Fault prediction method for refrigerating unit

Abstract

The invention relates to the technical field of fault prediction, and discloses a fault prediction method for a refrigerating unit, and the method comprises the steps: employing a trend weighted interpolation method and local trend correction based on a sine function to fill missing data; a multi-scale disturbance scoring index is constructed, deviations of variables in different time windows are quantified, asymmetric risk guidance prediction distribution is constructed based on disturbance scores, and directional distinguishing and risk self-adaptive adjustment of abnormal fluctuation are achieved; a dynamic structure is introduced to drive a state updating unit, and time dependence and risk evolution are captured; a joint optimization objective function is designed, and the prediction fitting degree and the disturbance sensitivity are balanced; a structural disturbance perception output fusion module is constructed, asymmetric disturbance characteristics are extracted through symmetric disturbance variable mapping, redundant dependence is suppressed in combination with a non-mutual information suppression attention mechanism, and periodic errors are dynamically corrected by adopting a frequency offset filtering fusion device; and finally, through fusion of the loss function, precise prediction and robust early warning of the fault of the refrigerating unit are realized.

Description

Technical Field

[0001] The present invention belongs to the field of fault prediction, and in particular relates to a fault prediction method for a refrigeration unit. Background Art

[0002] Refrigeration units are widely used in industrial production, food refrigeration, air conditioning and refrigeration, and other fields. They are important equipment to ensure production continuity and living comfort. However, during long-term operation, refrigeration units are prone to failure due to equipment aging, load changes, environmental factors and human operational errors. Once a failure occurs, it may lead to reduced refrigeration performance and increased energy consumption at the least, and may cause unit damage and production stoppage at the worst, resulting in economic losses and even safety hazards. Therefore, studying the fault prediction method of refrigeration units, realizing early fault warning of units, and timely maintenance intervention has important economic significance and social benefits.

[0003] Currently, many researchers are committed to using machine learning technology to predict refrigeration unit failures. The main methods used include support vector machines, random forests, neural networks and other models. Researchers usually construct machine learning prediction models based on the collected sensor data such as temperature, pressure, vibration, and current through feature extraction and selection. These methods rely on historical data for model training and can classify or regress the operating status of the unit to achieve prediction and early warning of potential failures.

[0004] Although the above methods have achieved certain results, the existing technology still has many shortcomings. First, most studies rely on single or static models, which fail to effectively consider the changes in the operating status of refrigeration units under complex dynamic conditions, resulting in insufficient prediction accuracy and generalization ability. Second, current machine learning models generally require a large amount of labeled historical fault data for training, but fault samples in actual industrial environments are scarce and unevenly distributed, making model training difficult. In addition, existing methods usually focus on short-term fault prediction and lack effective mechanisms for long-term trend fault risk estimation, which limits the effectiveness of prediction methods in actual industrial applications. Therefore, there is an urgent need to propose a more efficient, accurate and suitable method for refrigeration unit fault prediction in complex industrial environments. Summary of the Invention

[0005] The main purpose of the present invention is to provide a refrigeration unit fault prediction method, which aims to build a refrigeration unit fault prediction model. The refrigeration unit fault prediction uses a trend weighted interpolation method to fill in missing values, introduces a local trend correction term based on a sine function, models the change trend of the original data points before and after, and calculates the missing values; constructs a disturbance-driven asymmetric risk modeling, constructs a multi-scale disturbance scoring index, and quantifies the deviation of the refrigeration unit data set variables in different time windows; uses the disturbance score as a dynamic weight to construct an asymmetric risk-guided prediction distribution, realizes the directional distinction of abnormal fluctuations of variables and adaptive risk skew; introduces a dynamic structure-driven state update unit element, capturing time dependency and risk evolution process; designing a joint optimization objective function to comprehensively maximize the prediction distribution fit and the sensitivity of the disturbance response; constructing a structural disturbance perception output fusion module, which encodes the rising and falling disturbances of each input variable separately through a symmetric disturbance variable mapper to extract asymmetric disturbance features; utilizing the non-mutual information suppression attention mechanism, introducing a mutual information penalty term in the attention weight calculation to suppress redundant dependencies between variables; adopting a frequency offset filter fusion to perform periodic frequency correction on the prediction output and dynamically adjust the prediction result error; designing a joint optimization objective function for disturbance perception; and finally integrating the loss function to realize the fault prediction of the refrigeration unit.

