An industrial sensor noise intelligent detection method

Abstract

The application provides an industrial sensor noise intelligent detection method, relates to the technical field of industrial data processing and artificial intelligence, and comprises the following steps: firstly, a sensor knowledge graph is constructed, and the knowledge graph is embedded through TransE and then injected into a basic model; secondly, global time sequence features of data are extracted to generate prompt information, and a local feature vector is extracted through a sliding window; then, multi-source real data and artificially synthesized noise are combined, a high-quality training set is constructed through data enhancement; subsequently, the prompt, the features and the data are input into a model, the model capacity is dynamically adjusted according to the task complexity, and the LoRA strategy is adopted for efficient fine-tuning; finally, the to-be-detected data is input, and noise detection results containing position, type and confidence are output. The method combines knowledge guidance, feature expression and adaptive training, and is suitable for complex industrial scenes by taking into account the precision and generalization ability.

Description

Technical Field

[0001] This application belongs to the field of industrial data processing and artificial intelligence technology, specifically relating to an intelligent noise detection method for industrial sensors. Background Technology

[0002] In the process of modern industrial automation and intelligentization, industrial sensors, as core data acquisition devices, directly affect the reliability of key aspects such as equipment status monitoring, process control, fault prediction, and health management due to the accuracy of the data they collect. However, affected by multiple factors such as environmental interference, transmission errors, equipment aging, and electromagnetic noise, time-series data acquired by industrial sensors generally contain noise components. If this noise is not effectively identified and processed, it will seriously interfere with subsequent data analysis and decision-making systems, leading to misjudgments, misoperations, or even equipment downtime accidents.

[0003] Currently, noise detection in industrial sensors mainly relies on two technical approaches: one is traditional statistical methods and specialized models such as recurrent neural networks (e.g., LSTM), which, while having high detection accuracy in specific scenarios, are highly dependent on data training under specific sensor types or working conditions, have poor generalization ability, and are difficult to transfer to other devices or scenarios; the other is emerging large-scale time series models, which, while possessing good cross-scenario generalization ability, often have insufficient detection accuracy when faced with the complex noise patterns unique to industrial sensor data, making it difficult to meet the high accuracy requirements of industrial sites. Summary of the Invention

[0004] This application provides an intelligent noise detection method for industrial sensors to solve one of the aforementioned technical problems.

[0005] The technical solution adopted in this application is as follows:

[0006] This application provides an intelligent noise detection method for industrial sensors, including:

[0007] A knowledge graph of industrial sensors is constructed, and the TransE algorithm is used to embed the knowledge graph into the base model.

[0008] Global time-series features are extracted from industrial sensor time-series data. Targeted prompts are generated based on these global time-series features. The sliding window technique is used to divide the industrial sensor time-series data into multiple fixed-size windows. The time-series features of each window are calculated to generate a fixed-dimensional feature vector.

[0009] Collect multi-source industrial sensor data, generate simulated noise data through artificial synthesis methods, and perform data augmentation processing on the data to construct a training dataset;

[0010] The prompt information, feature vectors, and training dataset are input into the entity embedding base model for model training. The task complexity is evaluated based on the length and dimension of the industrial sensor data, the model capacity is dynamically adjusted, and / or the LoRA parameter fine-tuning strategy is used to train the model.

[0011] The industrial sensor data to be detected is input into the trained model, and the noise detection results are output.

[0012] According to one embodiment of this application, the knowledge graph of the industrial sensor includes sensor type, measured physical quantity, noise characteristics, normal range, and abnormal range.

[0013] According to one embodiment of this application, the global time series features include trend features, periodic features, volatility features, distribution features, and anomaly detection features.

[0014] According to one embodiment of this application, the size of the sliding window is dynamically determined based on the sampling frequency of industrial sensor data.

[0015] According to one embodiment of this application, the data augmentation process includes adding Gaussian noise, impulse noise, and periodic noise, and adjusting the signal strength.

[0016] According to one embodiment of this application, the dynamic adjustment of model capacity is based on dynamically activating the expert layer in a multi-expert model based on a task complexity threshold.

[0017] According to one embodiment of this application, the noise detection result includes noise location, noise type, and confidence level.

[0018] According to one embodiment of this application, the feature vector includes mean, standard deviation, minimum value, and maximum value.

[0019] A second aspect of this application provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps described in the method.

[0020] A third aspect of this application provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method as described.

[0021] Due to the adoption of the above technical solution, the beneficial effects achieved by this application are as follows:

[0022] This application constructs an industrial sensor knowledge graph and uses the TransE algorithm for embedding processing, injecting entity embedding representations into the basic model. This enables the model to call upon prior domain knowledge such as sensor type, measured physical quantity, and noise characteristics during the reasoning process, thereby enhancing the model's ability to understand industrial semantics, effectively distinguishing real signals from noise patterns, and improving the accuracy and interpretability of noise identification.

[0023] By extracting global temporal features and generating targeted prompts, and combining sliding window partitioning with local feature vector construction, the model can not only grasp the overall trend and periodicity of the data, but also capture subtle fluctuations and anomalies within the local window. This enhances the ability to perceive noise patterns in both macroscopic and microscopic dimensions, thereby improving detection sensitivity and robustness.

[0024] By collecting multi-source data, artificially synthesized noise, and data augmentation, a high-quality training dataset is constructed to ensure that the model is exposed to data covering a wide range of sensor types, noise patterns, and operating conditions during the training phase. This effectively improves the model's generalization ability in unknown scenarios and avoids overfitting or performance degradation caused by limited data.

[0025] By inputting prompts, feature vectors, and training data into a knowledge-embedded base model, and dynamically adjusting the model capacity and LoRA parameter fine-tuning strategy based on task complexity, the model's expressive power is ensured while effectively controlling computational resource consumption and overfitting risk. This enables the model to adapt to industrial data of different lengths and dimensions, achieving high-precision, lightweight, and transferable noise detection capabilities.

