Industrial process abnormal state intelligent early warning method based on LSTM auto-encoder

Abstract

The invention discloses an industrial process abnormal state intelligent early warning method based on an LSTM auto-encoder, and belongs to the technical field of industrial automation and intelligent manufacturing. The industrial process abnormal state intelligent early warning method based on the LSTM auto-encoder comprises the following steps: carrying out data preprocessing on an original multi-dimensional time sequence signal to obtain sample data containing a normal working condition; sample data containing normal working conditions are used to train a network model, an LSTM auto-encoder early warning model is obtained, the network model is an auto-encoder architecture built based on a TensorFlow framework and composed of multiple layers of stacked LSTM units, and a loss function of the network model is a mean square error; and carrying out monitoring and early warning on the data acquired in real time by using the LSTM auto-encoder early warning model. According to the method, full-process automation from data acquisition to abnormal early warning is realized, the manual monitoring burden is reduced, and the operation and maintenance efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of industrial automation and intelligent manufacturing technology, specifically to an intelligent early warning method for abnormal states in industrial processes based on an LSTM autoencoder. Background Technology

[0002] In modern industrial production, especially in automated production lines in fields such as shipbuilding, aerospace, and precision manufacturing, the stability and reliability of equipment operation are crucial. Taking processes such as automatic demolding of propellant cartridges and precision component assembly as examples, these processes involve the coupling effects of multiple physical fields, including mechanics and electricity, and require extremely high operational precision. Any slight deviation in process parameters, such as slow drifts or increased fluctuations in demolding force, static electricity, pressure, or temperature, can accumulate into significant faults, leading to product scrap, equipment damage, or even production stoppage.

[0003] Currently, safety and status monitoring of industrial production lines mainly relies on alarm systems based on fixed thresholds. These systems monitor sensor signals via PLCs (Programmable Logic Controllers), triggering an alarm when the signal value exceeds a preset upper or lower limit. However, this traditional monitoring method has several problems: First, it is a "reactive" mechanism, only issuing alarms when a fault has already occurred or parameters have deviated significantly, failing to provide proactive warnings and missing the optimal intervention window. Second, for slowly changing, trend-based anomalies (such as a gradual increase in demolding force due to equipment wear), the magnitude of change is insufficient to trigger the fixed threshold in the early stages, causing potential risks to be "ignored" and creating monitoring blind spots. Finally, traditional systems typically perform single-variable judgments, lacking the ability to analyze the complex correlations between multi-dimensional sensor data, and cannot identify "sub-healthy" states caused by the coupling of multiple factors. Summary of the Invention

[0004] This invention provides an intelligent early warning method for abnormal states in industrial processes based on LSTM autoencoders to solve the problem that traditional systems usually make single-variable judgments, lack the ability to analyze the complex correlations between multi-dimensional sensor data, and cannot identify "sub-healthy" state problems caused by the coupling of multiple factors.

[0005] In a first aspect of the present invention, an intelligent early warning method for abnormal states in industrial processes based on an LSTM autoencoder is provided, comprising:

[0006] The original multidimensional time series signal is preprocessed to obtain sample data containing normal operating conditions;

[0007] The network model was trained using sample data containing normal operating conditions to obtain an LSTM autoencoder early warning model. This network model is built on the TensorFlow framework and consists of an autoencoder architecture composed of multiple stacked LSTM units. The loss function of this network model is mean squared error.

[0008] An LSTM autoencoder early warning model is used to monitor and issue early warnings for real-time collected data.

[0009] In some optional embodiments of the present invention, the raw multidimensional time-series signals collected from the PLC and sensors are preprocessed to obtain sample data containing normal operating conditions, including:

[0010] Acquire timing data from each sensor within the same production cycle;

[0011] The time-series data is cleaned, aligned, and enhanced sequentially to obtain enhanced data.

[0012] The augmented data is standardized and segmented using a sliding window to obtain sample data that includes normal operating conditions.