[0006] To achieve the above objectives, the technical solution of the present invention is: a fault prediction method for a refrigeration unit, the method comprising: S1. Collect the dataset related to the refrigeration unit. The dataset consists of multiple data features and constitutes the original dataset. S2. Use trend-weighted interpolation to fill missing values. By introducing a local trend correction term based on the sine function, the changing trend of the original data points before and after is modeled and the missing values ​​are calculated. S3. Construct disturbance-driven asymmetric risk modeling and construct a multi-scale disturbance scoring index. Quantify the deviation of variables in the refrigeration unit dataset within different time windows to form a comprehensive score reflecting the local and global disturbance intensity. Use the disturbance score as a dynamic weight to construct an asymmetric risk-guided prediction distribution. By adjusting the weights of high- and low-risk quantiles, directional differentiation of abnormal variable fluctuations and adaptive risk skewness are achieved. A dynamic structure-driven state update unit is introduced to fuse current disturbance characteristics with historical state information to capture time dependence and risk evolution. Design a joint optimization objective function to comprehensively maximize the prediction distribution fit and the sensitivity of the disturbance response. S4. Construct a structural perturbation perception output fusion module. The structural perturbation perception output fusion module encodes the rising and falling perturbations of each input variable separately through a symmetric perturbation variable mapper to extract asymmetric perturbation features; utilizes a non-mutual information suppression attention mechanism, introduces a mutual information penalty term in the attention weight calculation, and suppresses redundant dependencies between variables; uses a frequency offset filter fusion to perform periodic frequency correction on the prediction output and dynamically adjust the prediction result error; and designs a joint optimization objective function for disturbance perception. S5. Construct a refrigeration unit fault prediction model. The refrigeration unit fault prediction model includes data preprocessing, disturbance-driven asymmetric risk modeling, structural disturbance perception output fusion module, and final loss function calculation and optimization. During the training process, a dynamic feedback adjustment mechanism and hyperparameter adjustment ensure that the model adaptively adjusts the weighting coefficients. Finally, the test set is input into the refrigeration unit fault prediction model, and the refrigeration unit failure probability is output to achieve refrigeration unit failure prediction.

[0007] Preferably, in step S2, the missing values ​​in the data are processed and the missing data in the original data set are filled using the trend weighted interpolation method. Missing values ​​at all times, using The data point before time and the subsequent data points Calculate the missing values, and the mathematical model of the trend weighted interpolation method is: ; Where, for The difference data of time, for Time data, for Time data, is the local trend adjustment factor, is a local trend correction term based on sinusoidal function modulation, and its mathematical model is: ; is the influence weight of the previous time point on the current missing value, and the mathematical model is: ; is the influence weight of the next time point on the current missing value, and the mathematical model is: .

[0008] Preferably, the trend weighted interpolation method introduces a local trend correction term based on the sine function to model the changing trend of the previous and subsequent original data points, and realizes dynamic adjustment in combination with the weight distribution of time and position; the trend weighted interpolation method not only retains the simplicity of traditional weighted interpolation, enhances the sensitivity to the trend of the original data, but also accurately restores the true value of the missing point in scenarios with slow changes, periodic fluctuations or sudden transitions. The trend weighted interpolation method significantly improves the quality of preprocessed data and improves the accuracy and robustness of the overall fault prediction system.

