A multidimensional feature analysis method and system for compressor airtightness testing

Abstract

This invention relates to the field of airtightness testing technology, specifically to a multi-dimensional feature analysis method and system for compressor airtightness testing. The method involves collecting data using a multi-sensor array at each testing stage of the compressor under test, obtaining multiple types of feature vectors, and then concatenating and normalizing them according to time windows to form a high-dimensional comprehensive feature vector. An ensemble learning model is trained using historical sample testing data. The high-dimensional comprehensive feature vector is input into the ensemble learning model to obtain the leakage probability and leakage inference decision, and the feature contribution confidence score is calculated. Based on the leakage probability and feature contribution confidence score, the airtightness determination result is obtained. This technical solution combines multiple sensors to collect multiple types of data and performs multi-dimensional feature fusion analysis, fundamentally improving the completeness and robustness of perception, resulting in high detection accuracy and high adaptability.

Description

Technical Field

[0001] This invention relates to the field of airtightness testing technology, specifically to a multi-dimensional feature analysis method and system for airtightness testing of compressors. Background Technology

[0002] As a core power component, the airtightness of the compressor is a key indicator for measuring product reliability and energy efficiency. Currently, the mainstream airtightness testing methods in the industry primarily rely on direct pressure measurement and its derivative technologies. The classic pressure drop method operates by filling the sealed compressor cavity with gas at a set pressure. After a pressure stabilization phase, a high-precision pressure sensor monitors the pressure change over time, and a preset threshold is used to determine if a leak exists. To improve detection sensitivity, the industry has developed the differential pressure method. This method reduces the interference of ambient temperature fluctuations on the test results by comparing the pressure difference between the tested workpiece and a standard reference container. In addition, some advanced production lines supplement the results with bubble leak detection or halogen leak detection methods; however, these methods have limitations such as low detection efficiency or the need for specific tracer gases and complex testing environments. In recent years, non-destructive testing technologies such as acoustic testing and infrared thermal imaging testing have been gradually explored and applied to online or offline testing scenarios of compressor air tightness. However, in practical applications, the pressure method is still the core basis for air tightness judgment, and other testing methods are only used as auxiliary verification means in the fault reproduction stage. Moreover, the data collected by various sensors are all processed independently, lacking in-depth spatiotemporal correlation analysis and multi-source information fusion mechanism.

[0003] While existing detection technologies offer diverse detection methods, they still exhibit significant limitations and technical problems in practical applications: First, the single pressure decay method is susceptible to interference from non-leakage factors such as ambient temperature fluctuations, compressor thermal expansion, and minor deformations of the test system, resulting in insufficient detection capability for slow or minor leaks. Second, multi-source detection information is isolated; even with the integration of multiple sensors, the readings or alarm statuses of each sensor are often displayed in parallel, failing to provide a comprehensive and three-dimensional characterization of leakage events or effectively distinguish between different leakage modes. Third, the detection system employs a design architecture with fixed thresholds and static models, making it unable to adapt to the drift in detection characteristics caused by different compressor models, production batch differences, sensor aging, and changes in environmental background noise. Frequent manual calibration and parameter adjustments are required to ensure detection effectiveness, leading to high maintenance costs and poor long-term system stability. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a multi-dimensional feature analysis method and system for compressor airtightness detection, so as to solve the problems of easy interference of single threshold judgment and isolated and insufficient adaptation of multi-source detection information in the prior art.

[0005] According to a first aspect of the present invention, a multidimensional feature analysis method for compressor airtightness testing is provided, applied to airtightness testing equipment, comprising:

[0006] Pressure data, acoustic data, thermodynamic data, and tracer gas concentration data are collected using a multi-sensor array during the preset charging detection stage, pressure holding detection stage, and pressure stabilization detection stage of the compressor under monitoring. The collected data are then integrated into a three-dimensional raw data matrix. Based on pressure data, we obtain pressure decay rate characteristic vector and spatial pressure field stability characteristic vector; based on acoustic data, we obtain acoustic leakage characteristic vector; based on thermodynamic data, we obtain thermodynamic anomaly characteristic vector; based on tracer gas concentration data, we obtain gas diffusion characteristic vector. All the obtained feature vectors are concatenated and normalized according to the time window to form a high-dimensional comprehensive feature vector. Historical sample detection data are integrated into a training sample library. Each historical sample detection data includes an airtightness status label and a high-dimensional comprehensive feature vector corresponding to the compressor sample. The pre-built ensemble learning model is trained using the training sample library. The high-dimensional integrated feature vector is input into the trained ensemble learning model to obtain the leakage probability and leakage inference decision; the feature contribution confidence score is calculated based on the contribution of each feature dimension to the current decision; and the airtightness determination result is obtained based on the leakage probability and the feature contribution confidence score.

[0007] Preferably, pressure data, acoustic data, thermodynamic data, and tracer gas concentration data are acquired using a multi-sensor array, including: The absolute pressure value inside the compressor cavity and the dynamic pressure gradient between different cavity regions are collected using a pressure sensor array. A broadband acoustic sensor is used to collect acoustic emission signals and ultrasonic signals from the compressor housing surface and the surrounding air medium. A sequence of thermal imagery of the key sealing surfaces and weld areas of the compressor was acquired using an infrared thermal imager at a fixed frame rate. The background concentration and potential leakage concentration of a specific tracer gas in the test environment are monitored in real time using a gas concentration sensor.