[0013] In some optional embodiments of the present invention, the enhanced data is standardized and segmented using a sliding window to obtain sample data containing normal operating conditions, including:

[0014] Robust scaling is used to standardize the augmented data, resulting in standardized data.

[0015] The standardized data is divided into fixed-length sample segments with temporal overlap by using a sliding window and stride of fixed length to obtain sample data containing normal operating conditions.

[0016] In some optional embodiments of the present invention, robust scaling is used to standardize the augmented data to obtain standardized data, including:

[0017] Scaling using the median and interquartile range of augmented data is insensitive to outliers in the data.

[0018] In some optional embodiments of the present invention, the training hyperparameters of the network model include:

[0019] The training rounds are 200, the batch size is 2, the learning rate is 0.0003, the optimizer is Adam, and the early stopping patience is 30.

[0020] In some optional embodiments of the present invention, an LSTM autoencoder early warning model is used to monitor and issue early warnings for real-time acquired data, including:

[0021] The real-time collected and processed data window is input into the trained LSTM autoencoder model to obtain the reconstructed output;

[0022] The reconstruction error is calculated, and a weighted fusion error strategy is used for weighted judgment.

[0023] In some optional embodiments of the present invention, a weighted fusion error strategy is employed for weighted judgment, including:

[0024] A weighted judgment is made by combining the results of static threshold, quantile threshold, local dynamic threshold and gradient detection;

[0025] When the overall score exceeds the preset fusion threshold, it is determined to be an abnormal state.

[0026] In a second aspect of the present invention, an intelligent early warning device for abnormal states of industrial processes based on an LSTM autoencoder is provided, comprising:

[0027] The data processing module is used to preprocess the raw multidimensional time-series signal to obtain sample data containing normal operating conditions.

[0028] The model training module is used to train the network model using sample data containing normal operating conditions, resulting in an LSTM autoencoder early warning model. This network model is built on the TensorFlow framework and consists of an autoencoder architecture composed of multiple stacked LSTM units. The loss function of this network model is the mean squared error.

[0029] The monitoring and early warning module is used to monitor and issue early warnings for real-time collected data using an LSTM autoencoder early warning model.

[0030] In a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the intelligent early warning method for abnormal industrial process states based on an LSTM autoencoder as described in any of the embodiments of the first aspect.

[0031] In a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the intelligent early warning method for abnormal industrial process states based on an LSTM autoencoder as described in any one of the embodiments of the first aspect.

[0032] The beneficial effects provided by this invention are:

[0033] The method described in this invention can detect subtle process deviations before a fault manifests, saving valuable time for preventative maintenance. It employs unsupervised learning, solving the problem of scarce real-world fault data in industrial applications and demonstrating strong engineering applicability. Through deep learning and fusion detection strategies, it exhibits higher detection sensitivity and lower false alarm rates for complex, slowly changing anomalies. This method automates the entire process from data acquisition to anomaly warning, reducing the burden of manual monitoring and improving operational efficiency.

[0034] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0035] Figure 1 A flowchart illustrating the intelligent early warning method for abnormal states in industrial processes based on an LSTM autoencoder provided in this embodiment of the invention.

[0036] Figure 2 This is a schematic diagram of the network structure of the LSTM autoencoder model provided in an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of sliding window segmentation provided in an embodiment of the present invention;

[0038] Figure 4 The flowchart for online abnormal state determination provided in the embodiments of the present invention is shown. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0040] like Figure 1 As shown in this embodiment of the invention, an intelligent early warning method for abnormal states in industrial processes based on an LSTM autoencoder is provided. This method mainly includes steps such as data preprocessing and feature construction, LSTM autoencoder early warning model construction, and online abnormal state determination, specifically including:

[0041] Step 1: Data Preprocessing and Feature Construction. This step aims to convert the raw multi-dimensional time-series signals (such as force, voltage, temperature, etc.) collected from PLCs and sensors into standardized inputs suitable for deep learning models. It includes three stages: data acquisition and alignment, data augmentation, and data standardization and windowing. The specific process is as follows:

[0042] (1) Data Acquisition and Alignment: Time-series data from each sensor within the same production cycle are acquired via OPC UA communication protocol or data export method. Taking the automatic demolding equipment for pharmaceutical columns as an example, the electrostatic force (dian) signal from 6 channels and the demolding force (li) signal from 6 channels are acquired to form 12-dimensional time-series data. The data from different channels are timestamped and data segments of the same length (e.g., 70 time steps) are extracted to ensure data frame synchronization.

