in the XYZ directions, 10,000 readings per second), attach a temperature sensor to the motor windings (collecting surface temperature once per second), and connect a current transformer to the power cable (sampling current waveform 5,000 times per second). Simultaneously, connect the equipment's PLC controller to the edge computing box (model: Raspberry Pi 4B+) via a network cable. After startup, the edge box automatically recognizes all sensors and displays a green "Equipment Ready" indicator light on the monitoring screen.

[0092] Step 2: Real-time cleaning and alignment of multi-source data aims to solve the problem of messy data in industrial fields and ensure its usability for subsequent analysis.

[0093] The edge computing box's built-in protocol parser unifies the conversion of data from different formats: for example, converting the rotation speed setting value transmitted via Modbus protocol into JSON format {"RPM_set":3000}, and converting the MQTT message from the temperature sensor into binary encoding. The clock synchronization engine timestamps all data according to the Precision Time Protocol (PTP) (error <100 nanoseconds) and interpolates and aligns data with different sampling rates (e.g., interpolating 1Hz temperature data to 5kHz to synchronize with vibration data). It automatically filters outliers: if a temperature reading suddenly changes to 230°C (normal range <150°C), the average of the preceding and following readings is used instead.

[0094] Step 3: Run a lightweight fusion model to extract features, the purpose of which is to transform complex sensor data into a machine-understandable "device health fingerprint".

[0095] Lightweight convergence model launched on edge devices:

[0096] Vibration / current processing: A one-dimensional convolutional neural network (64 kernels, stride 8) is used to extract frequency domain features and identify key frequency band features (e.g., the 6000Hz vibration component increases from 0.1g to 0.5g). Temperature analysis: A bidirectional recurrent neural network is used to identify trends (e.g., the temperature rises by 1.2°C per hour for 2 consecutive hours). Log parsing: Control commands are constructed into a relationship graph (e.g., a causal chain of "acceleration command → speed increase → load increase"). Feature compression: The above three types of results are concatenated and compressed into a feature vector consisting of 128 numbers (e.g., [0.76, -1.2, ..., 0.33]) through a fully connected layer.

[0097] Step 4: Upload to the cloud for privacy protection. The purpose is to protect the privacy of enterprise equipment data and prevent the leakage of production secrets.

[0098] Random noise (fluctuating within ±0.01 of the original value) is added to the feature vectors and uploaded to the cloud server via an encrypted channel. For example, if the original feature value of 0.76 changes to 0.758 or 0.763, the cloud cannot deduce the actual vibration data, but it can identify the overall pattern. Here, "overall pattern" refers to the statistical information contained in the feature vectors aggregated from a large number of devices, reflecting the common patterns and fault characteristics of the device group. Although the cloud cannot see the real, precise data of any single device, when it aggregates feature vectors from hundreds or thousands of devices that have undergone similar perturbations, the noise cancels out, and the true data patterns emerge.

[0099] Step 5: Federated learning aggregates global knowledge, with the aim of allowing all devices to share fault experience without leaking specific data from any single device.

[0100] Every 24 hours, the cloud performs aggregation: filtering 50 similar devices (e.g., all are CNC lathes processing aluminum alloy parts). "Similar" here refers to the similar operating conditions of these 50 devices. They collectively form a "similar device group." The mean of the feature vectors is calculated (e.g., the mean 6000Hz vibration of the 50 devices increases from 0.4g to 0.5g). Fault knowledge rules are injected (e.g., the knowledge base records "6000Hz vibration > 0.5g and temperature rise > 1°C / hour = bearing failure"). Global model parameters are updated (using gradient descent to adjust millions of internal parameters).

[0101] Step 6: Edge model adaptive optimization, the purpose of which is to enable the model to continuously adapt to new situations such as equipment aging and environmental changes.

[0102] After the edge device receives the new model: Monitor data changes: Check temperature distribution differences every 5 minutes (calculate the KL divergence value between the current data and the same period last week; if >1.5, trigger an alarm). Local fine-tuning: If an alarm is triggered, only the last 20% of the neural network layers of the model are retrained (preserving core knowledge). Dynamic slimming: Automatically remove redundant connections when memory is insufficient (e.g., remove the 1 million least important parameters from 5 million parameters).

[0103] Step 7: Dual-scale state prediction aims to detect potential faults in advance and avoid unplanned downtime.

