A fault diagnosis method for a carbon dioxide transcritical prying block refrigeration unit

Abstract

The application provides a carbon dioxide transcritical prying block refrigerating unit fault diagnosis method, and belongs to the technical field of fault diagnosis based on machine learning; the application is aimed at the differences in signal-to-noise ratio interference and parameter convergence speed in multi-dimensional heterogeneous physical field environment measurement, and proposes a multi-scale heterogeneous parameter steady-state working condition discrimination and tensor construction method, and innovatively introduces an independent discrimination and Boolean intersection joint locking mechanism; aiming at the defect that the multi-variable thermal coupling of the carbon dioxide transcritical prying block refrigerating unit leads to hidden faults that are difficult to quantify, a physical and data dual-driven implicit state analysis model is designed; aiming at the feature overlapping area and coupling interference misjudgment caused by the internal closed loop cascade of the system, a fault classification method depending on an enhanced graph attention network topology aggregation and integrated voting mechanism is proposed. The application realizes high-precision fault diagnosis of the carbon dioxide transcritical prying block refrigerating unit.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology based on machine learning, and particularly relates to a fault diagnosis method for transcritical carbon dioxide skid-mounted refrigeration units. Background Technology

[0002] In the field of refrigeration and cryogenic engineering, transcritical carbon dioxide skid-mounted chillers are widely used due to their environmental friendliness and high refrigeration efficiency. The operating state of these units is directly and profoundly affected by extremely high exhaust pressures and complex transcritical thermodynamic cycles, making fault diagnosis technology indispensable for ensuring the safe and stable operation of the system. The core components deployed within these units, such as compressors, air coolers, electronic expansion valves, and evaporators, are highly coupled in both physical space and thermodynamic mechanisms. The macroscopic health of the system is not determined by single-dimensional sensor readings, but is fundamentally constrained by the dynamic synergistic effect of multi-dimensional heterogeneous physical field parameters. This physical parameter system specifically involves multi-node fluid temperature, fluid pressure, and compressor electrical power, among other multi-scale characteristics.

[0003] However, in actual industrial operation and troubleshooting, fault diagnosis of transcritical CO2 generator units still heavily relies on traditional single-threshold alarms or purely data-driven paradigms lacking physical constraints. Researchers and engineers typically pre-set alarm upper and lower limits for a single physical quantity, or directly input raw sensor data into a basic deep neural network for black-box prediction. In this forward diagnostic mode, fault identification often relies on simple statistical mapping of massive amounts of historical fault data, making it difficult to form an efficient diagnostic path with underlying physical interpretability.

[0004] The aforementioned traditional diagnostic methods have revealed a series of significant shortcomings in the field of transcritical refrigeration. First, the complex industrial environment makes temperature, pressure, and power data highly susceptible to contamination from high-frequency switching harmonics and mechanical resonance noise from inverters. Traditional time-domain smoothing filters struggle to effectively eliminate strong interference in specific frequency domains while preserving the system's dynamic characteristics, resulting in generally low-quality source data input to the diagnostic model. Second, the evolution and convergence speeds of different physical field parameters within the unit vary greatly. For example, the temperature field changes slowly while the pressure field responds extremely rapidly. Traditional globally unified steady-state discrimination methods ignore this multi-scale heterogeneous characteristic, easily triggering false alarms and missed alarms during transient system adjustments or start-stop cycles. Third, highly nonlinear thermodynamic coupling effects are prevalent among the key components of the unit. Traditional black-box deep learning methods lack the ability to explicitly model the unidirectional physical flow of refrigerant and the underlying thermal cycle mechanism, failing to extract mechanistic residuals with clear physical meaning. Such pure data networks, lacking physical topological constraints, struggle to accurately pinpoint the root causes of deep-seated faults involving multiple variables, such as evaporator scaling, air cooler blockage, or compressor valve leakage, severely limiting their deployment capabilities. Summary of the Invention

[0005] To address the above problems, this invention proposes a fault diagnosis method for transcritical carbon dioxide skid-mounted refrigeration units, comprising the following steps: S1 collects temperature, pressure and actual operating power of the compressor at multiple nodes of the unit; S2, the collected data is preprocessed to obtain the temperature matrix, pressure matrix and compressor actual operating power sequence; then, based on the multiple sliding observation window and Boolean intersection joint locking mechanism, the Boolean sequence of temperature parameters, pressure parameters and compressor actual operating power under steady-state conditions is calculated, and spliced ​​along the physical flow direction to generate steady-state condition state tensor; S3 inputs the steady-state state tensor into the physical and data-driven implicit state analytical model. First, the core thermal parameters are calculated through the basic thermal feature extraction layer. Then, the above parameters are input into the global mechanism residual generation layer, and the difference operation is performed with the preset health tensor benchmark value. The mechanism residual representation sequence is constructed by directional splicing. Finally, the steady-state state tensor and the mechanism residual representation sequence are input into the enhanced graph attention network to analyze the high-order correlation state vector. S4 uses the high-order correlation state vector as the core feature input to the random forest classifier. It constructs multiple independent decision trees based on the bootstrap sampling algorithm and the Gini impurity criterion. It uses the ensemble voting mechanism of multiple decision trees and the maximum value function of the independent variable to remove the multivariate coupling interference and output a unique fault classification label.

[0006] Preferably, the data collected by S1 specifically includes: Sensors are deployed on the fluid inlet and outlet pipelines of the compressor, air cooler, and evaporator of the transcritical carbon dioxide skid-mounted refrigeration unit to collect the inlet temperature of the air cooler. With import pressure air cooler outlet temperature Export pressure Evaporator inlet temperature With import pressure Evaporator outlet temperature Export pressure compressor suction temperature With inhalation pressure compressor discharge temperature With exhaust pressure The actual operating power of the compressor is obtained through the power distribution cabinet. Ambient temperature is obtained through the system casing. .

[0007] Preferably, the preprocessing involves performing multidimensional heterogeneous physical field frequency domain adaptive filtering for noise reduction, using fast Fourier transform, low-pass cutoff frequency threshold and notch filter algorithm to remove high-frequency and switching harmonic noise, and obtaining temperature matrix, pressure matrix and compressor actual operating power sequence.

[0008] Preferably, the specific process for calculating the Boolean sequence of temperature parameters, pressure parameters, and actual operating electrical power of the compressor under steady-state conditions based on a multi-sliding observation window and a Boolean intersection joint locking mechanism is as follows: For temperature matrix Set length as The temperature matrix at each time step within the first sliding observation window is calculated using a first-order backward difference algorithm. The time derivatives of all elements in the array are used to obtain a temperature-time derivative matrix in the form of a two-dimensional array; the current time step is then determined. Is the maximum absolute value in the temperature-time derivative matrix lower than the preset temperature steady-state derivative threshold? If so, then at the current time step Output a boolean value that is 1 if it is true, otherwise output a value at the current time step. Output a Boolean value of 0; obtain a steady-state temperature Boolean sequence composed of the Boolean values ​​at each time step; For a second sliding observation window with a length of , the pressure matrix at each time step within the second sliding observation window is calculated using a first-order backward difference algorithm. The time derivatives of all elements in the matrix are used to obtain a pressure time derivative matrix in the form of a two-dimensional array; the current time step is then determined. Is the maximum absolute value in the downpressure time derivative matrix lower than the preset steady-state pressure derivative threshold? If so, then at the current time step... Output a boolean value that is 1 if it is true, otherwise output a value at the current time step. Output a Boolean value of 0; obtain a steady-state Boolean sequence of pressure values ​​for each time step; For a third sliding observation window of length , the fluctuation variance of the compressor's actual operating power sequence at each time step within the third sliding observation window is calculated, resulting in a power fluctuation variance sequence in the form of a one-dimensional array. Next, the current time step is determined. Is the current variance value in the power fluctuation variance sequence lower than the preset power steady-state variance threshold? If so, then at the current time step... Output a boolean value that is 1 if it is true, otherwise output a value at the current time step. Output a Boolean value of 0; obtain a steady-state Boolean sequence of electrical power composed of the Boolean values ​​at each time step.

