Industrial signal curve state monitoring method and device based on attention neural network model
Through the industrial signal monitoring method based on the attention neural network model, the problem of insufficient monitoring accuracy and adaptability of complex industrial signal data in the prior art is solved, and efficient and accurate signal monitoring and fault warning are achieved.
Patent Information
- Application Number
- CN202510126978.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-27
- Publication Date
- 2025-05-30
AI Technical Summary
The existing industrial signal monitoring technology is difficult to effectively identify abnormal characteristics in complex and changing industrial environments, and lacks adaptability and dynamic adjustment capabilities, resulting in insufficient monitoring accuracy and efficiency.
The industrial signal curve status monitoring method based on the attention neural network model is adopted, and the significance of signal characteristics is dynamically identified and a significance curve is generated through data preprocessing, multi-level feature extraction, attention mechanism calculation and real-time feedback mechanism.
It improves the monitoring accuracy and abnormal detection capabilities of complex industrial signal data, enhances the adaptability and reliability of the system, and can achieve efficient and accurate signal monitoring and fault warning in complex industrial environments.
Smart Images
Figure CN120067577A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of industrial model signal monitoring, and particularly to a method and device for monitoring the state of industrial signal curves based on an attention neural network model. Background Art
[0002] In a modern industrial environment, with the rapid development of industrial automation and informatization, a large amount of continuous signal data is generated by various industrial equipment and technological processes. These data cover key information such as the operating state of equipment, changes in technological processes, and product quality, and are an important basis for realizing intelligent industrial production. However, most traditional signal monitoring methods rely on simple threshold judgment or statistical analysis. When faced with complex and diverse changes in curve data features, these methods are often difficult to effectively respond to, unable to detect equipment abnormalities or process deviations in a timely manner, thus limiting the accuracy and efficiency of equipment status monitoring and quality control in the industrial production process.
[0003] In recent years, signal monitoring methods based on machine learning or deep learning have gradually emerged. Although these methods have improved the deficiencies of traditional methods to a certain extent, there are still many limitations in the existing technologies. For example, these methods usually rely on manually set rules in feature extraction and significance analysis, lacking self-adaptability. Due to the complex and variable characteristics of industrial signal data, the significance of many features varies in different scenarios. It is difficult to dynamically and accurately quantify the significance of signal features only relying on fixed feature extraction and monitoring methods. In addition, industrial signal data is often mixed with noise, which will interfere with the accuracy of feature extraction and further affect the reliability of monitoring results.
[0004] Therefore, the existing industrial signal monitoring technologies still cannot meet the requirements of efficient, accurate, and dynamic monitoring and anomaly detection of signal data when faced with a complex and changeable industrial environment. This not only restricts the improvement of industrial production efficiency but also increases the risk of equipment failure and production costs. To solve the above problems, there is an urgent need for an intelligent monitoring method that can adaptively extract features, quantify significance, and detect anomalies in industrial signal data to improve the intelligent level and reliability of industrial production.
[0005] In view of this, the present application is proposed. Summary of the Invention
[0006] The present invention provides a method and device for monitoring the state of industrial signal curves based on an attention neural network model, which can at least partially improve the above problems.
[0007] To achieve the above object, the present invention adopts the following technical solutions:
[0008] An industrial signal curve state monitoring method based on an attention neural network model, which includes:
[0009] Obtain industrial signal data collected by an acquisition module configured on industrial equipment, preprocess the industrial signal data to generate multi-level feature inputs;
[0010] Use the step detection module of a pre-trained attention neural network model to perform signal feature calculation processing on the multi-level feature inputs to obtain quantization marker data;
[0011] Call a pre-trained attention neural network model, calculate the attention weights of the quantization marker data based on the attention mechanism, and generate a saliency curve according to the attention weights;
[0012] Obtain artificial subjective marker data, review and modify the saliency curve according to the artificial subjective marker data to obtain the final saliency curve.
[0013] The present invention also provides an industrial signal curve state monitoring device based on an attention neural network model, which includes:
[0014] A preprocessing unit for obtaining industrial signal data collected by an acquisition module configured on industrial equipment, preprocessing the industrial signal data to generate multi-level feature inputs;
[0015] A quantization feature unit for using the step detection module of a pre-trained attention neural network model to perform signal feature calculation processing on the multi-level feature inputs to obtain quantization marker data;
[0016] A curve generation unit for calling a pre-trained attention neural network model, calculating the attention weights of the quantization marker data based on the attention mechanism, and generating a saliency curve according to the attention weights;
[0017] A review and modification unit for obtaining artificial subjective marker data, reviewing and modifying the saliency curve according to the artificial subjective marker data to obtain the final saliency curve.
