Artificial intelligence-based scroll compressor thrust bearing fault diagnosis method and system

Abstract

The application relates to the technical field of fault diagnosis, and discloses a scroll compressor thrust bearing fault diagnosis method and system based on artificial intelligence. The method comprises the following steps: collecting original vibration signals of a scroll compressor thrust bearing through a piezoelectric acceleration sensor array, and performing angle domain synchronization and modal decomposition on the original vibration signals to obtain demodulated vibration signals; extracting fault characteristic frequency spectrum maps of inner ring damage, outer ring cracks, rolling body peeling and cage deformation in the scroll compressor thrust bearing according to the demodulated vibration signals; extracting fault features from time domain, frequency domain and time-frequency domain based on the fault characteristic frequency spectrum maps, and constructing a fault feature vector; inputting the fault feature vector into an integrated learning model for fault recognition analysis to obtain a fault type and severity. The application improves the accuracy of scroll compressor thrust bearing fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, and in particular to a scroll compressor thrust bearing fault diagnosis method and system based on artificial intelligence. BACKGROUND

[0002] Scroll compressors are widely used in refrigeration and heat pump systems such as air conditioners, refrigerators, and heat pumps due to their high efficiency, low vibration, and low noise. As the core component of scroll compressors, the thrust bearing bears reciprocating axial force and high-frequency torque fluctuations, and its operating state directly affects the performance and service life of the compressor. However, due to the unique rotating mechanism of scroll compressors, the thrust bearing exhibits non-stationary and nonlinear vibration characteristics during operation, resulting in weak modulation characteristics and short-term impact responses in the fault signal, which makes it difficult for traditional vibration signal processing and fault diagnosis methods to effectively identify early weak faults.

[0003] Existing thrust bearing fault diagnosis techniques have obvious shortcomings when faced with the complex working conditions of scroll compressors. Traditional methods are difficult to capture weak modulation signals; existing diagnosis models have defects in handling unbalanced samples and difficult samples; the fault characteristics of bearings change significantly under different working conditions such as discharge pressure, evaporation temperature, and condensation temperature, and traditional diagnosis methods lack working condition adaptive ability; for fault feature extraction and identification under non-steady-state operating conditions such as start-stop transition conditions, existing technologies cannot effectively solve the start-stop transition problem. SUMMARY

[0004] The present application provides a scroll compressor thrust bearing fault diagnosis method and system based on artificial intelligence, which improves the accuracy of scroll compressor thrust bearing fault diagnosis.

[0005] In a first aspect, the present application provides a scroll compressor thrust bearing fault diagnosis method based on artificial intelligence, which comprises:

[0006] The original vibration signal of the scroll compressor thrust bearing is collected by a piezoelectric acceleration sensor array, and the original vibration signal is subjected to angle domain synchronization and modal decomposition to obtain a demodulated vibration signal;

[0007] According to the demodulated vibration signal, the fault feature spectrum of the inner ring damage, outer ring crack, rolling body peeling, and cage deformation in the scroll compressor thrust bearing is extracted;

[0008] Based on the fault feature spectrum, fault features are extracted from the time domain, frequency domain, and time-frequency domain, and a fault feature vector is constructed;

[0009] The fault feature vector is input into an integrated learning model for fault identification analysis, to obtain a fault type and severity.

[0010] In a second aspect, the present application provides an artificial intelligence-based scroll compressor thrust bearing fault diagnosis system, comprising:

[0011] A collection module is configured to collect original vibration signals of a scroll compressor thrust bearing through a piezoelectric acceleration sensor array, and perform angle domain synchronization and modal decomposition on the original vibration signals to obtain demodulated vibration signals.

[0012] An extraction module is configured to extract fault feature spectrum patterns of inner ring damage, outer ring crack, rolling body peeling and cage deformation in the scroll compressor thrust bearing according to the demodulated vibration signals.

[0013] A construction module is configured to extract fault features from time domain, frequency domain and time-frequency domain based on the fault feature spectrum patterns, and construct a fault feature vector.

[0014] An analysis module is configured to input the fault feature vector into an integrated learning model for fault identification analysis, to obtain a fault type and severity.

[0015] In the technical solution provided by the application, through the piezoelectric acceleration sensor and the use of adaptive band-pass filtering and angle domain synchronous resampling processing, interference signals from other parts of the compressor can be effectively filtered, the influence of speed fluctuation on fault feature extraction is eliminated, and high-quality preprocessed vibration signals are obtained. The signal is preprocessed by using adaptive morphological filtering and Hilbert-Huang transform enhancement processing function, combined with variational mode decomposition technology and IMF automatic selection algorithm, the fault feature impact response is highlighted, and the defects that the traditional method is difficult to capture the weak modulation signal and short impact response of the thrust bearing of the scroll compressor are effectively overcome. By establishing an accurate mathematical model reflecting the dynamic load of the bearing under scroll motion, combining with finite element analysis to calculate the stress distribution, the theoretical expression of four typical fault characteristic frequencies is derived, which provides a theoretical basis and data support for fault diagnosis, and significantly improves the pertinence and accuracy of diagnosis. Feature vectors are extracted from time domain, frequency domain and time-frequency domain, and feature optimization is performed through the maximum correlation and minimum redundancy principle, while adaptive synthetic oversampling and boundary sample synthesis technology are used to balance sample distribution, solve the sample imbalance problem, and enhance the recognition ability of a small number of fault sample classes. The deep residual network is used as a base learner, combined with the improved Focal loss function and Adaboost integrated framework, the Dropout-Bagging technology is used to increase the diversity of the base learner, the cost-sensitive matrix is introduced to quantify the cost of different types of judgment errors, which has high sensitivity to early weak faults and can accurately identify early faults such as micro-cracks. Through the working condition encoder, the working condition compensation network and the domain adversarial network architecture, the dynamic fusion of working condition features and fault features is realized, the working condition migration problem is solved, and the dynamic features under non-steady state working conditions are captured, so that the model maintains stable diagnosis performance under different operating conditions, especially in the start-up and shutdown process. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor based on these drawings.

[0017] Figure 1 An embodiment schematic diagram of the scroll compressor thrust bearing fault diagnosis method based on artificial intelligence in the embodiment of the present application;

[0018] Figure 2 An embodiment schematic diagram of the scroll compressor thrust bearing fault diagnosis system based on artificial intelligence in the embodiment of the present application. DETAILED DESCRIPTION

[0019] The embodiments of the present application provide a scroll compressor thrust bearing fault diagnosis method and system based on artificial intelligence. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the term "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0020] For ease of understanding, the specific flow of the embodiments of the present application is described below. Please refer to Figure 1 One embodiment of the scroll compressor thrust bearing fault diagnosis method based on artificial intelligence in the embodiments of the present application includes:

[0021] Step S101, the original vibration signal of the scroll compressor thrust bearing is collected by the piezoelectric acceleration sensor array, and the original vibration signal is angle domain synchronized and modal decomposed to obtain a demodulated vibration signal;

[0022] It can be understood that the execution subject of the present application can be a scroll compressor thrust bearing fault diagnosis system based on artificial intelligence, and can also be a terminal or a server, and the specific execution subject is not limited herein. The embodiments of the present application take the server as the execution subject for example.

