Intelligent noise detection method for industrial sensor

Abstract

The invention provides an industrial sensor noise intelligent detection method, and relates to the technical field of industrial data processing and artificial intelligence, and the method comprises the steps: firstly constructing a sensor knowledge graph, and injecting the sensor knowledge graph into a basic model after TransE embedding; secondly, extracting global time sequence features of the data to generate prompt information, and extracting local feature vectors through a sliding window; combining multi-source real data with artificially synthesized noise, and constructing a high-quality training set through data enhancement; then, prompts, features and data are input into the model, the model capacity is dynamically adjusted according to task complexity, and efficient fine adjustment is performed by adopting a LoRA strategy; and finally, inputting to-be-detected data, and outputting a noise detection result containing the position, the type and the confidence coefficient. The method integrates knowledge guidance, feature expression and adaptive training, considers both precision and generalization ability, and is suitable for complex industrial scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial data processing and artificial intelligence, and particularly relates to an industrial sensor noise intelligent detection method. BACKGROUND

[0002] In the process of modern industrial automation and intelligentization, as the core equipment of data acquisition, the accuracy of industrial sensor data acquisition is directly related to the reliability of key links such as equipment state monitoring, process control, fault prediction and health management. However, affected by multiple factors such as environmental interference, transmission error, equipment aging and electromagnetic noise, there are generally noise components in the time series data collected by industrial sensors. If these noises are not effectively identified and processed, they will seriously interfere with the subsequent data analysis and decision system, leading to misjudgment, misoperation and even equipment shutdown accidents.

[0003] Currently, industrial sensor noise detection mainly relies on two types of technical paths: one is traditional statistical methods and special models such as recurrent neural networks (such as LSTM), which have high detection accuracy in specific scenarios, but their generalization ability is poor, and it is difficult to migrate to other devices or scenes; the other is emerging large-scale time series models, which have good cross-scene generalization ability, but when faced with the complex noise patterns specific to industrial sensor data, the detection accuracy is often insufficient, and it is difficult to meet the requirements of high accuracy in industrial field. SUMMARY

[0004] The application provides an industrial sensor noise intelligent detection method to solve one of the above technical problems.

[0005] The technical scheme adopted by the application is: The embodiment of the application provides an industrial sensor noise intelligent detection method, comprising: Constructing a knowledge graph of an industrial sensor, using a TransE algorithm to perform embedding processing on the knowledge graph, and embedding the generated entity embedding representation into a basic model; Based on the industrial sensor time series data to extract global time sequence features, generating targeted prompt information based on the global time sequence features, using a sliding window technology to divide the industrial sensor time series data into a plurality of fixed-size windows, calculating the time sequence features of each window, and generating a fixed-dimension feature vector; Collecting multi-source industrial sensor data, generating simulated noise data through an artificial synthesis method, and performing data enhancement processing on the data to construct a training data set; The prompt information, the feature vector and the training data set are input into a base model of entity embedding 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 model is trained by using a LoRA parameter fine-tuning strategy; The industrial sensor data to be detected is input into the trained model, and a noise detection result is output.

[0006] According to an embodiment of the present application, the knowledge graph of the industrial sensor includes sensor type, measured physical quantity, noise characteristic, normal range and abnormal range. According to an embodiment of the present application, the global timing feature includes trend feature, periodicity feature, volatility feature, distribution feature and anomaly detection feature.

[0007] According to an embodiment of the present application, the size of the sliding window is dynamically determined according to the sampling frequency of the industrial sensor data.

[0008] According to an embodiment of the present application, the data enhancement processing includes adding Gaussian noise, impulse noise and periodic noise, and adjusting signal strength.

[0009] According to an embodiment of the present application, the dynamic adjustment of the model capacity is to dynamically activate the expert layer in the multi-expert model based on the task complexity threshold.

[0010] According to an embodiment of the present application, the noise detection result includes noise position, noise type and confidence.

[0011] According to an embodiment of the present application, the feature vector contains mean, standard deviation, minimum value and maximum value.

[0012] The second aspect embodiment of the present application provides a computer readable storage medium having a program stored thereon, the program being executed by a processor to implement the steps in the method.

[0013] The third aspect embodiment of the present application provides an electronic device, including a memory, a processor and a program stored on the memory and executable on the processor, and the processor executes the program to implement the steps in the method.

[0014] As the above technical solutions are adopted, the present application has the following beneficial effects: The present application constructs an industrial sensor knowledge graph and uses a TransE algorithm for embedding processing, injects entity embedding representation into a base model, so that the model can call prior domain knowledge such as sensor type, measured physical quantity and noise characteristic in the reasoning process, thereby enhancing the model's understanding ability of industrial semantics, effectively distinguishing real signals from noise patterns, and improving the accuracy and interpretability of noise recognition.

[0015] By extracting global timing features and generating targeted prompt information, combined with sliding window division and local feature vector construction, the model can grasp the overall trend and periodicity of the data, and capture the subtle fluctuations and abnormal points within the local window, thereby enhancing the perception of noise patterns in both macro and micro dimensions, and improving the detection sensitivity and robustness.

[0016] By collecting multi-source data, artificially synthesizing noise and data enhancement processing to construct high-quality training data set, it ensures that the model is exposed to data covering a wide range of sensor types, noise patterns and working conditions during the training stage, thereby effectively improving the generalization ability of the model in unknown scenarios and avoiding overfitting or performance degradation caused by single data.

[0017] By inputting prompt information, feature vectors and training data into the basic model embedded with knowledge, and combining task complexity to dynamically adjust model capacity and LoRA parameter fine-tuning strategy, the model can adapt to industrial data of different lengths and dimensions while effectively controlling the consumption of computing resources and the risk of overfitting, achieving high-precision, lightweight and transferable noise detection capability.