[0009] Preferably, in step S3, a disturbance-driven asymmetric risk modeling is constructed, a multi-scale disturbance scoring index is constructed, the deviation of the variables in the refrigeration unit data set in different time windows is quantified, and a comprehensive score reflecting the local and global disturbance intensity is formed; the disturbance score is used as a dynamic weight to construct an asymmetric risk-guided prediction distribution, and by adjusting the weights of high and low risk quantiles, the directional distinction of abnormal fluctuations of variables and adaptive risk skew are achieved; a dynamic structure-driven state update unit is introduced to fuse the current disturbance characteristics with historical state information to capture time dependence and risk evolution process; a joint optimization objective function is designed to comprehensively maximize the prediction distribution fit and the sensitivity of the disturbance response; the specific method is: S301, construct a multivariate input sequence, the training set input variable dimension is , the time length is , construct the input sequence: ; ; in, is the constructed input sequence, For the A monitoring change at all times The observed value of For the moment Combination input of training set input variables; S302. Construct a multi-scale perturbation scoring index to capture the relative degree of change of the target variable in multiple time windows and define the perturbation scale set , each is the span of the time window, defined at each scale The mathematical model of the local sliding deviation under φ is: ; Where, is the target variable value at the current moment, is the target variable value at the jth moment in the past, ranging from ts to t-1, s is the time step, is the arithmetic mean of the s observations before the current time t, It is the absolute deviation of the current value relative to the average value of the past s steps, reflecting the short-term, medium-term and long-term disturbance intensity. Then, by aggregating the multi-scale disturbance scores, the multi-scale disturbance score index is obtained. , The mathematical model is: ; Where, is the disturbance weighting factor, which controls the weight of each scale, ranging from 0 to 1, and satisfies , It is a multi-scale disturbance scoring index, which indicates the relative intensity of the disturbance at the current moment to the historical disturbance. S303, constructing an asymmetric risk-guided prediction distribution, wherein the perturbation scoring index reveals the variable trend and serves as a regulating factor for risk distribution modeling, the method comprising: The risk guidance factor is calculated through Sigmoid mapping, and the mathematical model of the risk guidance factor is: ; Where, To perturb the risk weight, it is used to control the offset direction of the asymmetric distribution. The sigmoid function is defined as , used to convert the disturbance scoring index Values ​​are mapped between 0 and 1 and are represented by states Construct low-high risk quantiles and achieve low-risk quantiles The mathematical model is: ; Where, is the state representation vector at time t, representing the historical disturbance comprehensive information. The state representation vector is calculated by the dynamic structure driven memory unit in step S304. is a symmetric weight matrix used to capture The covariation relationship between the dimensions in is the state representation vector The 2-norm of is used to quantify the overall state disturbance intensity, is the norm weight coefficient, which adjusts the influence of the norm term on the quantile. is the quantile bias term, which is used to determine the basic shift of the distribution center. is the low-risk quantile, which represents the lower bound of the prediction distribution and is used to construct an asymmetric prediction density function; Achieving high-risk quantiles The mathematical model is: ; Where, is a symmetric weight matrix, is the state representation vector The 1-norm of is the norm weight coefficient, is the quantile bias term, is the high-risk quantile, indicating the upper bound of the predicted distribution; when When the state is too concentrated, Suppressed by the two-norm, tends to be conservative, when Indicates that multiple variables are active. Pulled up by the 1 norm, focus on the upper tail prediction; finally predict the distribution model, its mathematical model is: ; Finally, a joint optimization objective function is constructed. The first part is the log-likelihood loss term, and the second part is the disturbance score response term. This enhances the model's ability to focus on the disturbance intensity. The risk adjustment coefficient is used to perform a weighted balance between the two parts. The mathematical model is: ; Where, is the risk adjustment coefficient, which is used to control the trade-off between prediction accuracy and abnormal sensitivity. is the overall loss function, is the set of all trainable parameters of the model, including the state update module, quantile prediction module, and perturbation score weight. Forecast distribution at the current moment in the true value The logarithmic probability density at ; Where, is the prediction target at time t The conditional probability density function of For is the lower tail truncated probability density of the benchmark, For is the upper tail truncated probability density of the benchmark; S304. By introducing a structural selection mechanism and a dynamic state fusion approach, we achieve modeling of long-term dependencies, sudden disturbance response, and dynamic memory adaptation during the operation of refrigeration unit equipment. We propose a state aggregation mechanism based on structural gating that automatically selects the optimal combination from multiple state update paths at each moment, enhancing the model's ability to represent complex patterns. The method includes: The input features, perturbation scores, and historical observations at the current time t are jointly encoded to form a structure-aware tensor , The mathematical model is: ; Where, is the target variable value at time step t-1, MLP is a multi-layer perceptron, which is used for nonlinear transformation and interaction modeling. It is a structure perception tensor that represents the structural response characteristics of the current state. Through the gating function softmax, the structure perception tensor is mapped to the weights of M structural subunits to calculate the generated structure selection weight. , The mathematical model is ; Where, is the structure mapping weight matrix, is the structure mapping bias vector, , and satisfies , each For the The contribution of each structural unit to the current state; then construct the structural sub-state update unit and introduce M independent state update fast response units , each used to model different types of state transitions, the output of each substructure is: ; Where, For the A new rapid response unit, is the transformation matrix of the input variables, is the transformation matrix of the historical target variable, is the bias term of the fast response unit, It is a fast response rate factor that controls the driving strength of the new input on the current state. The larger it is, the more sensitive the response is. The final state is a weighted fusion of the outputs of each substructure. Its mathematical model is: ; Where M is the number of state structure submodules.