[0026] The final output includes the detection results of noise location, type and confidence level, which not only meets the industrial site's need to determine whether noise "occurs", but also supports more refined analysis such as "when it occurs", "what kind of noise" and "how reliable it is", providing a reliable basis for subsequent data cleaning, equipment maintenance or control strategy adjustment. Attached Figure Description

[0027] 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:

[0028] Figure 1 A flowchart illustrating an intelligent noise detection method for industrial sensors provided in this application embodiment;

[0029] Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0030] Figure label:

[0031] 810, Processor; 820, Communication interface; 830, Memory; 840, Communication bus. Detailed Implementation

[0032] To more clearly illustrate the overall concept of this application, a detailed explanation is provided below with reference to the accompanying drawings.

[0033] Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below. It should be noted that, unless otherwise specified, the embodiments of this application and the features thereof can be combined with each other.

[0034] In this application, unless otherwise expressly specified and limited, the "above" or "below" of the second feature can mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples.

[0035] Example 1

[0036] like Figure 1 As shown, an intelligent noise detection method for industrial sensors includes:

[0037] A knowledge graph of industrial sensors is constructed, and the TransE algorithm is used to embed the knowledge graph into the base model.

[0038] As mentioned above, constructing an industrial sensor knowledge graph involves organizing entities (such as "temperature sensor" and "pressure sensor"), attributes (such as "range 0-100℃" and "output signal 4-20mA"), and relationships (such as "measurement → temperature" and "susceptibility to → electromagnetic interference") related to industrial sensors into a semantic network in the form of "entity-relationship-entity" triples. This graph covers key areas of knowledge such as sensor type, physical quantity unit, typical noise sources, normal operating range, and abnormal behavior patterns.

[0039] Embedding using the TransE algorithm involves mapping each entity and relation in a knowledge graph to a vector representation in a low-dimensional continuous vector space. The core idea of ​​the TransE algorithm is that if a triple (head entity, relation, tail entity) exists, then in the vector space, the sum of the head entity vector and the relation vector should approximately equal the tail entity vector. This method effectively preserves the semantic relationships and logical structure between entities.

[0040] Embedding entity representations into the base model means using the generated entity vectors as additional inputs or low-level parameters and fusing them with the original input layer or intermediate layer of the base model. This allows the model to synchronously access domain knowledge when processing time series data, helping to determine whether the current data point conforms to the expected behavior pattern of the sensor type, thereby improving the accuracy and interpretability of noise identification.

[0041] For example, when processing data collected by a "turbine bearing vibration sensor":

[0042] The knowledge graph contains the entity "turbine bearing vibration sensor", whose attributes include "normal vibration range 0.1–2.5 mm / s" and relationships include "susceptible to mechanical imbalance noise" and "associated with speed signal".

[0043] After being embedded by TransE, the "turbine bearing vibration sensor" is mapped to a specific vector, and the "mechanical imbalance noise" is also mapped to another vector. The two have a specific distance and direction relationship in the vector space.

[0044] When the model receives a time series data in which the vibration value suddenly jumps to 5.0 mm / s, since this value is far beyond the numerical range associated with the "normal vibration range" entity and is close to the "mechanical imbalance noise" vector in semantic space, the model can combine this embedded knowledge to more accurately determine that the sudden increase is "noise" rather than "normal operating condition fluctuation" and improve the confidence output.

[0045] For example, when the model processes data from "thermocouple temperature sensor", if periodic high-frequency fluctuations occur, and the "thermocouple" entity in the knowledge graph is associated with "noise characteristics → low-frequency drift, not easily affected by high-frequency interference", then the model can determine that the high-frequency fluctuations are most likely transmission noise or electromagnetic interference, rather than actual temperature changes.

[0046] It should be noted that, in specific implementation scenarios, the above solutions can be further supplemented with "equipment model-sensor pairing relationship", "environmental conditions-noise impact relationship" (such as "high temperature environment → thermal drift noise enhancement"), "historical fault cases-noise feature mapping", etc., so that the knowledge graph can cover more granular industrial scenarios and improve the model's adaptation accuracy on specific production lines or equipment.

[0047] In specific implementation scenarios, based on the above solutions, and without deviating from the core path of "knowledge graph → vector embedding → injection model", improved knowledge embedding algorithms such as TransR, TransH, and RotatE can be used, or graph neural networks (GNNs) can be combined to perform multi-hop relationship reasoning in order to capture complex semantic relationships more precisely.

[0048] In specific implementation scenarios, based on the above solutions, entity embedding representations can not only be injected into the underlying input layer of the model, but also serve as key-value pairs in the attention mechanism, as a prerequisite knowledge module for the prompting project, or as a conditional bias parameter in the intermediate layer of the model, so as to adapt to the knowledge fusion needs of different model architectures.

[0049] In specific implementation scenarios, based on the above solutions, after the model is deployed, the knowledge graph content can be dynamically expanded or corrected according to the misjudgment cases or new device types reported on site, and the online updates of the embedded vectors can be triggered to achieve continuous evolution of the model knowledge base and maintain long-term detection accuracy.

[0050] In specific implementation scenarios, based on the above solutions, non-textual modal information such as sensor installation location diagrams, equipment structure diagrams, and process flow diagrams can be aligned with knowledge graph entities through multimodal embedding technology to form an enhanced knowledge representation of "text-graph association," which helps the model understand the noise propagation path in the spatial or process dimensions.

[0051] In specific implementation scenarios, based on the above scheme, the embedded knowledge representation can be used to not only determine "whether it is noise", but also to further distinguish the noise type (such as "electromagnetic interference", "mechanical vibration coupling" and "transmission packet loss"), providing a classification basis for subsequent noise source tracing and suppression strategies.

[0052] Global temporal features are extracted from industrial sensor time series data. Targeted prompts are generated based on these global temporal features. The sliding window technique is used to divide the industrial sensor time series data into multiple fixed-size windows. The temporal features of each window are calculated to generate a fixed-dimensional feature vector.

[0053] As mentioned above, indicators reflecting the overall characteristics of the data can be identified and extracted from time-series data collected by industrial sensors. These global time-series characteristics can include trend features, periodic features, fluctuation features, distribution features, and anomaly detection features. For example, by analyzing temperature sensor data, we can calculate the average temperature change trend over a period of time, or identify periodic temperature change patterns within a specific time period.