[0009] Preferably, the step of splicing along the physical flow direction to generate the steady-state state tensor specifically involves: At the current time step, the values ​​in the steady-state Boolean sequences of temperature, pressure, and electrical power are read synchronously. It is then determined whether the current values ​​of these sequences are all equal to 1. If so, the arithmetic mean of the temperature parameters covered in the first sliding observation window, the arithmetic mean of the pressure parameters covered in the second sliding observation window, and the arithmetic mean of the actual operating electrical power of the compressor covered in the third sliding observation window are calculated. These are then sequentially concatenated according to the physical flow direction of the refrigerant to obtain a one-dimensional spatial physical feature vector of the refrigeration unit at the current time step. Finally, the one-dimensional spatial physical feature vectors of the refrigeration unit obtained from all time steps are stacked and concatenated row-wise according to the time dimension to obtain a steady-state state tensor in the form of a two-dimensional matrix.

[0010] Preferably, the basic thermal feature extraction layer specifically comprises: For each time step in the steady-state state tensor, the average compressor discharge pressure at that time step is used. Average compressor suction pressure Calculate the actual operating pressure ratio ; Utilizing the average compressor discharge temperature at this time step Average compressor suction temperature Calculate the actual temperature rise of the compressor ; The average actual operating power of the compressor at this time step Compared with actual operating pressure ratio Calculate the power consumption per unit pressure ratio ; The average inlet pressure of the air cooler at this time step Average outlet pressure of the air cooler Calculate the actual flow pressure drop of the air cooler ; The average outlet temperature of the air cooler at this time step With average ambient temperature Calculate the approximate temperature difference of the air cooler ; Utilizing the average evaporator inlet pressure at this time step Average evaporator outlet pressure Calculate the physical pressure drop on the evaporator side. ; Utilizing the average evaporator outlet temperature at this time step Average temperature of evaporator inlet Calculate the actual temperature rise on the evaporator side .

[0011] Preferably, the global mechanism residual generation layer specifically comprises: Using the parameters output from the basic thermal feature extraction layer as input, for each time step, the actual operating pressure ratio is... Compared with the pressure ratio health benchmark value The pressure ratio residual is obtained by subtraction, and the actual temperature rise of the compressor is calculated. Compared with the healthy baseline value of temperature rise The temperature rise residual is obtained by subtraction, and the power consumption per unit pressure is calculated. Power consumption health benchmark The power consumption residual is obtained by subtracting the values. The pressure ratio residual, temperature rise residual, and power consumption residual are then concatenated and concatenated to generate the compressor degradation residual vector. ; Secondly, the actual flow pressure drop of the air cooler Compared with the preset air cooler pressure drop reference constant The residual pressure drop of the air cooler is obtained by subtraction, which brings the air cooler closer to the temperature difference. Approaching the healthy baseline value of temperature difference The approximate temperature difference residual is obtained by subtracting the residuals. The air cooler voltage drop residual is then concatenated with the approximate temperature difference residual to generate the air cooler composite impedance residual vector. ; Next, the physical pressure drop on the evaporator side. With evaporator pressure drop healthy benchmark value The residual pressure drop of the evaporator is obtained by subtraction, and the actual temperature rise on the evaporator side is calculated. With evaporator temperature rise healthy benchmark value The evaporator temperature rise residual is obtained by subtraction. The evaporator pressure drop residual and the evaporator temperature rise residual are then concatenated in series to generate the evaporator control deviation residual vector. Finally, the compressor degradation residual vector at this time step is... Air cooler composite impedance residual vector Evaporator control deviation residual vector By concatenating and splicing the data, single-step mechanism deviation parameters are obtained. These parameters are then stacked row-by-row in chronological order to obtain a mechanism residual characterization sequence in the form of a two-dimensional matrix. .

[0012] Preferably, the enhanced graph attention network specifically comprises: For each time step, the compressor, air cooler, electronic expansion valve, and evaporator are mapped to four nodes of the graph structure data. The average values ​​of the air cooler outlet temperature, air cooler outlet pressure, evaporator inlet temperature, and evaporator inlet pressure at that time step in the basic steady-state operating condition tensor are connected in series to construct the initial features of the electronic expansion valve node. The average compressor suction temperature, average compressor suction pressure, average compressor discharge temperature, average compressor discharge pressure, and average actual operating power of the compressor are compared with the compressor degradation residual vector at this time step. Perform serial splicing to construct the initial features of the compressor node; The average values ​​of the air cooler inlet temperature, inlet pressure, outlet temperature, outlet pressure, and ambient temperature are compared with the residual vector of the air cooler's overall impedance at that time step. Serial splicing is performed to construct the initial features of the air cooler nodes; the average values ​​of evaporator inlet temperature, evaporator inlet pressure, evaporator outlet temperature, and evaporator outlet pressure are compared with the evaporator control deviation residual vector at that time step. Perform series splicing to construct the initial features of the evaporator nodes; The unidirectional physical flow direction of the refrigerant in the transcritical cycle is extracted. Based on the unidirectional physical flow direction, directed connection edges of the graph structure data are established. A first directed edge is established from the compressor node to the gas cooler node, a second directed edge is established from the gas cooler node to the electronic expansion valve node, a third directed edge is established from the electronic expansion valve node to the evaporator node, and a fourth directed edge is established from the evaporator node to the compressor node. The initial features of the electronic expansion valve node, the compressor node, the gas cooler node, and the evaporator node are combined into a graph node feature matrix. The first directed edge, the second directed edge, the third directed edge, and the fourth directed edge are combined into a graph adjacency matrix. The graph node feature matrix and the graph adjacency matrix are fused to construct the physical graph structure data at this time step. The physical graph structure data is input into the enhanced graph attention network layer to parse the graph node feature matrix and graph adjacency matrix in the physical graph structure data; for any target node in the graph adjacency matrix with directed connecting edges... Physically adjacent nodes Extract target nodes from the graph node feature matrix respectively. The corresponding initial feature vector of the target node Physically adjacent nodes The corresponding initial feature vector of the adjacent nodes Using the weight matrix Initial feature vector of the target node Initial feature vectors of adjacent nodes Spatial projection is performed, and the linear unit activation function with leakage correction is embedded after the projected features to calculate the dynamic attention coefficients between nodes. ; Using the normalized exponential function for the target node The dynamic attention coefficients of all physically adjacent nodes are globally normalized to generate dynamic attention weights. ; Based on dynamic attention weights Perform a weighted aggregation operation on the initial feature vectors of adjacent nodes, combined with a nonlinear activation function. Obtain updated target node features ; Finally, the N enhanced graph attention network layers are connected in series and stacked to construct a complete enhanced graph attention network.