[0008] Preferably, the collected data is integrated into a three-dimensional raw data matrix, including: Based on a hardware-triggered unified clock, time-domain labels are synchronously applied to all sensor data streams, and synchronization alignment is performed according to the time-domain labels. Kalman filtering is used to smooth fluctuations in pressure data, wavelet thresholding is used to denoise acoustic data to preserve leakage characteristic frequency bands, and mean filtering is used to eliminate transient noise in thermodynamic data. By statistically analyzing outliers, we can detect, identify, and eliminate abnormal points caused by transient interference. The processed data is integrated into a three-dimensional raw data matrix with three dimensions: time, sensor channels, and data values.

[0009] Preferably, the pressure decay rate feature vector is obtained based on the pressure data, including: Plot the pressure decay curve based on the absolute pressure value; Calculate the rate of change of the absolute pressure value over time, and calculate the polynomial fitting coefficient of the pressure decay curve during the pressure holding stage. The rate of change of the absolute pressure value over time and the polynomial fitting coefficients are integrated into a pressure decay rate feature vector. The stability eigenvector of the spatial pressure field is obtained based on the pressure data, including: In the spatial domain, the dynamic changes of the differential gradients between dynamic differential pressure gradients and their correlation coefficient matrices are analyzed to construct a spatial pressure field stability eigenvector for the uniformity of pressure distribution inside the compressor. The acoustic leakage feature vector is obtained from the acoustic data, including: Time-frequency analysis of acoustic data is performed. The energy distribution, power spectral density peak and frequency band entropy of ultrasonic leakage in a specific frequency band are extracted by short-time Fourier transform. The acoustic array signal is used to make a preliminary spatial localization estimate of potential leakage sources and form an acoustic leakage feature vector. Thermodynamic anomaly feature vectors are obtained from thermodynamic data, including: Feature mining is performed on the hot spot image sequence. The area growth rate of the temperature anomaly region, the movement trajectory of the highest temperature point, and the temperature gradient distribution are extracted by image difference to construct a thermodynamic anomaly feature vector. The gas diffusion feature vector is obtained based on the tracer gas concentration data, including: By coupling the rate of change of tracer gas concentration with environmental parameters, a gas diffusion characteristic vector is obtained.

[0010] Preferably, the ensemble learning model is a multi-channel feedforward neural network; For the input layer of a multi-channel feedforward neural network, a dedicated sub-network is designed for various feature vectors to perform preliminary abstraction and dimensionality reduction respectively. The intermediate hidden layers of the multi-channel feedforward neural network splice and interact the output features of each specialized sub-network to achieve deep feature-level fusion.

[0011] Preferably, the resulting leak inference decision includes: leak type and leak location result.

[0012] Preferably, the airtightness determination result is obtained based on the leakage probability and the confidence score of the characteristic contribution, including: When the leakage probability is higher than the first threshold and the feature contribution confidence score is higher than the second threshold, it is directly determined as a leakage and the leakage type and leakage location result are output. When the leakage probability is within the preset critical region, but the confidence score of the feature contribution shows that a certain type of feature is abnormally significant, a refined review and analysis process for that type of feature is triggered, and the corresponding sensor is called to re-execute data acquisition.

[0013] Preferably, the method further includes: The compressor to be monitored is used as a new compressor sample, and all data generated during the air tightness test is used as historical sample test data and stored in the training sample library. When a model retraining instruction is received, the ensemble learning model is retrained based on the updated training sample library to optimize the parameters of the ensemble learning model.

[0014] Preferably, the method further includes: Key feature curves are plotted based on sensor data collected by a multi-sensor array; The key feature curves, airtightness determination results, and the decision basis for obtaining the airtightness determination results are integrated into a graphical user interface to automatically generate a structured test report.

[0015] According to a second aspect of the present invention, a multi-dimensional feature analysis system for compressor airtightness detection is provided, comprising: The data acquisition module is used to collect pressure data, acoustic data, thermodynamic data and tracer gas concentration data using a multi-sensor array during the preset charging detection stage, pressure holding detection stage and pressure stabilization detection stage of the compressor under monitoring, and integrate the collected data into a three-dimensional raw data matrix. The multi-dimensional feature fusion module is used to obtain pressure decay rate feature vector and spatial pressure field stability feature vector from pressure data, acoustic leakage feature vector from acoustic data, thermodynamic anomaly feature vector from thermodynamic data, and gas diffusion feature vector from tracer gas concentration data; all the obtained feature vectors are spliced ​​and normalized according to time windows to form a high-dimensional comprehensive feature vector. The diagnostic model module is used to integrate historical sample detection data into a training sample library. Each historical sample detection data includes an airtightness status label and a high-dimensional comprehensive feature vector corresponding to the compressor sample. The pre-built ensemble learning model is trained using the training sample library. The adaptive decision module is used to input the high-dimensional integrated feature vector into the trained ensemble learning model to obtain the leakage probability and leakage inference decision; calculate the feature contribution confidence score according to the contribution of each feature dimension to the current decision; and obtain the airtightness judgment result based on the leakage probability and the feature contribution confidence score.