[0010] Preferably, the disturbance-driven asymmetric risk modeling introduces a disturbance-driven mechanism, so that the refrigeration unit fault prediction model can extract the disturbance characteristics of key variables from multiple time scales, and use it as the core guiding signal of dynamic risk offset, thereby realizing adaptive control of the predicted distribution morphology; at the same time, with the help of a structure-aware state update unit, the refrigeration unit fault prediction model can continuously memorize the disturbance evolution process and model time dependence; finally, by jointly optimizing the objective function to collaboratively optimize the distribution fitting and disturbance sensitivity, the refrigeration unit fault prediction model can effectively improve its response capability to the abnormal trend of the refrigeration unit and the prediction accuracy of the tail risk, thereby enhancing its stability and robustness in a complex environment with multiple working conditions.

[0011] Preferably, in step S4, a structural disturbance perception output fusion module is constructed, and the structural disturbance perception output fusion module encodes the rising and falling disturbances of each input variable separately through a symmetric disturbance variable mapper to extract asymmetric disturbance features; utilizes a non-mutual information suppression attention mechanism, introduces a mutual information penalty term in the attention weight calculation, and suppresses redundant dependencies between variables; uses a frequency deviation filter fusion device to perform periodic frequency correction on the prediction output and dynamically adjusts the prediction result error; designs a joint optimization objective function of disturbance perception; and the specific method includes: S401. Construct a symmetric disturbance mapper. By establishing a symmetric disturbance response channel for each variable, a separable feature representation is constructed under disturbances in different directions. First, a forward disturbance response channel is constructed to simulate the activation effect of the signal on the system state when the variable of the refrigeration unit dataset increases. The mathematical model of the forward disturbance response channel is: ; Where, is the learnable forward perturbation response channel weight, is the forward disturbance response channel bias term, is the response mapping value of the current variable under the disturbance in the rising direction, that is, the rising response, and tanh is the activation function; then the reverse disturbance response channel is constructed to input data Take the negative and map the descending disturbance into the ascending dual signal, so that the positive and negative directions use the same structure. The mathematical model is: ; Where, is the learnable reverse perturbation response channel weight, is the inverse disturbance response channel bias term, is the disturbance response when the variable decreases, that is, the decrease response; after calculating the values ​​of the forward disturbance response channel and the reverse disturbance response channel, a symmetric difference enhancement term is constructed and its structural residual is calculated to capture the directional structural change. The calculation method is to make a difference between the forward disturbance response channel and the reverse disturbance response channel value to obtain , which measures the difference in response of the same variable under two directional perturbations. The absolute value is used to measure the asymmetry of the directional response amplitude. The perturbation expression is fused to construct the final embedding, and its mathematical model is: ; Where, is the disturbance difference weight factor, which is used to dynamically adjust the importance of the asymmetric difference term. The final perturbation embedding result of the variable; finally construct the multivariate perturbation embedding sequence ; S402, using the output of step S401 Construct query Q, key K, and value vector V. Their mathematical models are: ; ; ; Where, 、 、 They are the mapping matrices of query, key, and value, and finally the vector representation corresponding to each variable is obtained. 、 、 ; At the same time, the mutual information matrix between the variable pairs is calculated, and its mathematical model is: ; Where, For variables With variables The joint probability density of For variables With variables The mutual information matrix between them; then the mutual information suppression term is introduced into the constructed attention mechanism, and the mathematical model of the attention mechanism with the mutual information suppression term is: ; Where, is a constant, is a hyperparameter that controls the strength of mutual information suppression; The mutual information penalty imposed on highly redundant variables is calculated, and finally the variable-level attention weighted aggregation result is output. Its mathematical model is: ; Where n is the number of variables; weighted aggregation is performed on all variables to obtain the aggregate representation , this aggregated representation serves as the main path input to the frequency offset fuser in step S403; S403: Construct a frequency deviation filter fusion device to extract the dominant frequency by analyzing the frequency components of the main path prediction results; construct a frequency modulation function to periodically adjust the prediction output to achieve dynamic correction of the prediction frequency deviation error; specifically, the following steps are included: The main path output of the prediction model is defined as: ; Where, is the output vector of the self-attention module, is the weight of the linear mapping, is the bias of the linear mapping, is the prediction result without frequency offset adjustment; then the frequency of the signal is estimated using the prediction output window at the most recent moment, and its mathematical model is: ; Where, is the frequency analysis function, k is the frequency estimation window size, which is selected according to the sampling rate and period characteristics; then the modulation function based on the estimated frequency is constructed, and its mathematical model is: ; Where, is the modulation function value, is the phase offset; finally, the frequency modulation function is superimposed on the main path prediction as a dynamic adjustment item, and its mathematical model is: ; Where, The frequency compensation amplitude controls the strength of the adjustment item and is usually a learnable parameter. is the frequency offset correction prediction value after final fusion; S404: Construct a joint disturbance perception optimization target. The overall optimization target considers the balance between output prediction accuracy and disturbance compressibility. The optimized mathematical model is: ; Where, is the loss weight coefficient, which is used to balance the relative importance of prediction error and perturbation regularization term, T is the sequence length, is the loss value of the structured perturbation-aware output fusion module.

[0012] Preferably, the structural disturbance perception output fusion module effectively captures the asymmetric disturbance features existing in multi-variable inputs through symmetric disturbance variable mapping, and introduces a non-mutual information suppression attention mechanism to effectively suppress redundant information interference between input variables, avoiding the model's excessive reliance on highly correlated variables with limited predictive value, thereby enhancing the model's ability to focus on key driving variables. The frequency deviation filter fusion device effectively adjusts the frequency offset problem caused by the periodic start and stop and load switching of the refrigeration unit by dynamically adjusting the frequency components of the prediction results, thereby significantly improving the stability and reliability of the refrigeration unit's predicted output.