[0104] Real-time prediction using edge models: Short-term early warning: Analyzes the status for the next hour (e.g., predicts that the motor temperature will reach 68°C in 50 minutes, exceeding the threshold of 65°C). Long-term assessment: Predicts the lifespan of critical components (e.g., outputs "Remaining lifespan of the spindle bearing: 15±2 days").

[0105] Step 8: Precise location of level three faults.

[0106] The Level 3 fault localization process in step 8 will only be triggered when the anomaly probability, predicted in real-time by the edge model, exceeds a preset threshold of 85%. When the predicted anomaly probability > 85%, the process involves: Primary matching: querying features in the fault database (e.g., "6000Hz vibration = 0.52g" matches bearing wear code F203). Causal reasoning: analyzing the event chain (e.g., "current fluctuation → temperature increase → increased vibration" points to unstable power supply voltage). Digital twin verification: simulating bearing replacement in a virtual model (showing the vibration value drop to 0.2g, confirming the diagnosis is correct). The aim is to locate the root cause from the surface phenomenon and avoid incorrect repairs.

[0107] Step 9: Generate an executable maintenance plan to guide maintenance personnel to operate successfully on the first attempt and reduce downtime.

[0108] The system output includes the following steps: "① Power off → ② Remove end cap bolts → ③ Remove the old bearing (model #B-2032) → ④ Install the new bearing → ⑤ Tighten with a torque wrench to 85 Nm".

[0109] Safety tip: "Wear cut-resistant gloves and confirm that the capacitor has discharged completely."

[0110] Resource scheduling: "Spare parts are located on the second floor of rack 3 in warehouse A, with 12 in stock. The estimated replacement time is 45 minutes."

[0111] Step 10: Credibility assurance mechanism, the purpose of which is to prevent misjudgment caused by edge computing errors.

[0112] Calculations after each prediction:

[0113] Confidence interval: For example, remaining life expectancy 15 ± 2 days (meaning there is a 90% probability that the true value is between 13 and 17 days).

[0114] Uncertainty score: The reliability is obtained through 10 random tests (automatic cloud review is requested when the score is <0.3).

[0115] Step 11: Intelligent network transmission optimization aims to adapt to fluctuations in the factory network and ensure data real-time performance.

[0116] Data transmission is dynamically adjusted: Light compression (70% of data retained) when network connectivity is good. Deep compression (data size reduced to 30%) when network latency > 200ms. Only the difference portion is retransmitted after a transmission interruption (e.g., only +0.1 is sent when the temperature changes from 60.1°C to 60.2°C).

[0117] Step 12: Knowledge closed-loop update, the purpose of which is to continuously evolve the system and make it more and more accurate with use.

[0118] After the repairs were completed:

[0119] The technician scans the equipment's QR code and enters the actual cause of the malfunction (such as "bearing seal ring rupture").

[0120] The system automatically generates a new rule: "Damaged sealing ring → Lubricating oil leakage → Change in vibration characteristics" and adds it to the knowledge base.

[0121] The model operation process of this invention is divided into two major stages: collaborative training and real-time application, forming a closed-loop self-optimizing intelligent system.

[0122] The model training in this embodiment is a continuous collaborative process led by the cloud and participated in by the edge. First, each edge computing node independently collects multi-source heterogeneous data from its local device, including vibration, temperature, current, and control logs. A lightweight multimodal fusion model (comprising 1D-CNN, Attention-LSTM, and GNN parallel structures) is used for preprocessing and feature extraction, generating a 128-dimensional device state feature vector. To protect data privacy, Gaussian noise is added to these feature vectors before uploading for differential privacy protection. The cloud server periodically (e.g., every 24 hours) initiates a federated learning aggregation process. Its core innovation lies in dynamically selecting edge nodes with high operating condition similarity (e.g., cosine similarity > 0.7) and good data quality, rather than randomly selecting devices, to form aggregation groups, thus ensuring the consistency of the learning objectives. The cloud aggregates these encrypted feature vectors and, crucially, injects historical fault knowledge graphs as graph embedding vectors as regularization terms into the global model training process, achieving a deep integration of data-driven and knowledge-driven approaches. Finally, an algorithm such as FedAvg is used to update the global model parameters and distribute them to the edge nodes. After receiving the new model, the edge device also performs local adaptive optimization through the drift detection module based on KL divergence, such as incrementally fine-tuning the end network layer, so that the model can continuously track the aging or changes in operating conditions of the device itself.