[0018] The final output includes noise position, type and confidence, not only meeting the judgment needs of industrial sites on whether noise occurs, but also further supporting detailed analysis such as when it occurs, what kind of noise, and how reliable it is, providing reliable basis for subsequent data cleaning, equipment maintenance or control strategy adjustment. BRIEF DESCRIPTION OF DRAWINGS

[0019] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A flowchart of an industrial sensor noise intelligent detection method provided by an embodiment of the present application is shown. Figure 2 A structural schematic diagram of an electronic device provided by an embodiment of the present application is shown.

[0020] Reference signs: 810, processor; 820, communication interface; 830, memory; 840, communication bus. DETAILED DESCRIPTION

[0021] In order to more clearly explain the overall concept of the present application, the following will be described in detail in an exemplary manner with reference to the drawings of the specification.

[0022] 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.

[0023] 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.

[0024] Example 1 like Figure 1 As shown, an intelligent noise detection method for industrial sensors includes: A knowledge graph of industrial sensors is constructed, and the TransE algorithm is used to embed the knowledge graph into the base model.

[0025] 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.

[0026] 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.

[0027] Embedding the entity embedding into the base model: refers to the above-mentioned generated entity vector as an additional input or bottom layer parameter, and is fused with the original input layer or intermediate layer of the base model, so that the model can call the domain knowledge synchronously when processing time series data, assist in judging whether the current data point conforms to the behavior mode of the sensor type, thereby improving the accuracy and interpretability of noise identification.

[0028] For example, when processing data collected by a "turbine bearing vibration sensor": The knowledge graph contains the entity "turbine bearing vibration sensor", whose attributes include "normal vibration range 0.1-2.5mm / s", and its relationships include "susceptible to mechanical imbalance noise" and "associated with speed signal".

[0029] After TransE embedding, "turbine bearing vibration sensor" is mapped to a specific vector, and "mechanical imbalance noise" is also mapped to another vector, which has a specific distance and direction relationship in the vector space.

[0030] When the model receives a time series data with a sudden jump in vibration value to 5.0mm / s, since the value is far beyond the value range associated with the "normal vibration range" entity, and is close to the "mechanical imbalance noise" vector in the semantic space, the model can combine this embedded knowledge to more accurately determine that the sudden increase is "noise" rather than "normal working condition fluctuation", and improve the confidence output.

[0031] For another example, when the model processes data of a "thermocouple temperature sensor", if periodic high-frequency fluctuations occur, and the knowledge graph associates "thermocouple" entity with "noise characteristics: low-frequency drift, not susceptible to high-frequency interference", the model can determine that the high-frequency fluctuations are most likely transmission noise or electromagnetic interference, rather than real temperature changes.

[0032] It should be noted that in specific implementation scenarios, the above-mentioned scheme can be further incorporated with "device model-sensor pairing relationship", "environmental condition-noise influence relationship" (such as "high temperature environment → thermal drift noise enhancement"), "historical fault case-noise feature mapping", etc., so as to cover more fine-grained industrial scenarios for the knowledge graph, and improve the adaptation accuracy of the model on a specific production line or device.

[0033] In specific implementation scenarios, the above-mentioned scheme can be further incorporated with "device model-sensor pairing relationship", "environmental condition-noise influence relationship" (such as "high temperature environment → thermal drift noise enhancement"), "historical fault case-noise feature mapping", etc., so as to cover more fine-grained industrial scenarios for the knowledge graph, and improve the adaptation accuracy of the model on a specific production line or device.

[0034] In specific implementation scenarios, on the basis of the above scheme, the entity embedding can not only be injected into the input layer of the model, but also be used as the key-value pair in the attention mechanism, as the prior knowledge module for prompting engineering, or as the conditional bias parameter of the intermediate layer of the model, to adapt to the needs of knowledge fusion of different model architectures.

[0035] In specific implementation scenarios, on the basis of the above scheme, the knowledge graph content can be dynamically expanded or corrected according to the misjudgment cases or newly added equipment types in the field after the model is deployed, and the online update of the embedding vector is triggered to realize the continuous evolution of the model knowledge base and maintain long-term detection accuracy.

[0036] In specific implementation scenarios, on the basis of the above scheme, non-textual modal information such as sensor installation position map, equipment structure map, and process flow chart can be aligned with the entities of the knowledge graph through multi-modal embedding technology to form an enhanced knowledge representation of "text-image association" to assist the model in understanding the noise propagation path in the spatial or process dimension.

[0037] In specific implementation scenarios, on the basis of the above scheme, using the embedded knowledge representation, we can not only determine "whether it is noise", but also further distinguish the noise type (such as "electromagnetic interference", "mechanical vibration coupling", "transmission packet loss") to provide classification basis for subsequent noise tracing and suppression strategies.

[0038] Based on the time series data of industrial sensors to extract global time series features, based on the global time series features to generate targeted prompt information, using sliding window technology to divide the industrial sensor time series data into multiple fixed-size windows, calculate the time series features of each window, and generate a fixed-dimension feature vector.

[0039] As mentioned above, indicators that can reflect the overall characteristics of the data are identified and extracted from the time series data collected from industrial sensors. These global time series features can include trend features, periodicity features, volatility features, distribution features, and anomaly detection features, etc. For example, by analyzing the data of a temperature sensor, we can calculate the average temperature change trend over a period of time, or identify the periodic change pattern of the temperature in a specific time period.

[0040] After extracting the global time series features, the next step is to generate targeted prompt information based on these features. The prompt information aims to help the model better understand and process the input data. For example, if it is detected that there is a significant seasonal variation in a certain segment of sensor data, the prompt information may emphasize this point, thereby guiding the model to pay more attention to this characteristic for subsequent prediction or classification tasks.