[0043] (2) Data augmentation, mainly through ① missing value handling, ② outlier suppression, and ③ noise injection, enhances the data to improve the robustness of the model. Detailed procedures are as follows:

[0044] ① Missing value handling: A combination of linear interpolation and forward / backward filling (ffill / bfill) is used to handle the small number of missing data points that may occur during the acquisition, ensuring the continuity of the data stream.

[0045] ② Outlier Suppression: To prevent sensor transients or interference from affecting the data distribution, a method based on interquartile range (IQR) is used to limit the data. The lower quartile Q1 and upper quartile Q3 of each channel's data are calculated, and data points exceeding the range [Q1-1.5*IQR, Q3+1.5*IQR] are corrected to boundary values.

[0046] ③ Noise Injection: During the model training phase, slight Gaussian white noise is injected into the normal training data to simulate real sensor fluctuations and environmental interference. The standard deviation of the noise is set to one-thousandth of the standard deviation of the original signal (noise_level = 0.001). This step enhances the model's tolerance to normal fluctuations, allowing it to focus more on identifying true anomalous patterns.

[0047] (3) Data standardization and windowing: To eliminate the influence of different physical dimensions and construct input samples suitable for time series models, the following operations are performed:

[0048] ① Data Standardization: Robust scaling is used to standardize the data. This method uses the median and interquartile range for scaling, is insensitive to outliers in the data, and has better stability compared to Z-score standardization. The calculation formula is as follows:

[0049] X_scaled = (X - Median) / IQR

[0050] ② Sliding window segmentation: such as Figure 3As shown, standardized continuous multidimensional time series data is segmented into a series of temporally overlapping, fixed-length sample segments by setting a fixed-length sliding window (WINDOW_SIZE=10) and a stride (Stride=1). Each segment has dimensions [10,12] and serves as an independent input sample for subsequent model training and inference.

[0051] Step 2: Construction of the LSTM autoencoder early warning model. Its core is to use the sample data generated in Step 1, which only contains normal working conditions, to train a deep neural network model in an unsupervised learning manner.

[0052] (1) Input sample format: The windowed samples with dimensions [BatchSize, 10, 12] are used as the unified input of the network model.

[0053] (2) Network Model Construction: Based on the TensorFlow framework, an autoencoder architecture consisting of multiple stacked LSTM units was built. The specific network structure used is shown in the table below:

[0054] Table 1 LSTM autoencoder network structure

[0055]

[0056] The loss function of the network model is the mean squared error (MSE), which aims to minimize the difference between the original input window X and the model's reconstructed output window X'.

[0057] L(X,X')=(1 / (W×F))*Σ(X ij -X' ij ) 2

[0058] (3) Network training parameters: The key hyperparameters used to train the model are shown in the table below:

[0059] Table 2 Training Hyperparameters

[0060] parameter value Training epochs 200 Batch Size 2 Learning Rate 0.0003 Optimizer Adam Stop early and be patient. 30

[0061] Based on the above parameters and network structure, the model is trained until the validation set loss no longer decreases (early stopping mechanism is triggered).

[0062] Step 3: Online anomaly detection. After the model training is completed, it is deployed in the actual production environment to monitor and issue early warnings for the real-time collected data. This includes two stages: reconstruction error calculation and fusion detection strategy.