[0018] In summary, the industrial signal curve state monitoring method based on the attention neural network model aims to solve the problems of insufficient abnormal feature recognition ability, lack of adaptability, and difficulty in dynamically adjusting monitoring standards in traditional signal monitoring technologies in complex industrial environments. By combining data acquisition, normalization preprocessing, wavelet transform, attention neural network saliency quantization, and a real-time feedback mechanism, this method can efficiently and accurately monitor and analyze the signal data generated during the operation of industrial equipment.
[0019] Specifically, first, the method eliminates the influence brought by different signal dimensions through the maximum-minimum normalization method and uses one-dimensional total variation denoising technology to remove high-frequency noise, ensuring that the main features of the signal are retained. Second, wavelet transform is used to decompose the signal into multiple scales to extract low-frequency trends and high-frequency detail features, providing rich input information for subsequent significance analysis. Finally, the core attention neural network module dynamically identifies the significance of signal features and generates a significance curve through multi-level feature input and multi-head self-attention mechanism for real-time monitoring and anomaly detection. In addition, the system also has a real-time feedback mechanism that can dynamically adjust the attention weight according to the operator's subjective judgment and on-site feedback data, thus realizing the adaptive optimization of the model.
[0020] Briefly, the industrial signal curve state monitoring method based on the attention neural network model uses the attention neural network to simulate human subjective judgment, dynamically adjusts the feature significance quantization standard, and effectively improves the monitoring accuracy and anomaly detection ability of complex industrial signal data. At the same time, through the visual presentation of the significance curve and the real-time feedback mechanism, operators can intuitively observe the signal change trend and adjust the monitoring standard according to actual needs, further enhancing the practicality and flexibility. In addition, this method is widely applicable to the real-time monitoring and fault warning of various industrial equipment and production processes, with important practical application value and broad application prospects. Brief Description of the Drawings
[0021] Figure 1 is a schematic flowchart of the industrial signal curve state monitoring method based on the attention neural network model provided by the first embodiment of the present invention;
[0022] Figure 2 is a schematic diagram of the curve detection interface provided by the embodiment of the present invention;
[0023] Figure 3 is an example diagram of the multi-head self-attention neural network provided by the embodiment of the present invention;
[0024] Figure 4 is a schematic diagram of the interface recognition effect provided by the embodiment of the present invention;
[0025] Figure 5 is a schematic diagram of the modules of the industrial signal curve state monitoring device based on the attention neural network model provided by the second embodiment of the present invention. Detailed Embodiments
[0026] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0027] ReferenceFigures 1 to 4 As shown in Figures 1 to 4 , the first embodiment of the present invention discloses an industrial signal curve state monitoring method based on an attention neural network model, which can be executed by an industrial signal curve state monitoring device based on an attention neural network model (hereinafter referred to as the monitoring device), and specifically, executed by one or more processors in the monitoring device to implement the following method:
[0028] S1. Obtain industrial signal data collected by an acquisition module configured on an industrial device, preprocess the industrial signal data, and generate a multi-level feature input;
[0029] Specifically, step S1 includes: obtaining industrial signal data collected by an acquisition module configured on an industrial device, where the industrial signal data includes vibration signals, current signals, and temperature signals;
[0030] Use a one-dimensional total variation denoising algorithm to denoise the industrial signal data, and the mathematical expression is: where x i is the denoised signal value, y i is the noisy observed signal, λ is a regularization parameter used to balance the smoothness of the signal and the fitting degree to the original signal. The larger λ is, the closer the resulting signal xi is to the original signal yi; the smaller it is, the higher the smoothness. N is the length of the signal data, used to balance the fidelity and smoothness of the signal. N is the total length of the signal, that is, the number of data points. ∣x i -x i-1 ∣ is the absolute difference between adjacent data points, and ∣x i -y i ∣ 2 is the square difference between the denoised signal x i and the original signal y i ;
[0031] Sort the denoised industrial signal data according to the time stamp, perform maximum-minimum normalization processing on the industrial signal data, and map the industrial signal data to a preset range, where the preset range is [0,1]. The mathematical formula for maximum-minimum normalization is: x′ is the normalized value, and the range is usually [0,1]. x is the value of the original data, x min is the minimum value in the original data x, and x max is the maximum value in the original data x;
[0032] Use a wavelet transform algorithm to perform multi-scale decomposition processing on the normalized industrial signal data to obtain a wavelet spectrum, and combine the original waveform of the industrial signal data with the wavelet spectrum to generate a multi-level feature input, where the wavelet spectrum includes low-frequency components and high-frequency components.
