A device fault early warning method and device based on SLIM knowledge distillation and a medium

Abstract

The application discloses a device fault early warning method based on SLIM knowledge distillation, a device and a medium. The method comprises the following steps: collecting multi-modal sensor data in the running process of the device, preprocessing and extracting features of the multi-modal sensor data to generate a multi-dimensional feature vector; inputting the multi-dimensional feature vector into a pre-trained teacher model to output soft labels containing each fault category, and training a lightweight student model through the soft labels based on the SLIM knowledge distillation technology; reasoning the multi-dimensional feature vector through the student model to obtain a fault type probability distribution corresponding to the device, and calculating a dynamic health value of the device according to a space-time attention mechanism and a device working condition; triggering an early warning strategy of a corresponding early warning level according to the fault type probability distribution and / or the dynamic health value; and generating a maintenance strategy corresponding to the device according to a preset constraint condition in response to the early warning strategy.

Description

Technical Field

[0001] This application relates to the field of industrial equipment operation and maintenance technology, specifically to a method, equipment, and medium for early warning of equipment faults based on SLIM knowledge distillation. Background Technology

[0002] In industrial production, the stable operation of equipment directly determines production efficiency, product quality, and operating costs. With the continuous development of industrial intelligence, traditional models based on manual inspection and periodic maintenance can no longer meet the reliability requirements of modern production. Predictive maintenance has become a core technological direction for improving equipment management. However, existing maintenance methods still face a series of technical bottlenecks, making it difficult to significantly improve equipment reliability and reduce total lifecycle maintenance costs while ensuring high precision and real-time performance.

[0003] First, regarding model deployment, large and complex models are often used in pursuit of prediction accuracy. These models are difficult to deploy to resource-constrained edge environments, resulting in insufficient real-time performance. This fails to meet the stringent requirements of millisecond-level real-time response in scenarios such as stamping lines and rotating machinery, thus missing the optimal time for fault intervention. Furthermore, at the health assessment and early warning triggering level, the widely used static early warning methods based on fixed physical thresholds cannot adapt to the dynamic changes in different operating conditions such as equipment startup, operation, and shutdown. During operating condition switching, a large number of false alarms or missed alarms are easily generated, significantly reducing the reliability of early warnings. Summary of the Invention

[0004] To address the aforementioned issues, this application proposes a device fault early warning method based on SLIM knowledge distillation, comprising: Collect multimodal sensor data during equipment operation, preprocess and extract features from the multimodal sensor data to generate multidimensional feature vectors; The multi-dimensional feature vector is input into a pre-trained teacher model to output soft labels containing each fault category, and a lightweight student model is trained using the soft labels based on SLIM knowledge distillation technology. The student model is used to reason about the multi-dimensional feature vector to obtain the probability distribution of the fault type corresponding to the device, and the dynamic health value of the device is calculated based on the spatiotemporal attention mechanism and the device's operating condition. Based on the probability distribution of the fault type and / or the dynamic health value, trigger the warning strategy corresponding to the warning level; In response to the warning strategy, a maintenance strategy corresponding to the device is generated based on preset constraints.

[0005] In one implementation of this application, a lightweight student model is trained using the soft labels output by the teacher model based on SLIM knowledge distillation technology, specifically including: Based on the SLIM knowledge distillation unit deployed at the edge layer, the soft labels output by the teacher model are loaded, and a lightweight student model is trained by optimizing the hybrid loss function based on the soft labels. The optimized hybrid loss function is expressed as follows:

[0006] Where zs represents the student model output, and y represents the true fault label. denoted by , represents the softmax function, LCE is the cross-entropy loss used to constrain the consistency between zs and y, ps is the probability distribution of the student model output, pt represents the soft label distribution, LKL is the KL divergence loss used to measure the difference between ps and pt, and T represents the temperature parameter.

[0007] In one implementation of this application, the dynamic health value of the device is calculated based on the spatiotemporal attention mechanism and the device's operating condition, specifically including: Based on the spatiotemporal attention mechanism, the multi-dimensional feature vectors are weighted to obtain spatiotemporally weighted feature vectors; Identify the current operating phase of the device; wherein the operating phase includes a startup phase, a steady-state phase, and a shutdown phase; Based on the operational phase, assign corresponding operating condition weights to each modal feature in the spatiotemporal weighted feature vector; Based on a preset exponential smoothing model, the standardized spatiotemporal weighted feature vector is weighted and fused according to the operating condition weights to obtain the dynamic health value corresponding to the device.