[0063] ① Reconstruction Error Calculation: The real-time acquired and processed data window is input into the trained LSTM autoencoder model to obtain the reconstructed output. A weighted fusion error is used to comprehensively evaluate the degree of deviation, as shown in the following formula:

[0064] Error=0.4*MSE+0.2*MAE+0.4*MaxChannelError

[0065] Where MSE is the mean squared error, MAE is the mean absolute error, and MaxChannelError is the mean squared error of the maximum error across all channels. This fusion strategy can simultaneously take into account overall bias, absolute bias, and local prominent bias, thereby improving detection sensitivity.

[0066] ② Fusion Detection and Early Warning: Instead of using a single fixed threshold, a multi-algorithm fusion detection strategy is designed to improve the accuracy and robustness of the judgment. This strategy uses a weighted average of the results from the following four detection algorithms:

[0067] Static threshold detection: Set a global threshold that is insensitive to outliers based on the median absolute deviation (MAD).

[0068] Quantile detection: Set a threshold based on the higher quantile (e.g., 98%) of the normal data error distribution.

[0069] Local dynamic threshold detection: Calculate the local mean and standard deviation within a sliding small time window, and set a dynamic threshold to adapt to the slow changes in the error baseline.

[0070] Gradient detection: Calculate the gradient of the error curve. When the absolute value of the gradient exceeds the threshold, it is judged as a sudden abnormality.

[0071] The outputs (0 or 1) of the four algorithms are weighted and summed according to preset weights (e.g., [0.25, 0.2, 0.25, 0.3]) to obtain a comprehensive score. When this score exceeds the final fusion threshold (e.g., fusion_threshold = 0.3), the system triggers an anomaly warning and performs post-processing on the result (e.g., removing isolated single-point alarms) to reduce false alarms.

[0072] In another embodiment of the invention, the implementation process begins with data acquisition. Taking an automatic demolding device for pharmaceutical columns as an example, real-time data from its demolding force sensor and electrostatic sensor are acquired via a PLC at a sampling frequency of 10Hz. The acquired data first undergoes processing in step one: missing points are filled using linear interpolation, and normalization is performed using robust scaling. Subsequently, the window length W = 10 (i.e., 1 second of data) and the step size S = 1 are set, and the data stream is divided into multiple 10x12 matrix samples.

[0073] As attached Figure 2As shown, proceed to step two to construct a network consisting of a 3-layer LSTM encoder and a 3-layer LSTM decoder, the detailed structure of which is shown in Table 1. Implement the model using the TensorFlow framework and employ the Adam optimizer. Train on a massive number of normal operating condition samples for 200 epochs until the validation set loss no longer decreases (early stopping mechanism triggered, patience value 30).

[0074] After training is complete, proceed to step three, the online application phase. (See attached image.) Figure 4 As shown, the real-time acquired data window is fed into the model to calculate the reconstruction error. This invention employs a weighted fusion error strategy, combining MSE, MAE, and maximum channel error. It is assumed that the mean error obtained on the validation set is 0.005, and the standard deviation is 0.002. Subsequently, a fusion detection strategy is used, combining the results of static threshold, quantile threshold, local dynamic threshold, and gradient detection for weighted judgment. When the comprehensive score exceeds a preset fusion threshold (e.g., 0.3), the system determines it to be in an abnormal state and immediately issues a warning message to the operator via the host computer interface: "Abnormal trend in process parameters; please check the equipment."

[0075] The methods described above can capture subtle process deviations before faults manifest, buying valuable time for preventative maintenance; they employ unsupervised learning, solving the problem of scarce real-world fault data in the industrial field, and have strong engineering applicability; they exhibit high sensitivity and robustness: through deep learning and fusion detection strategies, they have higher detection sensitivity and lower false alarm rates for complex, slowly changing anomalies; they achieve full automation from data acquisition to anomaly warning, reducing the burden of manual monitoring and improving operational efficiency.

[0076] The above-described embodiments overcome the limitations of traditional threshold alarm systems and provide a method for highly accurate and intelligent early warning of early, slowly changing abnormal states in industrial processes under fault-free conditions.

[0077] It should be understood that in the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this description, 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 may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate different embodiments or examples described in this specification, as well as some features of different embodiments or examples.