[0016] The technical solution provided by this invention may include the following beneficial effects: It is understood that the technical solution presented in this invention can collect data using a multi-sensor array at each detection stage of the compressor under monitoring, obtain multiple types of feature vectors, and splice and normalize them according to time windows to form a high-dimensional comprehensive feature vector; train an ensemble learning model using historical sample detection data; input the high-dimensional comprehensive feature vector into the ensemble learning model to obtain the leakage probability and leakage inference decision, and calculate the feature contribution confidence score; and obtain the airtightness judgment result based on the leakage probability and feature contribution confidence score. This technical solution combines multiple sensors to collect multiple types of data and performs multi-dimensional feature fusion analysis, fundamentally improving the completeness and robustness of perception, resulting in high detection accuracy and high adaptability.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0019] Figure 1 This is a schematic diagram illustrating the steps of a multidimensional feature analysis method for compressor airtightness detection according to an exemplary embodiment; Figure 2 This is a schematic block diagram of a multi-dimensional feature analysis system for compressor airtightness detection according to an exemplary embodiment. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0021] In one embodiment, Figure 1 This is a schematic diagram illustrating the steps of a multidimensional feature analysis method for compressor airtightness detection according to an exemplary embodiment. See also... Figure 1 This paper provides a multi-dimensional feature analysis method for compressor airtightness testing, applied to airtightness testing equipment, including: Step S11: During the pre-set charging detection stage, pressure holding detection stage, and pressure stabilization detection stage of the compressor to be monitored, pressure data, acoustic data, thermodynamic data, and tracer gas concentration data are collected using a multi-sensor array, and the collected data are integrated into a three-dimensional raw data matrix.

[0022] Step S12: Obtain the pressure decay rate characteristic vector and the spatial pressure field stability characteristic vector based on the pressure data; obtain the acoustic leakage characteristic vector based on the acoustic data; obtain the thermodynamic anomaly characteristic vector based on the thermodynamic data; and obtain the gas diffusion characteristic vector based on the tracer gas concentration data.

[0023] Step S13: Concatenate and normalize all the obtained feature vectors according to the time window to form a high-dimensional comprehensive feature vector.

[0024] Step S14: Integrate historical sample detection data into a training sample library. Each historical sample detection data includes the airtightness status label and high-dimensional comprehensive feature vector corresponding to the compressor sample. Use the training sample library to train the pre-built ensemble learning model.

[0025] Step S15: Input the high-dimensional integrated feature vector into the trained ensemble learning model to obtain the leakage probability and leakage inference decision; calculate the feature contribution confidence score based on the contribution of each feature dimension to the current decision; and obtain the airtightness judgment result based on the leakage probability and the feature contribution confidence score.

[0026] It is understood that the technical solution presented in this invention can collect data using a multi-sensor array at each detection stage of the compressor under monitoring, obtain multiple types of feature vectors, and splice and normalize them according to time windows to form a high-dimensional comprehensive feature vector; train an ensemble learning model using historical sample detection data; input the high-dimensional comprehensive feature vector into the ensemble learning model to obtain the leakage probability and leakage inference decision, and calculate the feature contribution confidence score; and obtain the airtightness judgment result based on the leakage probability and feature contribution confidence score. This technical solution combines multiple sensors to collect multiple types of data and performs multi-dimensional feature fusion analysis, fundamentally improving the completeness and robustness of perception, resulting in high detection accuracy and high adaptability.

[0027] In a preferred embodiment, step S11, which involves acquiring pressure data, acoustic data, thermodynamic data, and tracer gas concentration data using a multi-sensor array, includes: A pressure sensor array is used to acquire the absolute pressure value inside the compressor cavity and the dynamic pressure differential gradient between different cavity regions. In practice, the pressure sensor array includes one main channel absolute pressure sensor and three distributed differential pressure sensors to simultaneously acquire the absolute pressure value P inside the compressor cavity. absAnd the dynamic pressure gradients ΔP1, ΔP2, and ΔP3 between different cavity regions.

[0028] A broadband acoustic sensor is used to collect acoustic emission and ultrasonic signals from the compressor casing surface and the surrounding air medium. In practice, the broadband acoustic sensor collects acoustic signals at a sampling frequency of no less than 100kHz.

[0029] A sequence of thermal images of the key sealing surfaces and weld seams of the compressor was acquired using an infrared thermal imager at a fixed frame rate.

[0030] Using a gas concentration sensor to monitor in real time the background concentration and potential leakage concentration C of a specific tracer gas in the test environment. leak .

[0031] In a preferred embodiment, step S11, integrating the collected data into a three-dimensional raw data matrix, includes: Based on a hardware-triggered unified clock, time-domain labels are synchronously added to all sensor data streams, and synchronization alignment is performed according to these time-domain labels. This step aligns the timestamp data for sensors with different sampling rates and interpolates or resamples low-frequency data between timestamps, thereby ensuring that all observations are under the same time reference, laying a solid foundation for subsequent spatiotemporal correlation analysis.

[0032] Kalman filtering is used to smooth fluctuations in pressure data, wavelet thresholding is applied to acoustic data to preserve leakage characteristic frequency bands, and mean filtering is used to eliminate transient noise in thermodynamic data. This effectively separates the true physical signal from various noise interferences, significantly improving the signal-to-noise ratio and quality of the original data.

[0033] By statistically analyzing outliers, we can detect, identify, and eliminate anomalies caused by transient interference.

[0034] The processed data is integrated into a three-dimensional raw data matrix comprising time, sensor channels, and data values. This step establishes a unified, structured foundation for multimodal data to support subsequent complex feature engineering and model analysis.

[0035] In step S12 of a preferred embodiment, obtaining the pressure decay rate feature vector based on pressure data includes: plotting a pressure decay curve based on the absolute pressure value; and calculating the absolute pressure value P. abs The rate of change of pressure over time, dP / dt, is used to calculate the polynomial fitting coefficients of the pressure decay curve during the pressure holding phase. The rate of change of the absolute pressure value over time and the polynomial fitting coefficients are then integrated into a pressure decay rate feature vector. This pressure decay rate feature vector describes the overall decreasing trend and rate of pressure decline within the entire compressor cavity over time.