[0041] The original time series data is segmented using a sliding window technique, creating a series of fixed-size time windows. For each window, various features of the time series data contained within it are calculated (such as mean, standard deviation, etc.), which are then combined into a fixed-dimension 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 consistently processed.

[0042] For example, suppose we are processing data from a wind turbine vibration sensor: Extract global temporal features: In certain months of the year, due to changes in wind speed, the vibration intensity shows a clear upward trend. In addition, there may be some fixed periodic fluctuations, such as increased vibration at noon due to increased wind speed.

[0043] Generate targeted prompt information: Based on the above findings, the model can be prompted to pay special attention to annual and daily periodic changes when analyzing this period of data, and any behavior deviating from these expected patterns may be a sign of potential problems.

[0044] Sliding window and feature vectorization: Divide the vibration data of a whole year into 30-day sliding windows by day, calculate statistical features such as mean, variance, etc. for each window, and integrate these features into a fixed-dimension feature vector for subsequent model training or prediction process.

[0045] It should be noted that in specific implementation scenarios, the size of the sliding window can also be automatically adjusted based on the characteristics of different types of sensor data to more accurately capture the dynamic characteristics of the data. For example, for rapidly changing data streams, a shorter window can be used; for slowly changing data, a longer window can be selected.

[0046] In specific implementation scenarios, in addition to the above scheme, professional knowledge in the industrial field can be used to guide which time series features should be prioritized or excluded to improve the effectiveness and efficiency of feature extraction. For example, for some types of sensor data, certain features may not be relevant and can be selected to be ignored.

[0047] In specific implementation scenarios, in addition to the above scheme, in addition to single-scale sliding windows, multi-scale feature extraction methods can also be introduced, that is, features at different time scales are considered simultaneously. This can capture more levels of information and help improve the performance of the model.

[0048] In specific implementation scenarios, on the basis of the above scheme, after generating a fixed-dimension feature vector, a dimension reduction technique (such as PCA principal component analysis) or a feature selection algorithm can be used to reduce redundant information and improve the efficiency and effectiveness of subsequent model training.

[0049] In specific implementation scenarios, on the basis of the above scheme, for real-time monitoring scenarios, a mechanism can be designed to update the feature vector in real time when new data arrives, ensuring that the model always makes decisions based on the latest information.

[0050] Collect multi-source industrial sensor data, generate simulated noise data through artificial synthesis method, and perform data enhancement processing on the data to construct a training data set.

[0051] As mentioned above, collecting multi-source industrial sensor data means collecting raw sensor data from multiple different sources, including but not limited to different device types (such as fans, pumps, compressors), different sensor types (such as vibration, temperature, pressure, current), different working environments (such as high temperature, high humidity, strong electromagnetic interference), different manufacturers' equipment, different production lines or factories. By extensive collection, ensure that the training data has sufficient coverage in the dimensions of equipment, environment, and physical quantities, avoiding "domain drift" caused by single training data.

[0052] Generating simulated noise data through artificial synthesis method means artificially superimposing or injecting typical industrial noise patterns, such as Gaussian white noise, impulse noise, periodic interference, step jump, transmission packet loss, drift noise, etc., based on "clean" or "low noise" data collected in reality, to construct "noisy samples". This process can control noise type, intensity, duration, frequency, etc., to achieve controllable simulation of real-world noise, and to address the problem of insufficient or difficult-to-label real noise samples.

[0053] Performing data enhancement processing on the data means applying transformation operations on existing data (including real data and synthetic data) to expand the data volume and improve data diversity. For example, time stretching / compression, amplitude scaling, local inversion, adding random offset, segment splicing, etc. are performed on the original signal to generate new samples with consistent semantics but different forms, improving the robustness of the model to noise form changes.

[0054] Constructing a training data set means that after formatting, standardizing, and labeling (such as noise position, type label) the above-mentioned multi-source collected data, artificially synthesized noise data, and data enhancement samples, the training set, validation set, and test set are divided in proportion to form a complete data resource that can be used for model training and evaluation.

[0055] For example, assume that a training data set is being constructed for a "motor bearing vibration monitoring system": Multi-source data collection: vibration data was collected from three different factories, including ABB motors, Siemens motors, and a domestic brand motor; the collection environment includes normal temperature workshops, high temperature smelting workshops, and humid coastal factories; sensor types include ICP type accelerometers, MEMS vibration sensors, etc. This ensures that the model can be adapted to different brands, different environments, and different sensor types in the future.

[0056] Artificially synthesized noise: On a piece of "normal operation" clean vibration data, "50Hz power frequency interference" (simulating electromagnetic noise), "random pulse noise" (simulating mechanical impact), "slow drift noise" (simulating sensor aging), etc. are superimposed, and different signal-to-noise ratios are set for each noise to form 10 different noise intensity variants for training the model to identify different levels of interference.

[0057] Data enhancement processing: For the above noisy samples, random cutting on the time axis (retaining 80% length), random scaling of amplitude (±20%), local signal inversion (simulating installation direction difference), adding a small time offset (simulating sampling out of sync), etc. operations, so that a single original sample can derive 5-10 enhanced samples, greatly expanding the data volume while improving the diversity of the form.

[0058] Building a dataset: a dataset containing 20,000 samples is finally formed, each sample is labeled with "whether it contains noise", "noise type", "noise start and end position", and divided into training / validation / test sets in a ratio of 7:2:1 for model training and performance evaluation.

[0059] It should be noted that in specific implementation scenarios, on the basis of the above scheme, a "noise template library" of common industrial noise patterns can be established, such as "bearing wear noise template", "electromagnetic interference template", "cavitation pulse template", etc., which supports quick matching and injection of corresponding noise according to device type or fault mechanism, improving the physical authenticity and task relevance of synthetic data.