[0013] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: In the present invention, a fault prediction method for refrigeration units is proposed, which aims to construct a refrigeration unit fault prediction model. The refrigeration unit fault prediction uses a trend weighted interpolation method to fill in missing values, and by introducing a local trend correction term based on a sine function, the change trend of the original data points before and after is modeled, and the missing values ​​are calculated; a disturbance-driven asymmetric risk modeling is constructed, and a multi-scale disturbance scoring index is constructed to quantify the deviation of the variables in the refrigeration unit data set in different time windows; using the disturbance score as a dynamic weight, an asymmetric risk-guided prediction distribution is constructed to achieve directional differentiation of abnormal fluctuations of variables and adaptive risk skewness; a dynamic structure-driven state update unit is introduced to capture time dependence and risk evolution process; a joint optimization objective function is designed to comprehensively optimize the optimal Maximize the prediction distribution fitting and the sensitivity of disturbance response; construct a structural disturbance perception output fusion module, which encodes the rising and falling disturbances of each input variable separately through a symmetric disturbance variable mapper to extract asymmetric disturbance features; utilize the non-mutual information suppression attention mechanism, introduce the mutual information penalty term in the attention weight calculation, and suppress the redundant dependence between variables; use the frequency offset filter fusion device to perform periodic frequency correction on the prediction output and dynamically adjust the prediction result error; design a joint optimization objective function of disturbance perception; finally, fuse the loss function to realize the prediction of refrigeration unit faults; the refrigeration unit fault prediction model has improved the accuracy, robustness and early warning capability of refrigeration unit fault prediction as a whole, and has significant engineering application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The present invention is a flowchart of the steps of a fault prediction method for a refrigeration unit.

[0015] Figure 2 Diagram of the structure for modeling disturbance-driven asymmetric risk.

[0016] Figure 3 This is the structural diagram of the structured perturbation perception output fusion module.

[0017] Figure 4 This is the training structure diagram of the refrigeration unit fault prediction model.

[0018] Figure 5 This is the training graph for the refrigeration unit failure prediction model.

[0019] Figure 6 This is a comparison chart of the predicted value and the actual value of the refrigeration unit failure prediction model. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0021] See also Figure 1-Figure 5 The present invention provides a technical solution: a fault prediction method for a refrigeration unit, the method comprising the following steps: using a trend-weighted interpolation method to fill missing values, introducing a local trend correction term based on a sine function, modeling the change trend of the original data points before and after, and calculating the missing values; constructing a disturbance-driven asymmetric risk modeling, constructing a multi-scale disturbance scoring index, and quantifying the deviation of the variables in the refrigeration unit data set in different time windows; using the disturbance score as a dynamic weight, constructing an asymmetric risk-guided prediction distribution, realizing directional differentiation of abnormal fluctuations of variables and adaptive risk skewness; introducing a dynamic structure-driven state update unit to capture time dependence and risk evolution process; design a joint optimization objective function to comprehensively maximize the prediction distribution fitting and the sensitivity of the disturbance response; construct a structural disturbance perception output fusion module, which encodes the rising and falling disturbances of each input variable separately through a symmetric disturbance variable mapper to extract asymmetric disturbance features; utilize the non-mutual information suppression attention mechanism, introduce the mutual information penalty term in the attention weight calculation, and suppress the redundant dependence between variables; use the frequency deviation filter fusion to perform periodic frequency correction on the prediction output and dynamically adjust the prediction result error; design a joint optimization objective function for disturbance perception; and finally fuse the loss function to realize the prediction of refrigeration unit failure.

[0022] Please refer to Figure 1 As shown, a method for predicting a failure of a refrigeration unit in an embodiment of the present application comprises the following specific steps:

[0023] S1. Collect the data set related to the refrigeration unit. The data set consists of multiple data features and constitutes the original data set.

[0024] Furthermore, in step S1, the data features cover multi-dimensional data features during the operation of the refrigeration unit, mainly including: temperature, pressure, flow, liquid level and oil quality characteristics of the refrigerant and lubricating oil systems, which are used to monitor the working fluid status inside the unit; vibration, noise, bearing temperature and speed characteristics of the core components of the compressor, pump and fan, which are used to reflect the mechanical health status of the equipment; current, voltage, power, power factor, harmonic content, voltage fluctuation and power quality characteristics, which are used to identify electrical anomalies; ambient temperature, humidity and atmospheric pressure characteristics, which are used to evaluate the impact of the external environment on the performance of the unit; refrigerant leakage and unit air tightness characteristics, which are used for fault safety detection; the above multi-dimensional data can comprehensively and accurately reflect the operating status of the refrigeration unit, and multiple data features together constitute the original data set for refrigeration unit fault prediction.