[0078] Of course, those skilled in the art can make various corresponding changes and modifications based on the present invention without departing from its spirit and essence, but such changes and modifications should all fall within the protection scope of the claims of the present invention.

Claims

1. An intelligent early warning method for abnormal state of industrial process based on LSTM autoencoder, characterized in that, The method comprises the following steps: Data preprocessing is performed on the original multi-dimensional time series signal to obtain sample data containing normal working conditions; A LSTM autoencoder early warning model is obtained by training a network model using the sample data containing normal working conditions, the network model is built based on a TensorFlow framework and has a self-encoder architecture composed of multiple layers of stacked LSTM units, and a mean square error is used as the loss function of the network model; Real-time collected data is monitored and warned by using the LSTM autoencoder early warning model. 2.The intelligent early warning method of industrial process abnormal state based on LSTM auto-encoder according to claim 1, characterized in that, Data preprocessing is performed on the original multi-dimensional time series signal collected from a PLC and a sensor to obtain sample data containing normal working conditions, which comprises the following steps: Time series data of each sensor in the same production cycle is obtained; The time series data is sequentially cleaned, aligned and enhanced to obtain enhanced data; The enhanced data is standardized and divided into sliding windows to obtain sample data containing normal working conditions. 3.The intelligent early warning method of industrial process abnormal state based on LSTM auto-encoder according to claim 2, characterized in that, The enhanced data is standardized and divided into sliding windows to obtain sample data containing normal working conditions, which comprises the following steps: Robust scaling is used to standardize the enhanced data to obtain standardized data; The standardized data is divided into sample segments with time series overlap and fixed length by using a fixed length sliding window and a step to obtain sample data containing normal working conditions. 4.The intelligent early warning method of industrial process abnormal state based on LSTM auto-encoder according to claim 3, characterized in that, Robust scaling is used to standardize the enhanced data to obtain standardized data, which comprises the following steps: The median and interquartile range of the enhanced data are used for scaling, which is insensitive to outliers in the data. 5.The intelligent early warning method of industrial process abnormal state based on LSTM auto-encoder according to claim 1, characterized in that, The training hyperparameters of the network model comprise: The training round is 200, the batch size is 2, the learning rate is 0.0003, the optimizer is Adam, and the early stopping patience is 30. 6.The intelligent early warning method of industrial process abnormal state based on LSTM auto-encoder according to claim 1, characterized in that, Real-time collected data is monitored and warned by using the LSTM autoencoder early warning model, which comprises the following steps: The data window collected and processed in real time is input into the trained LSTM autoencoder model to obtain the reconstructed output; The reconstruction error is calculated, and a weighted fusion error strategy is used for weighted judgment.

7. The intelligent early warning method of industrial process abnormal state based on LSTM autoencoder according to claim 6, characterized in that, The weighted fusion error strategy is used for weighted judgment, which comprises the following steps: The results of the static threshold, the quantile threshold, the local dynamic threshold and the gradient detection are combined for weighted judgment; When the comprehensive score exceeds the preset fusion threshold, it is determined that the abnormal state occurs.

8. An intelligent early warning device for abnormal state of an industrial process based on LSTM autoencoder, characterized in that, The method comprises the following steps: A data processing module is configured to perform data preprocessing on the original multi-dimensional time series signal to obtain sample data containing normal working conditions; A model training module is configured to train a network model using the sample data containing normal working conditions to obtain a LSTM autoencoder early warning model, the network model is built based on a TensorFlow framework and has a self-encoder architecture composed of multiple layers of stacked LSTM units, and a mean square error is used as the loss function of the network model A monitoring and warning module is configured to monitor and warn real-time collected data by using the LSTM autoencoder early warning model.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the electronic device to perform the method of any one of claims 1-8. The processor executes the computer program to implement the steps of the LSTM autoencoder-based industrial process abnormal state intelligent early warning method according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the intelligent early warning method for abnormal states of an industrial process based on an LSTM autoencoder according to any one of claims 1-7.

Images