[0054] After extracting global temporal features, the next step is to generate targeted prompts based on these features. These prompts aim to help the model better understand and process the input data. For example, if a significant seasonal variation is detected in a segment of sensor data, the prompt might emphasize this, guiding the model to focus more on this characteristic for subsequent prediction or classification tasks.

[0055] The sliding window technique is used to segment the original time series data, creating a series of fixed-size time windows. For each window, various features of the time series data contained within it (such as mean, standard deviation, etc.) are calculated, and these features are then combined into a fixed-dimensional feature vector. This approach helps improve the model's ability to handle complex time series data and allows time series data of different lengths to be processed consistently.

[0056] For example, suppose we are processing data from a vibration sensor on a wind turbine:

[0057] Extracting global temporal features: In certain months of the year, vibration intensity shows a significant upward trend due to changes in wind speed. In addition, there may be some fixed periodic fluctuations, such as increased vibration at midday due to stronger winds.

[0058] Generate targeted prompts: Based on the above findings, prompts can be provided to the model to indicate that special attention should be paid to annual and daily periodic variations when analyzing data during this period, and that any deviation from these expected patterns may be a sign of potential problems.

[0059] Sliding window and feature vectorization: The vibration data of the whole year is divided into 30-day sliding windows. Statistical features such as mean and variance are calculated for each window, and these features are integrated into a fixed-dimensional feature vector for subsequent model training or prediction.

[0060] It should be noted that, in specific implementation scenarios, the size of the sliding window can be automatically adjusted based on the characteristics of different types of sensor data, in order to more accurately capture the dynamic characteristics of the data. For example, a shorter window can be used for rapidly changing data streams, while a longer window can be selected for slowly changing data.

[0061] In specific implementation scenarios, based on the above solutions, industry expertise can be leveraged to guide which time-series features should be prioritized or excluded, thereby improving the effectiveness and efficiency of feature extraction. For example, for certain types of sensor data, some features may be irrelevant and can therefore be ignored.

[0062] In specific implementation scenarios, in addition to the single-scale sliding window, a multi-scale feature extraction method can be introduced based on the above scheme, that is, features at different time scales can be considered simultaneously. This can capture more layers of information and help improve the performance of the model.

[0063] In specific implementation scenarios, based on the above solutions, after generating fixed-dimensional feature vectors, dimensionality reduction techniques (such as PCA principal component analysis) or feature selection algorithms can be used to reduce redundant information and improve the efficiency and effectiveness of subsequent model training.

[0064] In specific implementation scenarios, based on the above solutions, a mechanism can be designed for real-time monitoring scenarios to update feature vectors in real time when new data arrives, ensuring that the model always makes decisions based on the latest information.

[0065] Collect multi-source industrial sensor data, generate simulated noise data through artificial synthesis methods, and perform data augmentation processing on the data to construct a training dataset.

[0066] As mentioned above, collecting multi-source industrial sensor data refers to acquiring raw sensor data from multiple different sources, including but not limited to different equipment types (such as fans, pumps, and compressors), different sensor types (such as vibration, temperature, pressure, and current), different working environments (such as high temperature, high humidity, and strong electromagnetic interference), equipment from different manufacturers, and different production lines or factories. Through extensive collection, it ensures that the training data has sufficient coverage in terms of equipment, environment, and physical quantities, avoiding "domain bias" in the model due to a single training data source.

[0067] Generating simulated noise data through artificial synthesis involves artificially superimposing or injecting typical industrial noise patterns, such as Gaussian white noise, impulse noise, periodic interference, step transitions, transmission packet loss, and drift noise, onto real-world "clean" or "low-noise" data to construct "noisy samples." This process allows control over parameters such as noise type, intensity, duration, and frequency of occurrence, enabling controllable simulation of real-world noise and compensating for the lack of real noise samples or the difficulty in labeling them.

[0068] Data augmentation refers to applying transformation operations to existing data (including real and synthetic data) to expand the data volume and increase its diversity. For example, operations such as time stretching / compression, amplitude scaling, local inversion, adding random offsets, and segment splicing can be performed on the original signal to generate new samples with consistent semantics but different forms, thereby improving the model's robustness to changes in noise morphology.

[0069] Constructing a training dataset refers to uniformly formatting, standardizing, and labeling (such as noise location and type labels) the multi-source collected data, artificially synthesized noise data, and data-augmented samples, and then dividing them into training, validation, and test sets according to proportions to form a complete data resource that can be used for model training and evaluation.

[0070] For example, suppose we are building a training dataset for a "motor bearing vibration monitoring system":

[0071] Multi-source data collection: Vibration data were collected from three different factories, including ABB motors, Siemens motors, and a domestic brand motor; the collection environments included ambient temperature workshops, high-temperature smelting workshops, and humid coastal factories; the sensor types included ICP accelerometers and MEMS vibration sensors. This ensures that the model can be adapted to different brands, environments, and sensor types in the future.

[0072] Artificially synthesized noise: On a clean vibration data set of "normal operation", "50Hz power frequency interference" (simulating electromagnetic noise), "random pulse noise" (simulating mechanical shock), "slow drift noise" (simulating sensor aging) and other noises are superimposed. Each noise is set with a different signal-to-noise ratio to form 10 different noise intensity variants, which are used to train the model to identify different levels of interference.

[0073] Data augmentation processing: For the above noisy samples, random time axis cropping (retaining 80% of the length), random amplitude scaling (±20%), local signal inversion (simulating installation direction differences), and adding small time offsets (simulating asynchronous sampling) are performed to generate 5-10 enhanced samples from a single original sample, greatly expanding the amount of data and improving morphological diversity.

[0074] Dataset construction: The final dataset contains 20,000 samples. Each sample is labeled with "whether it contains noise", "noise type" and "noise start and end position", and is divided into training / validation / test sets in a 7:2:1 ratio for model training and performance evaluation.

[0075] It should be noted that, in specific implementation scenarios, a "noise template library" of common industrial noise patterns can be established based on the above solutions, such as "bearing wear noise template", "electromagnetic interference template", "cavitation pulse template", etc. This supports the rapid matching and injection of corresponding noise according to equipment type or fault mechanism, thereby improving the physical authenticity and task relevance of the synthesized data.