[0013] Preferably, the specific process of S4 is as follows: First, the higher-order correlation state vector As core sensitive input features, these are fed into a pre-built random forest classifier; the random forest classifier is set to contain a number of... Independent decision trees; based on high-order correlation state vectors The feature dimensions are used to perform random sampling with replacement using a bootstrap sampling algorithm to generate... Each training subset is a separate training subset, and an independent decision tree is built for each training subset; Secondly, for any independent decision tree in the random forest classifier, the high-order association state vector is... Input the root node of the independent decision tree, and perform feature splitting on each internal node according to the Gini impurity criterion; define the samples contained in the current node as belonging to the i-th... The probability of each hardware failure category is Calculate the Gini impurity of the node; based on the principle of minimizing Gini impurity, propagate the higher-order correlation state vector. Continue until the corresponding leaf node is reached, then output the single-tree fault prediction category of this independent decision tree; Again, summarizing all The single-tree fault prediction categories output by each independent decision tree are used to calculate the cumulative votes for each preset hardware fault classification label using an integrated voting mechanism of multiple decision trees; the first... Independent decision trees for the first The voting indicator variables for each hardware fault category are: When the predicted category matches the hardware failure category The value is 1 if it is not 1, otherwise the value is 0; calculate the first... Cumulative votes for each hardware failure category ; Finally, the cumulative vote count of all hardware fault categories is filtered using the maximum value function, and the target fault category with the highest cumulative vote count is extracted. A unique fault classification label pointing to evaporator scaling, air cooler blockage, or compressor valve leakage is output.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. To address the signal-to-noise ratio interference and parameter convergence speed differences in multi-dimensional heterogeneous physical field environment measurements, a multi-scale heterogeneous parameter steady-state condition discrimination and tensor construction method is proposed. This method utilizes Fast Fourier Transform to map the spectrum, forcibly sets high-frequency resonance noise to zero through determined temperature low-pass cutoff frequency thresholds and pressure cutoff frequency thresholds, and directly attenuates switching harmonics by calling a notch filter algorithm based on the actual operating power of the compressor. It innovatively introduces a joint locking mechanism of independent discrimination and Boolean intersection, requiring that the current value of the steady-state Boolean sequence corresponding to each parameter be synchronously equal to 1 before row-by-row stacking can be performed, thereby ensuring that the extracted steady-state condition tensor possesses absolute physical reliability and technical determinism. 2. To address the difficulty in quantifying hidden faults caused by multivariate thermodynamic coupling in transcritical CO2 skid-mounted chiller units, a dual-drive implicit state analysis model based on both physics and data is designed. The model's embedded basic thermodynamic feature extraction layer directly derives core indicators, including the actual operating pressure ratio, actual flow pressure drop of the air cooler, and actual temperature rise on the evaporator side, by analyzing the steady-state operating condition tensor. Furthermore, the global mechanism residual generation layer forces the above-mentioned real-time calculated parameters to perform subtraction operations with the health baseline values ​​under fault-free conditions, strictly concatenating and splicing them to generate compressor degradation residual vectors, air cooler comprehensive impedance residual vectors, and evaporator control deviation residual vectors. This design transforms the small nonlinear shifts of multidimensional variables into a physically unique residual representation sequence, completely eliminating the logical ambiguity caused by simply relying on the original numerical values. 3. To address the misjudgment caused by overlapping feature regions and coupling interference resulting from closed-loop cascading within the system, a fault classification method relying on enhanced graph attention network topology aggregation and integrated voting mechanism is proposed. This network maps four sets of core hardware as graph nodes, constructs the first to fourth directed edges based on the unidirectional physical flow of refrigerant, and completes the physical graph structure data construction. The algorithm dynamically generates attention weights and performs weighted aggregation operations using a linear unit activation function with leakage correction and a normalized exponential function, outputting a high-order relational state vector. Using this vector as the core sensitive input feature, multiple independent decision trees of a random forest classifier are generated using a bootstrap sampling algorithm and the principle of minimizing Gini impurity. Relying on the multi-tree cumulative vote competition mechanism, the multivariate coupling interference caused by single feature mutation is removed, and the unit fault classification label is output. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the following description is only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the overall steps of the visual rotational speed measurement method based on temporal feature analysis and dual-domain refinement estimation of the present invention.

[0017] Figure 2 This is a diagram of the enhanced graph attention network structure of the present invention.

[0018] Figure 3 This is a diagram illustrating the offline joint training and online deployment architecture of the diagnostic model of this invention.

[0019] Figure 4 This is a graph showing the comprehensive comparison of the fault identification accuracy of different diagnostic methods in this invention.

[0020] Figure 5 This is a comparison of the feature dimensionality reduction clustering distribution with and without mechanism residual guidance in the embodiments of the present invention.

[0021] Figure 6 The figures show the experimental results of robust attenuation curves of various methods under different sensor signal-to-noise ratio interference in the embodiments of the present invention. Detailed Implementation

[0022] To achieve high-precision fault diagnosis of transcritical carbon dioxide skid-mounted chiller units, this invention proposes a fault diagnosis method for transcritical carbon dioxide skid-mounted chiller units. The specific steps are as follows: Figure 1As shown: First, multi-dimensional heterogeneous physical field frequency domain adaptive filtering denoising and multi-scale steady-state tensor construction are performed. The collected unit temperature, pressure, and power data are mapped to the frequency domain via Fast Fourier Transform (FFT). Low-pass cutoff frequency thresholds and notch filter algorithms are applied for targeted denoising. The time-domain matrix and sequence are reconstructed via Inverse Fourier Transform (IFT). Based on multiple sliding observation windows, the time derivative matrix and fluctuation variance sequence are calculated respectively. A steady-state Boolean sequence is generated by comparing with preset thresholds. The steady-state conditions are established through a Boolean intersection joint locking mechanism. Then, the arithmetic mean is calculated and strictly concatenated along the physical flow direction, stacked row by row to construct the steady-state operating condition tensor. Second, feature extraction is performed based on a dual-driven implicit state analysis model of physics and data. The steady-state operating condition tensor is input into the basic thermal feature extraction layer to calculate core thermal parameters. Then, these parameters are input into the global mechanism residual generation layer, and the difference operation is performed with the preset health tensor benchmark value. The sequence is then constructed by targeted concatenation to build the mechanism residual representation sequence. Finally, an enhanced graph attention network is used to analyze higher-order correlation states. The steady-state state tensor and the mechanism residual representation sequence are input into the network. The core hardware of the unit is mapped as graph nodes, and directed edges are constructed along the unidirectional physical flow of refrigerant. The resulting physical graph structure data is fused. Dynamic attention weights are calculated using a weight matrix and activation function to aggregate and update the target node features. Then, a global average pooling operation is performed to reduce the time dimension and extract the high-order correlation state vector. Finally, a fault classification decision closed loop based on an integrated voting mechanism is constructed. The high-order correlation state vector is used as the core feature and input into a random forest classifier. Multiple independent decision trees are constructed based on the bootstrap sampling algorithm and the Gini impurity criterion. The integrated voting mechanism of multiple decision trees and the maximum value function of the independent variable are used to remove the multivariate coupling interference and output a unique fault classification label.