[0060] In specific implementation scenarios, on the basis of the above scheme, in the artificial synthesis stage, the noise signal can be generated in accordance with the physical law by combining the device dynamics model or signal propagation model (such as vibration transfer function, heat conduction equation), rather than simply mathematically superimposing, so that the synthetic data is closer to the real industrial scene.

[0061] In specific implementation scenarios, on the basis of the above scheme, during data enhancement, not only the signal form is transformed, but also the "working condition label" (such as speed, load, temperature) is transformed, generating "high load + strong noise", "low speed + weak drift" and other complex scene samples, to improve the noise recognition ability of the model under complex working condition coupling.

[0062] In a specific implementation scenario, on the basis of the above scheme, an adversarial sample generation technology can be introduced to inject a slight perturbation into the original data to make the model misjudge as "no noise" or "other type of noise", and then add such samples to the training set to improve the robustness of the model to adversarial interference or edge cases.

[0063] In a specific implementation scenario, on the basis of the above scheme, for part of the unlabeled industrial data, the model can be first "pseudo-labeled" using the preliminary trained model, and then incorporated into the training set after manual correction to form a closed loop of "model assisted labeling-human correction-retraining", thereby reducing the labeling cost while ensuring the quality.

[0064] In a specific implementation scenario, on the basis of the above scheme, part of the sample combination of "source device with label and target device without label" can be deliberately reserved in the data set, and the model can be trained through domain adaptation or meta-learning strategy to enable it to have the ability of "migrating from known device noise pattern to unknown device", thereby strengthening the generalization performance.

[0065] In a specific implementation scenario, on the basis of the above scheme, after the model is deployed, new data can be collected on site regularly, and samples with "low model confidence" or "manual feedback misjudgment" can be automatically identified and added to the noise synthesis and enhancement process to realize online iteration and continuous optimization of the training data set.

[0066] In a specific implementation scenario, on the basis of the above scheme, when building the data set, a "noise sample proportion control module" can be set to ensure that the proportions of various noises (such as impulse type, drift type, and periodic type) in the data set are balanced, thereby avoiding the model from being biased towards recognizing the noise type that appears frequently, and improving the balanced detection capability of multiple types of noise.

[0067] The prompt information, feature vector and training data set are input into the entity embedding base model for model training, the length and dimension of the industrial sensor data are evaluated to assess the task complexity, the model capacity is dynamically adjusted, and / or the LoRA parameter fine-tuning strategy is used to train the model.

[0068] As described above, the prompt information, feature vector and training data set are input into the entity embedding base model for model training, which means that the three types of key inputs, "prompt information" generated based on global time series features (used to guide the model to focus on data characteristics), "fixed dimension feature vector" generated by a sliding window (used to express local time series structure), and "labeled training data set" (containing real / synthetic noise samples) are input into the base model that has completed "industrial sensor knowledge graph embedding". During the training process, the model simultaneously utilizes domain knowledge, time series context and data diversity to learn how to accurately identify noise patterns.

[0069] Task complexity evaluation based on length and dimension of industrial sensor data, which refers to evaluating two core attributes of input data before or during training: one is the time series "length" (e.g., 1000 sampling points vs. 10000 sampling points), and the other is the "dimension" (single variable temperature data vs. multi-variable vibration + current + temperature joint data). The longer and higher the dimension of the data, the more complex the time series dependency and the more hidden the noise pattern, and the higher the task complexity.

[0070] Dynamic adjustment of model capacity, which refers to dynamically activating or suppressing the computing resources within the model based on the evaluation results. For example, more "expert layers" or "attention heads" are activated when facing high complexity tasks, and some modules are frozen to reduce computational overhead when facing low complexity tasks. The essence is to achieve "task-adaptive model structure" to avoid "overkill" or "underkill".

[0071] Training the model using the LoRA parameter fine-tuning strategy, which refers to not directly updating all the parameters of the base model (usually billions of levels) during training, but only training a set of low-rank matrices (Low-Rank Adaptation). These matrices are added as "bypass" to the original model weights, achieving lightweight fine-tuning. This strategy greatly reduces memory usage and training cost, while retaining the general ability of the base model, making it particularly suitable for deployment environments with limited data and resources in industrial scenarios.

[0072] The expression "and / or" indicates that dynamic adjustment of model capacity and LoRA fine-tuning can be used independently or in combination, forming a "structure-adaptive + parameter-efficient" dual optimization mechanism.

[0073] For example, suppose we are training a noise detection model for "multi-variable industrial pump set state monitoring": The input received by the model includes: prompt information "the pump set data contains strong periodicity + occasional impulse noise"; a 10-dimensional feature vector (mean, standard deviation, peak-to-peak value, etc.) generated by a sliding window; the training data set contains 20,000 samples from 5 pump types, 3 working conditions, and 4 types of artificial noise injection; the model has embedded knowledge entities such as "centrifugal pump - susceptible to cavitation noise" and "pressure sensor - range 0-10MPa".

[0074] The data of the current training batch is "10,000 sampling points length, 3 dimensions (pressure + vibration + temperature)", and the system determines it as a "high complexity task" because long sequences need to capture long-term dependencies and multiple dimensions need to model the coupling relationship between variables.

[0075] The system automatically activates 3 expert modules (only 1 activated by default) in the MoE (Mixture of Experts) architecture and enables all 12 layers of the Transformer attention mechanism to enhance the model's ability to model complex patterns.

[0076] During training, only the LoRA bypass matrix of each layer of attention matrix (rank=8) is updated by gradient, and more than 99% of the parameters of the original model are frozen. Training only requires 4GB of video memory, reduces the time consumption by 60%, and the noise detection F1-score on the test set reaches 94.2%, which is better than the 93.8% of full-parameter fine-tuning.

