A labyrinth piston compressor intake state monitoring method and system

Abstract

This application relates to the field of data processing technology, and in particular to a method and system for monitoring the intake status of a labyrinth piston compressor. The method includes: acquiring real-time multidimensional data containing pressure, temperature, and vibration information; inputting the real-time multidimensional data into a trained LSTM autoencoder model to obtain reconstructed data; calculating the difference between the real-time multidimensional data and the reconstructed data to obtain a smoothed residual signal; constructing a feature vector based on the smoothed residual signal; inputting the feature vector into a decision model trained based on a feature vector set under normal operating conditions to obtain an anomaly score; and, in response to the anomaly score exceeding a set threshold, determining that the compressor intake status is abnormal and issuing an alarm. This application effectively reduces the occurrence of missed or false detections in compressor intake status monitoring.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for monitoring the intake status of a labyrinth piston compressor. Background Technology

[0002] Reciprocating compressors, especially labyrinth piston compressors, are used in petrochemical and gas transportation industries due to their oil-free lubrication and high airtightness, serving as core power equipment in production processes. During operation, the state of the intake system directly affects the overall efficiency and operational safety of these compressors. Early, subtle faults such as broken intake valve plates, leaks caused by seal wear, or signs of impending liquid slugging, if not detected promptly, can easily escalate into serious accidents such as shutdowns or even equipment damage. Therefore, achieving highly sensitive, low-false-alarm real-time monitoring of the intake status of labyrinth piston compressors has been a crucial and long-standing focus within the industry.

[0003] In existing technologies, early static threshold alarm methods were used, which set fixed thresholds for raw signals such as pressure, temperature, and vibration to determine anomalies. While simple to implement, these methods are highly susceptible to fluctuations in normal operating conditions and environmental noise, resulting in a high false alarm rate. Furthermore, they lack sensitivity to weak, slowly developing fault signals, leading to a significant risk of missed alarms. Subsequently, researchers attempted to extract frequency or time-frequency domain features using signal processing techniques such as Fourier transform and wavelet analysis, then combined these with empirical rules or classifiers for diagnosis. However, these methods are highly dependent on expert experience, and the feature extraction process is complex. When fault signals are overwhelmed by strong background noise, the discriminative power of the extracted features decreases significantly, making reliable early warning difficult.

[0004] With the development of deep learning, autoencoder-based temporal anomaly detection methods are increasingly being applied in industrial monitoring scenarios. A typical example is the LSTM autoencoder method, which trains an LSTM encoder-decoder structure on normal operating data, enabling the model to learn the dynamic patterns of normal time series and using reconstruction error as the criterion for anomaly detection. These methods can achieve a certain degree of automated modeling, eliminating the need for tedious manual feature extraction, and have shown certain advantages in various temporal anomaly detection tasks.

[0005] However, this closest existing technology still has a significant problem: because the training objective is only to minimize the reconstruction error, the model reproduces high-frequency random noise in the input signal while learning normal patterns. This results in noise and real fault features being mixed in the obtained residual signal, leading to a low signal-to-noise ratio. In this case, downstream discrimination based on features such as residual amplitude or first-order difference is easily interfered with by noise, which may misjudge normal fluctuations as abnormalities or miss early fault signals with weak amplitude and slow development. Summary of the Invention

[0006] To reduce the occurrence of missed detections and misjudgments during compressor intake monitoring, this application provides a method and system for monitoring the intake status of a labyrinth piston compressor.

[0007] Firstly, this application provides a method for monitoring the intake status of a labyrinth piston compressor, employing the following technical solution: A method for monitoring the intake status of a labyrinth piston compressor includes: acquiring real-time multidimensional data containing pressure, temperature, and vibration information; inputting the real-time multidimensional data into a trained LSTM autoencoder model to obtain reconstructed data; calculating the difference between the real-time multidimensional data and the reconstructed data to obtain a smoothed residual signal; constructing a feature vector based on the smoothed residual signal, the feature vector including the following features: the amplitude of the smoothed residual signal; the absolute value of the first-order difference of the smoothed residual signal; the standard deviation of the smoothed residual signal within a set window; and the root mean square value of the smoothed residual signal within a set window; inputting the feature vector into an adjudication model trained based on a feature vector set under normal operating conditions to obtain an anomaly score; and, in response to the anomaly score being greater than a set threshold, determining that the compressor intake status is abnormal and issuing an alarm.