[0023] Specifically, a plurality of high-sensitivity piezoelectric acceleration sensors are arranged outside the thrust bearing of the scroll compressor, which are arranged equidistantly in the circumferential direction of the bearing to form a sensor array, for collecting high-frequency vibration signals of the compressor during operation, especially the micro-impact response and structural resonance characteristics generated at the thrust bearing position. The collected original vibration signals are converted into high-resolution digital signals by an analog-to-digital conversion module, which converts continuous analog signals into high-resolution digital signals. The digital vibration signals are selected by a band-pass filter, and the filter parameters are dynamically adjusted according to the current speed of the compressor and the geometric characteristics of the bearing, so as to focus on the frequency band of the bearing and effectively suppress the interference components brought by other mechanisms in the system, and obtain the preliminary filtered signal. The preliminary filtered signal is processed by noise reduction, and a multi-scale analysis method is used to decompose and reconstruct the signal at different levels, while preserving the weak abnormal signal as much as possible, suppressing high-energy broadband noise, and ensuring that the initial fault signal such as micro-crack and wear can be extracted. At the same time, by reading the real-time phase signal provided by the compressor rotary encoder, the corresponding relationship between time and rotation angle is established, which is used to solve the speed fluctuation caused by the compressor during startup, acceleration or variable load operation, so that the vibration signal can be uniformly represented and analyzed in the angle dimension, thereby ensuring the comparability and synchronicity of the signal characteristics between different working periods. After obtaining the angle-synchronous signal, the signal is divided into multiple data segments according to a certain length and overlap ratio, each of which covers one or more complete rotation periods, thereby constructing a representative signal sample set. The signal sample set is input into the demodulation processing module, and a nonlinear transformation technique is used to extract the envelope signal, which reveals the impact characteristics and damage indications hidden in the modulation information through the envelope form. The envelope signal is subjected to modal decomposition, which is decomposed into a group of sub-signals with different frequency characteristics and physical meanings, to obtain the demodulated vibration signal. Each modal component corresponds to a typical dynamic process in the bearing system, such as rolling body impact, cage disturbance or inner and outer ring deformation.

[0024] In this embodiment, the adaptive morphological filtering technique is used to enhance the pulse characteristics of the pre-processed vibration signal sample set. The filtering process identifies and enhances the transient components in the signal by designing a structural element that conforms to the typical impact response of the thrust bearing, thereby effectively retaining the mutation edges and local extrema while suppressing the background trend and low-frequency interference, highlighting the fault impact pulses caused by the rolling elements contacting the inner and outer rings or the cage movement, and forming a pulse-enhanced signal with significant time-domain mutation characteristics. The Hilbert-Huang transform enhancement processing function is used to demodulate the pulse characteristic enhanced signal. By constructing an analytical signal form composed of the real part signal and its Hilbert transform, the envelope amplitude of the signal at different angular positions is extracted, and combined with a set of weighted window functions with local time advantage, the weak periodic modulation information contained in the signal is enhanced, forming a demodulation initial signal reflecting the modulation behavior. The demodulation initial signal is input into the variational modal decomposition module, which automatically decomposes a number of intrinsic modal functions based on the structural characteristics of the signal itself without prior frequency information. These modal functions correspond to different frequency bands and physical processes hidden in the signal. Each group of modes has good local concentration in the frequency spectrum, and the decomposition method has non-overlapping properties, which helps independent analysis of subsequent fault feature components. For the modal functions obtained by decomposition, to evaluate their contribution to fault identification, the kurtosis value of each mode, i.e. the sharpness of the signal waveform, is calculated. The kurtosis can be used to judge the concentration intensity of the impact component in the signal, and the energy ratio of each mode near the preset fault feature frequency is also calculated, which reflects the correlation between the mode and the specific fault mechanism. After obtaining the evaluation indicators of all modes, the modal functions are sorted and selected by constructing a comprehensive scoring mechanism. The scoring mechanism weights and fuses the kurtosis and energy ratio, so that the modes that contain both high impact characteristics and fault-related frequency sets have an advantage in scoring. Finally, the mode function with the highest score is selected as the optimal mode set. Based on the optimal mode set, the signal is reconstructed to form a demodulation vibration signal containing obvious fault modulation characteristics. This signal exhibits clear characteristic frequency distribution and sideband modulation structure in the frequency domain, and shows stable and periodic impact behavior in the time domain.

[0025] In step S102, the demodulation vibration signal is used to extract the fault characteristic spectrum of the inner ring damage, outer ring crack, rolling element peeling and cage deformation in the thrust bearing of the scroll compressor.

[0026] Specifically, the geometric motion characteristics and load evolution law of the scroll compressor during operation are analyzed, wherein the rotation behavior of the scroll disc is taken as the core, the two-dimensional trajectory change of the scroll disc at different time nodes is calculated through kinematic modeling, the eccentric motion mode of the scroll disc relative to the fixed disc in the rotating state is determined, and a dynamic function model reflecting the change of the trajectory of the scroll disc with time is established. On this basis, combined with the structural parameters and real-time operating state of the scroll cavity, the gas pressure distribution function inside the scroll compressor cavity is solved, and the gas dynamic factor is included in the overall analysis of the resultant force of the thrust bearing, and a set of dynamic balance equations including pressure action, moment of inertia, centrifugal response and other factors is constructed. Through solving the equation set, the instantaneous axial load fluctuation of the scroll compressor under steady state or variable load state is accurately described, and the dynamic load distribution is obtained. The dynamic load distribution is mapped to the structure of the thrust bearing, including the inner ring, the outer ring, the rolling body and the cage, combined with the material properties, structural stiffness and bearing mode, the stress distribution and deformation trend of each part under various operating conditions are calculated by using the finite element analysis method or the semi-analytical elastic mechanics method, and a working condition stress distribution model is established. Relying on the model, from the perspective of the stress mechanism of the bearing structure, the response law of the corresponding inner ring damage, outer ring crack, rolling body spalling and cage deformation in the frequency domain is derived, forming a set of theoretical fault characteristic frequency models. These models take structural parameters, rotational speed, contact angle and load state as input conditions, and obtain modulation frequencies or harmonic frequencies related to fault types. Based on the fault characteristic frequency theory mathematical model, the Hilbert envelope demodulation of the demodulated vibration signal is carried out, the envelope signal containing the modulation component is extracted, and its form in the frequency spectrum is analyzed. By observing the spectrum line enhancement, sideband appearance and amplitude change near each type of fault characteristic frequency, the characteristic frequency band related to the inner ring, the outer ring, the rolling body or the cage is identified. In the frequency spectrum demodulation result, different fault types have different mechanical impact point positions, modulation modes and propagation paths due to their causes, forming distinguishable pattern characteristics in the frequency spectrum diagram, so the frequency characteristics are classified and arranged, and their corresponding fault mechanisms are labeled. According to the fault type, the frequency spectrum data is structured and integrated to construct a set of fault characteristic frequency spectrum atlas covering four typical states of inner ring damage, outer ring crack, rolling body spalling and cage deformation.

[0027] Step S103, extracting fault features from time domain, frequency domain and time-frequency domain based on the fault characteristic frequency spectrum atlas, and constructing a fault feature vector;

[0028] Specifically, time-domain statistical feature calculation is performed on the demodulated vibration signal. A series of statistical property operations are performed on the demodulated signal, including amplitude distribution, extreme value feature, waveform structure and fluctuation law comprehensive analysis, from which representative statistics such as peak factor, kurtosis, pulse index, margin factor and sample entropy are extracted. These parameters can effectively reflect the impact strength, pulse density, energy concentration and waveform complexity in the signal, forming a set of time-domain feature sub-vectors describing the time variation mode. The frequency spectrum characteristics of the demodulated vibration signal are aligned and analyzed with the aforementioned fault feature spectrum, to identify the energy distribution of the corresponding fault characteristic frequencies and their side frequency bands in the signal, and to calculate the proportion of the total energy occupied by the characteristic frequencies of inner ring damage, outer ring crack, rolling element spalling and cage deformation, as well as the energy coverage of the frequency spectrum region in different time periods, to construct a frequency-domain feature sub-vector composed of frequency energy ratio and frequency band energy proportion, describing the response strength and frequency domain pattern of the demodulated signal near the typical fault frequency. To capture the multi-scale and non-stationary characteristics of the fault signal, the demodulated signal is input into a time-frequency analysis module, the signal is decomposed into multiple frequency bands, the energy entropy of each frequency band is calculated, the complexity of the energy distribution in time and frequency is described, and the joint distribution entropy is extracted using the short-time Fourier transform method, the instantaneous energy gravity fluctuation of the signal in the frequency variation process is described, and a time-frequency analysis feature sub-vector with time sequence resolution is constructed. The above three sub-vectors are spliced and fused to construct an initial feature vector covering multiple statistical dimensions. Feature selection and optimization mechanism is used to screen the initial feature vector, a feature selection criterion considering correlation and non-redundancy is used to extract an optimized feature subset with high discriminability, which retains data components highly sensitive to different fault states and eliminates redundant or weakly discriminative invalid information. When dealing with the problem of sample imbalance, a hybrid sample balancing strategy is introduced, which is composed of adaptive synthetic oversampling technique and boundary sample synthesis technique. By identifying the distribution density and decision boundary position of the minority class samples in the optimized feature subset in the feature space, new sample points are generated according to the density-oriented and boundary-oriented principles, the spatial coverage of the minority class samples is expanded, and the recognition ability of the model for the easily confused categories is improved. All the newly generated samples are integrated with the original optimized feature subset, and are converted into a unified numerical representation format through the vector mapping process, forming the fault feature vector.