[0076] In specific implementation scenarios, based on the above solutions, during the artificial synthesis stage, noise signals that conform to physical laws can be generated by combining equipment dynamics models or signal propagation models (such as vibration transfer functions and heat conduction equations), rather than simple mathematical superposition, so that the synthesized data is closer to the real industrial scenario.

[0077] In specific implementation scenarios, based on the above scheme, during data augmentation, not only can the signal form be changed, but also the "operating condition labels" (such as speed, load, and temperature) can be changed in combination to generate composite scenario samples such as "high load + strong noise" and "low speed + weak drift", thereby improving the model's noise recognition capability under complex operating conditions.

[0078] In specific implementation scenarios, adversarial sample generation technology can be introduced on the basis of the above solutions to inject small perturbations into the original data, causing the model to misjudge it as "noise-free" or "other types of noise". Such samples are then added to the training set to improve the robustness of the model to adversarial interference or marginal cases.

[0079] In specific implementation scenarios, based on the above solutions, for some unlabeled industrial data, a preliminary trained model can be used for "pseudo-labeling" first, and then manually verified before being included in the training set, forming a closed loop of "model-assisted labeling → manual correction → retraining", which reduces labeling costs while ensuring quality.

[0080] In specific implementation scenarios, based on the above solutions, some sample combinations of "source devices with labels and target devices without labels" can be deliberately retained in the dataset. The model can be trained through domain adaptation or meta-learning strategies to enable it to "transfer from known device noise patterns to unknown devices" and enhance its generalization performance.

[0081] In specific implementation scenarios, based on the above solutions, new on-site data can be collected periodically after model deployment, and samples with "low model confidence" or "misjudgment by human feedback" can be automatically identified and added to the noise synthesis and enhancement process to achieve online iteration and continuous optimization of the training dataset.

[0082] In specific implementation scenarios, based on the above solution, a "noise sample proportion control module" can be set when constructing the dataset to ensure that the proportion of various types of noise (such as pulse type, drift type, and periodic type) is balanced in the dataset, avoid the model bias in recognizing the high-frequency noise type, and improve the balanced detection capability of multiple types of noise.

[0083] The prompt information, feature vectors, and training dataset are input into the entity embedding base model for model training. The task complexity is evaluated based on the length and dimension of the industrial sensor data, the model capacity is dynamically adjusted, and / or, a LoRA parameter fine-tuning strategy is used to train the model.

[0084] As described above, inputting the cue information, feature vectors, and training dataset into the basic model for entity embedding for model training refers to uniformly inputting three types of key inputs—"cue information" generated based on global temporal features (used to guide the model to focus on data characteristics), "fixed-dimensional feature vectors" generated by a sliding window (used to express local temporal structure), and the constructed "annotated training dataset" (containing real / synthetic noise samples)—into the basic model that has completed "industrial sensor knowledge graph embedding." During the training process, this model simultaneously utilizes domain knowledge, temporal context, and data diversity to learn how to accurately identify noise patterns.

[0085] Assessing task complexity based on the length and dimensionality of industrial sensor data involves evaluating two core attributes of the input data before or during training: first, the "length" of the time series (e.g., 1000 sampling points vs. 10000 sampling points); and second, the "dimensionality" (e.g., univariate temperature data vs. multivariate vibration + current + temperature combined data). Longer data and higher dimensionality indicate more complex temporal dependencies and more hidden noise patterns, thus increasing task complexity.

[0086] Dynamically adjusting model capacity refers to dynamically activating or suppressing computational resources within the model based on the aforementioned evaluation results. For example, when facing highly complex tasks, more "expert layers" or "attention heads" are activated; for low-complexity tasks, some modules are frozen to reduce computational overhead. Essentially, this achieves a "task-adaptive model structure," avoiding overcomplicating simple tasks or undercomplicating complex ones.

[0087] The LoRA parameter fine-tuning strategy for model training refers to training a model without directly updating all parameters of the base model (typically in the billions). Instead, only a set of low-rank matrices (Low-Rank Adaptation) are trained. These matrices are then superimposed on the original model weights as "bypasses," achieving lightweight fine-tuning. This strategy significantly reduces memory usage and training costs while retaining the general capabilities of the base model, making it particularly suitable for deployment environments with limited data and resources in industrial settings.

[0088] The use of "and / or" indicates that dynamic adjustment of model capacity and LoRA fine-tuning can be used independently or in combination to form a dual optimization mechanism of "structural adaptation + parameter efficiency".

[0089] For example, suppose we are training a noise detection model for "multivariable industrial pump set condition monitoring":

[0090] The model receives the following inputs: a prompt message “The pump data contains strong periodicity and occasional impulse noise”; a 10-dimensional feature vector (mean, standard deviation, peak-to-peak value, etc.) generated by a sliding window; a training dataset containing 20,000 samples from 5 pump types, 3 operating conditions, and 4 types of noise artificially injected; and knowledge entities such as “centrifugal pump - susceptible to cavitation noise” and “pressure sensor - range 0-10MPa” are embedded in the model's underlying layer.

[0091] The current training batch of data is "10,000 sampling points long, 3 dimensions (pressure + vibration + temperature)", which the system judges as a "high-complexity task" because long sequences need to capture long-term dependencies, and multi-dimensional variables need to model the coupling relationship between variables.

[0092] The system automatically activates three expert modules in the MoE (Mixture of Experts) architecture (only one is activated by default) and enables all 12 layers of Transformer attention mechanism to enhance the model's ability to model complex patterns.

[0093] During training, gradient updates are performed only on the LoRA bypass matrix (rank=8) of each attention matrix, keeping over 99% of the original model's parameters frozen. Training requires only 4GB of GPU memory, reducing training time by 60%, and achieving a noise detection F1 score of 94.2% on the test set, outperforming the 93.8% achieved by full parameter fine-tuning.

[0094] For example, when training a task of "short sequence univariate temperature sensor noise detection" (length=500, dimension=1):

[0095] The system is classified as "low complexity", so only one expert module is activated and the remaining 6 Transformer layers are disabled.