[0008] An LSTM autoencoder model is trained, and then healthy training data under normal operating conditions is input into the LSTM autoencoder model. The model reconstructs the healthy training data, and residuals are obtained from the reconstructed data. Feature vectors corresponding to the healthy training data are then extracted. For ease of understanding, the feature vectors obtained from the reconstructed healthy training data are used as healthy feature vectors. An isolated forest model is trained based on these healthy feature vectors, thus completing the preparation work for monitoring. During actual monitoring, real-time multidimensional data is input into the trained LSTM autoencoder model, and smoothed residual signals are obtained from the reconstructed data. Similarly, feature engineering is performed on the smoothed residual signals to extract feature vectors. These feature vectors are then input into the adjudication model to complete the monitoring of compressor intake. The feature vectors represent the amplitude of the smoothed residual signal; the absolute value of the first-order difference of the smoothed residual signal; the standard deviation of the smoothed residual signal within a set window; and the root mean square value of the smoothed residual signal within a set window. These features exhibit different sensitivities to different types of faults, thereby improving the accuracy and sensitivity of multi-type fault detection during the detection process.

[0009] Optionally, the LSTM autoencoder model is trained based on normal operating condition data, and its training process includes a smoothing constraint term that penalizes the second-order difference amplitude of the residual signal.

[0010] By combining the smoothing constraint term and the reconstruction error term into the loss function, the goal of noise suppression is integrated into the optimization process of model training. This design ensures that the model simultaneously considers both reconstruction accuracy and smoothed residual generation at each step of weight update, giving the final model an inherent noise reduction capability in its structure.

[0011] Optionally, the loss function during the training process of the LSTM autoencoder model is composed of a smoothing constraint term and a reconstruction error term, wherein the reconstruction error term is the mean square error between the normal data and the reconstructed data corresponding to the input training data.

[0012] Optionally, the loss function may also include weighting coefficients multiplied by the smoothing constraint term.

[0013] The introduction of adjustable weighting coefficients provides crucial flexibility for model optimization, allowing users to balance reconstruction accuracy with smoother residuals. For example, weighting coefficients can be increased in noisy environments to achieve smoother residuals, while they can be decreased in scenarios requiring the capture of minute, normal fluctuations, thus better adapting to different devices and operating conditions.

[0014] Optionally, the steps for obtaining the weight coefficients include: using the product of the standard deviation of the noise artificially added to the training data and the preset adjustment coefficient as the adjustment amount, and using the sum of the preset base weights and the adjustment amount as the weight coefficients.

[0015] Dynamic weight coefficients are generated based on the noise fluctuations in the training data to ensure the smoothness of the residuals and reduce the likelihood of learning fault modes. When the noise fluctuations in the training data are large, larger weight coefficients are used to impose more constraints on the residuals, preventing the model from learning artificially introduced noise from the training data.

[0016] Optionally, during the process of determining that the compressor intake state is abnormal and issuing an alarm, there are multiple alarm types, and the multiple alarm types are determined based on the magnitude of the abnormality score.

[0017] Alarms are categorized into multiple types, which allows for the quantification of the level of data anomalies and the determination of the urgency of repairs based on the level.

[0018] Optionally, there are three alarm types: a secondary alarm is triggered when the abnormal score is within the first threshold range; and a primary alarm is triggered when the abnormal score is greater than the upper limit of the first threshold range.

[0019] Optionally, when an alarm is triggered, a retrospective analysis is performed on the feature that contributes the most to the current anomaly score; based on the feature that contributes the most, preliminary diagnostic suggestions on the fault mode are provided to the operations and maintenance personnel.

[0020] When an anomaly occurs, the feature that contributes the most to the increase in the anomaly score is the backtracking analysis, which can help determine the type of intake anomaly. For example, if the increase in the anomaly score at a certain moment is mainly caused by the root mean square value of the smoothed residual signal in the feature vector, then the current fault may be a slowly accumulating fault.

[0021] Optionally, the feature vector may further include the smoothing consistency of the smoothed residual signal, where smoothing consistency is the ratio of the mean of the second-order difference of the smoothed residual signal to the preset baseline smoothness in the model.

[0022] The baseline smoothness represents the smoothness of the residual signal between the reconstructed signal and the normal signal under normal conditions. This smoothness is the normal condition, so using this smoothness as the baseline smoothness can capture abnormal changes in the smoothness of the residual signal during real-time detection, thereby improving the accuracy of detection.

[0023] Secondly, the labyrinth piston compressor intake status monitoring system disclosed in this application adopts the following technical solution: A labyrinth piston compressor intake status monitoring system includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the labyrinth piston compressor intake status monitoring method described above is implemented.