[0008] In one implementation of this application, based on a preset exponential smoothing model, the spatiotemporal weighted feature vector is weighted and fused according to the operating condition weights to obtain the dynamic health value corresponding to the device, specifically including: The dynamic health value of the device is obtained using the following formula:

[0009] in, This represents the historical health value decay coefficient. This represents the working condition weight corresponding to the i-th modal feature. This represents the standardized spatiotemporal weighted feature vector. and These represent the dynamic health values ​​at times t and t-1, respectively.

[0010] In one implementation of this application, a warning strategy corresponding to the warning level is triggered based on the fault type probability distribution and / or the dynamic health value, specifically including: If the fault type probability distribution indicates the corresponding fault type, or if the dynamic health value is lower than the first threshold, a level one warning is triggered, and an emergency shutdown and maintenance process is initiated. If the dynamic health value is between the first threshold and the second threshold, a level two warning is triggered, and a planned maintenance suggestion is generated. If the dynamic health value is between the second threshold and the third threshold, a level 3 warning is triggered, and the monitoring frequency is increased.

[0011] In one implementation of this application, a maintenance strategy corresponding to the device is generated based on preset constraints, specifically including: A mixed-integer programming model is constructed with the objective function of minimizing the total maintenance cost; wherein the total maintenance cost includes repair costs and downtime losses. Based on preset constraints, the mixed integer programming model is solved to obtain the optimal maintenance strategy under the constraints; wherein the constraints include at least spare parts inventory, personnel skills and production scheduling.

[0012] In one implementation of this application, after generating the maintenance strategy corresponding to the device, the method further includes: Collect newly generated data samples during the operation and maintenance of the equipment; The teacher model is incrementally trained using the data samples. Based on the incrementally trained teacher model, the student model deployed on the edge side is hot-updated using the SLIM knowledge distillation technique.

[0013] In one implementation of this application, the teacher model is a model built based on the Stacking ensemble learning framework. The multi-dimensional feature vector is input into the pre-trained teacher model to output soft labels containing each fault category, specifically including: By using multiple heterogeneous base learners, preliminary inference is performed on the multi-dimensional feature vectors to obtain a preliminary probability distribution; The initial probability distributions output by multiple base learners are fused using a meta-learner to generate soft labels containing each fault category.

[0014] This application provides a device fault early warning device based on SLIM knowledge distillation, the device comprising: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are then executed by the at least one processor to enable the at least one processor to perform a device fault early warning method based on SLIM knowledge distillation as described above.

[0015] This application provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as follows: The above describes a device fault early warning method based on SLIM knowledge distillation.

[0016] The equipment fault early warning method based on SLIM knowledge distillation proposed in this application can bring the following beneficial effects: By employing a knowledge distillation framework, a complex teacher model is used in the cloud to ensure high-precision diagnosis, while knowledge is successfully transferred to a lightweight student model. This allows for direct deployment on edge devices, significantly reducing parameter size and inference latency, meeting the real-time response requirements of device alerts, and resolving the challenge of balancing model accuracy and deployment efficiency. Secondly, by integrating a spatiotemporal attention mechanism with a condition-adaptive dynamic health value model, replacing traditional fixed thresholds, the model dynamically adjusts evaluation weights based on device operating stages and multimodal data characteristics. This significantly reduces false alarms and missed alarms caused by changes in operating conditions, greatly improving the reliability of alerts. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a device fault early warning method based on SLIM knowledge distillation, provided for an embodiment of this application; Figure 2 This is a schematic diagram of a device fault early warning device based on SLIM knowledge distillation, provided as an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Existing predictive maintenance technologies for equipment have the following prominent problems: Inefficient model deployment: Traditional deep learning models (such as LSTM and Transformer) typically have more than 50 million parameters, and inference latency on edge devices such as PLCs and industrial gateways often exceeds 2 seconds, failing to meet the real-time requirements of industrial scenarios. For example, fault prediction in the hydraulic system of an automotive stamping line needs to be analyzed within 100ms; otherwise, the optimal intervention time may be missed. Insufficient utilization of multimodal data: Existing systems often rely on single vibration signals for fault diagnosis, neglecting the complementary value of multimodal features such as temperature, current, and pressure. In actual industrial scenarios, early motor bearing faults often manifest as abnormal temperatures, while the vibration signal shows no significant change; gearbox tooth breakage faults are accompanied by a sudden increase in high-frequency vibration energy and current harmonic distortion, making it easy for single-modal data to lead to missed diagnoses. Static threshold warnings have poor reliability: Using fixed thresholds for warnings cannot adapt to the different characteristics of equipment during different operating stages such as startup, steady state, and shutdown. During equipment startup, the vibration amplitude can be 3-5 times that of the steady state, easily triggering false alarms; while during low-speed operation, excessive threshold redundancy leads to missed alarms. Statistics from a wind power company show that the false alarm rate of the static threshold method is as high as 15%. Maintenance decisions lack overall optimization: Traditional maintenance planning does not fully consider constraints such as production scheduling, spare parts inventory, and personnel skills, resulting in high maintenance costs and low efficiency.