[0096] LoRA fine-tuning is still used, but the rank is reduced to 4;

[0097] The model training speed is increased by 3 times, the inference latency is reduced by 70%, and the accuracy is still maintained above 92%, meeting the deployment requirements of edge devices.

[0098] It should be noted that, in specific implementation scenarios, in addition to "length" and "dimension," the above solutions can also incorporate indicators such as "noise density," "signal entropy," and "spectral complexity" to comprehensively assess the task difficulty and achieve more refined resource allocation. For example, high-entropy signals can activate more frequency domain processing modules.

[0099] In specific implementation scenarios, in addition to the MoE expert layer, the following can also be dynamically adjusted based on the above scheme: the number of Transformer layers; the number of attention heads; the width of the feedforward network; and the number / depth of convolutional kernels. All adjustments are triggered based on the task complexity threshold and fall under the category of "structural adaptation".

[0100] In specific implementation scenarios, based on the above scheme, the rank of the LoRA matrix can be dynamically adjusted according to the task complexity or data scarcity. High rank (e.g., rank=16) is used for complex tasks, while low rank (e.g., rank=4) is used for simple tasks or small samples, achieving "parameter efficiency adaptation".

[0101] In specific implementation scenarios, LoRA can be combined with other efficient fine-tuning methods, such as Adapter, Prefix-Tuning, and BitFit, to form a "multi-strategy collaborative fine-tuning" approach. For example, LoRA can handle the temporal modeling part, while the Adapter can handle the knowledge fusion part, improving the efficiency of modular training.

[0102] In specific implementation scenarios, based on the above approach, a low-capacity model can be used to converge quickly in the early stages of training. As the number of training rounds increases, more expert layers or attention heads can be gradually unlocked to achieve a stable training process from simple to complex, thus avoiding initial oscillations in high-complexity models.

[0103] In specific implementation scenarios, based on the above solutions, LoRA can be used to fine-tune a lightweight model on edge devices, while a full-capacity model can be used in the cloud for knowledge distillation or gradient aggregation, thus realizing an industrial AI training paradigm of "lightweight deployment + centralized optimization".

[0104] In specific implementation scenarios, based on the above solution, during the training data loading stage, the sampling weight of "high-complexity samples" can be dynamically adjusted according to the complexity of the current batch of tasks to ensure that the model prioritizes learning the most difficult samples during peak resource allocation times, thereby improving training efficiency.

[0105] In specific implementation scenarios, based on the above solutions, when deploying industrial edge devices, the maximum number of active experts or layers can be dynamically limited according to the device's remaining power or heat dissipation capacity, so as to achieve "optimal energy efficiency training / inference" while ensuring the bottom line of accuracy.

[0106] The industrial sensor data to be detected is input into the trained model, and the noise detection results are output.

[0107] As mentioned above, inputting the industrial sensor data to be detected into the trained model refers to receiving real-time or historical time-series data collected by sensors from actual industrial equipment (such as fans, pumps, compressors, CNC machine tools, etc.) during the model deployment phase. Before input, this data undergoes a preprocessing process consistent with that of the training phase, including sliding window partitioning, local feature vector extraction, global time-series feature calculation, and prompt information generation, to ensure that the input format is consistent with that during training, thereby activating the model's learned knowledge and pattern recognition capabilities.

[0108] Outputting noise detection results refers to the structured and actionable noise identification information output after the model inference is completed. This typically includes, but is not limited to: the presence of noise (binary classification or probability output); the time and location of noise occurrence (start point, end point); noise type (e.g., impulse noise, drift noise, periodic interference, etc.); noise confidence level (the reliability of the model's judgment); and noise intensity level (mild, moderate, severe). The output format can be structured data (JSON, CSV), visual markers (e.g., highlighting noise segments), or alarm signals (trigger thresholds).

[0109] This step emphasizes the model's "end-to-end usability" and "industry interpretability," requiring not only that it be "detectable," but also that it be "explainable" and "usable."

[0110] For example, suppose a blast furnace cooling water circulating pump unit in a steel plant is in operation, and its vibration sensor continuously collects data:

[0111] Before the 10-minute vibration data collected in real time (sampling rate 1kHz, a total of 600,000 points) is sent into the model, the system automatically performs the following steps: slide division in 5-second windows (5000 points); calculate 10-dimensional features such as mean, standard deviation, peak-to-peak value, etc. for each window; extract global features such as "overall entropy value is high" and "existence of 50Hz periodic component"; generate prompts such as "suspected mechanical loosening + power frequency interference"; and call the embedded "centrifugal pump - vibration and noise knowledge" entity.

[0112] The model, combining prompts, window features, and knowledge graphs, identified an abnormal high-frequency pulse lasting 1.2 seconds in window number 32 (corresponding to the 160th second of operation). Its shape highly matched the "bearing ball damage noise template" and deviated from the normal vibration range of this pump type.

[0113] Output result:

[0114] The model outputs structured results:

[0115] {

[0116] "Noise exists": true,

[0117] "Noise start time": "160.0s",

[0118] "Noise end time": "161.2s",

[0119] Noise type: Mechanical impact impulse noise

[0120] Confidence level: 0.96

[0121] Intensity Level: Severe

[0122] Related Information: This may be caused by partial detachment of bearing balls. It is recommended to stop the machine for inspection within 72 hours.

[0123] }

[0124] The result is automatically pushed to the equipment health management system and triggers a yellow alert. Based on this, maintenance engineers can arrange maintenance in advance to avoid unplanned downtime.

[0125] For example, in the cleanroom temperature control system of a semiconductor factory:

[0126] Input a temperature sensor data, the model identifies the slowly rising "drift noise", and combines it with the entity "PT100 sensor - prone to zero drift after long-term use" in the knowledge graph to output "suggest calibrating the sensor" to avoid misjudging it as "abnormal ambient temperature".

[0127] It should be noted that, in specific implementation scenarios, in addition to the above-mentioned solutions, multi-dimensional confidence indicators such as "consistency of voting among multiple modules within the model", "matching degree with knowledge graph", and "matching degree with historical similar cases" can be output to assist manual judgment on whether to adopt the model results.