[0036] In practice, after the compressor completes its charging and enters the pressure holding stage, the main channel absolute pressure sensor installed on the compressor cavity continuously collects and records the absolute pressure value within the cavity. Each pressure reading P collected throughout the entire pressure holding stage is recorded. abs It is correlated with its corresponding precise data acquisition time point t, with time t as the x-axis and real-time pressure value P as the y-axis. abs Using the vertical axis as the ordinate, the continuous trajectory drawn in the coordinate system constitutes a curve describing the natural decrease of pressure over time; this is the pressure decay curve.

[0037] In step S12 of a preferred embodiment, obtaining the spatial pressure field stability feature vector based on pressure data includes: analyzing the difference gradient ΔP between dynamic pressure difference gradients (ΔP1, ΔP2, ΔP3) in the spatial domain. i Based on the dynamic changes and their correlation coefficient matrices, a spatial pressure field stability eigenvector V of the compressor's internal pressure distribution uniformity is constructed. pressure .

[0038] The spatial pressure field stability eigenvector describes the relative pressure differences, balance relationships, and dynamic fluctuations between different regions inside the compressor. By analyzing the correlation coefficient matrix and dynamic changes between pressure differences, it is possible to determine whether the pressure distribution is uniform and whether there is a local imbalance.

[0039] In step S12 of a preferred embodiment, obtaining the acoustic leakage feature vector based on acoustic data includes: performing time-frequency analysis on the acoustic data, extracting the energy distribution, power spectral density peak, and frequency band entropy of ultrasonic leakage in a specific frequency band through short-time Fourier transform, and using acoustic array signals to perform preliminary spatial localization estimation of potential leakage sources to form an acoustic leakage feature vector V. acoustic The acoustic leakage eigenvector includes the obtained energy distribution, peak power spectral density, band entropy, and preliminary spatial location estimation.

[0040] In step S12 of a preferred embodiment, obtaining a thermodynamic anomaly feature vector based on thermodynamic data includes: performing feature mining on the hot spot image sequence, extracting the area growth rate of the temperature anomaly region, the movement trajectory of the highest temperature point, and the temperature gradient distribution through image difference, and constructing a thermodynamic anomaly feature vector V. thermal .

[0041] In step S12 of a preferred embodiment, obtaining the gas diffusion feature vector based on the tracer gas concentration data includes: converting the tracer gas concentration change rate dC... leak The gas diffusion characteristic vector V is obtained by coupling / dt with environmental parameters. gas .

[0042] After obtaining multiple feature vectors, step S13 is executed to concatenate and normalize all the obtained feature vectors according to the time window, forming a high-dimensional comprehensive feature vector F. total .

[0043] In step S14, a training sample library is constructed, which includes historical detection data and known leakage status labels. Each sample corresponds to a high-dimensional comprehensive feature vector F. total It obtains the actual airtightness state labels; then, it trains the ensemble learning model. The ensemble learning model is able to obtain F and its actual airtightness state labels. total Complete feature importance screening and classification, and output the comprehensive leakage probability P. leak And preliminary leak inference and decision-making.

[0044] Historical testing data comes from compressor samples confirmed through manual re-inspection or destructive testing during long-term production line testing. Each sample is associated with its complete multidimensional raw data and extracted high-dimensional comprehensive feature vector F. total The supervised learning sample library, along with precise labels confirmed by experts, forms the knowledge source for the model's intelligent diagnostic capabilities.

[0045] Preferably, the integrated learning model is a multi-channel feedforward neural network; the input layer of the multi-channel feedforward neural network is designed with dedicated sub-networks for various feature vectors to complete preliminary abstraction and dimensionality reduction respectively. This structure allows the model to learn the deep features of different physical modes in depth, avoiding early confusion and interference between features.

[0046] The intermediate hidden layers of the multi-channel feedforward neural network splice and interact the output features of each specialized sub-network to achieve deep feature-level fusion.

[0047] It also introduces gradient boosting decision trees to directly capture F total The nonlinear relationships between features are ranked by importance.

[0048] This multi-channel feedforward neural network model combines the powerful feature representation learning ability of neural networks with the excellent feature selection and interpretation ability of decision trees, thereby comprehensively improving classification accuracy and model robustness.

[0049] In practical applications, the data structures of pressure, acoustic, and thermal features are complex (time series, spectrograms, image sequences), requiring convolutional neural networks for deep, automatic feature abstraction and dimensionality reduction to extract spatial, temporal, or spectral patterns. Gas concentration features, on the other hand, are typically composed of scalars or low-dimensional vectors. For this type of highly structured, low-dimensional data, fully connected networks are used for processing, explicitly listed as the fourth parallel processing branch.

[0050] Feature importance selection is performed by the Gradient Boosting Decision Tree (GBDT) model. During training, the GBDT model automatically evaluates the F-value by calculating metrics such as information gain. total The system contains tens of thousands of feature dimensions. Classification is performed by a multi-channel feedforward neural network. Each specialized sub-network performs deep abstraction and transformation of the original features of different modalities such as pressure and acoustics. After fusion in the intermediate layer, a preliminary classification tendency is output. Finally, the outputs of the neural network and the GBDT base model are used as inputs to generate a comprehensive leakage probability P. leak And preliminary leak inference decision.