[0023] The present invention will be further described below with reference to embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0024] S1. Multiphysics Data Acquisition and Preprocessing for Transcritical Carbon Dioxide Skid-Mounted Refrigeration Units Multi-physics data acquisition and preprocessing for a transcritical carbon dioxide skid-mounted refrigeration unit: First, temperature, pressure, and actual operating power of the compressor at multiple nodes of the unit are collected. Second, multi-dimensional heterogeneous physical field frequency domain adaptive filtering and noise reduction are performed. High-frequency and switching harmonic noise are removed using fast Fourier transform, low-pass cutoff frequency threshold, and notch filter algorithms to obtain the temperature matrix, pressure matrix, and compressor actual operating power sequence. Finally, multi-scale heterogeneous parameter steady-state condition discrimination and tensor construction are performed. The time derivative and fluctuation variance are calculated based on the sliding observation window. The steady-state conditions are established by relying on the Boolean intersection joint locking mechanism, and the steady-state condition state tensor is generated by splicing along the physical flow direction.

[0025] S1-1 Data Acquisition: Sensors are deployed on the fluid inlet and outlet pipelines of the compressor, air cooler, and evaporator of the transcritical carbon dioxide skid-mounted refrigeration unit. Simultaneously, sensors are deployed on the system casing and electrical control cabinet for data acquisition. Specifically, the inlet temperature of the air cooler is collected. With import pressure air cooler outlet temperature Export pressure Evaporator inlet temperature With import pressure Evaporator outlet temperature Export pressure compressor suction temperature With inhalation pressure compressor discharge temperature With exhaust pressure The actual operating power of the compressor is obtained through the power distribution cabinet. Ambient temperature is obtained through the system casing. .

[0026] S1-2 Multidimensional Heterogeneous Physical Field Frequency Domain Adaptive Filtering for Noise Reduction: For temperature field data, firstly, seven types of temperature data... , , , , , and The spectra of seven temperature data points were obtained by mapping the data to the frequency domain using Fast Fourier Transform; subsequently, a low-pass cutoff frequency threshold for temperature was set. For the spectrum of a certain temperature data, frequencies greater than [a certain value] will be [selected]. The spectral amplitudes were forced to zero to obtain the filtered spectra of seven temperature data points. Finally, the filtered spectra of the seven temperature data points were restored to the time domain through inverse Fourier transform to obtain seven filtered temperature data points. The seven filtered temperature data points were then concatenated according to their feature dimensions to obtain the temperature matrix. ; For pressure field data, firstly, six types of pressure data... , , , , and The spectra of six pressure data points were obtained by mapping them to the frequency domain using Fast Fourier Transform; then, a pressure cutoff frequency threshold was set. This threshold is equal to three times the highest operating frequency of the compressor. For the spectrum of a certain pressure data, frequencies greater than [a certain value] will be considered. The spectral amplitudes were forced to zero to remove high-frequency resonance noise, resulting in filtered spectra for six pressure data types. Finally, inverse Fourier transform was used to restore the filtered spectra of the six pressure data types to the time domain, yielding six filtered pressure data types. These six filtered pressure data types were then concatenated according to their feature dimensions to obtain the pressure matrix. ; Based on the actual operating power of the compressor First, the power spectrum of the compressor during actual operation is obtained by mapping it to the frequency domain using a fast Fourier transform; second, because... The compressor is susceptible to contamination from high-frequency switching harmonics of the inverter. Therefore, a notch filter algorithm is used to directly attenuate the switching frequency and its integer harmonics of the actual operating power spectrum of the compressor, obtaining the filtered spectrum of the actual operating power of the compressor. Finally, the filtered spectrum of the actual operating power of the compressor is restored to the time domain by inverse Fourier transform to obtain the sequence of the actual operating power of the compressor. .

[0027] S1-3 Multi-scale heterogeneous parameter steady-state condition discrimination and tensor construction; to address the differences in convergence speed among different physical quantities within the system, a joint locking mechanism of independent discrimination and Boolean intersection is constructed, with the specific steps as follows: First, regarding the temperature matrix Set length as The temperature matrix at each time step within the first sliding observation window is calculated using a first-order backward difference algorithm. The time derivatives of all elements in the matrix are used to obtain a temperature-time derivative matrix in the form of a two-dimensional array; then, the current time step is determined. Is the maximum absolute value in the temperature-time derivative matrix lower than the preset temperature steady-state derivative threshold? This application sets If so, then at the current time step Output a boolean value that is 1 if it is true, otherwise output a value at the current time step. Output a Boolean value of 0; finally, obtain a steady-state temperature Boolean sequence composed of the Boolean values ​​at each time step.

[0028] First, for a second sliding observation window with a length of , the pressure matrix at each time step within the second sliding observation window is calculated using a first-order backward difference algorithm. The time derivatives of all elements in the matrix are used to obtain a pressure time derivative matrix in the form of a two-dimensional array; then, the current time step is determined. Whether the maximum absolute value in the pressure time derivative matrix is ​​lower than the preset pressure steady-state derivative threshold (this application sets the pressure steady-state derivative threshold to 1); if so, then at the current time step... Output a boolean value that is 1 if it is true, otherwise output a value at the current time step. Output a Boolean value of 0; finally, obtain the pressure steady-state Boolean sequence composed of the Boolean values ​​at each time step; First, a third sliding observation window of length is set for the actual operating power sequence of the compressor. The fluctuation variance of the actual operating power sequence of the compressor at each time step within the third sliding observation window is calculated to obtain a power fluctuation variance sequence in the form of a one-dimensional array. Second, the current time step is determined. Is the current variance value in the power fluctuation variance sequence lower than the preset power steady-state variance threshold? If so, then at the current time step... Output a boolean value that is 1 if it is true, otherwise output a value at the current time step. Output a Boolean value of 0; finally, obtain the steady-state Boolean sequence of electric power composed of the Boolean values ​​at each time step; Finally, at the current time step, the values ​​in the steady-state Boolean sequences of temperature, pressure, and electrical power are read synchronously. It is then determined whether the current values ​​of these sequences are all equal to 1. If so, the arithmetic mean of the seven temperature parameters covered in the first sliding observation window, the arithmetic mean of the six pressure parameters covered in the second sliding observation window, and the arithmetic mean of the actual operating electrical power of the compressor covered in the third sliding observation window are calculated. Then, these parameters are sequentially connected in series according to the physical flow direction of the refrigerant. The specific connection order is: compressor suction temperature... The average values ​​of the following parameters are collected: compressor suction pressure, compressor discharge temperature, compressor discharge pressure, compressor actual operating power, air cooler inlet temperature, air cooler inlet pressure, air cooler outlet temperature, air cooler outlet pressure, ambient temperature, evaporator inlet temperature, evaporator inlet pressure, evaporator outlet temperature, and evaporator outlet pressure. These are used to obtain a one-dimensional spatial physical feature vector of the refrigeration unit at each time step. The one-dimensional spatial physical feature vectors of the refrigeration unit obtained from all time steps are then stacked and concatenated row-wise according to the time dimension to obtain a steady-state state tensor in the form of a two-dimensional matrix.

[0029] The above data is collected in real time after the model is deployed. In addition, the above steps are required to obtain the steady-state state tensor and the variance of electrical power of the transcritical CO2 skid-mounted chiller unit during the fault-free operation phase before model deployment. The steady-state state tensor of the transcritical CO2 skid-mounted chiller unit during the fault-free operation phase is defined as the health tensor to distinguish it from the steady-state state tensor constructed after model deployment. The power steady-state variance threshold is the variance of electrical power.