[0029] Step S104, input the fault feature vector into the integrated learning model for fault recognition analysis, to obtain the fault type and severity.

[0030] Specifically, a base learner model with deep feature extraction and classification capabilities is constructed. This model employs a residual deep neural network architecture. The input layer receives optimized fault feature vectors, which are then processed through a series of convolutional layers to extract local features and connected to batch normalization layers to stabilize the training process. A ReLU activation function is then used to enhance the network's non-linear expressive power. Subsequently, the network passes through two residual blocks containing skip connections. This structural design effectively avoids the gradient vanishing and feature degradation problems common in deep networks and extracts high-level features with multi-scale structural information. The model introduces a global average pooling layer to compress spatial dimensions and aggregate feature representations. A fully connected layer then maps the features to the classification space. A Dropout layer is added to improve the model's robustness and generalization ability during training. The output layer generates the corresponding prediction results. During training, to enhance the model's ability to identify imbalanced sample distributions and difficult-to-classify samples, a Focal loss function is introduced as the training objective function. This loss function dynamically adjusts the loss response by reducing the weights of easily classified samples and increasing the weights of difficult-to-classify samples, allowing the model to focus more on the learning process of minority classes and critical samples. Based on this base learner model, the Adaboost ensemble training framework is constructed. The entire training process iterates through multiple cycles. In each iteration, a base learner is trained according to the current sample distribution weights. The weighted error rate of this learner on the training set is calculated, and its weight in the final ensemble is determined accordingly. Simultaneously, the weight distribution of all samples is updated, so that misclassified samples receive higher attention in the next round, thus gradually forming a set of complementary and diverse trained base learners. To enhance the model's generalization ability and structural diversity, Dropout-Bagging technology is applied to each trained base learner. By randomly discarding some feature dimensions during training, the focus of different base learners in the feature space becomes different, thereby improving the discrimination ability and stability of the ensemble model on complex samples. After all base learners are constructed, they are combined into an ensemble learning model. The fault feature vectors extracted in the previous iteration are input into this ensemble model, and each base learner outputs an initial prediction result. After obtaining multiple independent judgments, these initial results are integrated through strategies such as weighted voting or confidence fusion to form a unified ensemble prediction result. A dynamic decision threshold adjustment strategy is applied to the integrated prediction results. A cost-sensitive matrix is ​​constructed to quantify the cost of different judgment errors. By assigning different penalty weights to different types of errors, such as misclassifying normal conditions as faults or ignoring early micro-faults as normal, the classification boundary of the prediction decision is dynamically adjusted. This makes the model more inclined to improve the fault detection rate in early sensitive states, especially maintaining high sensitivity to early signs such as microcracks and local fatigue. Based on this mechanism, specific fault type discrimination results are output, and the severity level of the current fault is calculated by combining the prediction confidence and the threshold mechanism.

[0031] In this embodiment, a plurality of key parameters representing the operating state of the scroll compressor, including discharge pressure, evaporation temperature, condensation temperature, and the current speed of the compressor, are normalized and combined to form a uniform format of working condition parameter vector, which reflects the current thermodynamic cycle conditions, load level and working condition change characteristics of mechanical response in numerical form. The working condition parameter vector is input into the working condition encoder, and the encoder extracts a stable and discriminative working condition feature vector through a multi-layer fully connected structure and a nonlinear activation function. The vector has the ability to describe the numerical values of the original working condition variables and integrates the coupling relationship and potential nonlinear structure between the variables, which can be used as an effective working condition expression to participate in the subsequent compensation and migration reasoning process. Based on the force response difference of the thrust bearing of the scroll compressor under different operating states, combined with the nonlinear variation law of the load change with the working condition, the working condition feature vector and the fault type and its severity are coupled and analyzed, the working condition compensation mapping relationship is constructed, and the compensation feature representation containing the working condition adaptation factor is generated. The representation contains the fault information directly related to the structural damage in the original diagnostic vector and integrates the modulation effect of the fault expression under the current operating state, thereby converting the static classification result into a more engineering-adaptive dynamic compensation result. The domain adversarial network architecture based on separable convolution performs working condition migration processing on the working condition compensation feature. By inputting the working condition compensation feature into the feature extractor, fault classifier and working condition discriminator at the same time, an optimization strategy with an adversarial objective is constructed in the training process, which promotes the feature extractor to learn neutral features that can be used to distinguish fault types without working condition differences, and obtains working condition adaptive features. Dynamic feature analysis is performed on the working condition adaptive features, and a time series feature extraction module containing a bidirectional recurrent neural network structure is constructed for the non-steady state vibration characteristics presented during the start, stop or speed fluctuation process of the scroll compressor. The module is used to frame-by-frame encode and sequence model the working condition adaptive features in the continuous time period, to mine the internal relevance and evolution trend between different time slices, and to extract time series fault features representing the structural health change process. Based on the time series feature results, health state evaluation is performed, the health state of the thrust bearing is diagnosed by comparing the similarity of historical fault patterns and current time series features, and the change rate and amplitude trend, and the health state diagnosis result reflecting the actual use state is output.

[0032] In the embodiment of the present application, by using a piezoelectric acceleration sensor and adopting adaptive band-pass filtering and angle domain synchronous resampling processing, interference signals from other parts of the compressor can be effectively filtered, the influence of speed fluctuation on fault feature extraction can be eliminated, and high-quality preprocessed vibration signals can be obtained. By using adaptive morphological filtering and Hilbert-Huang transform enhancement processing function for signal preprocessing, combining variational mode decomposition technology and IMF automatic selection algorithm, the fault feature impact response is highlighted, and the defects of traditional methods that are difficult to capture the weak modulation signal and short impact response of the thrust bearing of the scroll compressor are effectively overcome. By establishing an accurate mathematical model reflecting the dynamic load of the bearing under scroll motion, combining finite element analysis to calculate the stress distribution, and deducing the theoretical expressions of four typical fault feature frequencies, a theoretical basis and data support are provided for fault diagnosis, which significantly improves the pertinence and accuracy of diagnosis. Feature vectors are extracted from time domain, frequency domain and time-frequency domain, and feature optimization is performed through the maximum correlation and minimum redundancy principle. At the same time, adaptive synthetic oversampling and boundary sample synthesis techniques are used to balance sample distribution, solve the sample imbalance problem, and enhance the recognition ability of minority class fault samples. A deep residual network is used as a base learner, combined with an improved Focal loss function and an Adaboost integrated framework, the diversity of base learners is increased through Dropout-Bagging technology, and a cost-sensitive matrix is introduced to quantify the cost of different types of judgment errors. It has high sensitivity to early weak faults and can accurately identify early faults such as micro-cracks. Through the working condition encoder, working condition compensation network and domain adversarial network architecture, the dynamic fusion of working condition features and fault features is realized, the working condition migration problem is solved, and the dynamic features under non-steady state working conditions are captured. The model maintains stable diagnostic performance under different operating conditions, especially during startup and shutdown.