[0077] For example, when training a "short sequence univariate temperature sensor noise detection" task (length=500, dimension=1): The system determines that it is "low complexity" and only activates 1 expert module, and turns off the last 6 layers of Transformer; LoRA fine-tuning is still used, but the rank is reduced to 4; The model training speed is improved by 3 times, the inference delay is reduced by 70%, and the accuracy remains above 92%, meeting the deployment requirements of edge devices.

[0078] It should be noted that in specific implementation scenarios, in addition to "length" and "dimension", "noise density", "signal entropy value", "spectral complexity" and other indicators can be introduced to comprehensively evaluate the task difficulty based on the above scheme, to achieve more fine-grained resource allocation. For example, high-entropy signals activate more frequency domain processing modules.

[0079] In specific implementation scenarios, in addition to the MoE expert layer, the number of Transformer layers, the number of attention heads, the width of the feedforward network, and the number / depth of convolution kernels can also be dynamically adjusted based on the above scheme. All adjustments are triggered based on task complexity thresholds and belong to the "structure adaptive" category.

[0080] In specific implementation scenarios, in addition to the MoE expert layer, the "rank" size of the LoRA matrix can also be dynamically adjusted based on the task complexity or data scarcity. High rank (e.g. rank=16) for complex tasks and low rank (e.g. rank=4) for simple tasks or small samples to achieve "parameter efficiency adaptation".

[0081] In specific implementation scenarios, in addition to the above scheme, LoRA can be combined with other efficient fine-tuning methods such as Adapter, Prefix-Tuning, BitFit, etc. to form "multi-strategy collaborative fine-tuning". For example, LoRA is responsible for time series modeling, and Adapter is responsible for knowledge fusion, improving the efficiency of modular training.

[0082] In a specific implementation scenario, on the basis of the above scheme, a low-capacity model can be used to quickly converge at the beginning of training, and more expert layers or attention heads can be gradually unlocked as the training round increases, to realize a stable training process from simple to complex and avoid initial shock of high-complexity models.

[0083] In a specific implementation scenario, on the basis of the above scheme, a light version of the model can be fine-tuned on the edge device using LoRA, and a full-capacity model can be used for knowledge distillation or gradient aggregation in the cloud, to realize an industrial AI training paradigm of "light deployment + centralized optimization".

[0084] In a specific implementation scenario, on the basis of the above scheme, 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 task, to ensure that the model prioritizes learning the most difficult samples during the resource allocation peak, thereby improving training efficiency.

[0085] In a specific implementation scenario, on the basis of the above scheme, when deploying on an industrial edge device, the maximum number of activated experts or layers can be dynamically limited according to the remaining power or heat dissipation capacity of the device, to realize "energy-efficient training / inference" while ensuring the accuracy baseline.

[0086] The industrial sensor data to be detected is input into the trained model, and a noise detection result is output.

[0087] As described above, inputting the industrial sensor data to be detected into the trained model means that during the model deployment stage, real-time or historical time series data collected by sensors of actual industrial devices (such as fans, pumps, compressors, numerical control machine tools, etc.) is received. Before input, the data will go through the same preprocessing process as in the training stage, including sliding window division, local feature vector extraction, global time series feature calculation, prompt information generation, etc., to ensure that the input format is consistent with that during training, thereby activating the knowledge and pattern recognition ability learned by the model.

[0088] Outputting the noise detection result means that after the model inference is completed, structured and operable noise identification information is output, which usually includes but is not limited to: whether there is noise (binary classification or probability output); the time position of noise occurrence (start point, end point); noise type (such as impulse noise, drift noise, periodic interference, etc.); noise confidence (the degree of confidence in the model's judgment); noise intensity level (mild, moderate, severe). The output form can be structured data (JSON, CSV), visual markers (such as highlighted noise segments), or alarm signals (trigger threshold).

[0089] This step emphasizes the "end-to-end availability" and "industrial interpretability" of the model, not only requiring "detection", but also requiring "clear explanation" and "usefulness".

[0090] For example, assume that a steel mill's blast furnace cooling water circulating pump group is running, and its vibration sensor is continuously collecting data: Before the 10-minute vibration data collected in real time (1 kHz sampling rate, 600,000 points in total) is sent into the model, the system automatically performs the following operations: division by sliding window with a 5-second window (5000 points); calculation of 10-dimensional features such as mean value, standard deviation, and peak-to-peak value for each window; extraction of global features such as "overall entropy value is high" and "there is a 50 Hz periodic component"; generation of a prompt "suspected mechanical looseness + power frequency interference"; and simultaneous calling of the embedded "centrifugal pump-vibration noise knowledge" entity.

[0091] The model, in combination with the prompt information, window features, and knowledge graph, identifies an abnormal high-frequency pulse lasting 1.2 seconds in the 32nd window (corresponding to the 160th second of operation). The pulse has a shape that highly matches the "bearing ball damage noise template" and deviates from the normal vibration range of the pump type.

[0092] Output result: Model output structured result: { "noise exists": true, "noise start time": "160.0s", "noise end time": "161.2s", "noise type": "mechanical impact type pulse noise", "confidence level": 0.96, "intensity level": "severe", "associated knowledge": "may be caused by bearing ball local peeling, suggest checking within 72 hours" } The result is automatically pushed to the equipment health management system and triggers a yellow warning, and the maintenance engineer arranges maintenance in advance based on this, avoiding unplanned downtime.

[0093] For another example, in a clean room temperature control system of a semiconductor factory: Input a segment of temperature sensor data, and the model identifies a slowly rising "drift noise". In combination with the "PT100 sensor-easy to have zero drift after long-term use" entity in the knowledge graph, the model outputs a "suggest calibrating the sensor" prompt to avoid misjudgment as "environmental temperature anomaly".