[0030] S2, Physical and Data-Driven Implicit State Analysis Model Design This invention, based on the fundamental steady-state operating condition state tensor obtained from S1, designs a physical and data-driven implicit state analysis model. This model sequentially includes a fundamental thermodynamic feature extraction layer, a global mechanism residual generation layer, and an enhanced graph attention network. First, the steady-state operating condition state tensor is input into the fundamental thermodynamic feature extraction layer to calculate the actual operating pressure ratio, actual compressor temperature rise, power consumption per unit pressure ratio, actual flow pressure drop of the air cooler, approximate temperature difference of the air cooler, physical pressure drop on the evaporator side, and actual temperature rise on the evaporator side. Second, the global mechanism residual generation layer subtracts the above parameters from the baseline values ​​of the health tensor and concatenates them to construct a mechanism residual representation sequence. Finally, the enhanced graph attention network maps the compressor, air cooler, electronic expansion valve, and evaporator as graph nodes and establishes directed connections along the unidirectional physical flow direction. Dynamic attention weights are used to aggregate and update the target node features, and global average pooling is used to perform time dimension reduction, outputting a high-order associated state vector. The specific calculation process is as follows: S2-1 Basic Thermodynamic Feature Extraction Layer; First, the steady-state operating condition tensor is input into the basic thermodynamic feature extraction layer; For each time step in the steady-state operating condition tensor, the average compressor discharge pressure at that time step is used. Average compressor suction pressure Calculate the actual operating pressure ratio The calculation formula is:

[0031] Utilizing the average compressor discharge temperature at this time step Average compressor suction temperature Calculate the actual temperature rise of the compressor The calculation formula is:

[0032] Secondly, the average actual operating power of the compressor at this time step is used. Compared with actual operating pressure ratio Calculate the power consumption per unit pressure ratio The calculation formula is:

[0033] The average inlet pressure of the air cooler at this time step Average outlet pressure of the air cooler Calculate the actual flow pressure drop of the air cooler The calculation formula is:

[0034] Next, the average outlet temperature of the air cooler at this time step is used. With average ambient temperature Calculate the approximate temperature difference of the air cooler The calculation formula is:

[0035] Utilizing the average evaporator inlet pressure at this time step Average evaporator outlet pressure Calculate the physical pressure drop on the evaporator side. The calculation formula is:

[0036] Finally, the average evaporator outlet temperature at this time step is used. Average temperature of evaporator inlet Calculate the actual temperature rise on the evaporator side The calculation formula is:

[0037] Similarly, the health tensor is input into the basic thermodynamic feature extraction layer to calculate the pressure ratio health baseline value. Health benchmark for temperature rise Power consumption health benchmark healthy benchmark value for air cooler pressure drop Approaching the healthy baseline value for temperature difference Evaporator pressure drop health benchmark With evaporator temperature rise healthy benchmark value .

[0038] S2-2 Global Mechanism Residual Generation Layer; First, the parameters output from the basic thermodynamic feature extraction layer are input into the global mechanism residual generation layer; For each time step, the actual operating pressure ratio is... Compared with the pressure ratio health benchmark value The pressure ratio residual is obtained by subtraction, and the actual temperature rise of the compressor is calculated. Compared with the healthy baseline value of temperature rise The temperature rise residual is obtained by subtraction, and the power consumption per unit pressure is calculated. Power consumption health benchmark The power consumption residual is obtained by subtracting the values. The pressure ratio residual, temperature rise residual, and power consumption residual are then concatenated and concatenated to generate the compressor degradation residual vector. ; Secondly, the actual flow pressure drop of the air cooler Compared with the preset air cooler pressure drop reference constant The residual pressure drop of the air cooler is obtained by subtraction, which brings the air cooler closer to the temperature difference. Approaching the healthy baseline value of temperature difference The approximate temperature difference residual is obtained by subtracting the residuals. The air cooler voltage drop residual is then concatenated with the approximate temperature difference residual to generate the air cooler composite impedance residual vector. ; Next, the physical pressure drop on the evaporator side. With evaporator pressure drop healthy benchmark value The residual pressure drop of the evaporator is obtained by subtraction, and the actual temperature rise on the evaporator side is calculated. With evaporator temperature rise healthy benchmark value The evaporator temperature rise residual is obtained by subtraction. The evaporator pressure drop residual and the evaporator temperature rise residual are then concatenated in series to generate the evaporator control deviation residual vector. Finally, the compressor degradation residual vector at this time step is... Air cooler composite impedance residual vector Evaporator control deviation residual vector By concatenating and splicing the data, single-step mechanism deviation parameters are obtained. These parameters are then stacked row-by-row in chronological order to obtain a mechanism residual characterization sequence in the form of a two-dimensional matrix. .

[0039] S2-3 Enhanced Graph Attention Network; like Figure 2 As shown, firstly, the steady-state state tensor and the residual representation sequence are... Input an enhanced graph attention network; for each time step, map the compressor, air cooler, electronic expansion valve, and evaporator to the four nodes of the graph structure data respectively; The average values ​​of the air cooler outlet temperature, air cooler outlet pressure, evaporator inlet temperature, and evaporator inlet pressure at that time step in the basic steady-state operating condition tensor are connected in series to construct the initial features of the electronic expansion valve node. The average compressor suction temperature, average compressor suction pressure, average compressor discharge temperature, average compressor discharge pressure, and average actual operating power of the compressor are compared with the compressor degradation residual vector at this time step. Perform serial splicing to construct the initial features of the compressor node; The average values ​​of the air cooler inlet temperature, inlet pressure, outlet temperature, outlet pressure, and ambient temperature are compared with the residual vector of the air cooler's overall impedance at that time step. Serial splicing is performed to construct the initial features of the air cooler nodes; the average values ​​of evaporator inlet temperature, evaporator inlet pressure, evaporator outlet temperature, and evaporator outlet pressure are compared with the evaporator control deviation residual vector at that time step. Perform series splicing to construct the initial features of the evaporator nodes; The unidirectional physical flow of refrigerant in the transcritical cycle is extracted, and the specific physical flow is defined as: from the compressor discharge port to the air cooler inlet, from the air cooler outlet to the electronic expansion valve inlet, from the electronic expansion valve outlet to the evaporator inlet, and from the evaporator outlet back to the compressor suction port. Based on the unidirectional physical flow, directed edges of the graph structure data are established, creating a first directed edge from the compressor node to the air cooler node, a second directed edge from the air cooler node to the electronic expansion valve node, a third directed edge from the electronic expansion valve node to the evaporator node, and a fourth directed edge from the evaporator node to the compressor node. The initial features of the electronic expansion valve node, compressor node, air cooler node, and evaporator node are combined into a graph node feature matrix. The first, second, third, and fourth directed edges are combined into a graph adjacency matrix. The graph node feature matrix and the graph adjacency matrix are then fused to construct the physical graph structure data for that time step. Secondly, the physical graph structure data is input into the enhanced graph attention network layer to parse the graph node feature matrix and graph adjacency matrix in the physical graph structure data; for any target node in the graph adjacency matrix that has a directed connection edge. Physically adjacent nodes Extract target nodes from the graph node feature matrix respectively. The corresponding initial feature vector of the target node Physically adjacent nodes The corresponding initial feature vector of the adjacent nodes Using the weight matrix Initial feature vector of the target node Initial feature vectors of adjacent nodes Spatial projection is performed, and the linear unit activation function with leakage correction is embedded after the projected features to calculate the dynamic attention coefficients between nodes. The calculation formula is:

[0040] Next, the normalized exponential function is used to evaluate the target node. The dynamic attention coefficients of all physically adjacent nodes are globally normalized to generate dynamic attention weights. The calculation formula is:

[0041] Based on dynamic attention weights Perform a weighted aggregation operation on the initial feature vectors of adjacent nodes, combined with a nonlinear activation function. Obtain updated target node features The calculation formula is:

[0042] Finally, the N enhanced graph attention network layers are concatenated and stacked to construct a complete enhanced graph attention network. The features of all target nodes, after being passed and updated through multiple enhanced graph attention network layers at each time step, are concatenated to form the single-step global graph representation vector for that time step. The single-step global graph representation vectors of all time steps are then concatenated and expanded sequentially according to the time series. Global average pooling is then used for feature extraction and temporal dimensionality reduction to obtain a high-order association state vector in the form of a one-dimensional array. .

[0043] S3. Construction of a fault classification module based on an integrated voting mechanism This invention constructs a fault classification module based on an ensemble voting mechanism, utilizing the independent space partitioning and ensemble evaluation capabilities of multiple decision trees to remove multivariate coupling interference. First, the high-order correlation state vector is input into a random forest classifier, and a bootstrap sampling algorithm is used to extract a training subset and construct multiple independent decision trees. Second, feature splitting is performed at internal nodes according to the Gini impurity criterion, outputting single-tree fault prediction categories. Finally, the ensemble voting mechanism of multiple decision trees is used to calculate the cumulative vote count for each hardware fault classification label, and the target fault category is selected by taking the maximum value of the independent variable function, outputting fault classification labels pointing to specific hardware. The specific classification execution steps are as follows: First, the higher-order correlation state vector As core sensitive input features, these are fed into a pre-built random forest classifier; the random forest classifier is set to contain a number of... Independent decision trees; based on high-order correlation state vectors The feature dimensions are used to perform random sampling with replacement using a bootstrap sampling algorithm to generate... Each training subset is a separate training subset, and an independent decision tree is built for each training subset; Secondly, for any independent decision tree in the random forest classifier, the high-order association state vector is... Input the root node of the independent decision tree, and perform feature splitting on each internal node according to the Gini impurity criterion; define the samples contained in the current node as belonging to the i-th... The probability of each hardware failure category is The Gini impurity of this node is calculated using the following formula:

[0044] In the above formula Represents the total number of hardware fault categories; based on the principle of minimizing Gini impurity, it propagates the higher-order correlation state vector. Continue until the corresponding leaf node is reached, then output the single-tree fault prediction category of this independent decision tree; Again, summarizing all The single-tree fault prediction categories output by each independent decision tree are used to calculate the cumulative votes for each preset hardware fault classification label using an integrated voting mechanism of multiple decision trees; the first... Independent decision trees for the first The voting indicator variables for each hardware fault category are: When the predicted category matches the hardware failure category The value is 1 if it is not 1, otherwise the value is 0; calculate the first... Cumulative votes for each hardware failure category The calculation formula is:

[0045] Finally, the cumulative vote count of all hardware fault categories is filtered using the maximum value function, the target fault category with the highest cumulative vote count is extracted, the multivariate coupling interference caused by single feature mutation is removed and a unique fault classification label pointing to evaporator scaling, air cooler blockage or compressor valve leakage is output.

[0046] S4. Offline Joint Training and Online Deployment of Diagnostic Models This invention constructs a supervised learning mechanism based on historical operational data to achieve joint training and field deployment of the enhanced graph attention network and the fault classification module based on an ensemble voting mechanism in the implicit state parsing model; such as Figure 3 As shown: First, an offline training sample set and a temporary end-to-end training network are constructed. The cross-entropy loss function is calculated, and the weight matrix of the locked enhanced graph attention network is updated via backpropagation. Second, the sample set is input again into the locked enhanced graph attention network. Based on the principle of minimizing Gini impurity, the node splitting and tree structure generation of the independent decision tree of the fault classification module are completed. Finally, the trained enhanced graph attention network and the fault classification module based on the ensemble voting mechanism are packaged into an executable deployment file and deployed to the field main control equipment to realize the real-time output of unit fault classification labels. The specific execution steps are as follows: First, collect full-dimensional sensor data of the transcritical carbon dioxide skid-mounted chiller unit under fault-free and various physical fault conditions during its historical operating cycle. Based on the data processing logic in step one, the sensor data is converted into steady-state condition tensors, and each set of steady-state condition tensors is labeled with the corresponding real fault category label to construct an offline training sample set. Secondly, a temporary end-to-end training network containing a single-layer linear classifier is constructed. The offline training sample set is input into the enhanced graph attention network to extract high-order correlation state vectors, and the high-order correlation state vectors are input into the single-layer linear classifier to output the predicted fault probability distribution. The cross-entropy loss function is calculated by combining the predicted fault probability distribution with the real fault category label. The gradient descent optimization algorithm is called to backpropagate and update the weight matrix in the enhanced graph attention network until the cross-entropy loss function converges. This completes the independent training of the enhanced graph attention network and locks all the weight matrices inside the enhanced graph attention network. Next, the offline training sample set is input again into the enhanced graph attention network with the weight matrix locked, and the high-order correlation state vectors corresponding to all training samples are extracted. The extracted high-order correlation state vectors are used as input features, and the corresponding real fault category labels are used as target outputs. These are input into the fault classification module based on the ensemble voting mechanism. Based on the principle of minimizing Gini impurity and the bootstrap sampling algorithm, the internal node splitting and tree structure generation of all independent decision trees are completed, and the trained fault classification module based on the ensemble voting mechanism is obtained. Finally, the trained enhanced graph attention network and fault classification module are packaged into an executable deployment file and deployed to the field main control equipment of the transcritical carbon dioxide skid-mounted chiller unit. During the online operation of the unit, the field main control equipment collects multi-dimensional heterogeneous physical field sensing data in real time and outputs the unit fault classification labels in real time for operation and maintenance personnel to refer to.

[0047] S5. Experimental Verification and Analysis Figure 4 This paper presents a comparison of the accuracy of the algorithm of this invention in fault diagnosis of transcritical carbon dioxide units. The horizontal axis represents the deep network model configuration, and the vertical axis represents the evaluation indicators of overall fault diagnosis accuracy and recall rate under extreme operating conditions. After processing by the enhanced graph attention network and global mechanism residual generation layer of this invention, the model achieves an overall diagnostic accuracy of 92.5% for three typical faults: evaporator scaling, air cooler blockage, and compressor valve leakage. The accuracy rates of traditional support vector machines and basic long short-term memory networks using single feature extraction are 80% and 86%, respectively. These objective data demonstrate that the physical and data-driven implicit state analysis model can extract root cause features of faults and reduce the misclassification rate caused by multivariate thermodynamic coupling in transcritical systems.