[0033] In this embodiment, first, industrial signal data is obtained through an acquisition module configured on industrial equipment or in a process flow. Among them, the acquisition module uses high-precision sensors and data acquisition devices to ensure that the sampling frequency meets the requirements of industrial signal monitoring. These signal data cover key information during the operation of the equipment, including vibration signals, current signals, temperature signals, etc. The acquisition of these signal data is the basis for subsequent intelligent monitoring and anomaly detection, and their diversity and continuity provide rich monitoring dimensions for the system.
[0034] After obtaining the signal data, noise reduction processing is immediately carried out on it. This is because there are various interference factors in the industrial environment, and the signal data is often mixed with high-frequency noise, which will interfere with subsequent feature extraction and analysis. Therefore, this method uses a one-dimensional total variation (1D Total Variation) denoising algorithm to process the signal data; by minimizing the above total variation expression, the algorithm can effectively remove the high-frequency noise in the signal while retaining the main change trend of the signal. The advantage of this denoising method is that it can significantly improve the signal-to-noise ratio of the signal without losing important features, providing a more accurate input for subsequent feature extraction and analysis.
[0035] Subsequently, the signal data is sorted according to the time stamp to ensure the temporal consistency of the data; the signal data is subjected to maximum-minimum normalization processing to map the data to a preset range [0, 1]. Among them, through the normalization processing, the differences brought by different signal dimensions can be eliminated, and the consistency of the data can be enhanced, thus laying a foundation for subsequent feature analysis. This normalization method not only simplifies the calculation process but also improves the robustness of the system, enabling it to adapt to signal data with different dimensions and ranges.
[0036] Finally, after completing the normalization processing, the wavelet transform algorithm is used to perform multi-scale decomposition processing on the signal data. Wavelet transform is a tool that can analyze signals in both the time domain and the frequency domain simultaneously and can effectively extract multi-scale features of signals. Through wavelet transform, the signal data is decomposed into low-frequency components and high-frequency components. The low-frequency components reflect the long-term trend changes of the signal, such as the aging process of the equipment or the slow rise of temperature; while the high-frequency components capture the short-term violent fluctuations of the signal, such as sudden failures or instantaneous changes caused by external interferences. The normalized original signal waveform is combined with the wavelet spectrum to generate a multi-level feature input. This multi-level feature input not only retains the original information of the signal but also introduces multi-scale frequency domain features through wavelet transform, providing richer information for subsequent significance quantification analysis.
[0037] S2. Use the step detection module of the pre-trained attention neural network model to perform signal feature calculation and processing on the multi-level feature input to obtain quantization marker data;
[0038] Specifically, step S2 includes: The step detection module includes a step significance quantization unit, an up / down trend significance quantization unit, a periodicity significance quantization unit, a mutation significance quantization unit, and a stationarity significance quantization unit. The quantization marker data includes a step significance quantization value, a trend significance quantization value, a periodicity significance quantization value, a significance quantization value of a sudden rise or fall of the signal, and a stationarity significance quantization value.
[0039] The formula for the periodicity significance quantization value is: The formula for the significance quantization value of a sudden rise or fall of the signal is: The formula for the stationarity significance quantization value is: where τ is the time offset between two signals, t is the index of the current time point, x t is the value of the time series signal at time t, x t+τ is the value of the time series signal at time t + τ, that is, the signal value after a delay of τ, a(t) is the second derivative of the industrial signal, v(t) is the value of the signal at time t, v(t - Δt) is the value of the signal at time t - Δt, Δt is the time interval, γ is the mutation significance threshold, S jump (t) is the mutation significance quantization value, x i is the i-th sample value in the dataset, and μ is the value at the center of the dataset, reflecting the overall stationary state.
[0040] In this embodiment, the industrial signal data after preprocessing and multi-level feature extraction is input into the pre-trained attention neural network model, especially its core module - the step detection module. This module is the key to realizing signal feature calculation and processing. It analyzes and quantifies various features of the signal through multiple significance quantization units, thereby generating quantization marker data for subsequent analysis.