[0024] This application has the following technical effects: By introducing a smoothing constraint term into the loss function during model training, the problems of low signal-to-noise ratio and confusion between noise and fault features in the residual signal are reduced. This method can generate high-purity smooth residual signals, thereby highlighting early and weak fault characteristics. Combined with a multi-dimensional feature vector constructed based on this signal and an isolated forest model for intelligent adjudication, this significantly improves the sensitivity and reliability of detecting early and weak faults in the compressor intake system. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a method for monitoring the intake status of a labyrinth piston compressor, as provided in an embodiment of this application. Detailed Implementation

[0026] This application discloses a method for monitoring the intake state of a labyrinth piston compressor, which constructs a specially designed Long Short-Term Memory (LSTM) autoencoder model. This model reconstructs the input signal, thereby refining a smooth residual signal with a high signal-to-noise ratio. The high-purity residual signal is then subjected to deep, multi-dimensional analytical extraction of feature vectors. These feature vectors are input into an adjudication model, and the compressor intake state is determined based on the anomaly score output by the adjudication model.

[0027] Reference Figure 1A method for monitoring the intake status of a labyrinth piston compressor includes steps S1-S4.

[0028] S1: Acquire real-time multidimensional data containing pressure, temperature, and vibration information.

[0029] First, sensors are deployed at key measuring points of the labyrinth piston compressor. Specifically, high-frequency dynamic pressure sensors, fast-response temperature sensors, and broadband vibration acceleration sensors are simultaneously deployed at locations such as the compressor's intake manifold and the intake chambers of each cylinder. These sensors can comprehensively capture dynamic information about the compressor's operating status from multiple physical dimensions, including pressure, temperature, and vibration.

[0030] To ensure precise data alignment over time, a high-precision synchronous data acquisition card (DAQ) is used to synchronously acquire signals from all sensors. As a preferred approach, the sampling frequency of the sensors ranges from 10kHz to 50kHz. When the sampling frequency is below 10kHz, high-frequency characteristic information related to early, subtle faults, such as the subtle pressure pulsations that precede liquid hammer, may be lost due to non-compliance with the Nyquist sampling theorem. When the sampling frequency is above 50kHz, a large amount of redundant data is generated, increasing the burden of subsequent computation and storage while offering limited improvement in monitoring effectiveness. Therefore, setting the sampling frequency within this preferred range ensures both information integrity and computational efficiency.

[0031] The acquired raw multidimensional time series data is fed into a high-performance edge computing industrial PC (IPC) for caching, alignment, and standardization preprocessing. Standardization can be achieved using methods such as Z-score standardization to eliminate dimensional differences between data from different sensors, thereby obtaining real-time multidimensional data.

[0032] S2: Input the real-time multidimensional data into the trained LSTM autoencoder model to obtain reconstructed data; calculate the difference between the real-time multidimensional data and the reconstructed data to obtain a smoothed residual signal.

[0033] The LSTM autoencoder model is pre-trained, employing a Long Short-Term Memory (LSTM)-based autoencoder structure during training. This network consists of an encoder and a decoder. The encoder compresses the high-dimensional time-series input data into a low-dimensional hidden state vector, while the decoder attempts to reconstruct the original input sequence from this hidden state vector. The input, forget, and output gate structures within the LSTM unit enable it to effectively learn long-term dependencies in time-series data; this structure is a standard technique in the field and will not be elaborated upon further.

[0034] The LSTM autoencoder model is trained based on pre-collected training data. This training data includes normal data acquired during the normal operation of the equipment and simulated fault data created by artificially adding noise to the normal data. The overall training objective is for the LSTM autoencoder model to reconstruct normal data regardless of whether the input training data is simulated fault data or normal data. The loss function during the LSTM autoencoder model training process includes: a smoothing constraint term, a reconstruction error term, and weight coefficients multiplied by the smoothing constraint term.

[0035] Specifically, the loss function during the training process of the LSTM autoencoder model can be expressed as: In the formula, Represented as the total loss function; The reconstruction error term is used to ensure the basic predictive ability of the model, that is, to make the reconstructed data approximate the normal signal. In this embodiment, the reconstruction error term is the mean square error between the normal data and the reconstructed data corresponding to the input training data. The calculation of the mean square error is a conventional technique in this field and will not be described in detail here. These are the weighting coefficients; This is a smoothing constraint term.

[0036] In the formula, To facilitate description and differentiation, the residual signal generated by reconstructing the data during training is used as the training residual signal for the smoothing constraint term.

[0037] Calculate the sum of squares of the second-order differences of the training residual signal at each time step, and use this sum of squares as the second-order difference penalty term.