[0051] Preferably, the leakage probability can be generated using a stacking method. First, two base learners, a multi-channel neural network and a gradient boosting decision tree, are independently trained using a training sample library. The original training set is divided into K parts, and the base models are trained in turn using K-1 parts. The trained models are then used to predict the remaining part. In this way, each training sample will obtain a prediction value Pnn and Pgbdt made by the two base models that were not involved in their training.

[0052] The predictions of these base models (Pnn and Pgbdt) are used as new features, and combined with the true qualified / leaked labels of the samples to form a new training set, which is used to train a simple learner.

[0053] When used in online applications, real-time data F total First, two base models predict Pnn and Pgbdt respectively. Then, these two values ​​are input into a pre-trained learner, which outputs the final, comprehensive leakage probability P. leak .

[0054] After the model training is complete, the online detection stage can begin, as in step S15, where the comprehensive feature vector F obtained through real-time acquisition and processing is used. total The ensemble learning model, after real-time input training, yields the initial leakage probability P. leak And leakage inference decision. At the same time, dynamic decision is introduced to analyze the contribution w of each feature dimension in the model to the current decision, and calculate the feature contribution confidence score C accordingly. conf Then set by P leak and C conf A jointly determined two-dimensional judgment region is used to output the airtightness judgment result. The resulting leak inference decision includes: leak type and leak location result.

[0055] The leak location results are obtained by first estimating the initial location using multimodal methods, and generating preliminary spatial clues using data from different sensors: for example, using acoustic array signal processing and thermal imaging for preliminary location, analyzing infrared thermal image sequences to extract temperature anomaly areas (hot spots), and using image processing techniques to determine their two-dimensional pixel coordinates on the thermal image.

[0056] In the process of analyzing the contribution of each feature dimension in the model to the current decision (w), each feature dimension refers to the contribution of each feature dimension to the current real-time decision (F). total The decision-making process relies on hundreds or thousands of specific feature dimensions, such as: for example, in V acoustic In the text, the "energy value of a specific ultrasonic frequency band" is one characteristic dimension, the "peak value of the power spectrum at 38kHz" is another, and the "value of the frequency band entropy" and the "estimated positioning angle" are also independent characteristic dimensions. Similarly, in V... pressure In this context, "the instantaneous value of dP / dt" and "the fitting coefficient a of the second-order polynomial" are also feature dimensions.

[0057] The formula for calculating the confidence score of the feature contribution is: C conf =γ·α+(1 γ)·β The confidence score for feature contribution is calculated based on the contribution *w* of each feature dimension to the current decision. The contribution vector *w* is normalized, and then its information entropy or Gini impurity is calculated. Finally, the entropy value is mapped to the [0,1] interval to obtain α. α is the decision concentration, a scalar between 0 and 1, used to measure the concentration and clarity of the basis upon which the model's decision depends. The closer its value is to 1, the more the decision depends on a few features, and the clearer the basis; the closer it is to 0, the more the contribution is dispersed across many feature dimensions, and the ambiguity of the decision basis.

[0058] β is the significance of key feature anomalies, a scalar between 0 and 1, used to measure the key feature dimension that contributes the most to the decision. Its value itself deviates significantly from the historical normal range. The closer the value is to 1, the more significant the anomalies of all key features are; the lower the value, the more the key features may be within the normal fluctuation range.

[0059] γ is the balancing weight, a preset constant between 0 and 1, used to balance the relative importance of decision concentration (α) and key feature anomaly significance (β) in the final confidence score.

[0060] In a preferred embodiment, the airtightness determination result is obtained based on the leakage probability and the feature contribution confidence score, including: when the leakage probability is higher than a first threshold and the feature contribution confidence score is higher than a second threshold, a leak is directly determined and the leak type and leak location result are output. This achieves a rapid and accurate response to high-confidence leakage events.

[0061] The leak location results output in this step have been validated by the leak probability and feature contribution confidence score. A high leak probability and a high feature contribution confidence score indicate that the model not only judges the possibility of a leak to be high, but also that this judgment is mainly based on a few unusually significant features. At this point, the system will check whether the acoustic localization and thermal imaging localization results are spatially consistent; if they are consistent, the system will determine the coordinates of this intersection area or hot spot centroid as the final leak point and assign it the highest confidence level.

[0062] When the leakage probability is within the preset critical region, but the confidence score of the feature contribution shows that a certain type of feature is abnormally significant, a refined review and analysis process for that type of feature is triggered, and the corresponding sensor is called to re-execute data acquisition.

[0063] The determination of a significant anomaly in a particular feature category is made through procedural analysis of feature contribution, primarily based on two quantifiable criteria. The first is a contribution percentage threshold: calculating the proportion of the sum of the contributions of a particular feature category to the total contribution; if this proportion exceeds a preset significance threshold, it is determined that this feature category dominates the decision-making process. The second is anomaly in the values ​​of key sub-features: examining the specific values ​​of the feature dimensions with the highest contribution within this feature category (from F...). total Whether it deviates significantly from its historical normal baseline, only when the contribution is high and its own numerical value is abnormal at the same time, will a refined review of this type of feature be triggered.

[0064] A certain type of feature is significantly anomalous; for example, for pressure field features: P leak It is at a critical value (e.g., 0.6), but C conf Analysis shows that the contribution is highly concentrated in the dramatic fluctuations in variance of ΔP2 (pressure difference in the end cap region). The pressure decay curve (dP / dt) itself, however, is relatively flat. Regarding acoustic characteristics: P leak The value is moderately suspicious (e.g., 0.55), but the contributions of "80kHz ultrasonic band energy" and "sound source positioning direction angle (pointing to the weld)" in the contribution spectrum are abnormally prominent, and the energy value of this band far exceeds the historical baseline. Regarding thermodynamic characteristics: P leak Slightly above the threshold (e.g., 0.65), the contribution comes almost entirely from the "hotspot area growth rate" and "local temperature gradient," with pressure and acoustic features contributing very little. For gas concentration characteristics: P leak Critical threshold (e.g., 0.58), but contribution indicates "the rate of change of concentration dC of a specific tracer gas (e.g., helium)". leakThe contribution of " / dt" is extremely high, and the concentration value continues to rise in specific ventilation dead zones.