[0123] The trained and optimized model operates as a real-time diagnostic process driven by the edge and assisted by the cloud. During equipment operation, edge nodes utilize locally deployed lightweight models to perform high-speed feature extraction and status prediction on real-time collected data, directly outputting short-term (e.g., within one hour) predicted equipment status parameters or long-term (e.g., over 24 hours) health trend assessments. The entire process can be completed within 100 milliseconds, meeting industrial real-time requirements. When the predicted anomaly probability exceeds a preset threshold (e.g., 85%), the system automatically triggers a multi-level fault tracing mechanism: from quickly matching fault feature libraries to initiating a causal inference engine to analyze the fault propagation chain and locate the root cause, and finally calling digital twin simulation to verify the maintenance strategy and generate an executable detailed solution. Furthermore, the model credibility assessment module monitors prediction uncertainty in real time, automatically requesting collaborative inference from the cloud when confidence is insufficient to ensure decision reliability. Post-maintenance results are entered into the knowledge base through a feedback mechanism, forming a complete closed loop from "precise early warning" to "deep diagnosis" to "radical treatment feedback," enabling the system to continuously evolve and become increasingly accurate with use.

[0124] The equipment status prediction device provided in the embodiments of the present invention will be described below. The equipment status prediction device described below can be referred to in correspondence with the equipment status prediction method described above.

[0125] Please refer to the details. Figure 6 , Figure 6A schematic diagram of the structure of the device state prediction device provided in the embodiments of the present invention may include:

[0126] The device status feature vector determination module 100 is used to acquire device data of a preset type of target device, determine the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, and align each feature vector to obtain the device status feature vector.

[0127] The global update module 200 is used to upload the device state feature vector to the server based on the federated learning framework, so that the server can update the device state prediction model using cross-device group knowledge without sharing device data.

[0128] The device abnormal state determination module 300 is used to determine the device abnormal state by predicting based on the device state feature vector using the latest device state prediction model; wherein, the latest device state prediction model is a model based on global updates of the server and local incremental updates of edge nodes.

[0129] Furthermore, based on any of the above embodiments, the device state feature vector determination module 100 may include:

[0130] The frequency domain feature vector extraction unit is used to extract the frequency domain feature vectors of vibration signals and current waveform data.

[0131] The trend feature vector extraction unit is used to extract the trend feature vector of temperature time series data;

[0132] The graph embedding feature vector extraction unit is used to model the operational logic relationships in the device control log data through a graph neural network to obtain graph embedding feature vectors; wherein, the preset types of device data include the vibration signal, the current waveform data, the temperature time series data, and the device control log data;

[0133] The device state feature vector extraction unit is used to compress and fuse the frequency domain feature vector, the trend feature vector, and the graph embedding feature vector to obtain the device state feature vector.

[0134] Furthermore, based on any of the above embodiments, the global update module 200 may include:

[0135] The global update unit is used to add random noise to the device state feature vector and upload the noise-added feature vector to the server through an encrypted channel. This allows the server to dynamically select the device state feature vectors corresponding to the edge nodes participating in the aggregation based on the similarity of device operating conditions and data quality. After aggregation, the device state prediction model is updated. During the update process, the knowledge graph constructed from historical failure cases is embedded into the model as a knowledge enhancement mechanism for regularization.

[0136] Furthermore, based on any of the above embodiments, the above-mentioned device state prediction device may further include:

[0137] The judgment module is used to determine whether the probability of the abnormal state of the device exceeds a preset abnormal threshold.

[0138] The three-level traceability module is used to output the fault type, fault cause and maintenance strategy based on the multi-level fault traceability mechanism when the anomaly probability exceeds the preset anomaly probability threshold.

[0139] Furthermore, the three-level traceability module may include:

[0140] A primary tracing unit is used to determine the fault type based on the abnormal state of the equipment using a fault mode matching library;

[0141] The secondary tracing unit is used to analyze the dependencies between abnormal features related to the fault type through a causal reasoning engine to determine the cause of the fault.

[0142] The three-level traceability unit is used to call the digital twin simulation module to verify the cause of the fault and generate the maintenance strategy containing operating specifications.

[0143] Furthermore, based on any of the above embodiments, the latest device state prediction model includes a temporal convolutional network and a spatiotemporal graph neural network, and the device abnormal state determination module 300 may include:

[0144] The short-term prediction unit is used to analyze the evolution law of the individual device state using the temporal convolutional network to obtain the short-term abnormal state of the device; wherein, the evolution law of the individual device state is the inherent pattern and trend of the device's own health status parameters changing over time; the temporal convolutional network adopts a set number of dilated convolutional layers and a sequence of dilation factors.