[0025] S2. Preprocess the original data set, including missing value processing and normalization operations, and divide the preprocessed data into a training set and a test set. The specific steps are as follows: S201, process the missing values ​​in the data, use the trend weighted interpolation method to fill the missing data in the original data set, Missing values ​​at all times, using The data point before time and the subsequent data points Calculate the missing values, and the mathematical model of the trend weighted interpolation method is: ; Where, for The difference data of time, for Time data, for Time data, is the local trend adjustment factor, and the local trend adjustment factor is set to 0.2. is a local trend correction term based on sinusoidal function modulation, and its mathematical model is: ; is the influence weight of the previous time point on the current missing value, and the mathematical model is: ; is the influence weight of the next time point on the current missing value, and the mathematical model is: ; S202. Use the Min-Max normalization method to uniformly map each feature data to the interval [0,1] to eliminate dimensional differences. The mathematical model of Min-Max normalization is: ; Where, is the value to be normalized, for The maximum value in the column. for The minimum value in the column. for The normalized value of The value range is from 0 to 1; all values ​​in the original data are normalized to obtain the preprocessed data set; S203. The preprocessed dataset is divided into a training set and a test set according to the time series. The training set is used for model training, accounting for 70% of the dataset and having 7,000 samples. The test set is used for model performance verification, accounting for 30% of the dataset and having 3,000 samples.