[0128] In specific implementation scenarios, based on the above solutions and the output noise type, the "noise source-equipment component" mapping relationship in the knowledge graph can be combined to automatically recommend possible root causes of faults (such as "pulse noise → bearing damage" and "50Hz interference → poor inverter grounding"), thereby enhancing the engineering guidance value of the results.

[0129] In specific implementation scenarios, based on the above solutions, and according to the noise intensity, duration, and frequency of occurrence, combined with the equipment operating parameters, an "impact assessment on equipment lifespan" or "interference level on control accuracy" can be output to support risk classification management.

[0130] In specific implementation scenarios, based on the above solution, noise interval markers, type color codes, and confidence transparency can be superimposed on the original time series curve to generate a "noise heat map," which makes it easier for engineers to quickly locate the problem period.

[0131] In specific implementation scenarios, the output format can be automatically switched based on the usage scenario, in addition to the above solution: outputting lightweight alarm signals on edge devices; outputting structured JSON in the central control system; and outputting report documents with traceability analysis on the expert workstation.

[0132] In specific implementation scenarios, based on the above solution, operators can be allowed to "confirm / correct" the model output results and send the correction results back to the training data pool to trigger online model fine-tuning or dataset updates, thereby achieving "human-machine collaborative continuous optimization".

[0133] In specific implementation scenarios, based on the above solution, when multiple related sensor data (such as vibration + temperature + current) are input simultaneously, the model can output "cross-sensor noise correlation analysis", such as "vibration pulses and current spikes occur synchronously, pointing to sudden changes in motor load", thereby improving system-level diagnostic capabilities.

[0134] In specific implementation scenarios, based on the above solutions, subsequent actions can be automatically recommended according to the type and severity of noise, such as "automatic filtering," "triggering equipment to reduce load," "generating maintenance work orders," and "notifying on-duty engineers," thus achieving a closed loop from "detection" to "decision-making" to "execution."

[0135] According to one embodiment of this application, the knowledge graph of the industrial sensor includes sensor type, measured physical quantity, noise characteristics, normal range, and abnormal range.

[0136] As mentioned above, sensor type: used to identify the type or model of a sensor, such as temperature sensor, pressure sensor, vibration sensor, current sensor, displacement sensor, etc. Different types of sensors differ in structure, working principle and applicable scenarios.

[0137] The measured physical quantity refers to the specific physical parameter measured by the sensor, such as a temperature sensor measuring "temperature in Celsius", a pressure sensor measuring "pressure in megapascals", and a vibration sensor measuring "acceleration or displacement". It usually includes the corresponding physical unit, such as °C, MPa, or mm / s. 2 Or mm, etc.

[0138] Noise characteristics: Describe the common noise sources and manifestations of this type of sensor in actual operation, such as "temperature sensors are susceptible to thermal drift noise", "vibration sensors are susceptible to power frequency interference in strong electromagnetic environments", and "pressure sensors are prone to step noise under rapid impact", etc., to help the model identify whether the noise conforms to the typical behavior of the sensor.

[0139] Normal range: refers to the reasonable range of values ​​that the sensor's output data should fall within when the equipment is operating normally, such as "the normal value range of a certain bearing vibration sensor is 0.5–2.0 mm / s", which is used as a benchmark for judging whether the data is abnormal.

[0140] Abnormal range: refers to the range of values ​​in which the sensor data exceeds normal operating conditions, which may indicate equipment failure or data distortion. For example, "vibration values ​​that are consistently higher than 5.0 mm / s are considered serious abnormalities." This is used to help the model distinguish noise from real fault signals, or to determine whether the noise has reached a level that requires intervention.

[0141] According to one embodiment of this application, the global time series features include trend features, periodic features, volatility features, distribution features, and anomaly detection features.

[0142] As mentioned above, trend characteristics reflect the overall direction of data changes over time, such as whether the data shows an upward, downward, or stable trend. Commonly used indicators include autocorrelation coefficient, mean, median, slope estimate, area under the curve, entropy value, etc., which are used to determine whether the data has systematic drift or slow changes.

[0143] Periodicity characteristics: These reflect whether there are recurring, regular fluctuations in the data, such as periodic patterns caused by equipment start-up and shutdown cycles, power frequency interference, and diurnal temperature variations. Commonly used indicators include peak-to-peak distance, quartile range, dominant frequency components of the spectrum, and autocorrelation peak spacing, which are used to identify regular signals or interference related to the normal operating rhythm.

[0144] Volatility characteristics: These reflect the degree of drastic change or stability of data over time, such as whether the data fluctuates frequently or remains stable. Commonly used indicators include standard deviation, variance, absolute deviation of the mean, absolute deviation of the median, sum of absolute differences, and energy distribution after wavelet transform, which are used to measure noise intensity or signal stability.

[0145] Distribution characteristics: Reflect the distribution pattern of data values ​​in the numerical space, such as whether it is concentrated, skewed, or has a long tail. Commonly used indicators include skewness, kurtosis, maximum value, minimum value, mode, quartiles, area under the curve, etc., used to determine whether the data conforms to the numerical distribution pattern that should be expected for this type of sensor.

[0146] Anomaly detection features: These reflect whether there are points or segments in the data that significantly deviate from the normal pattern, such as sudden pulses, step jumps, or continuous shifts. Commonly used indicators include maximum difference, maximum absolute difference, local extremum deviation, and entropy mutation value, which are used to initially locate possible noise areas or abnormal events.

[0147] According to one embodiment of this application, the size of the sliding window is dynamically determined based on the sampling frequency of industrial sensor data.

[0148] As mentioned above, when the sampling frequency is high (e.g., a vibration sensor collects 10,000 points per second), the data changes rapidly and is rich in detail, the sliding window should be set to a shorter time span (e.g., 0.1 seconds or 500 sampling points) to avoid the data within the window being too dense and the features becoming blurred, thereby capturing transient noise or high-frequency interference more accurately.