[0094] It should be noted that in specific implementation scenarios, in addition to a single confidence level, the above-mentioned scheme can also output "model internal multi-module voting consistency", "matching degree with knowledge graph", and "agreement degree with historical similar cases" as multi-dimensional confidence indicators to assist manual judgment of whether to adopt the model result.

[0095] In a specific implementation scenario, on the basis of the above scheme, the possible fault sources (such as "impulse noise → bearing damage" and "50Hz interference → poor grounding of the frequency converter") can be automatically recommended based on the output noise type and the "noise source-device component" mapping relationship in the knowledge graph, thereby improving the engineering guidance value of the result.

[0096] In a specific implementation scenario, on the basis of the above scheme, the "impact assessment on the service life of the device" or the "interference level on the control accuracy" can be output based on the noise intensity, duration and frequency, combined with the device operation parameters, thereby supporting risk classification management.

[0097] In a specific implementation scenario, on the basis of the above scheme, the noise interval label, type color coding and confidence transparency can be superimposed on the original time series curve to generate a "noise heat map", thereby facilitating engineers to quickly locate the problem period.

[0098] In a specific implementation scenario, on the basis of the above scheme, the output form can be automatically switched according to the use scenario: output a lightweight alarm signal on an edge device; output a structured JSON on a central control system; and output a report document with traceability analysis on an expert workstation.

[0099] In a specific implementation scenario, on the basis of the above scheme, the operator can confirm or correct the model output result, and the correction result can be fed back to the training data pool to trigger online fine-tuning of the model or update of the data set, thereby realizing "human-machine collaborative continuous optimization".

[0100] In a specific implementation scenario, on the basis of the above scheme, when multiple associated sensor data (such as vibration + temperature + current) are input at the same time, the model can output "cross-sensor noise correlation analysis", for example, "vibration pulse and current spike occur synchronously, indicating a sudden change in motor load", thereby improving the system-level diagnosis capability.

[0101] In a specific implementation scenario, on the basis of the above scheme, according to the noise type and severity, subsequent actions such as "automatic filtering processing", "triggering device load reduction operation", "generating a maintenance work order" and "notifying the on-duty engineer" can be automatically recommended, thereby realizing a closed loop from "detection" to "decision" to "execution".

[0102] According to an embodiment of the present application, the knowledge graph of the industrial sensor includes a sensor type, a measured physical quantity, a noise characteristic, a normal range and an abnormal range. As described above, the sensor type is used to identify the category or model of the sensor, for example, a temperature sensor, a pressure sensor, a vibration sensor, a current sensor, a displacement sensor, etc. Different types of sensors differ in structure, working principle and applicable scenarios.

[0103] Measured physical quantity: refers to the specific physical parameter measured by the sensor, such as temperature sensor measures "temperature in Celsius", pressure sensor measures "pressure in MPa", vibration sensor measures "acceleration or displacement", and usually contains corresponding physical units such as ℃, MPa, mm / s, etc. 2 Or mm, etc.

[0104] Noise characteristics: describes the common noise sources and forms of this type of sensor in actual operation, such as "temperature sensor is susceptible to thermal drift noise", "vibration sensor is susceptible to power frequency interference in strong electromagnetic environment", "pressure sensor is susceptible to step noise under rapid impact", etc., which is used to assist the model in identifying whether the noise conforms to the typical behavior of the sensor.

[0105] Normal range: refers to the reasonable value interval of the output data of the sensor under normal operation of the equipment, such as "the normal value range of a certain bearing vibration sensor is 0.5-2.0 mm / s", which is used as a reference basis for judging whether the data is abnormal.

[0106] Abnormal range: refers to the value interval of the sensor data that exceeds the normal working condition and may indicate equipment failure or data distortion, such as "vibration value continuously higher than 5.0 mm / s is considered as a serious anomaly", which is used to assist the model in distinguishing noise from real fault signals or judging whether the noise has reached the degree of intervention.

[0107] According to one embodiment of the present application, the global timing features include trend features, periodicity features, volatility features, distribution features, and anomaly detection features.

[0108] As described above, the trend feature: reflects the overall trend of data changes over time, such as whether the data presents an upward, downward, or stable trend. Common indicators include autocorrelation coefficient, mean, median, slope estimate, area under the curve, entropy value, etc., which are used to judge whether there is a systematic drift or slow change in the data.

[0109] Periodicity feature: reflects whether there are regular fluctuations that repeat in the data, such as periodic patterns caused by device start-stop cycles, power frequency interference, diurnal temperature difference, etc. Common indicators include peak-to-peak distance, interquartile range, spectral main frequency component, autocorrelation peak spacing, etc., which are used to identify regular signals or interference related to normal operation rhythm.

[0110] Volatility feature: reflects the degree of change or stability of data in the time dimension, such as whether the data frequently jumps or remains stable. Common indicators include standard deviation, variance, mean absolute deviation, median absolute deviation, absolute difference sum, energy distribution after wavelet transform, etc., which are used to measure noise intensity or signal stability.

[0111] Distribution feature: reflects the distribution pattern of data values in the numerical space, such as whether concentrated, skewed or with long tail. Common indicators include skewness, kurtosis, maximum value, minimum value, mode, quartile, area under the curve, etc., used to determine whether the data conforms to the numerical distribution rule of the sensor type.

[0112] Abnormality detection feature: reflects whether there are points or segments in the data that significantly deviate from the normal mode, such as sudden pulses, step jumps, sustained deviations, etc. Common indicators include maximum difference, maximum absolute difference, local extreme deviation, entropy mutation value, etc. to preliminarily locate the possible noise occurrence area or abnormal event.