[0020] Based on this, this application utilizes SLIM knowledge distillation technology to efficiently transfer knowledge from the teacher model to a lightweight student model, effectively reducing inference latency at the edge device and meeting industrial real-time requirements. Simultaneously, a spatiotemporal attention mechanism is designed to dynamically allocate weights for multimodal features such as vibration, temperature, and current, fully leveraging the complementary value of different physical signals and resolving the problem of false positives and false negatives caused by single sensor data. Furthermore, a dynamic health value calculation model based on equipment operation stages is constructed, adaptively adjusting the weights of evaluation indicators according to operating conditions such as startup, steady state, and shutdown, replacing traditional static thresholds and reducing false alarm rates. Moreover, based on a mixed-integer programming algorithm, considering constraints such as production scheduling, spare parts inventory, and personnel skills, a maintenance strategy that minimizes total cost is generated, achieving globally optimal allocation of maintenance resources.

[0021] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0022] like Figure 1 As shown in the embodiment of this application, a device fault early warning method based on SLIM knowledge distillation is provided, including: S101: Collects multimodal sensor data during equipment operation, preprocesses and extracts features from the multimodal sensor data to generate multidimensional feature vectors.

[0023] The edge layer, as the core processing unit on the device side, is deployed in the industrial field, directly connecting to equipment sensors and actuators. It features low latency and high reliability, and mainly comprises three core components: a data acquisition module, a lightweight inference engine, and a local early warning device. The data acquisition module integrates multiple types of industrial sensors to achieve comprehensive perception of equipment operating status. These sensors include vibration accelerometers, infrared thermal imagers, and current transformers. Vibration monitoring uses a 10kHz sampling rate triaxial accelerometer, installed in key areas such as the bearing housing at the equipment drive end, accurately capturing vibration waveform characteristics. Temperature monitoring is equipped with a 1Hz sampling rate infrared thermal imager, acquiring the surface temperature field distribution of equipment such as motors and gearboxes in a non-contact manner. Current monitoring uses a 1kHz sampling rate Hall current transformer to collect three-phase stator current signals. All sensors support the Modbus / TCP industrial communication protocol and are connected to the edge computing platform via shielded twisted-pair cables. Data transmission latency is strictly controlled within 50ms to ensure the timeliness of the raw data.

[0024] The edge layer data acquisition module synchronously collects multimodal sensor data during equipment operation, including vibration, temperature, and current signals, through sensor groups deployed at key parts of the equipment. The collected raw data requires preprocessing for noise reduction. First, wavelet thresholding is used to denoise the vibration signal, effectively filtering out high-frequency electromagnetic interference and mechanical noise above 300Hz. Temperature and current signals are filtered using moving averages to eliminate random fluctuations and ensure data stability. Missing values ​​are filled using linear interpolation, and outliers are marked and corrected. After the above noise reduction operations, the min-max normalization method is used to eliminate dimensional differences between different data types, mapping vibration RMS, temperature values, and current harmonic distortion rate to the [0,1] interval. The specific calculation formula is as follows: , This represents the maximum value of the feature. The minimum, maximum, and minimum values ​​represent features determined statistically based on historical datasets and support dynamic updates to adapt to long-term changes in equipment performance. After preprocessing, the multimodal sensor data needs to have fault-sensitive feature parameters extracted based on the physical characteristics of different modal data to generate corresponding multi-dimensional feature vectors. Specifically, vibration signals extract time-domain features such as kurtosis factor, peak factor, and impulse factor, as well as frequency-domain features such as first harmonic energy, second harmonic energy, and high-frequency band energy; temperature signals extract features such as mean, slope, fluctuation rate, and maximum deviation; and current signals extract features such as total harmonic distortion rate, odd harmonic content, and even harmonic content. Ultimately, the multi-dimensional multimodal feature vectors comprehensively characterize the equipment's operating status. S102: Input multi-dimensional feature vectors into a pre-trained teacher model to output soft labels containing each fault category, and train a lightweight student model using soft labels based on SLIM knowledge distillation technology.