[0149] When the sampling frequency is low (e.g., a temperature sensor collects data at 1 point per minute), the data changes slowly and shows a strong trend. In this case, the sliding window should be set to a longer time span (e.g., 30 minutes or 60 sampling points) to ensure that the window contains enough data points to reflect low-frequency noise characteristics such as trend changes or slow drift.

[0150] This dynamic determination mechanism can be implemented through preset rules, such as: window size = reference time length × sampling frequency, where the "reference time length" can be set according to the sensor type or industrial scenario (e.g., 0.1 seconds for vibration reference, 10 minutes for temperature reference), thereby maintaining the consistent physical meaning of the window in the time dimension under different sampling rates.

[0151] In this way, the sliding window can adapt to the data characteristics of different sensors, and ensure the stability and effectiveness of subsequent feature extraction, thereby improving the model's generalization ability and detection accuracy in various industrial scenarios.

[0152] According to one embodiment of this application, the data augmentation process includes adding Gaussian noise, impulse noise, and periodic noise, and adjusting the signal strength.

[0153] As mentioned above, adding Gaussian noise refers to superimposing random noise that follows a normal distribution onto the original data to simulate continuous, small-amplitude random fluctuations in the sensor caused by thermal disturbances of electronic components, environmental electromagnetic interference, etc. The intensity of this noise can be controlled by adjusting the standard deviation, such as low intensity (σ=0.01), medium intensity (σ=0.05), and high intensity (σ=0.1).

[0154] Adding impulse noise: This refers to inserting instantaneous large-amplitude spikes at random or specific positions in a data sequence to simulate sudden interference caused by mechanical shocks, switching transients, poor contact, etc. Impulse noise can be set with different amplitudes, widths, and frequencies, such as single-point large-amplitude pulses or continuous multi-point small-amplitude pulses.

[0155] Adding periodic noise: This refers to superimposing sinusoidal or square wave components with fixed frequency and amplitude onto the original signal to simulate regular interference such as power frequency interference (e.g., 50Hz / 60Hz) and periodic vibrations caused by rotating parts of equipment. Different variations of periodic noise can be generated by adjusting parameters such as frequency, phase, and amplitude.

[0156] Adjusting signal strength: This refers to scaling the overall signal amplitude (e.g., amplifying by 1.2 times or attenuating by 0.8 times) after adding the aforementioned noise or based on the original data, to simulate sensor range drift, gain changes, signal attenuation, and other operating conditions, so that the model can adapt to noise performance under different signal strengths.

[0157] According to one embodiment of this application, the dynamic adjustment of model capacity is based on dynamically activating the expert layer in a multi-expert model based on a task complexity threshold.

[0158] As mentioned above, task complexity assessment: Before inputting data into the model, the system first analyzes the data characteristics of the task, mainly including the time series length (e.g., 1000 points vs. 10000 points) and data dimensions (univariate temperature vs. multivariate vibration + current + pressure), and can combine auxiliary indicators such as noise density and spectral complexity to comprehensively calculate a "task complexity score".

[0159] Complexity threshold setting: The system presets multiple complexity threshold ranges, such as "low complexity (score 0-30)", "medium complexity (score 31-70)" and "high complexity (score 71-100)", with each range corresponding to a different number of expert layer activations.

[0160] Dynamically activate the expert layer:

[0161] If the task complexity score falls into the "low complexity" range, only one expert layer will be activated to complete the reasoning with minimal computational overhead.

[0162] If it falls into the "medium complexity" range, then activate 2-3 expert layers to enhance the model's ability to recognize medium-difficulty noise patterns.

[0163] If it falls into the "high complexity" range, then all available expert layers (e.g., 4-8) are activated to fully utilize the model's capacity and cope with complex scenarios such as long sequences, multiple variables, and strong coupling noise.

[0164] Expert layer mechanism: The expert layer is an independent sub-network module in a multi-expert model (such as MoE, Mixture of Experts). Each expert layer can focus on processing specific types of data patterns or noise features. The dynamic activation mechanism enables the model to run lightweightly when facing simple tasks and automatically expand when facing complex tasks, achieving "on-demand allocation and elastic computing".

[0165] According to one embodiment of this application, the noise detection result includes noise location, noise type, and confidence level.

[0166] As mentioned above, noise location refers to the specific time point or time period in the original time series data where the noise occurs, usually expressed as sampling point number, timestamp, or start and end interval. For example, "the noise starts at sampling point 1250 and ends at sampling point 1280" or "the noise occurs between 14:32:15 and 14:32:18". This information is used to accurately locate abnormal data segments, facilitating data extraction, visualization, or local cleaning.

[0167] Noise type refers to the category of noise identified by the model, which is classified based on preset or learned noise patterns, such as "Gaussian white noise," "impulse noise," "periodic interference," "drift noise," and "transmission packet loss." This classification result can be matched with noise characteristics in a knowledge graph to help users understand the source or cause of noise, providing a basis for subsequent noise suppression or equipment maintenance.

[0168] Confidence level refers to the degree of certainty with which a model makes a judgment about current noise, and is usually expressed as a numerical value or percentage between 0 and 1 (such as 0.92 or 92%). Confidence level reflects the certainty of the model's internal judgment. High confidence level indicates that the model is highly consistent in feature matching, knowledge retrieval, and pattern recognition, while low confidence level suggests the possibility of marginal cases or data ambiguity, requiring manual verification.

[0169] The above three elements together constitute the basic output structure of noise detection results, ensuring that the results are not only "conclusive" but also "locatable, categorizable, and assessable," meeting the requirements of interpretability, traceability, and operability in industrial scenarios. This output structure can be further expanded, but it should at least include these three core pieces of information to support effective decision-making.

[0170] According to one embodiment of this application, the feature vector includes mean, standard deviation, minimum value, and maximum value.

[0171] As mentioned above, the mean reflects the average level or central trend of the data within the window and is used to determine whether the overall data deviates from the normal baseline.

[0172] Standard deviation: Reflects the degree of dispersion of data points around the mean within a window. It is used to measure the volatility or stability of data. The larger the standard deviation, the more unstable the data is, and there may be noise or anomalies.

[0173] Minimum value: Reflects the lower limit of the data within this window, used to capture sudden negative pulses, signal loss, or abnormal sensor lower limit.