[0113] According to an embodiment of the present application, the size of the sliding window is dynamically determined according to the sampling frequency of the industrial sensor data.

[0114] As mentioned above, the sampling frequency is high (such as 10,000 points per second for a vibration sensor), the data changes quickly and is rich in details, at this time the sliding window should be set to a shorter time span (such as 0.1 seconds or 500 sampling points), to avoid the data in the window being too dense and the features being blurred, so as to more accurately capture transient noise or high-frequency interference.

[0115] The sampling frequency is low (such as 1 point per minute for a temperature sensor), the data changes slowly and is strong in trend, at this time the sliding window should be set to a longer time span (such as 30 minutes or 60 sampling points), to ensure that the window contains enough data points to reflect the trend change or slow drift of low-frequency noise characteristics.

[0116] This dynamic determination mechanism can be realized through preset rules, for example: window size = baseline time length × sampling frequency, where the "baseline time length" can be set according to the sensor type or industrial scene (such as 0.1 seconds for vibration type and 10 minutes for temperature type), so that the window has the same physical meaning in the time dimension under different sampling rates.

[0117] In this way, the sliding window can adapt to the data characteristics of different sensors, and also ensure the stability and effectiveness of subsequent feature extraction, improving the generalization ability and detection accuracy of the model in multiple types of industrial scenes.

[0118] According to an embodiment of the present application, the data enhancement processing includes adding Gaussian noise, impulse noise and periodic noise, and adjusting the signal strength.

[0119] As mentioned above, adding Gaussian noise: refers to superimposing random noise obeying normal distribution on the original data, simulating the continuity and small amplitude random fluctuations caused by electronic component thermal disturbance, environmental electromagnetic interference, etc. The noise intensity can be controlled by adjusting the standard deviation, such as low intensity (σ=0.01), medium intensity (σ=0.05), high intensity (σ=0.1).

[0120] Adding pulse noise: refers to inserting transient large amplitude spikes at random or specific positions in the data sequence, simulating sudden interference caused by mechanical impact, switching transient, poor contact, etc. The pulse noise can be set to different amplitude, width and occurrence frequency, such as single point large amplitude pulse, continuous multi-point small amplitude pulse, etc.

[0121] Adding periodic noise: refers to superimposing sine or square wave components with fixed frequency and amplitude in the original signal, simulating regular interference such as power frequency interference (such as 50Hz / 60Hz), periodic vibration caused by rotating parts of equipment, etc. Periodic noise can be generated in different variants by adjusting parameters such as frequency, phase, amplitude, etc.

[0122] Adjusting signal strength: refers to scaling the overall signal amplitude after adding the above noise, or on the basis of the original data (such as amplifying 1.2 times or attenuating 0.8 times), simulating sensor range drift, gain change, signal attenuation, etc. Make the model adapt to the noise performance under different signal strength.

[0123] According to one embodiment of the present application, 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.

[0124] As mentioned above, task complexity evaluation: before the data is input into the model, the system first analyzes the data characteristics of the task, mainly including time series length (such as 1000 points vs. 10000 points) and data dimension (single variable temperature vs. multi-variable vibration + current + pressure), and can combine noise density, spectral complexity, etc. Auxiliary indicators, a “task complexity score” is calculated.

[0125] Complexity threshold setting: the system presets multiple complexity threshold intervals, for example, “low complexity (score 0-30)”, “medium complexity (score 31-70)”, “high complexity (score 71-100)”, each interval corresponds to a different number of activated expert layers.

[0126] Dynamic activation of expert layers: If the task complexity score falls into the “low complexity” interval, only 1 expert layer is activated to complete the inference with the smallest calculation cost; If it falls into the “medium complexity” interval, 2-3 expert layers are activated to enhance the model's ability to recognize medium difficulty noise patterns; If it falls into the "high complexity" interval, all available expert layers (such as 4-8) are activated, fully utilizing the model's capacity to handle complex scenarios such as long sequences, multiple variables, and strong coupling noise.

[0127] Expert layer mechanism: Expert layers are independent sub-network modules 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 characteristics. The dynamic activation mechanism allows the model to run lightly when facing simple tasks and automatically expand when facing complex tasks, achieving "on-demand allocation and flexible computing".

[0128] According to an embodiment of the present application, the noise detection result includes noise position, noise type and confidence.

[0129] As mentioned above, noise position refers to the specific time point or time period at which noise occurs in the original time series data, usually represented in the form of sample point number, timestamp or start-end interval. For example, "noise starts at the 1250th sample point and ends at the 1280th sample point" or "noise occurs between 14:32:15 and 14:32:18". This information is used to accurately locate the abnormal data segment, facilitating data cutting, visual marking or local cleaning.

[0130] Noise type refers to the category of noise identified by the model, which is classified based on pre-set or learned noise patterns, such as "Gaussian white noise", "impulse noise", "periodic interference", "drift noise", "transmission packet loss", etc. The classification result can be matched and output in combination with the noise characteristics in the knowledge graph, helping users understand the source or cause of the noise and providing a basis for subsequent noise suppression or equipment maintenance.

[0131] Confidence refers to the credibility of the model's judgment result, usually represented in the form of a value between 0 and 1 or a percentage (such as 0.92 or 92%). Confidence reflects the model's internal certainty about the judgment, with high confidence indicating high consistency in feature matching, knowledge invocation, pattern recognition, etc., and low confidence indicating the presence of edge cases or data ambiguity, which requires human intervention for review.

[0132] The above three items together constitute the basic output structure of the noise detection result, ensuring that the detection result not only has a conclusion, but also is locatable, classifiable and assessable, meeting the requirements of industrial scenarios for explainability, traceability and operability. This output structure can be further extended, but at least it should include these three core information to support effective decision-making.