[0041] Among them, the step detection module contains five main significance quantification units, which calculate and process different features of the signal respectively. These units include the step significance quantification unit, the rising / falling trend significance quantification unit, the periodic significance quantification unit, the mutation significance quantification unit, and the stationary significance quantification unit. Each unit focuses on a specific feature of the signal and generates corresponding significance quantification values through precise calculation formulas. These quantification values not only reflect the intensity of the signal features but also provide important bases for subsequent anomaly detection and trend analysis. Specifically, the step significance quantification unit is used to detect the amplitude mutation points in the signal and quantify the change amplitude and duration; this kind of change usually indicates a sudden change in the device state, such as the switching of the operating state or a sudden failure. The generated significance quantification value marks a specific state transition or event. The rising / falling trend significance quantification unit is used to distinguish continuous unidirectional changes, which may be a gradually rising or gradually falling trend. It generates a trend significance quantification value based on the first-order derivative of the signal time series; this usually means that the device is gradually entering a certain state, such as the temperature gradually rising or the current gradually increasing.
[0042] The periodic significance quantification unit is used to distinguish the regular periodic fluctuations of the signal, which usually indicates the existence of some oscillation processes in the system; it extracts periodic features using the autocorrelation function and generates periodic significance quantification values; this kind of fluctuation may represent the normal working cycle of the device or irregular vibrations caused by external disturbances (such as electrical interference). The mutation significance quantification unit is used to detect a sudden drastic change at a certain moment, usually a sudden rise or fall phenomenon. According to the set change rate threshold, it generates the significance quantification value of the sudden rise or fall of the signal; this state is often an indication signal of a fault, such as the device suddenly stopping, sensor failure, external shock, etc. The stationary significance quantification unit is used for the signal to remain at a certain fixed level within a period of time with very small changes, showing a stationary state; it calculates the standard deviation of the signal through a sliding window and generates a stationary significance quantification value; this usually means that the device is operating in a stable environment and the signal has no abnormal fluctuations.
[0043] In this embodiment, through the above steps, the step detection module can comprehensively analyze various features of industrial signal data and generate corresponding quantified marker data. These data not only include the step significance quantification value, the trend significance quantification value, the periodic significance quantification value, the significance quantification value of the sudden rise or fall of the signal, and the stationary significance quantification value, but also ensure the accuracy and reliability of feature analysis through precise calculation formulas. The beneficial effect of this process is that it can convert complex signal features into quantifiable data. The significance quantification marker data generation formula remains in the form provided in the original document, providing high-quality input for subsequent attention neural network analysis, thus significantly improving the monitoring accuracy and anomaly detection ability of the system.
[0044] S3. Call the pre-trained attention neural network model, calculate the attention weights of the quantization marker data based on the attention mechanism, and generate a saliency curve according to the attention weights;
[0045] Specifically, step S3 includes: extracting local features of the quantization marker data using a one-dimensional convolutional kernel of the attention neural network model, and calculating the attention weights in a dot product manner. The calculation formula is: Q is the query matrix, K is the key matrix, V is the value matrix, K T is the transpose of the key matrix, d k is the dimension of the key matrix, where the attention neural network model adopts a multi-head self-attention mechanism;
[0046] Generate a saliency curve according to the attention weights. The generation formula of the saliency curve is: where α i is the attention weight, f i (t) is the eigenvalue, and H is the number of eigen-components.
[0047] In this embodiment, the attention mechanism is used to dynamically calculate the attention weights of the quantization marker data, and saliency curves are generated based on these weights, thereby providing an intuitive and accurate analysis tool for real-time monitoring and anomaly detection of industrial signals. Among them, the saliency curve is composed of a step saliency curve, a trend saliency curve, a periodic saliency curve, a mutation saliency curve, and a fluctuation saliency curve using a weighted sum formula. The formula is: α i is the dynamically adjusted attention head weight, head i (t) is the output of different heads. Among them, the step saliency curve is used to strengthen the attention of the low-frequency attention head to the step change points, and assign the saliency weight to the trend mutation area; the trend saliency curve is generated by the output of the low-frequency attention head, reflects the long-term trend change, and is used for real-time monitoring and anomaly detection; the periodic saliency curve is used to strengthen the capture of periodic fluctuations by the high-frequency attention head, and assign the saliency weight to the periodic area, dynamically adjusting the weight of the high-frequency head; the mutation saliency curve is output according to the mutation head weight, marking the mutation point and the change amplitude; the fluctuation saliency curve: output by the high-frequency head weight, quantifying short-term local fluctuations.