[0048] Figure 5The experimental results of module effectiveness ablation of the method of this invention are presented. In the figure, the horizontal and vertical axes represent the first and second principal components of the high-order correlated state vector after dimensionality reduction, respectively. The experimental results show that, using the data-driven baseline model without introducing mechanistic residuals, the sample clusters representing evaporator fouling and air cooler blockage have overlapping regions in 2D space. After introducing the global mechanistic residual generation layer, by inputting the compressor degradation residual vector and the air cooler integrated impedance residual vector into the network, the feature points of the three fault categories are distributed into three non-overlapping clusters, and the geometric distance between the clusters increases. The above clustering results show that the global mechanistic residual generation layer can achieve multivariate fault decoupling and produce a quantitative superposition effect with the enhanced graph attention network.

[0049] Figure 6 The graph shows the accuracy decay curves of various methods under different sensor signal-to-noise ratios (SNRs). The horizontal axis represents the SNR value, and the vertical axis represents the fault diagnosis accuracy. When the SNR drops to 10, the accuracy of the basic network without physical topology constraints drops to 70%; the model of this invention, based on a joint low-pass and notch filter front-end and a dynamic weight redistribution mechanism, maintains an accuracy of 85% or higher. Combined with the fault classification rules of the integrated voting mechanism, the diagnostic sequence does not exhibit classification jumps that violate thermodynamic principles. The above test results demonstrate that the method of this invention meets the accuracy and logical reliability requirements of industrial diagnostic applications and is suitable for intelligent operation and maintenance and health management systems for refrigeration units.

[0050] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0051] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for a transcritical carbon dioxide skid-mounted refrigeration unit, characterized in that, The process includes the following: S1 collects temperature, pressure and actual operating power of the compressor at multiple nodes of the unit; S2, the collected data is preprocessed to obtain the temperature matrix, pressure matrix and compressor actual operating power sequence; then, based on the multiple sliding observation window and Boolean intersection joint locking mechanism, the Boolean sequence of temperature parameters, pressure parameters and compressor actual operating power under steady-state conditions is calculated, and spliced ​​along the physical flow direction to generate steady-state condition state tensor; S3 inputs the steady-state state tensor into the physical and data-driven implicit state analytical model. First, the core thermal parameters are calculated through the basic thermal feature extraction layer. Then, the above parameters are input into the global mechanism residual generation layer, and the difference operation is performed with the preset health tensor benchmark value. The mechanism residual representation sequence is constructed by directional splicing. Finally, the steady-state state tensor and the mechanism residual representation sequence are input into the enhanced graph attention network to analyze the high-order correlation state vector. S4 uses the high-order correlation state vector as the core feature input to the random forest classifier. It constructs multiple independent decision trees based on the bootstrap sampling algorithm and the Gini impurity criterion. It uses the ensemble voting mechanism of multiple decision trees and the maximum value function of the independent variable to remove the multivariate coupling interference and output a unique fault classification label.

2. The fault diagnosis method for a transcritical carbon dioxide skid-mounted refrigeration unit as described in claim 1, characterized in that: The data collected by S1 specifically includes: Sensors are deployed on the fluid inlet and outlet pipelines of the compressor, air cooler, and evaporator of the transcritical carbon dioxide skid-mounted refrigeration unit to collect the inlet temperature of the air cooler. With import pressure air cooler outlet temperature Export pressure Evaporator inlet temperature With import pressure Evaporator outlet temperature Export pressure compressor suction temperature With inhalation pressure compressor discharge temperature With exhaust pressure The actual operating power of the compressor is obtained through the power distribution cabinet. Ambient temperature is obtained through the system casing. .

3. A fault diagnosis method for a transcritical carbon dioxide skid-mounted refrigeration unit as described in claim 1 or 2, characterized in that: The preprocessing involves performing multidimensional heterogeneous physical field frequency domain adaptive filtering for noise reduction. It utilizes fast Fourier transform, low-pass cutoff frequency threshold, and notch filter algorithms to remove high-frequency and switching harmonic noise, thereby obtaining the temperature matrix, pressure matrix, and compressor actual operating power sequence.

4. A fault diagnosis method for a transcritical carbon dioxide skid-mounted refrigeration unit as described in claim 1 or 2, characterized in that: The specific process for calculating the Boolean sequence of temperature parameters, pressure parameters, and actual operating electrical power of the compressor under steady-state conditions based on a joint locking mechanism of multiple sliding observation windows and Boolean intersection is as follows: For temperature matrix Set length as The temperature matrix at each time step within the first sliding observation window is calculated using a first-order backward difference algorithm. The time derivatives of all elements in the array are used to obtain a temperature-time derivative matrix in the form of a two-dimensional array; the current time step is then determined. Is the maximum absolute value in the temperature-time derivative matrix lower than the preset temperature steady-state derivative threshold? If so, then at the current time step Output a boolean value that is 1 if it is true, otherwise output a value at the current time step. Output a Boolean value that is 0; Obtain the steady-state temperature Boolean sequence consisting of the Boolean values ​​at each time step; For a second sliding observation window with a length of , the pressure matrix at each time step within the second sliding observation window is calculated using a first-order backward difference algorithm. The time derivatives of all elements in the matrix are used to obtain a pressure time derivative matrix in the form of a two-dimensional array; the current time step is then determined. Is the maximum absolute value in the downpressure time derivative matrix lower than the preset steady-state pressure derivative threshold? If so, then at the current time step... Output a boolean value that is 1 if it is true, otherwise output a value at the current time step. Output a Boolean value that is 0; Obtain the steady-state Boolean sequence of pressure values ​​at each time step; For a third sliding observation window of length , the fluctuation variance of the compressor's actual operating power sequence at each time step within the third sliding observation window is calculated, resulting in a power fluctuation variance sequence in the form of a one-dimensional array. Next, the current time step is determined. Is the current variance value in the power fluctuation variance sequence lower than the preset power steady-state variance threshold? If so, then at the current time step... Output a boolean value that is 1 if it is true, otherwise output a value at the current time step. Output a Boolean value of 0; obtain a steady-state Boolean sequence of electrical power composed of the Boolean values ​​at each time step.

5. The fault diagnosis method for a transcritical carbon dioxide skid-mounted refrigeration unit as described in claim 4, characterized in that: The process of splicing along the physical flow direction to generate the steady-state state tensor is as follows: At the current time step, the values ​​in the steady-state Boolean sequences of temperature, pressure, and electrical power are read synchronously. It is then determined whether the current values ​​of these sequences are all equal to 1. If so, the arithmetic mean of the temperature parameters covered in the first sliding observation window, the arithmetic mean of the pressure parameters covered in the second sliding observation window, and the arithmetic mean of the actual operating electrical power of the compressor covered in the third sliding observation window are calculated. These are then sequentially concatenated according to the physical flow direction of the refrigerant to obtain a one-dimensional spatial physical feature vector of the refrigeration unit at the current time step. Finally, the one-dimensional spatial physical feature vectors of the refrigeration unit obtained from all time steps are stacked and concatenated row-wise according to the time dimension to obtain a steady-state state tensor in the form of a two-dimensional matrix.