[0145] The long-term prediction unit is used to model the synergistic effect between equipment clusters using the spatiotemporal graph neural network to obtain long-term abnormal equipment states for a set long-term prediction; wherein, the synergistic effect between equipment clusters refers to the mutual influence between multiple devices in a production system or production line due to physical connections, process flow, or control logic; wherein, the time period of the set short-term prediction is shorter than the time period of the set long-term prediction.

[0146] Furthermore, based on any of the above embodiments, the above-mentioned device state prediction device may further include:

[0147] The incremental optimization module is used to monitor changes in the operating condition distribution of the target equipment in real time. When distribution drift is detected, the module uses the latest equipment data to locally retrain the key layers of the model and dynamically compresses the size of the equipment condition prediction model to obtain an incrementally optimized equipment condition prediction model.

[0148] It should be noted that the order of the modules and units in the above-mentioned device status prediction device can be changed without affecting the logic.

[0149] An embodiment of the present invention provides a device status prediction apparatus, which may include: a device status feature vector determination module 100, used to acquire device data of a preset type of target device, determine the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, and align each feature vector to obtain a device status feature vector; a global update module 200, used to upload the device status feature vector to a server based on a federated learning framework, so that the server can update the device status prediction model using cross-device group knowledge without sharing device data; and a device abnormal status determination module 300, used to predict based on the device status feature vector using the latest device status prediction model to determine the device abnormal status; wherein, the latest device status prediction model is a model based on global updates by the server and local incremental updates by edge nodes.

[0150] The following is a description of a device state prediction device provided by an embodiment of the present invention. The device state prediction device described below and the device state prediction method described above can be referred to in correspondence.

[0151] Please refer to Figure 7 , Figure 7 A schematic diagram of a device state prediction device provided in an embodiment of the present invention may include:

[0152] Memory 10 is used to store computer programs;

[0153] The processor 20 is used to execute a computer program to implement the device state prediction method described above.

[0154] The memory 10, processor 20, and communication interface 30 all communicate with each other through the communication bus 40.

[0155] In this embodiment of the invention, the memory 10 is used to store one or more programs. The programs may include program code, which includes computer operation instructions. In this embodiment of the invention, the memory 10 may store programs for implementing the following functions:

[0156] Obtain device data of preset types of target devices, determine the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, and align each feature vector to obtain the device status feature vector;

[0157] Based on the federated learning framework, device state feature vectors are uploaded to the server so that the server can update the device state prediction model using cross-device group knowledge without sharing device data.

[0158] Based on the device status feature vector, the latest device status prediction model is used to predict and determine the abnormal status of the device; the latest device status prediction model is a model based on global updates from the server and local incremental updates from edge nodes.

[0159] In one possible implementation, the memory 10 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; and the data storage area may store data created during use.

[0160] Furthermore, memory 10 may include read-only memory and random access memory, providing instructions and data to the processor. A portion of the memory may also include NVRAM. The memory stores operating systems and operating instructions, executable modules, or data structures, or subsets thereof, or extended sets thereof, wherein the operating instructions may include various operating instructions for implementing various operations. The operating system may include various system programs for implementing various basic tasks and handling hardware-based tasks.

[0161] Processor 20 can be a central processing unit (CPU), an application-specific integrated circuit, a digital signal processor, a field-programmable gate array, or other programmable logic device. Processor 20 can be a microprocessor or any conventional processor. Processor 20 can call programs stored in memory 10.

[0162] The communication interface 30 can be an interface for the communication module, used to connect with other devices or systems.

[0163] Of course, it should be noted that, Figure 7 The structure shown does not constitute a limitation on the device state prediction device in the embodiments of the present invention. In practical applications, the device state prediction device may include more than Figure 7 More or fewer components as shown, or combinations of certain components.

[0164] The computer-readable storage medium provided in the embodiments of the present invention is described below. The computer-readable storage medium described below can be referred to in correspondence with the device state prediction method described above.

[0165] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described device state prediction method.

[0166] The computer-readable storage medium may include 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.

[0167] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0168] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. 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 implementations should not be considered beyond the scope of this invention.

[0169] Finally, it should be noted that in this document, relationships such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0170] The above provides a detailed description of the device state prediction method, apparatus, device, and readable storage medium provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting equipment status, characterized in that, include: Obtain device data of a preset type for the target device, determine the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, and align each feature vector to obtain the device status feature vector; Based on the federated learning framework, the device state feature vector is uploaded to the server so that the server can update the device state prediction model using cross-device group knowledge without sharing device data. Based on the device status feature vector, the latest device status prediction model is used to predict and determine the abnormal state of the device; wherein, the latest device status prediction model is a model based on global updates of the server and local incremental updates of edge nodes.