[0033] In a specific embodiment, the process of step S101 can specifically include the following steps:

[0034] A piezoelectric acceleration sensor array is arranged on the thrust bearing of the scroll compressor, and the original vibration signal of the thrust bearing of the scroll compressor is acquired through the piezoelectric acceleration sensor array;

[0035] The original vibration signal is subjected to analog-to-digital conversion to obtain a digital vibration signal, and the digital vibration signal is subjected to filtering processing to obtain a preliminary filtered signal;

[0036] The preliminary filtered signal is subjected to noise reduction processing to obtain a noise-reduced signal, and a time-angle mapping relationship is established according to a phase signal output by a rotary encoder of the scroll compressor;

[0037] The noise-reduced signal is subjected to rotational angle synchronous resampling processing based on the time-angle mapping relationship to obtain an angle domain synchronous signal, and a preprocessed vibration signal sample set containing multiple working cycles is constructed based on the angle domain synchronous signal;

[0038] The pre-processed vibration signal sample set is subjected to Hilbert transform and modal decomposition to obtain demodulated vibration signals.

[0039] Specifically, a high-sensitivity piezoelectric acceleration sensor array is uniformly arranged on the periphery of the bearing structure. The sensor array is composed of multiple sensors distributed along the circumference of the thrust bearing. It can sense the vibration changes of the bearing at different angular positions, thereby improving the spatial sampling density and directional resolution, and enhancing the detection capability of small impact sources or local structural abnormalities. During the operation of the compressor, these sensors record the vibration response of the bearing surface caused by load changes or structural damage in real time and continuously output the original vibration data in the form of analog signals. This data contains complex non-stationary components caused by rolling element impact, cage sway, or abnormalities in the inner and outer rings. All analog vibration signals from the sensors are simultaneously connected to an analog-to-digital conversion module. This module uses a high-precision A / D converter to convert continuous analog signals into high-time-resolution digital vibration signals. The digital signals are subjected to preliminary filtering to remove mixed environmental interference components and noise background caused by non-target mechanical structures. An adaptive band-pass filter is selected, and its filtering parameters are dynamically calculated and adjusted in combination with the current operating speed of the compressor and the thrust bearing structure parameters, ensuring that the filter can always cover the bearing impact frequency band and shield irrelevant signals generated by other parts of the system under different operating conditions. Through this stage of processing, the low-frequency baseline disturbance and high-frequency electromagnetic interference in the signal that are unrelated to the bearing characteristic frequency are effectively suppressed, forming a preliminary filtered signal with more focused content. The preliminary filtered signal is subjected to noise reduction processing. A noise reduction technique based on multi-scale decomposition is used. Through wavelet threshold processing of the preliminary filtered signal, the local impact features in the signal are highlighted, while non-structural random disturbances are suppressed. In this process, a wavelet basis function with both time-frequency locality and resolution ability is selected, and the decomposition level is set to adapt to the signal variation trend under different time scales. Then, an adaptive threshold method is used to select the noise reduction boundary, achieving the preservation of small abnormal features in the signal and the removal of redundant information. At the same time, in order to eliminate the characteristic frequency drift and modulation spectrum ambiguity caused by compressor speed fluctuations, the phase signal output by the rotary encoder installed on the compressor main shaft is simultaneously read. This signal is used to accurately depict the instantaneous angular velocity variation of the rotor. After obtaining the continuous angular velocity curve, the mapping relationship between time and angle is established through integration operation, i.e., the mapping function from the corresponding rotation angle coordinates of any time point is established. Based on this mapping relationship, the vibration signal sampled in the time domain is remapped to the angle domain, so that the data points at the same angle within the same rotation period remain consistent in different operating stages. After completing the time-angle mapping, the noise reduction signal is subjected to a rotating angle synchronous resampling operation. In the angle coordinate, the sampling points are extracted at equal angle intervals, so that the signals of different operating periods maintain a unified structural standard, and a set of angle domain synchronous vibration signals with good comparability and consistent period are constructed.On the basis of these signals, a plurality of sample segments are divided according to the length of the working cycle, each segment of data covers at least one complete rotation cycle, and a sliding window strategy with a fixed length and an overlap ratio is used to extract the data segment to generate a preprocessed vibration signal sample set covering different working conditions and multiple cycles. The preprocessed sample set is input into the demodulation processing flow, and a Hilbert transform operation is performed to convert each vibration signal into a corresponding analytic signal, and the envelope amplitude is calculated to capture the periodic modulation intensity and impact amplitude fluctuation characteristics. This operation can reveal the energy modulation law caused by local damage and is suitable for identifying periodic impact signals caused by early unstable contact. The demodulated envelope signal is input into the modal decomposition module, and a variational modal decomposition method is used to decompose the signal into a group of intrinsic mode functions with specific frequency bandwidth. These modal functions reflect the components of different modulation modes and impact characteristics in the signal, and can effectively separate the mixed responses generated by multiple sources such as rolling body impact, cage swing or harmonic modulation, to obtain the demodulated vibration signal.

[0040] In a specific embodiment, the process of performing Hilbert transform and modal decomposition on the preprocessed vibration signal sample set to obtain the demodulated vibration signal can specifically include the following steps:

[0041] Adaptive morphological filtering is performed on the preprocessed vibration signal sample set to obtain a pulse feature enhanced signal;

[0042] A Hilbert-Huang transform enhancement function is used to demodulate the pulse feature enhanced signal to obtain a demodulated initial signal;

[0043] A variational modal decomposition technique is used to decompose the demodulated initial signal to obtain a plurality of intrinsic mode functions (IMFs);

[0044] Kurtosis indicators and fault feature frequency energy ratios are calculated for the plurality of intrinsic mode functions (IMFs) to obtain IMF evaluation indicators;

[0045] The plurality of intrinsic mode functions (IMFs) are comprehensively scored and optimally selected according to the IMF evaluation indicators to obtain an optimal IMF set, and the optimal IMF set is integrated into a demodulated vibration signal containing obvious fault features.

[0046] Specifically, the adaptive morphological filtering is applied to the preprocessed vibration signal sample set to highlight the non-stationary impact response caused by local structural damage. Based on the essential advantage of morphological operation, i.e. the erosion and dilation operations are performed on the signal by constructing a structure element matching a specific impact morphology, and then the two are combined to form open and close operation functions, thereby realizing high-fidelity extraction of local mutation points, sharp extreme values and short-time impact characteristics in the original signal. In practical applications, in order to make the structure element more accurately adapt to the irregular vibration waveform generated by the scroll compressor in the running state, an adaptive parameter optimization mechanism is adopted to automatically generate a structure function by analyzing the pulse period, amplitude envelope and modulation morphology of the sample signal, so that the morphological filtering not only has the ability to filter out the trend item, but also can maximize the enhancement effect of abnormal impact points, and output a pulse feature enhanced signal with significant edge characteristics and pulse contrast. The above pulse enhanced signal is input into the demodulation processing module, and by constructing a specially designed Hilbert-Huang transform enhancement processing function, the envelope change information reflecting damage development or structural periodic disturbance in the modulated signal is extracted. In this process, the Hilbert transform is applied to the signal to obtain an analytical signal containing real and imaginary parts, and then the envelope amplitude and instantaneous frequency change trajectory are calculated, and the amplitude weighted processing is applied to the envelope curve by the Gaussian window function, so that the local amplitude change highlighted area in the envelope is further amplified in the final output. Since the Hilbert-Huang transform itself has the advantage of processing non-stationary and nonlinear signals, combined with the Gaussian window function designed for local mutation enhancement, the transform function can effectively capture the periodic modulation changes caused by the local shedding of the rolling body, the deformation or crack propagation of the cage, so as to separate the weak components with low energy but high correlation from the background, and form a demodulation initial signal containing impact period, amplitude mutation and energy focusing area. The variational mode decomposition technique is used to decompose the demodulation initial signal to obtain a plurality of intrinsic mode functions, each mode function has good frequency concentration and amplitude consistency in a specific bandwidth range. Variational mode decomposition automatically determines the center frequency and bandwidth distribution of each mode by solving the variational optimization problem, so that the decomposition result is more stable, non-aliasing, and suitable for processing complex signals with overlapping harmonics and non-stationary impact. In the implementation process, the number of modes is dynamically set according to the spectral energy density and non-stationary fluctuation characteristics of the demodulation initial signal, and the decomposition optimization function is solved iteratively to output a plurality of intrinsic mode functions that cover the entire signal frequency domain but do not overlap with each other, each mode reflects a frequency component or modulation response in the signal, and has clear physical meaning and structural directionality. The kurtosis index and fault feature frequency energy ratio are calculated for each mode, where the kurtosis is used to reflect the impact degree and peak prominence of the signal, and a higher kurtosis value means that the mode contains strong pulse information, and the energy ratio is used to evaluate the energy proportion of the mode near the fault feature frequency, so as to judge whether it is highly related to the structural damage mode.The two indexes of all modes are standardized and combined by setting weight coefficients to construct a comprehensive score function, the performances of various modes in diagnostic effectiveness are uniformly quantified, and the optimal IMF combination is screened out according to the score high-low order. The optimal modal set not only presents a clear impact cycle structure in the time domain, but also has obvious characteristic frequency band energy concentration phenomenon in the frequency domain, which can accurately reflect the multi-dimensional response under typical fault conditions such as inner ring damage, outer ring crack, rolling body spalling or cage offset. The optimal modal set is synthesized to reconstruct a group of complete demodulation vibration signals by linear superposition of multiple modal functions with diagnostic direction. The reconstructed signal has the characteristics of significant impact response, clear modulation sideband and clear spectral structure, and represents a highly purified and diagnostic feature signal form after multiple processing of complex original vibration data.