6. The fault diagnosis method for a transcritical carbon dioxide skid-mounted refrigeration unit as described in claim 2, characterized in that: The basic thermal feature extraction layer is specifically as follows: For each time step in the steady-state state tensor, the average compressor discharge pressure at that time step is used. Average compressor suction pressure Calculate the actual operating pressure ratio ; Utilizing the average compressor discharge temperature at this time step Average compressor suction temperature Calculate the actual temperature rise of the compressor ; The average actual operating power of the compressor at this time step Compared with actual operating pressure ratio Calculate the power consumption per unit pressure ratio ; The average inlet pressure of the air cooler at this time step Average outlet pressure of the air cooler Calculate the actual flow pressure drop of the air cooler ; The average outlet temperature of the air cooler at this time step With average ambient temperature Calculate the approximate temperature difference of the air cooler ; Utilizing the average evaporator inlet pressure at this time step Average evaporator outlet pressure Calculate the physical pressure drop on the evaporator side. ; Utilizing the average evaporator outlet temperature at this time step Average temperature of evaporator inlet Calculate the actual temperature rise on the evaporator side .

7. The fault diagnosis method for a transcritical carbon dioxide skid-mounted refrigeration unit as described in claim 6, characterized in that: The global mechanism residual generation layer is specifically as follows: Using the parameters output from the basic thermal feature extraction layer as input, for each time step, the actual operating pressure ratio is... Compared with the pressure ratio health benchmark value The pressure ratio residual is obtained by subtraction, and the actual temperature rise of the compressor is calculated. Compared with the healthy baseline value of temperature rise The temperature rise residual is obtained by subtraction, and the power consumption per unit pressure is calculated. Power consumption health benchmark The power consumption residual is obtained by subtracting the values. The pressure ratio residual, temperature rise residual, and power consumption residual are then concatenated and concatenated to generate the compressor degradation residual vector. ; Secondly, the actual flow pressure drop of the air cooler Compared with the preset air cooler pressure drop reference constant The residual pressure drop of the air cooler is obtained by subtraction, which brings the air cooler closer to the temperature difference. Approaching the healthy baseline value of temperature difference The approximate temperature difference residual is obtained by subtracting the residuals. The air cooler voltage drop residual is then concatenated with the approximate temperature difference residual to generate the air cooler composite impedance residual vector. ; Next, the physical pressure drop on the evaporator side. With evaporator pressure drop healthy benchmark value The residual pressure drop of the evaporator is obtained by subtraction, and the actual temperature rise on the evaporator side is calculated. With evaporator temperature rise healthy benchmark value The evaporator temperature rise residual is obtained by subtraction. The evaporator pressure drop residual and the evaporator temperature rise residual are then concatenated in series to generate the evaporator control deviation residual vector. Finally, the compressor degradation residual vector at this time step is... Air cooler composite impedance residual vector Evaporator control deviation residual vector By concatenating and splicing the data, single-step mechanism deviation parameters are obtained. These parameters are then stacked row-by-row in chronological order to obtain a mechanism residual characterization sequence in the form of a two-dimensional matrix. .

8. The fault diagnosis method for a transcritical carbon dioxide skid-mounted refrigeration unit as described in claim 7, characterized in that: The enhanced graph attention network is specifically as follows: For each time step, the compressor, air cooler, electronic expansion valve, and evaporator are mapped to four nodes of the graph structure data. The average values ​​of the air cooler outlet temperature, air cooler outlet pressure, evaporator inlet temperature, and evaporator inlet pressure at that time step in the basic steady-state operating condition tensor are connected in series to construct the initial features of the electronic expansion valve node. The average compressor suction temperature, average compressor suction pressure, average compressor discharge temperature, average compressor discharge pressure, and average actual operating power of the compressor are compared with the compressor degradation residual vector at this time step. Perform serial splicing to construct the initial features of the compressor node; The average values ​​of the air cooler inlet temperature, inlet pressure, outlet temperature, outlet pressure, and ambient temperature are compared with the residual vector of the air cooler's overall impedance at that time step. Perform series splicing to construct the initial features of the air cooler nodes; The average values ​​of evaporator inlet temperature, evaporator inlet pressure, evaporator outlet temperature, and evaporator outlet pressure are compared with the residual vector of evaporator control deviation at that time step. Perform series splicing to construct the initial features of the evaporator nodes; The unidirectional physical flow direction of the refrigerant in the transcritical cycle is extracted. Based on the unidirectional physical flow direction, directed connection edges of the graph structure data are established. A first directed edge is established from the compressor node to the gas cooler node, a second directed edge is established from the gas cooler node to the electronic expansion valve node, a third directed edge is established from the electronic expansion valve node to the evaporator node, and a fourth directed edge is established from the evaporator node to the compressor node. The initial features of the electronic expansion valve node, the compressor node, the gas cooler node, and the evaporator node are combined into a graph node feature matrix. The first directed edge, the second directed edge, the third directed edge, and the fourth directed edge are combined into a graph adjacency matrix. The graph node feature matrix and the graph adjacency matrix are fused to construct the physical graph structure data at this time step. The physical graph structure data is input into the enhanced graph attention network layer to parse the graph node feature matrix and graph adjacency matrix in the physical graph structure data; for any target node in the graph adjacency matrix with directed connecting edges... Physically adjacent nodes Extract target nodes from the graph node feature matrix respectively. The corresponding initial feature vector of the target node Physically adjacent nodes The corresponding initial feature vector of the adjacent nodes Using the weight matrix Initial feature vector of the target node Initial feature vectors of adjacent nodes Spatial projection is performed, and the activation function of the linear unit with leakage correction is embedded after the projected features to calculate the dynamic attention coefficients between nodes. ; Using the normalized exponential function for the target node The dynamic attention coefficients of all physically adjacent nodes are globally normalized to generate dynamic attention weights. ; Based on dynamic attention weights Perform a weighted aggregation operation on the initial feature vectors of adjacent nodes, combined with a nonlinear activation function. Obtain updated target node features ; Finally, the N enhanced graph attention network layers are connected in series and stacked to construct a complete enhanced graph attention network.

9. A fault diagnosis method for a transcritical carbon dioxide skid-mounted refrigeration unit as described in claim 1 or 2, characterized in that: The specific process of S4 is as follows: First, the higher-order correlation state vector As core sensitive input features, these are fed into a pre-built random forest classifier; the random forest classifier is set to contain a number of... Independent decision trees; based on high-order correlation state vectors Based on the feature dimensions, a bootstrap sampling algorithm is used to perform random sampling with replacement to generate... Each training subset is a separate training subset, and an independent decision tree is constructed for each training subset; Secondly, for any independent decision tree in the random forest classifier, the high-order association state vector is... Input the root node of the independent decision tree, and perform feature splitting on each internal node according to the Gini impurity criterion; define the samples contained in the current node as belonging to the i-th... The probability of each hardware failure category is Calculate the Gini impurity of the node; based on the principle of minimizing Gini impurity, propagate the higher-order correlation state vector. Continue until the corresponding leaf node is reached, then output the single-tree fault prediction category of this independent decision tree; Again, summarizing all The single-tree fault prediction categories output by each independent decision tree are used to calculate the cumulative votes for each preset hardware fault classification label using an integrated voting mechanism of multiple decision trees; the first... Independent decision trees for the first The voting indicator variables for each hardware fault category are: When the predicted category matches the hardware failure category The value is 1 if it is not 1, otherwise the value is 0; calculate the first... Cumulative votes for each hardware failure category ; Finally, the cumulative vote count of all hardware fault categories is filtered using the maximum value function, and the target fault category with the highest cumulative vote count is extracted. A unique fault classification label pointing to evaporator scaling, air cooler blockage, or compressor valve leakage is output.

Images