[0025] The cloud layer, deployed in a public or private cloud environment, leverages large-scale computing resources and data storage capabilities to provide model support and decision-making services to the edge layer. It primarily comprises a global model training platform, a knowledge distillation service, and a maintenance decision engine. The global model training platform, built on a GPU cluster, provides a distributed training environment supporting parallel processing of large-scale industrial datasets. It employs an automated machine learning process to automate data cleaning, feature engineering, and model training end-to-end, ensuring the stability and generalization ability of the teacher model. The knowledge distillation service, acting as a crucial link between the teacher and student models, enables knowledge transfer from complex to lightweight models. This service module includes three main functional units: a distillation strategy optimizer, a model compressor, and a version manager. The distillation strategy optimizer maximizes knowledge transfer efficiency by dynamically adjusting the temperature parameter T and the loss function weights. The model compressor uses channel pruning and INT8 quantization techniques to compress the number of parameters in the teacher model. The version manager supports full lifecycle management of the model, recording information such as training time, performance metrics, and deployment devices, providing precise model update services to the edge layer. The maintenance decision engine integrates a mixed-integer programming solver with an industrial data integration interface to achieve global optimization of maintenance strategies. The engine connects to enterprise business systems through API interfaces to obtain data such as production scheduling, spare parts inventory, and personnel skills in real time. Based on a mixed-integer programming model, with the goal of minimizing total maintenance costs, it quickly solves the optimal maintenance plan under the conditions of spare parts inventory constraints, personnel skill matching constraints, and production scheduling constraints. The decision results are output in two forms: visual reports and structured instructions, which provide decision-making references for managers and clear guidance for edge layer maintenance operations.

[0026] The teacher model inference unit is deployed in the cloud. The teacher model is built on the Stacking ensemble learning framework and is responsible for executing the Stacking ensemble learning inference process. That is, by inputting multi-dimensional feature vectors into the pre-trained teacher model, it can output soft labels containing various fault categories. The Shift-Learnable Instance Matching (SLIM) knowledge distillation framework enables lightweight and high-precision inference. Through the knowledge transfer mechanism guided by the teacher model and learned by the student model, it significantly reduces model complexity while ensuring fault identification accuracy, meeting the deployment requirements of industrial edge devices. A lightweight student model is obtained through soft label training based on SLIM knowledge distillation technology.

[0027] In one embodiment, the underlying base learner comprises two gradient boosting tree models: LightGBM and XGBoost. LightGBM sets the number of decision trees to 150 and the maximum depth to 6, using the histogram algorithm to accelerate feature splitting and efficiently capture the nonlinear relationships of multimodal features such as vibration, temperature, and current. XGBoost uses a learning rate of 0.05 and controls model complexity through regularization terms to enhance generalization ability. The two heterogeneous base learners perform independent preliminary inference on the multi-dimensional feature vectors, outputting preliminary probability distributions of fault types. The meta-learner uses a random forest, fusing the outputs of the base learners as input features to generate a fine-grained probability distribution of fault types, such as a bearing wear probability of 0.82, a gear breakage probability of 0.15, and a normal operation probability of 0.03, which serves as soft labels for the student model distillation. In this embodiment, the student model employs a hybrid architecture of 1D-CNN and BiLSTM, achieving lightweighting while maintaining feature extraction capabilities. The 1D-CNN part contains 32 convolutional units with a kernel size of 3. It captures local temporal patterns such as the impact characteristics of vibration signals and the abrupt changes in temperature signals through local receptive fields. The BiLSTM part has 64 hidden units with a hidden layer dimension of 128. It effectively captures long-sequence dependencies before equipment failure using a bidirectional recurrent structure, such as the gradual increase in high-frequency vibration energy before a gearbox failure. To further compress the model size, channel pruning is used to remove redundant convolutional channels with a mean absolute value of weights <0.01. Combined with INT8 quantization perceptual training, the model parameters are converted from floating-point numbers to integers, ultimately achieving a significant reduction in the number of parameters to meet the memory and computing power constraints of edge devices. In one embodiment, the SLIM knowledge distillation unit is deployed at the edge layer to learn the reasoning ability of the teacher model through knowledge distillation. First, soft labels generated by the teacher model are loaded, and a 1D-CNN / BiLSTM hybrid model is trained using the Adam optimizer: the 1D-CNN layer extracts local temporal features, the BiLSTM layer captures long-sequence dependencies, and the output layer generates a smooth probability distribution using a softmax function with a temperature parameter T=3. Then, based on the soft labels, knowledge transfer is achieved by optimizing the hybrid loss function, resulting in a lightweight student model. The optimized hybrid loss function is expressed as follows:

[0028] Where zs represents the student model output, and y represents the true fault label. denoted by , represents the softmax function, LCE is the cross-entropy loss used to constrain the consistency between zs and y, ps is the probability distribution of the student model output, pt represents the soft label distribution, LKL is the KL divergence loss used to measure the difference between ps and pt, and T represents the temperature parameter. The weight ratio of cross-entropy loss to KL divergence loss in the above hybrid loss function is only an example. The specific weight values ​​can be set according to actual needs to fully utilize the soft label knowledge of the teacher model while ensuring classification accuracy and improving the generalization performance of the student model.

[0029] S103: The student model is used to reason about the multi-dimensional feature vectors to obtain the probability distribution of the fault type corresponding to the equipment, and the dynamic health value of the equipment is calculated based on the spatiotemporal attention mechanism and the equipment operating conditions.

[0030] The lightweight inference engine, as the core computing unit of the edge layer, is equipped with a student model derived from SLIM knowledge distillation. This student model undergoes quantization, pruning, and acceleration optimization using the TensorRT deep learning optimization library, allowing direct deployment on industrial PLC controllers or industrial gateways to meet the stringent real-time requirements of industrial scenarios. By inferring from multi-dimensional feature vectors using the student model, the probability distribution of fault types corresponding to the equipment can be obtained. To address the limitations of single-modal data, highlight fault-sensitive information, and improve the model's adaptability to complex industrial scenarios, this application embodiment employs multi-modal spatiotemporal attention fusion technology. Based on the spatiotemporal attention mechanism and equipment operating conditions, it calculates the dynamic health value of the equipment, achieving a quantitative assessment of the equipment's health status.

[0031] In one embodiment, a spatial attention mechanism is used to dynamically learn the importance weights of different modal features, enabling adaptive fusion of multi-source data such as vibration, temperature, and current. Specifically, a fully connected layer scores the fault contribution of each modal feature, and then a softmax function is used to calculate the weights. For example, in a motor fault diagnosis scenario, the system automatically assigns a weight of 0.62 to the vibration signal (because it is sensitive to faults such as bearing wear and rotor imbalance), a weight of 0.25 to the temperature signal (reflecting problems such as winding overheating and poor lubrication), and a weight of 0.13 to the current signal (identifying electrical faults such as inter-turn short circuits and phase loss operation). The weight values ​​are dynamically adjusted according to the equipment type and fault mode. For example, in transformer monitoring, the temperature weight is increased to 0.4 to ensure that the fused features focus on the physical signals with the most diagnostic value. The temporal attention mechanism focuses on critical time windows preceding fault occurrence, addressing the issue of redundant information interference in long-term time-series data. After encoding temporal features using BiLSTM, the attention layer calculates the weight distribution for each time step, allocating higher weights to fault-sensitive periods. For example, in gear fracture fault diagnosis, the system automatically identifies stages where the energy of high-frequency vibration components continuously increases, increasing the weight of the last four time steps (corresponding to the 30 minutes before the fault occurs) to ensure the model focuses on feature changes during this period. Conversely, for stable data during steady-state operation, lower weights are assigned to reduce interference. The introduction of the temporal attention mechanism enables the model to accurately locate fault precursor information from historical data, improving the sensitivity of early warning.

[0032] Based on the aforementioned spatiotemporal attention mechanism, multi-dimensional feature vectors are weighted to obtain spatiotemporally weighted feature vectors. These spatiotemporally weighted feature vectors can more effectively reflect abnormal device states.

[0033] Based on spatiotemporal weighted feature vectors, the dynamic health value of the equipment is calculated in real time. The dynamic health value can overcome the limitations of traditional static thresholds and achieve accurate health assessment of the equipment throughout its entire life cycle by adjusting the weights of operating conditions.