[0038] In this embodiment, the second-order difference penalty term can be directly used as the smoothing constraint term. When the residual signal fluctuates drastically, the value of its second-order difference will be very large, causing the smoothing constraint term to increase, thereby imposing a penalty during model training and forcing the model to reconstruct a smooth signal.

[0039] , which is a weighting coefficient, is mainly used to balance the reconstruction accuracy and the smoothness of the residual.

[0040] In one embodiment, this value can be set by someone skilled in the art based on experience, and its range can be from 0.1 to 10. When When the value is too small (e.g., less than 0.1), the smoothing constraint term has little effect, and the generated residual signal will still contain a lot of noise. When its value is too large (e.g., greater than 10), the model will excessively suppress all dynamic changes in order to minimize the smoothing constraint term, which may lead to a decrease in sensitivity to real fault signals. Therefore, methods such as cross-validation are used to... Setting it to, for example, around 1.0, can ensure reasonable reconstruction accuracy while obtaining the smoothest and purest residual signal.

[0041] In another embodiment, dynamic weight coefficients can be obtained based on the noise fluctuations of the input training data.

[0042] For example, for any input training data, if the training data is simulated fault data with added noise, the weight coefficients are adjusted according to the standard deviation of the added noise. Specifically, the adjustment amount can be obtained through the standard deviation of the noise. For normal data without added noise, the standard deviation of the noise in the training data is directly set to 0. The sum of the preset base weight and the adjustment amount is used as the weight coefficient of the total loss function for this training. In this embodiment, the base weight is set to 0.2 based on the experience of the staff.

[0043] In this embodiment, the product of the standard deviation and the preset adjustment coefficient is used as the adjustment amount. The preset adjustment coefficient is set by those skilled in the art based on experience. In this embodiment, it is set to 0.05, which is mainly used to adjust the sensitivity of the influence of the standard deviation on the weighting coefficient.

[0044] After the LSTM autoencoder model is trained, real-time multidimensional data is input into the trained LSTM autoencoder model. The LSTM autoencoder model reconstructs the input data to obtain reconstructed data; the difference between the real-time multidimensional data and the reconstructed data is calculated to obtain the reference at each time step, and then the smoothed residual signal composed of the residuals at each time step is obtained.

[0045] S3: Construct a feature vector based on the smoothed residual signal. The feature vector includes the following features: the amplitude of the smoothed residual signal, the absolute value of the first-order difference of the smoothed residual signal, the standard deviation of the smoothed residual signal within a set window, and the root mean square value of the smoothed residual signal within a set window.

[0046] The amplitude of the smoothed residual signal is most sensitive to impact-type and sudden faults (such as instantaneous pressure anomalies caused by valve plate breakage). The absolute value of the first-order difference of the smoothed residual signal, that is, the difference between the residuals of two adjacent moments, is more sensitive to creep-type and trend-type faults (such as gradual leakage caused by seal wear), thus enabling the capture of slowly changing faults.

[0047] The standard deviation of the smoothed residual signal within the set window refers to the standard deviation of the residual signal within a relatively short set window, for example, the length of the set window could be 10 sampling times. This feature is used to capture intermittent, unstable faults (such as system oscillations caused by controller parameter drift). When the device state intermittently switches between normal and minor abnormalities, the standard deviation of the value will increase.

[0048] The root mean square (RMS) value of the smoothed residual signal within the set window is used to measure the cumulative effect and persistence of the fault. Compared to the instantaneous amplitude, the RMS value better reflects the cumulative level of abnormal energy over a period of time.

[0049] In other embodiments, the feature vector may further include a smoothing consistency term representing the smoothness of the smoothed residual signal and the smoothing constraint term of the model's intrinsic parameters. The formula for calculating the smoothing consistency can be expressed as: In the formula, Represents time in the smoothed residual signal Smooth consistency; Indicated by time Based on the average of the second-order differences of the residual signals at each time step within a preset fixed window (the window length is set by those skilled in the art based on experience); This represents the baseline smoothness in the model. In actual implementation, this value will not be 0. Of course, to prevent the denominator from being 0, hyperparameters can be set in the denominator and added to the smoothness constraint term. A value close to 1 indicates that the smoothness of the current residual is consistent with the smoothness under normal operating conditions.

[0050] The steps to obtain the baseline smoothness include: after training is completed, recording the mean of the second-order difference penalty term of the model on the normal working condition dataset of different training batches, and using it as the baseline smoothness.

[0051] Its calculation formula can be expressed as: in, Indicates the baseline smoothness; The total number of training batches. For the first The second-order difference penalty value corresponds to each training batch. The baseline smoothness value represents the best smoothness consistency baseline that the LSTM model can achieve under normal operating conditions.