[0047] In a specific embodiment, the process of performing step S102 can specifically include the following steps:

[0048] Based on the kinematics principle of the scroll disc, the scroll motion trajectory of the scroll disc in the rotation process is calculated;

[0049] According to the scroll motion trajectory, the scroll compressor gas cavity pressure distribution equation and the dynamics balance equation are solved to obtain the dynamic load distribution;

[0050] According to the dynamic load distribution, the stress distribution and deformation characteristics of each component of the thrust bearing under different working conditions are calculated, and a stress distribution model is established;

[0051] According to the stress distribution model, the theoretical mathematical model of the fault characteristic frequency of the inner ring damage, outer ring crack, rolling body spalling and cage deformation in the scroll compressor thrust bearing is derived;

[0052] Based on the theoretical mathematical model of the fault characteristic frequency, the demodulation vibration signal is Hilbert envelope demodulated to obtain the spectral characteristics of each fault type;

[0053] The spectral characteristics of each fault type are classified and arranged according to the inner ring damage, outer ring crack, rolling body spalling and cage deformation to construct a fault characteristic spectrum atlas.

[0054] Specifically, the kinematic principle of scroll is taken as the theoretical basis for modeling and analysis. During the operation of scroll compressor, the orbit of orbiting scroll is determined by its eccentricity, angular velocity and phase angle, which is not free rotation but constrained by the forced trajectory, thus forming the continuous compression of closed gas cavity. In this process, the orbit of orbiting scroll is a typical circular trajectory or a combination of trajectories, which is constantly changing with time in two-dimensional plane. Based on the above trajectory model, the pressure distribution equation of gas cavity at each time is established. Due to the periodic change of gas cavity volume with the movement of scroll, the gas will produce significant pressure gradient during compression, which not only acts on the wall of gas cavity, but also transfers dynamic load to the thrust bearing through the scroll structure. Therefore, under the premise of known trajectory function, the pressure evolution law in the corresponding thermal process is solved, and the gas cavity pressure is mapped to the axial load acting between the orbiting scroll and the thrust bearing. On the basis of load modeling, through the dynamic balance analysis, the dynamic load distribution model is established, which comprehensively considers the pressure force, inertia force, reaction support force and friction force, etc. This model reflects the time-varying load fluctuation characteristics of thrust bearing under different operating conditions, and can reveal the coupling response of axial and radial forces with the change of scroll movement state. After obtaining the dynamic load distribution, these time-varying loads are mapped to the internal structure of the thrust bearing, and combined with the geometric parameters, material properties and boundary conditions of the bearing, the stress distribution model of the key components such as bearing inner ring, outer ring, rolling body and cage is constructed. In the modeling process, the stress concentration area, deformation gradient and material response of the structure under periodic load are calculated systematically to reflect the evolution trend of local area prone to crack initiation, surface spalling or fatigue damage. Especially for the typical working conditions such as start-up, shutdown, low temperature and high pressure, the stress distribution map under each state is calculated respectively, and the potential fault sensitive parts and fault causes are identified. On this basis, combined with the existing bearing theory and signal modulation mechanism, the characteristic frequency theoretical model of typical fault modes such as inner ring damage, outer ring crack, rolling body spalling and cage deformation is derived. These frequency models explain from the perspective of dynamics why different faults will form independent frequency bandwidth and modulation form in the vibration spectrum, for example, the inner ring damage shows modulation components related to rotational speed frequency, the outer ring crack causes periodic impact of spatial fixed frequency due to fixed contact position, the rolling body spalling causes high frequency modulation in the middle and high frequency band, and the cage deformation forms low frequency harmonic due to its periodic disturbance. Combined with the structure and rotational speed parameters of the bearing, the fault frequency prediction system is constructed. After the theoretical model is constructed, it is applied to the demodulation of vibration signal analysis in real time, so the Hilbert envelope demodulation analysis of the demodulated vibration signal extracted in the early stage is carried out based on the fault characteristic frequency model. This process converts the periodic modulation behavior into significant characteristic frequency components and sideband structure by extracting the frequency spectrum structure of envelope signal.The system focuses on the frequency region predicted by the theoretical model, identifies whether there are fault characteristic behaviors such as frequency peak enhancement, bandwidth expansion, harmonic overlap or side frequency symmetry, and performs amplitude positioning, spectrum band separation and characteristic energy evaluation on the spectrum form of each fault to ensure that the corresponding fault type in the current vibration signal is accurately identified from the spectrum level. After identifying the spectrum characteristics of multiple faults, the extracted spectrum structure, main frequency amplitude, side frequency energy and frequency band distribution and other multi-dimensional information are summarized and archived, and a fault characteristic spectrum atlas is constructed. In this atlas, different fault types have independent spectrum characteristic patterns and identification parameters, which are convenient for subsequent use as input labels or reference templates in intelligent identification models.

[0055] In a specific embodiment, the process of performing step S103 can specifically include the following steps:

[0056] Performing time-domain statistical feature calculation on the demodulated vibration signal to obtain a time-domain feature sub-vector;

[0057] Based on the fault characteristic spectrum atlas, calculating the fault characteristic frequency energy ratio and the frequency band energy distribution characteristics to obtain a frequency-domain feature sub-vector;

[0058] Decomposing the demodulated vibration signal into multiple frequency bands, calculating the energy entropy and the time-frequency joint distribution entropy of the short-time Fourier transform of different frequency bands to obtain a time-frequency analysis feature sub-vector;

[0059] Combining the time-domain feature sub-vector, the frequency-domain feature sub-vector and the time-frequency analysis feature sub-vector to obtain an initial feature vector, and performing feature optimization on the initial feature vector to obtain an optimized feature subset;

[0060] Using a hybrid sample balancing strategy of adaptive synthetic oversampling and boundary sample synthesis technology, identifying the sample distribution density and the decision boundary position of the optimized feature subset, generating a new feature subset, and performing vector mapping on the new feature subset to obtain a fault feature vector.