[0034] Specifically, the system identifies the current operating stage of the equipment, which includes the startup stage, steady-state stage, and shutdown stage. Based on the characteristics of each operating stage, the system dynamically adjusts the corresponding operating condition weights for each modal feature in the spatiotemporal weighted feature vector. For example, if the equipment has a large vibration amplitude but is sensitive to temperature changes, the temperature weight can be set to 0.6, the vibration weight to 0.3, and the current weight to 0.1, focusing on monitoring whether the motor winding temperature rise is normal. The steady-state stage generally refers to the period from 5 minutes after startup to 5 minutes before shutdown. During this stage, the equipment operates stably, and the vibration signal contains the richest fault information; therefore, the vibration weight is increased to 0.7, the temperature weight to 0.2, and the current weight to 0.1. The shutdown stage refers to the last 5 minutes, where the decreasing trend of current and temperature better reflects the equipment status; therefore, the current weight is set to 0.5, the temperature weight to 0.4, and the vibration weight to 0.1. Operating condition identification is achieved through a rule engine, which can automatically determine the operating stage based on current amplitude and speed fluctuations.

[0035] After identifying the operating condition weights, the dynamic health value of the equipment is updated using exponential smoothing. Specifically, based on a preset exponential smoothing model, the standardized spatiotemporal weighted feature vector is weighted and fused according to the operating condition weights to obtain the corresponding dynamic health value of the equipment. This can be expressed as the following formula:

[0036] in, This represents the historical health value decay coefficient. This represents the working condition weight corresponding to the i-th modal feature. This represents the standardized spatiotemporal weighted feature vector. and These represent the dynamic health values ​​at times t and t-1, respectively.

[0037] Generally, the historical health value decay coefficient can be set to 0.85 to ensure that the dynamic health value reflects the gradual changes in equipment status. The dynamic health value ranges from 0 to 1; the closer the value is to 1, the healthier the equipment is, and the closer it is to 0, the higher the risk of failure. For example, when a wind turbine gearbox is operating normally, the HI value is stable at around 0.85. When early wear occurs, the vibration characteristics gradually increase, and the HI value begins to slowly decrease. After decreasing to a certain level, it will trigger an early warning at the corresponding level, enabling early detection of the fault.

[0038] It should be noted that the inference engine supports hot model updates, which can be achieved online through incremental model packages pushed from the cloud without interrupting device operation.

[0039] The edge-layer local storage unit continuously records equipment operation data, early warning records, maintenance results, and other information, forming a complete equipment operation and maintenance archive. Data is periodically uploaded to the cloud database via incremental synchronization to expand the training dataset. The cloud periodically evaluates the student model's accuracy, recall, F1-score, and other metrics at fixed intervals. When these metrics are abnormal, a model update mechanism is automatically triggered. Simultaneously, false positives and false negatives are analyzed to pinpoint model defects. The cloud collects newly generated data samples during equipment operation and maintenance and initiates incremental training based on these new data samples, updating only parameters in the teacher model related to newly added fault types. Based on the incrementally trained teacher model, an incremental model package is generated using SLIM knowledge distillation technology. The edge-layer model hot update function then performs hot updates on the student model deployed at the edge, enabling online upgrades of the student model without interrupting equipment operation. The system automatically records the model version iteration history and supports rollback to the optimal version.

[0040] S104: Trigger the warning strategy corresponding to the warning level based on the probability distribution of the fault type and / or the dynamic health value.

[0041] This application embodiment sets up a multi-level early warning mechanism to achieve refined management of fault risks. Therefore, based on the probability distribution of fault types and / or dynamic health values, the corresponding early warning strategy can be triggered to achieve closed-loop management from fault early warning to maintenance execution, and reduce operation and maintenance costs through hierarchical response and global optimization.

[0042] In one embodiment, if the probability distribution of the fault type indicates the corresponding fault type and the probability of the fault type exceeds the preset warning threshold, or the dynamic health value is lower than the first threshold, it indicates that the equipment has a serious fault risk and needs to be shut down immediately. At this time, a first-level warning needs to be triggered and an emergency shutdown maintenance process needs to be started. A high-frequency alarm is issued by activating the audible and visual alarm, and an emergency work order is pushed to the maintenance supervisor, including the fault location, possible causes and emergency handling steps, and the emergency spare parts allocation process is automatically started.

[0043] If the dynamic health value is between the first and second thresholds, it indicates that there is a potential fault in the equipment. At this time, a level two warning should be triggered, and a planned maintenance suggestion should be generated. The corresponding maintenance plan should be arranged in conjunction with the production schedule, and the required spare parts should be pre-allocated to the site at the same time to avoid delays in emergency procurement.