[0052] S4: Input the feature vector into the adjudication model trained based on the feature vector set under normal operating conditions to obtain the abnormal score; in response to the abnormal score being greater than a set threshold, determine that the compressor intake state is abnormal and issue an alarm.

[0053] In this embodiment, the adjudication model is the isolated forest model, which has been pre-trained.

[0054] The data used in the training of the isolated forest model is the feature vector generated by the trained LSTM autoencoder model from healthy training data. The acquisition of the feature vector is the same as in step S3 above, and will not be repeated here.

[0055] The feature vectors corresponding to the real-time multidimensional data are input into the isolated forest model to obtain anomaly scores, and the compressor status is determined based on the anomaly scores.

[0056] Specifically, in this embodiment, multiple alarm thresholds are set to achieve tiered alarms. For example, a level-two alarm is triggered when the abnormal score is within the first threshold range; a level-one alarm is triggered when the abnormal score is greater than the upper limit of the first threshold range. For instance, a level-two alarm is triggered when 0.7 ≤ abnormal score < 0.85, and a level-one alarm is triggered when the abnormal score ≥ 0.85. Simultaneously, when an alarm occurs, the system can retrospectively analyze the main characteristic dimensions that led to the increased abnormal score, thereby providing maintenance personnel with preliminary diagnostic suggestions regarding the fault mode, greatly shortening the fault investigation time.

[0057] For example, if the increase in the final abnormal score is mainly due to the standard deviation of the smoothed residual signal within the set window, then the fault may be a capture intermittent, unstable fault.

[0058] This application also discloses a labyrinth piston compressor intake status monitoring system, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, a labyrinth piston compressor intake status monitoring method according to this application is implemented.

[0059] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0060] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method of monitoring the intake condition of a labyrinth piston compressor, characterized in that include: Real-time multidimensional data containing pressure, temperature, and vibration information is acquired; the real-time multidimensional data is input into a trained LSTM autoencoder model to obtain reconstructed data; the difference between the real-time multidimensional data and the reconstructed data is calculated to obtain a smoothed residual signal; a feature vector is constructed based on the smoothed residual signal, the feature vector containing the following features: the amplitude of the smoothed residual signal, the absolute value of the first-order difference of the smoothed residual signal, the standard deviation of the smoothed residual signal within a set window, and the root mean square value of the smoothed residual signal within a set window; the feature vector is input into an adjudication model trained based on a feature vector set under normal operating conditions to obtain an anomaly score; in response to the anomaly score being greater than a set threshold, the compressor intake state is determined to be abnormal and an alarm is issued; The LSTM autoencoder model is trained based on pre-collected training data, which includes normal data acquired during normal equipment operation and simulated fault data formed by artificial noise in the normal data. The training process includes a smoothing constraint term that penalizes the second-order difference amplitude of the residual signal. The loss function in the training process of the LSTM autoencoder model consists of the sum of the smoothing constraint term and the reconstruction error term, where the reconstruction error term is the mean square error between the normal data and the reconstructed data corresponding to the input training data. The loss function also includes weighting coefficients multiplied by the smoothing constraint term; The steps for obtaining the weight coefficients include: using the product of the standard deviation of the noise artificially added to the training data and the preset adjustment coefficient as the adjustment amount, and using the sum of the preset base weights and the adjustment amount as the weight coefficients.

2. A labyrinth piston compressor suction condition monitoring method according to claim 1, characterized in that, In the process of determining that the compressor intake status is abnormal and issuing an alarm, there are multiple alarm types, and the determination of multiple alarm types is based on the magnitude of the abnormality score.

3. A labyrinth piston compressor suction condition monitoring method according to claim 2, characterized in that, There are two types of alarms: a level 2 alarm is triggered when the abnormal score is within the first threshold range; and a level 1 alarm is triggered when the abnormal score is greater than the upper limit of the first threshold range.

4. A labyrinth piston compressor suction condition monitoring method according to claim 1, characterized in that, When an alarm is triggered, the feature that contributes the most to the current anomaly score is analyzed back; based on the feature that contributes the most, preliminary diagnostic suggestions about the fault mode are provided to the operation and maintenance personnel.

5. A labyrinth piston compressor suction condition monitoring method according to claim 1, characterized in that, The feature vector also includes the smoothing consistency of the smoothed residual signal, where smoothing consistency is the ratio of the mean of the second-order difference of the smoothed residual signal to the preset baseline smoothness in the model.

6. A labyrinth piston compressor suction condition monitoring system characterized by, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement a method for monitoring the intake status of a labyrinth piston compressor according to any one of claims 1-5.

Images