[0061] Specifically, statistical time-domain feature analysis is performed on the demodulated vibration signal samples, and multiple key indicators reflecting the impact intensity, amplitude distribution, fluctuation amplitude and nonlinear evolution characteristics are extracted, including the ratio between the maximum amplitude and the root mean square value to describe the peak factor, the ratio of the fourth central moment to the variance to measure the kurtosis characteristics of the signal, and parameters such as pulse factor, margin factor and sample entropy to represent the non-stationary complexity of the signal. These statistics can capture the periodic impact or amplitude mutation caused by rolling contact structure damage, and reflect the local concentration of energy in the early stage of micro-crack initiation, so they are integrated to form a set of time-domain feature sub-vectors reflecting the waveform shape change in the time domain. Combined with the previously constructed fault feature spectrum, energy distribution analysis is performed on the frequency regions related to various typical faults in the demodulated signal. In this analysis process, the main characteristic frequency positions corresponding to inner ring damage, outer ring crack, rolling element spalling and cage deformation in the spectrum are identified, and the energy proportion concentrated around these frequency points or their sidebands is calculated to form multiple fault frequency energy ratios. Combined with the key frequency band division scheme in the spectrum, the total energy of a specific frequency band and its relative proportion in the entire signal spectrum are extracted to form the frequency band energy distribution characteristics. These features can accurately depict the energy distribution change of the vibration signal in the frequency spectrum space under different fault conditions, and are used to reveal the modulation effect of local damage on the system resonance structure. All these frequency domain features are finally collected into a frequency domain feature sub-vector, which is complementary to the time domain feature to build a multi-domain feature space. The demodulated vibration signal is input into the time-frequency analysis module, which divides the original signal into several frequency sub-bands through multi-scale filtering, then calculates the energy entropy value of each sub-band to measure the non-uniformity of energy distribution in each frequency band; at the same time, the short-time Fourier transform method is used to convert the demodulated signal into a two-dimensional time-frequency matrix, and the entropy value of the joint probability distribution in the matrix is calculated to form the time-frequency joint distribution entropy index, which describes the joint change characteristics of the signal in the time and frequency axes. This type of time-frequency analysis feature is suitable for capturing the transient modulation behavior caused by cage eccentric disturbance, rolling element repeated impact or local shedding, and is particularly effective under non-steady-state working conditions. Finally, these analysis results are integrated to form a time-frequency analysis feature sub-vector. The time-domain feature sub-vector, frequency-domain feature sub-vector and time-frequency analysis feature sub-vector are combined to obtain the initial feature vector. The initial feature vector is subjected to subset screening by using a feature optimization mechanism. This optimization process combines the correlation maximization and redundancy minimization principles, calculates the mutual information strength between each feature and the fault type and the mutual redundancy degree between features, selects the optimal feature combination in terms of information contribution and independence, and forms an optimized feature subset. A hybrid sample balancing strategy using adaptive synthetic oversampling and boundary sample synthesis techniques is used to identify the sample distribution density and decision boundary position of the optimized feature subset.The adaptive oversampling technique identifies the difficult classification region according to the local density of samples in the feature space, and interpolates new samples in the adjacent space to improve the local distribution of the minority class samples; and the boundary sample synthesis technique focuses on the generation of samples near the decision boundary, identifies the potential weak discrimination area by comparing and analyzing the existing edge samples, and synthesizes more diverse boundary samples accordingly, so as to improve the class separation degree and reduce the misjudgment probability. All the newly generated samples and the optimized feature subset jointly constitute an enhanced feature set with balanced structure, and are standardized and vector-mapped to form a feature matrix with uniform format, which forms the final fault feature vector for model training, reasoning and evaluation.

[0062] In a specific embodiment, the process of performing step S104 can specifically include the following steps:

[0063] The base learner model includes an input layer, a convolutional layer, a batch normalization layer, a ReLU activation function, two residual blocks, a global average pooling layer, a fully connected layer, a Dropout layer, and an output layer.

[0064] The base learner model is trained using a Focal loss function, and T base learner models are iteratively trained using an Adaboost ensemble framework. In each iteration, the weighted error rate and base learner weight are calculated, the sample weight is updated, and multiple trained base learners are obtained.

[0065] The Dropout-Bagging technique is applied to the multiple trained base learners to obtain an ensemble learning model.

[0066] The fault feature vector is input into the ensemble learning model for fault recognition analysis, a plurality of initial prediction results are obtained, and the plurality of initial prediction results are fused to obtain an ensemble prediction result.

[0067] The decision threshold dynamic adjustment strategy is applied to the ensemble prediction result, a cost-sensitive matrix is introduced to quantify the cost of different types of judgment errors, and the fault type and severity are obtained.

[0068] Specifically, the base learner model is constructed, and the prediction results of multiple models are fused through an ensemble mechanism to improve the robustness and anti-disturbance ability of diagnosis. In the model architecture design stage, the base learner model is constructed as a neural network with convolutional structure and deep residual mechanism. The input layer is used to receive the optimized fault feature vector, which maintains the same shape and feature dimension. Then, a set of convolutional layers with local perception ability is provided to extract the spatial correlation and local structure pattern in the input features. Through multiple small receptive field kernel functions, fine-grained feature mapping is achieved. After convolution processing, a batch normalization layer is introduced to adjust the mean and variance of the intermediate activation values during training, thereby accelerating the convergence speed and suppressing gradient oscillation, making the network training more stable. Then, a nonlinear activation function ReLU is connected, which introduces nonlinear expression ability in feature mapping and effectively enhances the model's fitting ability for complex boundaries. To deepen the network structure without causing degradation problems, two residual blocks are embedded in the main network, each containing two convolutional layers and a jump connection branch. Through the residual connection structure, the gradient can be directly propagated to the shallow parameters in the network, thereby maintaining the trainability of the deep network and improving the learning efficiency. After convolution and residual feature fusion, a global average pooling layer is connected to compress the parameter dimension and enhance the feature aggregation effect. This structure integrates spatial dimension information into channel-level feature abstraction while preserving the core semantic of the feature map. The pooled output enters the fully connected layer for high-dimensional feature to class label conversion. The Dropout layer is used to implement the partial node random inactivation strategy, enhancing the model's generalization ability and anti-overfitting ability. The final output layer uses the Softmax function to map the network output to a multi-class probability distribution, thereby completing the basic function design of fault type and discrimination output. In the model training stage, to effectively deal with the difficulties of sample class imbalance and overlapping distribution of different fault classes in feature space, the Focal loss function is selected as the training target. This loss function adjusts the difficulty weight of samples to guide the model to pay more attention to minority classes and boundary samples. The focal factor introduced in its structure can compress the influence of easy-to-classify samples on the total loss while enhancing the contribution of difficult samples in gradient propagation, thereby improving the recognition ability of weak faults, initial abnormalities, and other complex samples. On this basis, to construct an integrated model with generalization ability, the Adaboost ensemble framework is used to iteratively train multiple base learner models. In each iteration, the sample weights are redistributed according to the error distribution of the previous round of model on the training set, so that the misclassified samples receive more attention in subsequent training, thereby gradually enhancing the learning ability of the integrated model for difficult-to-classify samples.Within the Adaboost framework, after each round of iteration, the weighted error rate of each base learner is calculated according to the classification accuracy of the base learner on the training set, and the voting weight value of the base learner is determined according to the weighted error rate. Meanwhile, the sample distribution weight is dynamically updated according to the consistency between the output result of the base learner and the real label, and a set of heterogeneous weak classifier clusters trained under the constraint of sample distribution is formed. When the T base learner models are trained, in order to avoid the problems of feature redundancy, structural convergence or overfitting to the same pattern among different models, the Dropout-Bagging mechanism is introduced, and the input feature subset of each base learner is randomly suppressed in the integration stage. Through the hierarchical scattering of feature dimension, the diversity of input is formed, so as to improve the model difference of the ensemble learning system, so that each learner makes a decision with independent judgment path in different feature subspace, and finally forms a stable structure and complementary performance of the integrated classifier system. After the optimized fault feature vector is input into the integrated model, all base learners make independent judgment on it respectively, and output multiple initial prediction results. These results are combined and summarized through the weighted fusion mechanism set in the system to form the final integrated prediction result. In the fusion stage, the training performance and voting weight of each base learner are considered, and the sensitivity of each base learner in different fault categories is adjusted to adjust the final confidence value distribution, so that the model output has higher global consistency and classification stability. On this basis, in order to improve the decision-making adaptability of the model in different misjudgment situations, the decision threshold dynamic adjustment strategy is introduced and the cost sensitive matrix is constructed to quantify the risk difference caused by the judgment error of different categories. In the matrix construction, the two types of errors, "misjudging the normal state as fault" and "misjudging the fault as normal", are marked with high weight, especially for the missed judgment of early damage such as micro crack and rolling body peeling, a high penalty factor is set to ensure that the model is more cautious when dealing with sensitive categories, so as to maximize the diagnostic accuracy while avoiding the risk of costly misjudgment. According to the optimal threshold value output by the cost function, the fusion result is re-calibrated and judged, and the corresponding fault type label is output, and the severity level of the current fault is determined according to the model internal confidence analysis result.