[0174] Maximum value: Reflects the upper limit of the data within this window, used to identify positive spikes, over-range signals, or transient impact noise.

[0175] These four statistics together constitute a concise and effective local feature description, capable of stably representing the basic shape and range of variation of data within a window without relying on complex calculations. This feature vector has a unified structure and fixed dimensions, facilitating batch processing and model input, while retaining sufficient information to enable the model to help determine the presence of noise based on local statistical characteristics.

[0176] A second aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any of the embodiments of the first aspect above.

[0177] Figure 2 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 2 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 may call logical instructions in the memory 830 to execute the method in any of the embodiments of the first aspect described above, the method including:

[0178] A knowledge graph of industrial sensors is constructed, and the TransE algorithm is used to embed the knowledge graph into the base model.

[0179] Global time-series features are extracted from industrial sensor time-series data. Targeted prompts are generated based on these global time-series features. The sliding window technique is used to divide the industrial sensor time-series data into multiple fixed-size windows. The time-series features of each window are calculated to generate a fixed-dimensional feature vector.

[0180] Collect multi-source industrial sensor data, generate simulated noise data through artificial synthesis methods, and perform data augmentation processing on the data to construct a training dataset;

[0181] The prompt information, feature vectors, and training dataset are input into the entity embedding base model for model training. The task complexity is evaluated based on the length and dimension of the industrial sensor data, the model capacity is dynamically adjusted, and / or the LoRA parameter fine-tuning strategy is used to train the model.

[0182] The industrial sensor data to be detected is input into the trained model, and the noise detection results are output.

[0183] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0184] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer being able to perform the methods provided by the above methods, the method comprising:

[0185] A knowledge graph of industrial sensors is constructed, and the TransE algorithm is used to embed the knowledge graph into the base model.

[0186] Global time-series features are extracted from industrial sensor time-series data. Targeted prompts are generated based on these global time-series features. The sliding window technique is used to divide the industrial sensor time-series data into multiple fixed-size windows. The time-series features of each window are calculated to generate a fixed-dimensional feature vector.

[0187] Collect multi-source industrial sensor data, generate simulated noise data through artificial synthesis methods, and perform data augmentation processing on the data to construct a training dataset;

[0188] The prompt information, feature vectors, and training dataset are input into the entity embedding base model for model training. The task complexity is evaluated based on the length and dimension of the industrial sensor data, the model capacity is dynamically adjusted, and / or the LoRA parameter fine-tuning strategy is used to train the model.

[0189] The industrial sensor data to be detected is input into the trained model, and the noise detection results are output.

[0190] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the methods provided by the above methods, the method comprising:

[0191] A knowledge graph of industrial sensors is constructed, and the TransE algorithm is used to embed the knowledge graph into the base model.

[0192] Global time-series features are extracted from industrial sensor time-series data. Targeted prompts are generated based on these global time-series features. The sliding window technique is used to divide the industrial sensor time-series data into multiple fixed-size windows. The time-series features of each window are calculated to generate a fixed-dimensional feature vector.

[0193] Collect multi-source industrial sensor data, generate simulated noise data through artificial synthesis methods, and perform data augmentation processing on the data to construct a training dataset;

[0194] The prompt information, feature vectors, and training dataset are input into the entity embedding base model for model training. The task complexity is evaluated based on the length and dimension of the industrial sensor data, the model capacity is dynamically adjusted, and / or the LoRA parameter fine-tuning strategy is used to train the model.

[0195] The industrial sensor data to be detected is input into the trained model, and the noise detection results are output.

[0196] For any parts not mentioned in this application, existing technologies may be used or referenced.

[0197] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0198] 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. An industrial sensor noise intelligent detection method, characterized in that, Comprise: A knowledge graph of industrial sensors is constructed, and a TransE algorithm is used to embed the knowledge graph, and the generated entity embedding representation is embedded into a base model, specifically: the generated entity embedding representation is used as additional input or bottom layer parameters, and is fused with the original input layer or intermediate layer of the base model, so that the base model can call domain knowledge synchronously when processing time series data, and assist in judging whether the current data point conforms to the behavior pattern of the sensor type; The base model is a pre-trained model based on a Transformer architecture and using a hybrid expert structure; Global time series features are extracted based on industrial sensor time series data, and targeted prompt information is generated based on the global time series features, a sliding window technique is used to divide the industrial sensor time series data into multiple fixed-size windows, the time series features of each window are calculated, and a fixed-dimension feature vector is generated; Multiple source industrial sensor data is collected, synthetic noise data is generated by an artificial synthesis method, and data enhancement processing is performed on the data to construct a training data set; The prompt information, feature vector and training data set are input into the entity-embedded base model for model training, the task complexity is evaluated according to the length and dimension of the industrial sensor data, the model capacity is dynamically adjusted, and / or the LoRA parameter fine-tuning strategy is used to train the model; The prompt information is used to guide the model to focus on data characteristics; The industrial sensor data to be detected is input into the trained model, and a noise detection result is output.

2. The method of claim 1, wherein, The knowledge graph of the industrial sensor includes sensor type, measured physical quantity, noise characteristics, normal range and abnormal range.

3. The method of claim 1, wherein, The global time series features include trend features, periodicity features, volatility features, distribution features and anomaly detection features.

4. The method of claim 1, wherein, The size of the sliding window is dynamically determined according to the sampling frequency of the industrial sensor data.

5. The method of claim 1, wherein, The data enhancement processing includes adding Gaussian noise, impulse noise and periodic noise, and adjusting the signal strength.

6. The method of claim 1, wherein, The dynamic adjustment of the model capacity is based on the task complexity threshold to dynamically activate the expert layers in the multi-expert model.

7. The method of claim 1, wherein, The noise detection result includes noise position, noise type and confidence.

8. The method of claim 1, wherein, The feature vector contains mean, standard deviation, minimum value and maximum value.

9. A computer-readable storage medium having stored thereon a program, characterized in that, The program is executed by the processor to implement the steps in the method of any one of claims 1-8.

10. An electronic device comprising a memory, a processor, and a program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps in the method of any one of claims 1-8.

Images