[0044] If the dynamic health value is between the second and third thresholds, it indicates that the device status is slightly abnormal. At this time, a level 3 warning is triggered. By increasing the monitoring frequency, strengthening the data collection density, and generating a health report daily, the trend of status changes can be tracked without immediate shutdown.

[0045] A local early warning device is installed at the edge, including an audible and visual alarm unit and a communication unit, for real-time feedback on abnormal equipment status. The audible and visual alarm unit uses LED warning lights and a buzzer, emitting audible and visual signals of different frequencies according to the warning level. The communication unit supports dual-link redundant transmission of 4G cellular network and industrial Ethernet, and can push warning information to the mobile terminals of maintenance personnel via SMS, email, industrial APP, etc. It also has a local data caching function, which can save critical data when the network is interrupted and automatically upload it after the network is restored.

[0046] S105: In response to the early warning strategy, generate the corresponding maintenance strategy for the equipment based on preset constraints.

[0047] Upon receiving an early warning, the cloud-based maintenance decision engine responds to the warning strategy by retrieving real-time data from the business system via API, obtaining constraints such as production scheduling, spare parts inventory, and personnel skill matching. Based on a mixed-integer programming model and the aforementioned constraints, it solves for the corresponding global maintenance strategy for the equipment. The edge layer receives the cloud-based maintenance plan and displays maintenance steps, spare parts lists, and risk warnings through visualization units to guide maintenance personnel in performing the operations. After maintenance is completed, the system automatically records data such as actual maintenance costs, downtime, and causes of failure, serving as a basis for subsequent model optimization and decision improvement.

[0048] In one embodiment, a mixed-integer programming model is constructed with the objective function of minimizing the total maintenance cost. The total maintenance cost includes repair costs and downtime losses. The mixed-integer programming model is specifically represented as follows: ,in, Indicates repair costs, Indicates downtime losses. This represents the loss coefficient, which can be dynamically adjusted according to the production plan, increasing during peak order periods. Constraints include at least spare parts inventory, personnel skills, and production scheduling. The spare parts inventory constraint is Sk ≥ Dk, where Sk is the quantity of spare parts in inventory and Dk is the quantity required for maintenance. This constraint ensures sufficient spare parts are available for maintenance. The personnel skills matching constraint is expressed as... Where m represents the number of tasks. The variable is 0-1, indicating whether personnel i matches task j, ensuring that maintenance personnel have the corresponding skills; the production scheduling constraint means that the maintenance time must not exceed the equipment idle window to avoid affecting normal production.

[0049] Based on the above constraints, by solving the mixed integer programming model, the optimal maintenance strategy for the equipment can be obtained. For example, a shutdown window of 2-4 am the next day can be arranged for the first-level early warning equipment, 5 sets of sealing rings can be allocated, and 2 senior maintenance workers can be assigned to perform maintenance.

[0050] The above are embodiments of the methods proposed in this application. Based on the same idea, some embodiments of this application also provide devices and non-volatile computer storage media corresponding to the above methods.

[0051] Figure 2 This is a schematic diagram of a device fault early warning device based on SLIM knowledge distillation, provided as an embodiment of this application. Figure 2 As shown, it includes: At least one processor; and, At least one processor-communication-connected memory; wherein, The memory stores instructions that can be executed by at least one processor, and the instructions, when executed by at least one processor, enable at least one processor to: Implement a device fault early warning method based on SLIM knowledge distillation as described above.

[0052] This application provides a non-volatile computer storage medium storing computer-executable instructions. When the computer executes the executable instructions, it implements the device fault early warning method based on SLIM knowledge distillation as described above.

[0053] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0054] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0055] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0056] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0057] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0058] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0059] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0060] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0061] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0062] It should also be noted that the terms "comprising," "including," or any other variations thereof 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 a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0063] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for early warning of equipment faults based on SLIM knowledge distillation, characterized in that, The method includes: Collect multimodal sensor data during equipment operation, preprocess and extract features from the multimodal sensor data to generate multidimensional feature vectors; The multi-dimensional feature vector is input into a pre-trained teacher model to output soft labels containing each fault category, and a lightweight student model is trained using the soft labels based on SLIM knowledge distillation technology. The student model is used to reason about the multi-dimensional feature vector to obtain the probability distribution of the fault type corresponding to the device, and the dynamic health value of the device is calculated based on the spatiotemporal attention mechanism and the device's operating condition. Based on the probability distribution of the fault type and / or the dynamic health value, trigger the warning strategy corresponding to the warning level; In response to the warning strategy, a maintenance strategy corresponding to the device is generated based on preset constraints.