[0069] In a specific embodiment, the artificial intelligence-based scroll compressor thrust bearing fault diagnosis method further comprises the following steps:

[0070] Combining the exhaust pressure, evaporation temperature, condensation temperature and rotating speed into a working condition parameter vector;

[0071] Inputting the working condition parameter vector into a working condition encoder to extract working condition features and obtain a working condition feature vector;

[0072] According to the nonlinear relationship between the thrust bearing load of the scroll compressor and the working condition change, the working condition compensation analysis is performed on the working condition feature vector and the fault type and severity, and the working condition compensation feature is obtained.

[0073] The domain adversarial network architecture based on separable convolution performs working condition migration processing on the working condition compensation features to obtain working condition adaptive features;

[0074] The working condition adaptive features are subjected to dynamic feature analysis to obtain time sequence fault features, and the time sequence fault features are subjected to health state evaluation to obtain health state diagnosis results.

[0075] Specifically, the key operating condition variables related to the operation of the scroll compressor are monitored in real time, and the exhaust pressure, evaporation temperature, condensation temperature and rotating speed are combined into an operating condition parameter vector. The operating condition parameter vector is input into an operating condition encoder for operating condition feature extraction. The encoder takes a fully connected neural network as the main architecture, and contains multiple layers of nonlinear transformation modules. The input layer receives the operating condition parameter vector, and then the potential coupling relationship between variables, nonlinear evolution characteristics and implicit rules are extracted layer by layer through multiple hidden layers with activation functions, and an operating condition feature vector containing high-dimensional abstract features is output. The vector contains the expression of the original physical quantity, and also integrates the interaction between operating conditions and trend change information, effectively depicting the modulation ability of different operating condition states on the response mode of the mechanical system. According to the nonlinear relationship between the thrust bearing load of the scroll compressor and the operating condition change, operating condition compensation analysis is performed on the operating condition feature vector and the fault type and severity. The compensation process is based on the fault type and severity labels extracted from the actual operating data. The operating condition feature vector is fused with the fault labels output by the intelligent classification model, and a joint input vector is constructed. The joint input vector is then input into the operating condition compensation network, and a set of operating condition compensation features is output. The features represent the modulation state or expression variation of the original fault features under the current operating condition. This mechanism solves the problem of large differences in the same type of fault signal under different rotating speeds, loads, temperatures and other conditions, thereby improving the stability and consistency of the model in cross-condition environments. A domain adversarial network architecture based on separable convolution is used to perform operating condition migration processing on the operating condition compensation features. The network includes three functional sub-modules: a feature extractor, a fault classifier and a domain discriminator. The feature extractor performs hierarchical extraction of the compensation features while maintaining computational efficiency through multiple layers of separable convolution operations; the classifier uses the extracted shared features to complete fault type discrimination and output; and the introduction of the domain discriminator enables the entire network to have adversarial ability. Through the introduction of the gradient reversal mechanism during training, the feature extractor is forced to minimize the distribution difference between operating condition domains while maintaining classification discrimination ability, so that the extracted features have cross-condition sharing attributes. The domain adversarial training mechanism enables the model to maintain fault recognition performance when facing unseen operating conditions, thereby generating operating condition adaptive features with invariant operating conditions. The operating condition adaptive features are input into a time series feature analysis module to model the fault evolution trend during compressor operation. The module uses a bidirectional recurrent neural network architecture combined with a short-term memory unit to model the features in the continuous time window frame by frame, capturing the diagnostic features at the current time and combining context information to understand the historical evolution trajectory and future trend of the fault state, thereby extracting structured time series fault features. Through this process, the model identifies whether there is a continuously intensifying damage pattern, a periodic recurrence feature or a transient severe disturbance behavior in the diagnosis results, providing dynamic reference for health state analysis.According to the evolution behavior and distribution characteristics of the time sequence fault features in different dimensions, combined with multiple criteria such as diagnostic threshold, feature fluctuation trend and time continuity, a comprehensive health state evaluation is carried out, a fault state is mapped to a multi-level health state label by constructing an evaluation model, and the running health grade of the thrust bearing at the current time is output.

[0076] The above describes the method for diagnosing the thrust bearing fault of the scroll compressor based on artificial intelligence in the embodiment of the application, and the following describes the system for diagnosing the thrust bearing fault of the scroll compressor based on artificial intelligence in the embodiment of the application, please refer to Figure 2 An embodiment of the system for diagnosing the thrust bearing fault of the scroll compressor based on artificial intelligence in the embodiment of the application includes:

[0077] The acquisition module 201 is configured to acquire original vibration signals of the thrust bearing of the scroll compressor through a piezoelectric acceleration sensor array, and perform angle domain synchronization and modal decomposition on the original vibration signals to obtain demodulated vibration signals.

[0078] The extraction module 202 is configured to extract fault feature spectrum of inner ring damage, outer ring crack, rolling body peeling and cage deformation in the thrust bearing of the scroll compressor according to the demodulated vibration signals.

[0079] The construction module 203 is configured to extract fault features from time domain, frequency domain and time-frequency domain based on the fault feature spectrum, and construct a fault feature vector.

[0080] The analysis module 204 is configured to input the fault feature vector into an integrated learning model to perform fault recognition analysis, and obtain a fault type and severity.

[0081] Through the cooperation of each component, through the piezoelectric acceleration sensor and the use of adaptive band-pass filtering and angle domain synchronous resampling processing, the interference signals from other parts of the compressor can be effectively filtered, the influence of speed fluctuation on fault feature extraction is eliminated, and high-quality preprocessed vibration signals are obtained. Adaptive morphological filtering and Hilbert-Huang transform enhancement processing function are used for signal preprocessing, combined with variational mode decomposition technology and IMF automatic selection algorithm, the fault feature impact response is highlighted, and the defects of traditional methods that are difficult to capture the weak modulation signal and short impact response of the thrust bearing of the scroll compressor are effectively overcome. By establishing an accurate mathematical model reflecting the bearing dynamic load under scroll motion, combined with finite element analysis to calculate the stress distribution, the theoretical expression of four typical fault characteristic frequencies is derived, providing a theoretical basis and data support for fault diagnosis, significantly improving the pertinence and accuracy of diagnosis. Feature vectors are extracted from time domain, frequency domain and time-frequency domain, and feature optimization is performed through the maximum correlation and minimum redundancy principle, while adaptive synthetic oversampling and boundary sample synthesis techniques are used to balance sample distribution, solve the sample imbalance problem, and enhance the recognition ability of minority class fault samples. Deep residual network is used as a base learner, combined with improved Focal loss function and Adaboost integrated framework, through Dropout-Bagging technology to increase the diversity of base learners, and by introducing a cost-sensitive matrix to quantify the cost of different types of judgment errors, it has high sensitivity to early weak faults and can accurately identify early faults such as micro-cracks. Through the working condition encoder, working condition compensation network and domain adversarial network architecture, the dynamic fusion of working condition features and fault features is realized, the working condition migration problem is solved, and the dynamic features under non-steady-state working conditions are captured, so that the model maintains stable diagnostic performance under different operating conditions, especially during startup and shutdown.