[0048] Specifically, in this embodiment, a one-dimensional convolutional kernel of the attention neural network model is used to process the quantized labeled data. The role of the one-dimensional convolutional kernel is to extract local features in the data. These local features can capture subtle changes and patterns in the signal data, providing a basis for further analysis. Through the convolution operation, the model can efficiently process sequence data, reduce the computational complexity, and improve the running efficiency of the model. The beneficial effect of this local feature extraction method is that it can retain the key information in the signal data while removing redundant information, making the subsequent calculation of attention weights more accurate and efficient.
[0049] Subsequently, the model uses the multi-head self-attention mechanism to calculate the attention weights. The multi-head self-attention mechanism allows the model to concurrently focus on different parts of the signal data, thereby being able to capture various features such as long-term trends, short-term fluctuations, and sudden changes simultaneously. This mechanism decomposes the query matrix (Q), key matrix (K), and value matrix (V) into multiple "heads", with each head responsible for processing hierarchical features of different aspects of the data, thus significantly enhancing the model's ability to understand complex signal data and reducing the computational complexity. Among them, the low-frequency part focuses on the changes in long-term trends, and the high-frequency part focuses on short-term fluctuations, thereby comprehensively improving the recognition accuracy of different signal features. The beneficial effect of this step is that it can dynamically assign weights to different signal features, enabling the model to adaptively focus on the most informative parts, thereby improving the detection sensitivity to abnormal signals.
[0050] Subsequently, based on the calculated attention weights, the model further generates a saliency curve; the saliency curve reflects the significant changes in signal features through weighted eigenvalues, which can be used to help the system judge the abnormality degree of the signal and assist in subsequent event detection and fault prediction. It provides a visual analysis tool for industrial signal monitoring, enabling complex signal data to be presented in an intuitive manner, facilitating real-time monitoring and decision-making by operators.
[0051] In this embodiment, the attention neural network model can not only dynamically adjust the degree of attention to signal features but also generate a saliency curve, providing strong support for the monitoring and anomaly detection of industrial signals. The realization of this process not only improves the accuracy and efficiency of signal monitoring but also enhances the adaptability of the model to different industrial scenarios through the adaptive attention mechanism. The generation of the saliency curve further improves the interpretability of the system, enabling operators to more intuitively understand the change trend of signal data, thereby achieving precise monitoring and fault warning of the operating state of industrial equipment.
[0052] Preferably, before calling the pre-trained attention neural network model and calculating the attention weights of the quantized labeled data based on the attention mechanism, it further includes:
[0053] Using a supervised learning algorithm, the initial attention neural network model is trained by quantifying the generated significance training values according to the formula and the marked tags.
[0054] During the training process, the mean squared error loss function is used as the loss function of the attention neural network model to calculate the difference between the significance prediction value generated by the model and the significance training value generated by the formula quantification and marked tags, and the model parameters are adjusted according to this difference.
[0055] Obtain the manually marked data, combine the manually marked data with the significance training values generated by the formula quantification and marked tags, and perform iterative optimization on the initial attention neural network model.
[0056] When it is judged that the significant event window and features automatically recognized by the initial attention neural network model reach the preset standard, the training is ended, and a trained attention neural network model is obtained.
[0057] In this embodiment, a supervised learning algorithm is used to train the initial attention neural network model. This process is based on the significance training values generated by the formula quantification and marked tags, which are obtained by quantifying and analyzing the features of industrial signal data and can provide clear learning objectives for the model. The supervised learning algorithm uses these significance training values as the "true values" to guide the model to learn how to accurately identify the significant features from the input quantified and marked data. Moreover, during the training process, the mean squared error loss function (MSE) is used as the loss function of the model. Among them, the mean squared error loss function can effectively calculate the difference between the significance prediction value generated by the model and the significance training value generated by the formula quantification and marked tags. By minimizing this difference, the model parameters are gradually adjusted and optimized. The beneficial effect of this process is that the mean squared error loss function provides a clear optimization direction for the model, enabling the model to converge quickly and continuously improve the recognition accuracy of the signal feature significance during the training process.