[0065] Similarly, when the confidence level is low or the features conflict (inconsistent acoustic and thermal localization results), it indicates that the leak localization results are inaccurate and a review analysis process needs to be triggered.

[0066] The verification and analysis process involves using high-resolution thermal imaging for local scanning and extending the acoustic signal acquisition time at specific locations for secondary confirmation. By investing more sensing resources for targeted and in-depth exploration, the process ensures both overall detection efficiency and careful handling and high-precision confirmation of boundary conditions, thereby reducing false alarms and false negatives while maintaining detection speed.

[0067] The final leak location result will be mapped to the corresponding coordinates of the 3D compressor digital model by the system and displayed intuitively in the graphical user interface with highlights and markings.

[0068] In a preferred embodiment, the method further includes: plotting key feature curves based on sensor data collected by a multi-sensor array; and automatically generating a structured detection report by integrating the key feature curves, the airtightness determination results, and the decision basis for obtaining the airtightness determination results into a graphical user interface.

[0069] Key characteristic curves refer to time series or spectra that are directly calculated or visualized from raw data and most intuitively reflect changes in the airtightness of the compressor. They are the visual representation of the most important chain of evidence between sensor data and the final conclusion, and typically include: Pressure decay curve: The "absolute pressure-time" curve during the pressure holding stage is the most direct graphical evidence for judging the overall leakage trend.

[0070] Acoustic time-spectrum graph (sound spectrum): A two-dimensional image with time on the horizontal axis and frequency on the vertical axis, using color intensity to represent the intensity of sound energy. It can visually show the appearance and evolution of ultrasonic signals (such as specific high-frequency bands) generated by leakage over time.

[0071] Thermal imaging sequence / hotspot evolution animation: A series of infrared thermal images arranged in chronological order, or an animation generated from them that highlights the dynamic changes in areas of abnormal temperature, used to visually display the process of local temperature field changes caused by leaks.

[0072] Gas concentration change curve: A curve showing the change of tracer gas concentration over time, directly indicating the accumulation of leaked gas.

[0073] The decision-making basis refers to the data-driven presentation of the quantitative analysis and reasoning process that drives the intelligent system to make the final "qualified / leaking" judgment, including: Feature contribution distribution map: clearly shows the contribution weight (w) of various features (or feature dimensions) such as pressure, acoustics, heat, and gas mentioned in the steps to this decision. Leakage probability score (P) leak ) and confidence score (C conf The numerical values ​​and relationships of P calculated by the model are clearly displayed. leak Value, and the confidence level C of the decision. conf And show them in "P" leak -C conf The position within the "two-dimensional determination region"; Based on the above values ​​and charts, an automatically generated text description is provided, for example, "Because of P..." leak =0.95 and C conf =0.92, falling into the 'high confidence leakage' region, so leakage was directly determined; the main basis was the abnormal surge of 80kHz ultrasonic energy at the weld (contribution 65%) and the corresponding local hot spot formation (contribution 22%).

[0074] The decision formula for determining the airtightness can be expressed as:

[0075] Here, Result represents the final airtightness assessment result, such as "qualified / leaking" and its location. D is the dynamic decision function, whose input is the leakage probability P. leak With confidence score C conf It outputs the judgment result based on the two-dimensional judgment region. ensemble This is a heterogeneous ensemble learning model, with the input being the comprehensive feature vector F. total The output is P. leak and C conf Norm refers to data normalization standardization, a necessary preprocessing step that transforms heterogeneous feature data from multiple sources and with multiple dimensions into a standardized numerical vector that can be efficiently and fairly processed by machine learning models. Concat is a concatenation function for multi-dimensional feature structure fusion. It connects two or more feature vectors end-to-end to form a longer single vector. It does not change the order or value of elements within each vector; it only performs a physical connection. Only through this step can the state of the concatenation function be expressed as a mathematical entity that can be understood and processed by complex artificial intelligence models. p F a F t F g : These are the feature extraction functions for pressure, acoustics, thermodynamics, and gas diffusion, respectively. M is the three-dimensional raw data matrix, obtained from the raw sensor data after synchronization and preprocessing. θ is the ensemble learning model M. ensembleThe set of trainable parameters, whose updates are performed by an incremental learning function based on the encrypted new data F. total drive.

[0076] In practical applications, when automatically generating structured inspection reports within a graphical user interface (GUI), the GUI employs a layered design. The main view displays real-time airtightness assessment results, a leak probability dashboard, and highlighted suspected leak points on the 3D compressor model. Secondary windows retrieve pressure decay curves during the pressure holding phase and characteristic curves from the acoustic signal time-spectrum graph. Users can zoom, compare, and overlay historical data for analysis, providing multi-level, interactive, and visual insights from macroscopic conclusions to microscopic data, greatly facilitating status monitoring and result evaluation for operators. The structured inspection report is output in PDF format, with chapters including a summary of inspection conclusions, snapshots of all original sensor data, key parameters for feature extraction, numerical values ​​of feature vectors input to the model, bar charts of feature contribution during the decision-making process, and explanations of the final judgment criteria. Report data is directly retrieved from the database and cache, achieving comprehensive and traceable digital archiving of the inspection process and results, providing a complete chain of evidence for quality analysis, process improvement, and potential audits.