2. The equipment fault early warning method based on SLIM knowledge distillation according to claim 1, characterized in that, Based on SLIM knowledge distillation technology, a lightweight student model is trained using the soft labels output by the teacher model, specifically including: Based on the SLIM knowledge distillation unit deployed at the edge layer, the soft labels output by the teacher model are loaded, and a lightweight student model is trained by optimizing the hybrid loss function based on the soft labels. The optimized hybrid loss function is expressed as follows: Where zs represents the student model output, and y represents the true fault label. denoted by , represents the softmax function, LCE is the cross-entropy loss used to constrain the consistency between zs and y, ps is the probability distribution of the student model output, pt represents the soft label distribution, LKL is the KL divergence loss used to measure the difference between ps and pt, and T represents the temperature parameter.

3. The equipment fault early warning method based on SLIM knowledge distillation according to claim 1, characterized in that, Based on the spatiotemporal attention mechanism and device operating conditions, the dynamic health value of the device is calculated, specifically including: Based on the spatiotemporal attention mechanism, the multi-dimensional feature vectors are weighted to obtain spatiotemporally weighted feature vectors; Identify the current operating phase of the device; wherein the operating phase includes a startup phase, a steady-state phase, and a shutdown phase; Based on the operational phase, assign corresponding operating condition weights to each modal feature in the spatiotemporal weighted feature vector; Based on a preset exponential smoothing model, the standardized spatiotemporal weighted feature vector is weighted and fused according to the operating condition weights to obtain the dynamic health value corresponding to the device.

4. The equipment fault early warning method based on SLIM knowledge distillation according to claim 3, characterized in that, Based on a preset exponential smoothing model, the spatiotemporal weighted feature vector is weighted and fused according to the operating condition weights to obtain the dynamic health value corresponding to the device, specifically including: The dynamic health value of the device is obtained using the following formula: in, This represents the historical health value decay coefficient. This represents the working condition weight corresponding to the i-th modal feature. This represents the standardized spatiotemporal weighted feature vector. and These represent the dynamic health values ​​at times t and t-1, respectively.

5. The equipment fault early warning method based on SLIM knowledge distillation according to claim 1, characterized in that, Based on the probability distribution of the fault type and / or the dynamic health value, a warning strategy corresponding to the warning level is triggered, specifically including: If the fault type probability distribution indicates the corresponding fault type, or if the dynamic health value is lower than the first threshold, a level one warning is triggered, and an emergency shutdown and maintenance process is initiated. If the dynamic health value is between the first threshold and the second threshold, a level two warning is triggered, and a planned maintenance suggestion is generated. If the dynamic health value is between the second threshold and the third threshold, a level 3 warning is triggered, and the monitoring frequency is increased.

6. The equipment fault early warning method based on SLIM knowledge distillation according to claim 1, characterized in that, Based on preset constraints, a maintenance strategy corresponding to the device is generated, specifically including: A mixed-integer programming model is constructed with the objective function of minimizing the total maintenance cost; wherein the total maintenance cost includes repair costs and downtime losses. Based on preset constraints, the mixed integer programming model is solved to obtain the optimal maintenance strategy under the constraints; wherein the constraints include at least spare parts inventory, personnel skills and production scheduling.

7. The equipment fault early warning method based on SLIM knowledge distillation according to claim 1, characterized in that, After generating the maintenance policy corresponding to the device, the method further includes: Collect newly generated data samples during the operation and maintenance of the equipment; The teacher model is incrementally trained using the data samples. Based on the incrementally trained teacher model, the student model deployed on the edge side is hot-updated using the SLIM knowledge distillation technique.

8. The equipment fault early warning method based on SLIM knowledge distillation according to claim 1, characterized in that, The teacher model is a model built based on the Stacking ensemble learning framework. The multi-dimensional feature vectors are input into the pre-trained teacher model to output soft labels containing each fault category, specifically including: By using multiple heterogeneous base learners, preliminary inference is performed on the multi-dimensional feature vectors to obtain a preliminary probability distribution; The initial probability distributions output by multiple base learners are fused using a meta-learner to generate soft labels containing each fault category.

9. A device fault early warning device based on SLIM knowledge distillation, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform a device fault early warning method based on SLIM knowledge distillation as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are set as follows: A device fault early warning method based on SLIM knowledge distillation as described in any one of claims 1-8.

Images