[0058] Furthermore, to enhance the adaptability and robustness of the model, artificially marked data is introduced. These data are marked by experienced operators according to the signal characteristics in the actual industrial scenario and can reflect the significant events and characteristics in the real environment. By combining the artificially marked data with the significant training values generated by formula quantization marking, the initial attention neural network model is iteratively optimized. This process not only enriches the learning samples of the model but also enables the model to better adapt to the complex and changing industrial environment and improve its ability to identify signal characteristics in different scenarios. During the training process, the model continuously adjusts its internal parameters to reduce the difference between the predicted value and the training value. At the same time, by introducing artificially marked data, the model can learn more significant features closer to the actual operation requirements. The beneficial effect of this process is that the model can not only learn based on the training values generated by mathematical formulas but also combine artificial experience, thus showing higher accuracy and practicality in actual applications. Finally, when the automatic recognition ability of the model reaches the preset standard, that is, the model can accurately identify the significant event window and features and its prediction results are highly consistent with the artificially marked data, the training process ends. At this time, the obtained attention neural network model already has the ability to efficiently process industrial signal data and can generate reliable significant curves. The completion of this training process indicates that the model is ready for actual industrial signal monitoring tasks.
[0059] S4. Obtain artificially subjective marked data, and review and modify the significant curve according to the artificially subjective marked data to obtain the final significant curve.
[0060] Specifically, step S4 includes: the artificially subjective marked data includes confirming whether the significant events marked by the system are reasonable and adding or modifying the significant event marks not captured by the system.
[0061] In this embodiment, artificial subjective marking data is obtained. These data are marked by experienced operators according to the signal characteristics and equipment operation status in the actual industrial production process. The operators review the saliency curves automatically generated by the system to confirm whether the saliency events marked by the system are reasonable. This review process is based on the professional knowledge and on-site experience of the operators and can effectively identify saliency events that the model may miss or misjudge. For example, the operator can determine whether a certain signal change truly represents an abnormal state of the equipment, or whether a signal marked as abnormal actually belongs to normal fluctuations. During the review process, the operator not only confirms the events marked by the system but also adds or modifies the saliency event marks that the system fails to capture according to the actual situation. The introduction of this artificial intervention mechanism enables the saliency curve to more accurately reflect the actual characteristics and important changes of industrial signals. For example, if the operator discovers that a certain equipment has an abnormal signal under specific working conditions but the model fails to mark it as a significant event, the operator can manually add this event mark to enrich the saliency recognition scope of the model.
[0062] Reviewing and modifying the saliency curve through artificial subjective marking data not only improves the accuracy and reliability of the saliency curve but also enhances the adaptive ability of the system. This method of combining artificial experience with model calculation enables the saliency curve to better reflect the real situation in industrial production and provides stronger support for equipment condition monitoring and fault warning. The finally obtained saliency curve not only includes the efficient calculation ability of the model but also integrates the professional judgment of the operator, thus demonstrating higher practicality and accuracy in the complex and changeable industrial environment.
[0063] Preferably, it further includes: obtaining the current artificial marking data in real time and adjusting the parameters of the attention neural network model according to the current artificial marking data.
[0064] In this embodiment, to further improve the adaptability and accuracy of the attention neural network model in industrial signal monitoring, this method not only introduces artificial marking data in the initial training stage but also obtains artificial marking data in real time during the actual operation process and dynamically adjusts the model parameters according to these data. This process is the key link to realizing the adaptive optimization of the model and can ensure that the system always maintains high-efficiency and accurate monitoring capabilities in the face of complex and changeable industrial environments. During the actual use process, the operator provides artificial marking data according to the industrial signal data monitored in real time and the visualization results of the saliency curve. These marking data reflect the operator's subjective judgment on the current signal characteristics, such as confirming whether a certain signal change belongs to an abnormal event or pointing out potential problems that the model fails to identify. These artificial marking data not only provide additional supervision information for the system but also help the model better adapt to the dynamic changes in the actual production environment.
[0065] After obtaining these manually marked data in real time, they are immediately compared and analyzed with the saliency curves automatically generated by the model. If a difference is found between the manual markings and the model predictions, a parameter adjustment mechanism will be triggered. This mechanism dynamically adjusts the parameters of the attention neural network model by analyzing the deviation between the manually marked data and the model predictions. This adjustment process is based on the idea of reinforcement learning. By continuously optimizing the attention weights and feature recognition capabilities of the model, it can more accurately reflect the actual characteristics and important changes of industrial signals. It can be seen that by obtaining manually marked data in real time and dynamically adjusting the model parameters, this method can achieve continuous optimization of the saliency curves. The beneficial effect of this process is that it not only improves the adaptability of the model to complex industrial signals, but also enables the system to continuously learn and improve during operation, so as to better meet the changing needs in industrial production. For example, when the operating state of industrial equipment changes or new fault modes appear, the system can quickly adjust the model parameters through the manually marked data to ensure that the saliency curves always accurately reflect the characteristic changes of the signals.