[0077] In a preferred embodiment, the method further includes: using the compressor to be monitored as a new compressor sample, storing all data generated during the airtightness testing process as historical sample testing data in a training sample library; and when a model retraining instruction is received, retraining the ensemble learning model according to the updated training sample library to optimize the parameters of the ensemble learning model.

[0078] The technical solution shown in this embodiment, in actual implementation, transmits the currently detected and manually verified valid data packets to the secure storage area through an encrypted channel. The data packets are encrypted using an asymmetric encryption algorithm and bound to the compressor's unique serial number, timestamp, and operator ID. After their feature vectors and labels are structured, they are appended to the version-controlled training sample library, and the sample distribution statistics are updated. This process ensures the secure, orderly, and traceable accumulation of new knowledge and is a key link in the data closed loop for the system to achieve continuous learning.

[0079] When the accumulation of new samples reaches a preset threshold or a fixed period, the incremental learning process is automatically triggered. When fine-tuning the neural network parameters, constraints are imposed on the parameters corresponding to important historical samples. At the same time, the gradient boosting decision tree performs additional iterations at a limited depth using new samples as incremental data. The incremental learning strategy enables the model to absorb new knowledge and adapt to new changes (such as component aging and process changes), while effectively preventing catastrophic forgetting of existing knowledge. This achieves autonomous and stable evolution of the detection system's performance and autonomous evolution against feature drift.

[0080] This technical solution uses multi-dimensional feature fusion analysis to spatiotemporally align the dynamic distribution of the pressure field, broadband acoustics, infrared thermal imaging, and gas concentration information to form a comprehensive feature vector describing the leakage state. This fundamentally improves the completeness and robustness of the perception. Furthermore, it introduces an intelligent diagnostic model and dynamic decision-making mechanism based on ensemble learning, which not only outputs the leakage probability but also analyzes the contribution of each feature. Based on the two-dimensional judgment region composed of probability and confidence, it makes flexible judgments and triggers targeted and refined reviews. This achieves a shift from single threshold judgment to multi-evidence chain judgment, reducing the risk of missed detections and false alarms under complex operating conditions.

[0081] In another embodiment, see Figure 2 A multi-dimensional feature analysis system for compressor airtightness detection is provided, comprising: The data acquisition module 101 is used to collect pressure data, acoustic data, thermodynamic data and tracer gas concentration data using a multi-sensor array during the preset charging detection stage, pressure holding detection stage and pressure stabilization detection stage of the compressor to be monitored, and integrate the collected data into a three-dimensional raw data matrix. The multi-dimensional feature fusion module 102 is used to obtain the pressure decay rate feature vector and the spatial pressure field stability feature vector based on the pressure data, the acoustic leakage feature vector based on the acoustic data, the thermodynamic anomaly feature vector based on the thermodynamic data, and the gas diffusion feature vector based on the tracer gas concentration data; and to splice and normalize all the obtained feature vectors according to the time window to form a high-dimensional comprehensive feature vector. The diagnostic model module 103 is used to integrate historical sample detection data into a training sample library. Each historical sample detection data includes an airtightness status label and a high-dimensional comprehensive feature vector corresponding to the compressor sample. The pre-built ensemble learning model is trained using the training sample library. The adaptive decision module 104 is used to input the high-dimensional integrated feature vector into the trained ensemble learning model to obtain the leakage probability and leakage inference decision; calculate the feature contribution confidence score according to the contribution of each feature dimension to the current decision; and obtain the airtightness judgment result based on the leakage probability and the feature contribution confidence score.

[0082] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0083] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0084] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0085] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0086] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0087] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0088] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0089] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0090] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A multidimensional feature analysis method for compressor airtightness detection, characterized in that, Applications in airtightness testing equipment, including: Pressure data, acoustic data, thermodynamic data, and tracer gas concentration data are collected using a multi-sensor array during the preset charging detection stage, pressure holding detection stage, and pressure stabilization detection stage of the compressor under monitoring. The collected data are then integrated into a three-dimensional raw data matrix. Based on pressure data, we obtain pressure decay rate characteristic vector and spatial pressure field stability characteristic vector; based on acoustic data, we obtain acoustic leakage characteristic vector; based on thermodynamic data, we obtain thermodynamic anomaly characteristic vector; based on tracer gas concentration data, we obtain gas diffusion characteristic vector. All the obtained feature vectors are concatenated and normalized according to the time window to form a high-dimensional comprehensive feature vector. Historical sample detection data are integrated into a training sample library. Each historical sample detection data includes an airtightness status label and a high-dimensional comprehensive feature vector corresponding to the compressor sample. The pre-built ensemble learning model is trained using the training sample library. The high-dimensional integrated feature vector is input into the trained ensemble learning model to obtain the leakage probability and leakage inference decision; the feature contribution confidence score is calculated based on the contribution of each feature dimension to the current decision; and the airtightness determination result is obtained based on the leakage probability and the feature contribution confidence score.