2. The equipment status prediction method according to claim 1, characterized in that, Obtain device data of a preset type for the target device; determine the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements; align each feature vector to obtain a device status feature vector, including: Extract the frequency domain feature vectors from vibration signals and current waveform data; Extract the trend feature vector from temperature time series data; The operational logic relationships in the device control log data are modeled using a graph neural network to obtain graph embedding feature vectors; wherein, the preset types of device data include the vibration signal, the current waveform data, the temperature time series data, and the device control log data; The frequency domain feature vector, the trend feature vector, and the graph embedding feature vector are compressed and fused to obtain the device state feature vector.

3. The equipment status prediction method according to claim 1, characterized in that, The device state feature vector is uploaded to the server based on a federated learning framework, enabling the server to update the device state prediction model using cross-device group knowledge without sharing device data. This includes: Random noise is added to the device state feature vector, and the noise-added feature vector is uploaded to the server through an encrypted channel. The server dynamically selects the device state feature vectors corresponding to the edge nodes participating in the aggregation based on the similarity of device operating conditions and data quality. After aggregation, the device state prediction model is updated. During the update process, the knowledge graph constructed from historical failure cases is embedded as a regularization term into the model as a knowledge enhancement mechanism.

4. The equipment status prediction method according to any one of claims 1 to 3, characterized in that, After determining the abnormal state of the equipment by using the latest equipment state prediction model based on the equipment state feature vector, the process further includes: Determine whether the probability of the abnormal state of the device exceeds a preset abnormal threshold; When the anomaly probability exceeds the preset anomaly probability threshold, the fault type, fault cause, and maintenance strategy are output based on the multi-level fault tracing mechanism.

5. The equipment status prediction method according to claim 4, characterized in that, When the anomaly probability exceeds the preset anomaly probability threshold, the fault type, fault cause, and maintenance strategy are output based on a multi-level fault tracing mechanism, including: The fault type is determined using a fault mode matching library based on the abnormal state of the equipment. The cause of the failure is determined by analyzing the dependencies between abnormal features related to the failure type using a causal reasoning engine. The digital twin simulation module is invoked to verify the cause of the fault and to generate the maintenance strategy containing operating procedures.

6. The equipment status prediction method according to claim 1, characterized in that, The latest equipment status prediction model includes a temporal convolutional network and a spatiotemporal graph neural network. Based on the equipment status feature vector, the latest equipment status prediction model is used to predict and determine the abnormal equipment status, including: For short-term prediction, the temporal convolutional network is used to analyze the evolution law of individual device states to obtain short-term abnormal device states; wherein, the evolution law of individual device states is the inherent pattern and trend of the device's own health status parameters changing over time; the temporal convolutional network adopts a set number of dilated convolutional layers and a sequence of dilation factors; The spatiotemporal neural network is used to model the synergistic effect between equipment clusters for setting long-term predictions, thereby obtaining long-term abnormal equipment states. The synergistic effect between equipment clusters refers to the mutual influence between multiple devices in a production system or production line due to physical connections, process flows, or control logic. The time period for setting short-term predictions is shorter than the time period for setting long-term predictions.

7. The equipment status prediction method according to claim 1, characterized in that, After determining the abnormal state of the equipment by using the latest equipment state prediction model based on the equipment state feature vector, the process further includes: The system monitors changes in the operating conditions of target equipment in real time. When a distribution drift is detected, it uses the latest equipment data to locally retrain the key layers of the model and dynamically compresses the size of the equipment condition prediction model to obtain an incrementally optimized equipment condition prediction model.

8. A device for predicting equipment status, characterized in that, include: The device status feature vector determination module is used to obtain device data of a preset type of target device, determine the feature vector corresponding to each type of device data based on the coupling principle of data modality characteristics and business requirements, and align each feature vector to obtain the device status feature vector. The global update module is used to upload the device state feature vector to the server based on the federated learning framework, so that the server can update the device state prediction model using cross-device group knowledge without sharing device data. The device abnormal state determination module is used to predict and determine the device abnormal state based on the device state feature vector using the latest device state prediction model; wherein, the latest device state prediction model is a model based on global updates of the server and local incremental updates of edge nodes.

9. A device for predicting device status, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the device state prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the device state prediction method as described in any one of claims 1 to 7.

Images