[0066] In addition, this real-time feedback and dynamic adjustment mechanism also enhances the robustness of the industrial signal curve state monitoring method based on the attention neural network model. In the actual industrial environment, signal data is often affected by various factors, such as equipment aging, environmental interference, or changes in operating conditions. By obtaining manually marked data in real time and dynamically adjusting the model parameters, it can quickly adapt to these changes, reduce the possibility of false alarms and missed detections, and thus provide a more reliable guarantee for the stable operation of industrial equipment.
[0067] In summary, the industrial signal curve state monitoring method based on the attention neural network model realizes the efficient monitoring and dynamic analysis of complex industrial signals through steps such as preprocessing of industrial signal data, feature extraction, saliency quantification, and real-time feedback optimization. The aim is to improve the monitoring accuracy and anomaly detection ability of the operating state of industrial equipment through intelligent signal processing technology.
[0068] Briefly speaking, the industrial signal curve state monitoring method based on the attention neural network model combines advanced signal processing techniques, the dynamic analysis ability of the attention neural network, and the wisdom of artificial experience, providing an efficient, accurate, and adaptive solution for industrial signal monitoring. Through the visual presentation of the saliency curve, operators can intuitively understand the operating state of the equipment, timely discover potential problems, and take corresponding measures. In addition, the real-time feedback and dynamic optimization mechanism of this method further enhance its adaptability and reliability in complex industrial environments, providing strong guarantees for the intelligentization of industrial production and the efficient operation of equipment. In summary, this method not only improves the technical level of industrial signal monitoring but also opens up a new path for the fault warning and state monitoring of industrial equipment, with important practical application value and broad development prospects.
[0069] Please refer to Figure 5 , the second embodiment of the present invention provides an industrial signal curve state monitoring device based on the attention neural network model, which includes:
[0070] A preprocessing unit 201, configured to obtain industrial signal data collected by an acquisition module configured on an industrial device, preprocess the industrial signal data, and generate a multi-level feature input;
[0071] A quantization feature unit 202, configured to perform signal feature calculation and processing on the multi-level feature input by using a step detection module of a pre-trained attention neural network model to obtain quantization marker data;
[0072] A curve generation unit 203, configured to call a pre-trained attention neural network model, calculate the attention weights of the quantization marker data based on the attention mechanism, and generate a saliency curve according to the attention weights;
[0073] An audit and modification unit 204, configured to obtain artificial subjective marker data, audit and modify the saliency curve according to the artificial subjective marker data, and obtain a final saliency curve.
[0074] The above is the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.
Claims
1. A method for monitoring industrial signal curve status based on an attention neural network model, characterized in that: include: Acquire industrial signal data collected by a collection module configured on industrial equipment, pre-process the industrial signal data, and generate multi-level feature input; Using a pre-trained attention neural network model's step detection module to perform signal feature calculation processing on the multi-level feature input to obtain quantitative labeling data; Calling a pre-trained attention neural network model, calculating the attention weight of the quantized labeled data based on the attention mechanism, and generating a significance curve according to the attention weight; Obtaining artificial subjective marking data, and reviewing and modifying the significance curve according to the artificial subjective marking data to obtain a final significance curve.
2. The industrial signal curve state monitoring method based on the attention neural network model according to claim 1 is characterized in that: The industrial signal data collected by the collection module configured on the industrial equipment is obtained, and the industrial signal data is preprocessed to generate a multi-level feature input, specifically: Acquire industrial signal data collected by a collection module configured on the industrial equipment, wherein the industrial signal data includes a vibration signal, a current signal, and a temperature signal; The one-dimensional total variation denoising algorithm is used to denoise the industrial signal data. The mathematical expression is: Among them, x i is the signal value after denoising, y i is the observed signal with noise, λ is the regularization parameter used to balance the fidelity and smoothness of the signal, N is the total length of the signal, that is, the number of data points, |x i -x i-1 ∣ is the absolute difference between adjacent data points, ∣x i -y i ∣ 2 is the denoised signal x i and the original signal y i The square difference of The denoised industrial signal data is sorted according to the timestamp, and the industrial signal data is normalized to the maximum and minimum values, and the industrial signal data is mapped to a preset range, wherein the preset range is [0, 1], and the mathematical formula for the maximum and minimum normalization is: x′ is the normalized value, usually in the range of [0,1], x is the value of the original data, and x min is the minimum value in the original data x, x max is the maximum value in the original data x; The normalized industrial signal data is subjected to multi-scale decomposition using a wavelet transform algorithm to obtain a wavelet spectrum, and the original waveform of the industrial signal data is combined with the wavelet spectrum to generate a multi-level feature input, wherein the wavelet spectrum includes a low-frequency component and a high-frequency component.