[0133] According to an embodiment of the present application, the feature vector includes mean, standard deviation, minimum value and maximum value.

[0134] As described above, the mean reflects the average level or central tendency of the data within the window, used to determine whether the overall data deviates from the normal benchmark.

[0135] The standard deviation reflects the degree of dispersion of data points around the mean within the window, used to measure data volatility or stability, the larger the standard deviation, the more unstable the data, which may exist noise or abnormality.

[0136] The minimum value reflects the lower limit of the data within the window, used to capture sudden negative impulse, signal loss or sensor lower limit abnormality.

[0137] The maximum value reflects the upper limit of the data within the window, used to identify positive peaks, excessive range signals or transient impact noise.

[0138] The four statistics together constitute a concise and effective local feature description, which can stably represent the basic form and change range of the data within the window without relying on complex calculations. The feature vector structure is unified and fixed in dimension, facilitating batch processing and model input, while retaining sufficient information to enable the model to assist in determining whether there is noise based on local statistical characteristics.

[0139] The second aspect embodiment of the application provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to realize the method in any of the embodiments of the first aspect.

[0140] Figure 2 An example of an entity structure diagram of an electronic device is shown in Figure 2 As shown, the electronic device can 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 through the communication bus 840. The processor 810 can invoke the logical instructions in the memory 830 to execute the method in any of the embodiments of the first aspect, which includes: Building a knowledge graph of industrial sensors, using the TransE algorithm to embed the knowledge graph, and embedding the generated entity embedding into the base model; Based on the industrial sensor time series data to extract global time series features, generating targeted prompt information based on the global time series features, using a sliding window technique to divide the industrial sensor time series data into multiple fixed-size windows, calculating the time series features of each window, and generating a fixed-dimension feature vector; Collecting multi-source industrial sensor data, generating simulated noise data through artificial synthesis method, and performing data enhancement processing on the data to build a training data set; The prompt information, the feature vector and the training data set are input into a base model of entity embedding 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 model is trained by using a LoRA parameter fine-tuning strategy; The industrial sensor data to be detected is input into the trained model, and a noise detection result is output.

[0141] In addition, the logical instructions in the memory 830 described above can be implemented in the form of a software function unit and sold or used as an independent product, and can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk, and various media that can store program codes.

[0142] On the other hand, the present application also provides a computer program product, the computer program product includes a computer program, the computer program can be stored on a non-transitory computer readable storage medium, when the computer program is executed by a processor, the computer can execute the method provided by the above-mentioned methods, the method includes: A knowledge graph of the industrial sensor is constructed, a TransE algorithm is used to embed the knowledge graph, and the generated entity embedding representation is embedded into a base model; Based on the industrial sensor time series data, global time sequence features are extracted, targeted prompt information is generated based on the global time sequence features, a sliding window technology is used to divide the industrial sensor time series data into a plurality of fixed-size windows, time sequence features of each window are calculated, and a fixed-dimension feature vector is generated; Multiple industrial sensor data sources are collected, simulated noise data is generated by an artificial synthesis method, data enhancement processing is performed on the data, and a training data set is constructed; The prompt information, the feature vector and the training data set are input into a base model of entity embedding 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 model is trained by using a LoRA parameter fine-tuning strategy; The industrial sensor data to be detected is input into the trained model, and a noise detection result is output.

[0143] In yet another aspect, the present application also provides a non-transitory computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the method provided by any of the above methods, and the method comprises: A knowledge graph of the industrial sensor is constructed, and the knowledge graph is embedded using a TransE algorithm, and the generated entity embedding representation is embedded into a base model; Based on the time series data of the industrial sensor, global time sequence features are extracted, and targeted prompt information is generated based on the global time sequence features, the time series data of the industrial sensor is divided into a plurality of fixed-size windows using a sliding window technology, the time sequence features of each window are calculated, and a fixed-dimension feature vector is generated; Collecting multi-source industrial sensor data, generating simulated noise data by artificial synthesis method, and performing data enhancement processing on the data to construct a training data set; The prompt information, the feature vector and the 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 industrial sensor data to be detected is input into the trained model, and a noise detection result is output.

[0144] The places not described in the present application can be realized by using or referring to the existing technology.

[0145] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other, and each embodiment mainly describes the difference from other embodiments.

[0146] The above only describes the embodiments of the present application and does not limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the scope of the claims of the present application.

Claims

1. A method for intelligent noise detection of industrial sensors, characterized in that, include: A knowledge graph of industrial sensors is constructed, and the TransE algorithm is used to embed the knowledge graph into the base model. 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. 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; 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. The industrial sensor data to be detected is input into the trained model, and the noise detection results are output.

2. The method according to claim 1, characterized in that, The knowledge graph of the industrial sensors includes sensor type, measured physical quantity, noise characteristics, normal range, and abnormal range.

3. The method according to claim 1, characterized in that, The global time series features include trend features, periodic features, volatility features, distribution features, and anomaly detection features.

4. The method according to claim 1, characterized in that, The size of the sliding window is dynamically determined based on the sampling frequency of the industrial sensor data.

5. The method according to claim 1, characterized in that, The data augmentation process includes adding Gaussian noise, impulse noise, and periodic noise, and adjusting the signal strength.

6. The method according to claim 1, characterized in that, The dynamic adjustment of model capacity is based on the dynamic activation of the expert layer in the multi-expert model according to the task complexity threshold.

7. The method according to claim 1, characterized in that, The noise detection results include noise location, noise type, and confidence level.

8. The method according to claim 1, characterized in that, The feature vector includes the mean, standard deviation, minimum value, and maximum value.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-8.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-8.

Images