2. The method according to claim 1, characterized in that, Pressure data, acoustic data, thermodynamic data, and tracer gas concentration data are acquired using a multi-sensor array, including: The absolute pressure value inside the compressor cavity and the dynamic pressure gradient between different cavity regions are collected using a pressure sensor array. A broadband acoustic sensor is used to collect acoustic emission signals and ultrasonic signals from the compressor housing surface and the surrounding air medium. A sequence of thermal imagery of the key sealing surfaces and weld areas of the compressor was acquired using an infrared thermal imager at a fixed frame rate. The background concentration and potential leakage concentration of a specific tracer gas in the test environment are monitored in real time using a gas concentration sensor.

3. The method according to claim 1, characterized in that, The collected data is integrated into a three-dimensional raw data matrix, including: Based on a hardware-triggered unified clock, time-domain labels are synchronously applied to all sensor data streams, and synchronization alignment is performed according to the time-domain labels. Kalman filtering is used to smooth fluctuations in pressure data, wavelet thresholding is used to denoise acoustic data to preserve leakage characteristic frequency bands, and mean filtering is used to eliminate transient noise in thermodynamic data. By statistically analyzing outliers, we can detect, identify, and eliminate abnormal points caused by transient interference. The processed data is integrated into a three-dimensional raw data matrix with three dimensions: time, sensor channels, and data values.

4. The method according to claim 2, characterized in that, The pressure decay rate feature vector is obtained from the pressure data, including: Plot the pressure decay curve based on the absolute pressure value; Calculate the rate of change of the absolute pressure value over time, and calculate the polynomial fitting coefficient of the pressure decay curve during the pressure holding stage. The rate of change of the absolute pressure value over time and the polynomial fitting coefficients are integrated into a pressure decay rate feature vector. The stability eigenvector of the spatial pressure field is obtained based on the pressure data, including: In the spatial domain, the dynamic changes of the differential gradients between dynamic differential pressure gradients and their correlation coefficient matrices are analyzed to construct a spatial pressure field stability feature vector for the uniformity of pressure distribution inside the compressor. The acoustic leakage feature vector is obtained from the acoustic data, including: Time-frequency analysis of acoustic data is performed. The energy distribution, power spectral density peak and frequency band entropy of ultrasonic leakage in a specific frequency band are extracted by short-time Fourier transform. The acoustic array signal is used to make a preliminary spatial localization estimate of potential leakage sources and form an acoustic leakage feature vector. Thermodynamic anomaly feature vectors are obtained from thermodynamic data, including: Feature mining is performed on the hot spot image sequence. The area growth rate of the temperature anomaly region, the movement trajectory of the highest temperature point, and the temperature gradient distribution are extracted by image difference to construct a thermodynamic anomaly feature vector. The gas diffusion feature vector is obtained based on the tracer gas concentration data, including: By coupling the rate of change of tracer gas concentration with environmental parameters, a gas diffusion characteristic vector is obtained.

5. The method according to claim 1, characterized in that, The ensemble learning model is a multi-channel feedforward neural network; For the input layer of a multi-channel feedforward neural network, a dedicated sub-network is designed for various feature vectors to perform preliminary abstraction and dimensionality reduction respectively. The intermediate hidden layers of the multi-channel feedforward neural network splice and interact the output features of each specialized sub-network to achieve deep feature-level fusion.

6. The method according to claim 1, characterized in that, The resulting leak inference decision includes: leak type and leak location results.

7. The method according to claim 6, characterized in that, Based on the leakage probability and the confidence score of the characteristic contribution, the airtightness determination result is obtained, including: When the leakage probability is higher than the first threshold and the feature contribution confidence score is higher than the second threshold, it is directly determined as a leakage and the leakage type and leakage location result are output. When the leakage probability is within the preset critical region, but the confidence score of the feature contribution shows that a certain type of feature is abnormally significant, a refined review and analysis process for that type of feature is triggered, and the corresponding sensor is called to re-execute data acquisition.

8. The method according to claim 1, characterized in that, Also includes: The compressor to be monitored is used as a new compressor sample, and all data generated during the air tightness test is used as historical sample test data and stored in the training sample library. When a model retraining instruction is received, the ensemble learning model is retrained based on the updated training sample library to optimize the parameters of the ensemble learning model.

9. The method according to claim 1, characterized in that, Also includes: Key feature curves are plotted based on sensor data collected by a multi-sensor array; The key feature curves, airtightness determination results, and the decision basis for obtaining the airtightness determination results are integrated into a graphical user interface to automatically generate a structured test report.

10. A multi-dimensional feature analysis system for compressor airtightness detection, characterized in that, include: The data acquisition module is used to collect pressure data, acoustic data, thermodynamic data and tracer gas concentration data using a multi-sensor array during the preset charging detection stage, pressure holding detection stage and pressure stabilization detection stage of the compressor under monitoring, and integrate the collected data into a three-dimensional raw data matrix. The multi-dimensional feature fusion module is used to obtain pressure decay rate feature vector and spatial pressure field stability feature vector from pressure data, acoustic leakage feature vector from acoustic data, thermodynamic anomaly feature vector from thermodynamic data, and gas diffusion feature vector from tracer gas concentration data; all the obtained feature vectors are spliced ​​and normalized according to time windows to form a high-dimensional comprehensive feature vector. The diagnostic model module is used to integrate historical sample detection data into a training sample library. Each historical sample detection data includes an airtightness status label and a high-dimensional comprehensive feature vector corresponding to the compressor sample. The pre-built ensemble learning model is trained using the training sample library. The adaptive decision module is used to input the high-dimensional integrated feature vector into the trained ensemble learning model to obtain the leakage probability and leakage inference decision; calculate the feature contribution confidence score according to the contribution of each feature dimension to the current decision; and obtain the airtightness judgment result based on the leakage probability and the feature contribution confidence score.

Images