3. The industrial signal curve state monitoring method based on the attention neural network model according to claim 2 is characterized in that: The step detection module includes a step significance quantification unit, an ascending / descending trend significance quantification unit, a periodic significance quantification unit, a mutation significance quantification unit, and a stationary significance quantification unit. The quantitative marking data includes a step significance quantification value, a trend significance quantification value, a periodic significance quantification value, a signal sudden rise or fall significance quantification value, and a stationary significance quantification value.
4. The industrial signal curve state monitoring method based on the attention neural network model according to claim 3 is characterized in that: The formula for the periodic significance quantification value is: The formula for the significance quantification value of the signal surge or dip is: The formula for the stationary significance quantification value is: Among them, τ is the time offset between the two signals, t is the index of the current time point, and x t is the value of the time series signal at time t, x t+τ is the value of the time series signal at time t+τ, that is, the signal value after delay τ, a(t) is the second-order derivative of the industrial signal, v(t) is the value of the signal at time t, v(t-Δt) is the value of the signal at time t-Δt, Δt is the time interval, γ is the mutation significance threshold, S jump (t) is the quantitative value of mutation significance, x i is the i-th sample value in the data set, and μ is the value at the center of the data set.
5. The industrial signal curve state monitoring method based on the attention neural network model according to claim 1 is characterized in that: Call the pre-trained attention neural network model, calculate the attention weight of the quantized labeled data based on the attention mechanism, and generate a significance curve according to the attention weight, specifically: The one-dimensional convolution kernel of the attention neural network model is used to extract the local features of the quantized labeled data, and the attention weight is calculated by the dot product method. The calculation formula is: Q is the query matrix, K is the key matrix, V is the value matrix, K T is the transpose of the key matrix, d k is the dimension of the key matrix, wherein the attention neural network model adopts a multi-head self-attention mechanism; According to the attention weight, a saliency curve is generated, and the generation formula of the saliency curve is: Among them, α i is the attention weight, f i (t) is the eigenvalue, and H is the number of eigencomponents.
6. The industrial signal curve state monitoring method based on the attention neural network model according to claim 1 is characterized in that: Before calling the pre-trained attention neural network model and calculating the attention weight of the quantized labeled data based on the attention mechanism, the method further includes: A supervised learning algorithm is used to train the initial attention neural network model based on the saliency training values generated by the formula quantification mark; During the training process, a mean square error loss function is used as a loss function of the attention neural network model to calculate the difference between the significance prediction value generated by the model and the significance training value generated by the formula quantization mark, and the model parameters are adjusted according to the difference; Obtain manually labeled data, combine the manually labeled data with the saliency training values generated by the formula quantization labeling, and iteratively optimize the initial attention neural network model; When it is determined that the significant event windows and features automatically identified by the initial attention neural network model meet the preset standards, the training is terminated to obtain a trained attention neural network model.
7. The industrial signal curve state monitoring method based on the attention neural network model according to claim 1 is characterized in that: The artificial subjective marking data includes confirming whether the significant events marked by the system are reasonable, and adding or modifying the significant event marks that are not captured by the system.
8. The industrial signal curve state monitoring method based on the attention neural network model according to claim 1 is characterized in that: Also includes: The current manually labeled data is acquired in real time, and the parameters of the attention neural network model are adjusted according to the current manually labeled data.
9. An industrial signal curve state monitoring device based on an attention neural network model, characterized in that: include: A preprocessing unit, used to obtain industrial signal data collected by a collection module configured on the industrial equipment, preprocess the industrial signal data, and generate a multi-level feature input; A quantization feature unit, used to perform signal feature calculation processing on the multi-level feature input using a step detection module of a pre-trained attention neural network model to obtain quantization label data; A curve generating unit, used to call a pre-trained attention neural network model, calculate the attention weight of the quantized labeled data based on the attention mechanism, and generate a significance curve according to the attention weight; The review and modification unit is used to obtain artificial subjective marking data, review and modify the significance curve according to the artificial subjective marking data, and obtain a final significance curve.