[0082] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, system and unit can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0083] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a vortex compressor thrust bearing fault diagnosis device based on artificial intelligence (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0084] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for diagnosing thrust bearing faults in a scroll compressor based on artificial intelligence, characterized in that, include: The original vibration signal of the thrust bearing of the scroll compressor is acquired by a piezoelectric accelerometer array, and the original vibration signal is synchronized in the angular domain and decomposed in the mode to obtain the demodulated vibration signal. Based on the demodulated vibration signal, fault characteristic spectrum maps of inner ring damage, outer ring cracks, rolling element spalling, and cage deformation in the thrust bearing of the scroll compressor are extracted, including: calculating the scroll motion trajectory during rotation based on the kinematic principle of the scroll disk; solving the pressure distribution equation and dynamic equilibrium equation of the scroll compressor chamber based on the scroll motion trajectory to obtain the dynamic load distribution; calculating the stress distribution and deformation characteristics of each component of the thrust bearing under different operating conditions based on the dynamic load distribution, and establishing a stress distribution model; deriving the fault characteristic frequency theoretical mathematical model of inner ring damage, outer ring cracks, rolling element spalling, and cage deformation in the thrust bearing of the scroll compressor based on the stress distribution model; performing Hilbert envelope demodulation on the demodulated vibration signal based on the fault characteristic frequency theoretical mathematical model to obtain the spectrum characteristics of each fault type; classifying and organizing the spectrum characteristics of each fault type according to inner ring damage, outer ring cracks, rolling element spalling, and cage deformation to construct a fault characteristic spectrum map; Based on the fault feature spectrum, fault features are extracted from the time domain, frequency domain, and time-frequency domain, and a fault feature vector is constructed. The fault feature vector is input into an ensemble learning model for fault identification and analysis to obtain the fault type and severity. Combine exhaust pressure, evaporation temperature, condensation temperature, and speed into a vector of operating parameters; The working condition parameter vector is input into the working condition encoder for working condition feature extraction to obtain the working condition feature vector. The encoder is based on a fully connected neural network and includes multiple nonlinear transformation modules. The input layer receives the working condition parameter vector and then extracts the potential coupling relationship between variables, nonlinear evolution characteristics and implicit laws layer by layer through multiple hidden layers with activation functions, and outputs a working condition feature vector containing high-dimensional abstract features. Based on the nonlinear relationship between the thrust bearing load of the scroll compressor and the changes in operating conditions, operating condition compensation analysis is performed on the operating condition feature vector and the fault type and severity to obtain operating condition compensation features. Operating condition migration processing is then performed on the operating condition compensation features based on a domain adversarial network architecture with separable convolutions to obtain operating condition adaptive features. Dynamic feature analysis is then performed on the operating condition adaptive features to obtain time-series fault features, and a health status assessment is performed on the time-series fault features to obtain a health status diagnosis result.

2. The artificial intelligence-based method for diagnosing thrust bearing faults in a scroll compressor according to claim 1, characterized in that, The process involves acquiring the original vibration signal of the thrust bearing of the scroll compressor using a piezoelectric accelerometer array, and performing angular domain synchronization and modal decomposition on the original vibration signal to obtain a demodulated vibration signal, including: An array of piezoelectric acceleration sensors is installed on the thrust bearing of a scroll compressor, and the original vibration signal of the scroll compressor thrust bearing is collected through the array of piezoelectric acceleration sensors. The original vibration signal is converted from analog to digital to obtain a digital vibration signal, and the digital vibration signal is filtered to obtain a preliminary filtered signal. The preliminary filtered signal is subjected to noise reduction processing to obtain a noise-reduced signal. At the same time, a time-angle mapping relationship is established based on the phase signal output by the rotary encoder of the scroll compressor. Based on the time-angle mapping relationship, the noise-reduced signal is subjected to rotation angle synchronization resampling processing to obtain an angle domain synchronization signal, and a preprocessed vibration signal sample set containing multiple working cycles is constructed based on the angle domain synchronization signal. The preprocessed vibration signal sample set is subjected to Hilbert transform and mode decomposition to obtain the demodulated vibration signal.

3. The artificial intelligence-based method for diagnosing thrust bearing faults in a scroll compressor according to claim 2, characterized in that, The step of performing Hilbert transform and mode decomposition on the preprocessed vibration signal sample set to obtain the demodulated vibration signal includes: Adaptive morphological filtering is performed on the preprocessed vibration signal sample set to obtain a pulse feature enhancement signal; The pulse feature enhancement signal is demodulated using the Hilbert-Huang transform enhancement processing function to obtain the demodulated initial signal; The demodulated initial signal is decomposed using variational mode decomposition technology to obtain multiple sets of intrinsic mode functions (IMFs). The kurtosis index and fault characteristic frequency-energy ratio are calculated for the multiple sets of intrinsic mode functions (IMFs) to obtain the IMF evaluation index. The multiple sets of intrinsic mode functions (IMFs) are comprehensively scored and optimally selected based on the IMF evaluation index to obtain the optimal IMF set, and the optimal IMF set is then synthesized into a demodulated vibration signal containing obvious fault characteristics.

4. The artificial intelligence-based method for diagnosing thrust bearing faults in a scroll compressor according to claim 1, characterized in that, The step of extracting fault features from the time domain, frequency domain, and time-frequency domain based on the fault feature spectrum and constructing a fault feature vector includes: The demodulated vibration signal is subjected to time-domain statistical feature calculation to obtain a time-domain feature sub-vector; Based on the fault feature spectrum, calculate the frequency energy ratio and frequency band energy distribution characteristics of each fault feature to obtain the frequency domain feature sub-vector; The demodulated vibration signal is decomposed into multiple frequency bands, and the energy entropy and time-frequency joint distribution entropy of the short-time Fourier transform of different frequency bands are calculated to obtain the time-frequency analysis feature vector. The time-domain feature vector, the frequency-domain feature vector, and the time-frequency analysis feature vector are merged to obtain an initial feature vector, and the initial feature vector is then optimized to obtain an optimized feature subset. A hybrid sample balancing strategy combining adaptive synthetic oversampling and boundary sample synthesis techniques is employed to identify the sample distribution density and decision boundary location of the optimized feature subset, generating a new feature subset. The new feature subset is then vector-mapped to obtain the fault feature vector.

5. The artificial intelligence-based method for diagnosing thrust bearing faults in a scroll compressor according to claim 1, characterized in that, The step of inputting the fault feature vector into an ensemble learning model for fault identification and analysis to obtain the fault type and severity includes: Construct a base learner model, which includes an input layer, a convolutional layer, a batch normalization layer, a ReLU activation function, two residual blocks, a global average pooling layer, a fully connected layer, a Dropout layer, and an output layer. The base learner model is trained using the Focal loss function, and T base learner models are iteratively trained using the Adaboost ensemble framework. In each iteration, the weighted error rate and base learner weights are calculated, and the sample weights are updated to obtain multiple trained base learners. The Dropout-Bagging technique is applied to the multiple trained base learners to obtain an ensemble learning model; The fault feature vector is input into the ensemble learning model for fault identification and analysis to obtain multiple initial prediction results, and the multiple initial prediction results are fused to obtain an ensemble prediction result. The integrated prediction results are applied to a decision threshold dynamic adjustment strategy, and a cost-sensitive matrix is ​​introduced to quantify the cost of different types of judgment errors, thereby obtaining the fault type and severity.

6. A fault diagnosis system for thrust bearings of a scroll compressor based on artificial intelligence, characterized in that, For implementing the artificial intelligence-based scroll compressor thrust bearing fault diagnosis method as described in any one of claims 1 to 5, the artificial intelligence-based scroll compressor thrust bearing fault diagnosis system comprises: The acquisition module is used to acquire the original vibration signal of the thrust bearing of the scroll compressor through a piezoelectric accelerometer array, and to perform angular domain synchronization and modal decomposition on the original vibration signal to obtain the demodulated vibration signal; The extraction module is used to extract the fault feature spectrum of the inner ring damage, outer ring crack, rolling element spalling and cage deformation in the thrust bearing of the scroll compressor based on the demodulated vibration signal. A construction module is used to extract fault features from the time domain, frequency domain, and time-frequency domain based on the fault feature spectrum, and to construct a fault feature vector; The analysis module is used to input the fault feature vector into the integrated learning model for fault identification and analysis, and to obtain the fault type and severity.

Images