[0026] S3. Construct disturbance-driven asymmetric risk modeling, construct multi-scale disturbance scoring indicators, quantify the deviations of refrigeration unit dataset variables in different time windows, and form a comprehensive score reflecting the local and global disturbance intensity; use disturbance scores as dynamic weights to construct asymmetric risk-guided prediction distribution, and achieve directional differentiation of abnormal variable fluctuations and adaptive risk skew by adjusting the weights of high and low risk quantiles; introduce a dynamic structure-driven state update unit to fuse current disturbance characteristics with historical state information to capture time dependence and risk evolution process; design a joint optimization objective function to comprehensively maximize the prediction distribution fit and the sensitivity of the disturbance response. The disturbance-driven asymmetric risk modeling structure is as follows: Figure 2 As shown, the specific steps are: S301, construct a multivariate input sequence, the training set input variable dimension is , the time length is , construct the input sequence: ; ; in, is the constructed input sequence, For the A monitoring change at all times The observed value of For the moment Combination input of training set input variables; S302. Construct a multi-scale perturbation scoring index to capture the relative degree of change of the target variable in multiple time windows and define the perturbation scale set , each is the span of the time window, and its value is , defined at each scale The mathematical model of the local sliding deviation under φ is: ; Where, is the target variable value at the current moment, is the target variable value at the jth moment in the past, ranging from ts to t-1, s is the time step, is the arithmetic mean of the s observations before the current time t, It is the absolute deviation of the current value relative to the average value of the past s steps, reflecting the short-term, medium-term and long-term disturbance intensity. Then, by aggregating the multi-scale disturbance scores, the multi-scale disturbance score index is obtained. , The mathematical model is: ; Where, is the disturbance weighting factor, which controls the weight of each scale, ranging from 0 to 1, and satisfies , It is a multi-scale disturbance scoring index, which indicates the relative intensity of the disturbance at the current moment to the historical disturbance. S303, constructing an asymmetric risk-guided prediction distribution, the disturbance score index not only reveals the variable trend, but also serves as a regulating factor for risk distribution modeling, the implementation method includes: The risk guidance factor is calculated through Sigmoid mapping, and the mathematical model of the risk guidance factor is: ; Where, To perturb the risk weight, it is used to control the offset direction of the asymmetric distribution. The sigmoid function is defined as , used to convert the disturbance scoring index Values ​​are mapped between 0 and 1 and are represented by states Construct low-high risk quantiles and achieve low-risk quantiles The mathematical model is: ; Where, is the state representation vector at time t, representing the historical disturbance comprehensive information. The state representation vector is calculated by the dynamic structure driven memory unit in step S304. is a symmetric weight matrix used to capture The covariation relationship between the dimensions in is the state representation vector The 2-norm of is used to quantify the overall state disturbance intensity, is the norm weight coefficient, which adjusts the influence of the norm term on the quantile. is the quantile bias term, which is used to determine the basic shift of the distribution center. is the low-risk quantile, which represents the lower bound of the prediction distribution and is used to construct an asymmetric prediction density function; Achieving high-risk quantiles The mathematical model is: ; Where, is a symmetric weight matrix, is the state representation vector The 1-norm of is the norm weight coefficient, is the quantile bias term, is the high-risk quantile, indicating the upper bound of the predicted distribution; when When the state is too concentrated, Suppressed by the two-norm, tends to be conservative, when Indicates that multiple variables are active. Pulled up by the 1 norm, focus on the upper tail prediction; finally predict the distribution model, its mathematical model is: ; Finally, a joint optimization objective function is constructed. The first part is the log-likelihood loss term, and the second part is the disturbance score response term. This enhances the model's ability to focus on the disturbance intensity. The risk adjustment coefficient is used to perform a weighted balance between the two parts. The mathematical model is: ; Where, is the risk adjustment coefficient, which is used to control the trade-off between prediction accuracy and abnormal sensitivity, and its value range is 0 to 1. is the overall loss function, is the set of all trainable parameters of the model, including the state update module, quantile prediction module, and perturbation score weight. Forecast distribution at the current moment in the true value The logarithmic probability density at ; Where, is the prediction target at time t The conditional probability density function of For is the lower tail truncated probability density of the benchmark, For is the upper tail truncated probability density of the benchmark; S304. By introducing a structural selection mechanism and dynamic state fusion, we achieve modeling of long-term dependencies, sudden disturbance response, and dynamic memory adaptation during the operation of refrigeration unit equipment. We propose a state aggregation mechanism based on structural gating, which automatically selects the optimal combination from multiple state update paths at each moment, enhancing the model's ability to represent complex patterns. Specific methods include: The input features, perturbation scores, and historical observations at the current time t are jointly encoded to form a structure-aware tensor , The mathematical model is: ; Where, is the target variable value at time step t-1, MLP is a multi-layer perceptron, which is used for nonlinear transformation and interaction modeling. It is a structure perception tensor that represents the structural response characteristics of the current state. Through the gating function softmax, the structure perception tensor is mapped to the weights of M structural subunits to calculate the generated structure selection weight. , The mathematical model is ; Where, is the structure mapping weight matrix, is the structure mapping bias vector, , and satisfies , each For the The contribution of each structural unit to the current state; then construct the structural sub-state update unit and introduce M independent state update fast response units , each used to model different types of state transitions, the output of each substructure is: ; Where, For the A new rapid response unit, is the transformation matrix of the input variables, is the transformation matrix of the historical target variable, is the bias term of the fast response unit, It is a fast response rate factor that controls the driving strength of the new input on the current state. The larger it is, the more sensitive the response is. Its value range is 0 to 1. The final state is a weighted fusion of the outputs of each substructure. Its mathematical model is: ; Where M is the number of state structure submodules.

[0027] S4. Construct a structural disturbance perception output fusion module. The structural disturbance perception output fusion module encodes the rising and falling disturbances of each input variable separately through a symmetric disturbance variable mapper to extract asymmetric disturbance features; uses the non-mutual information to suppress the attention mechanism, introduces the mutual information penalty term in the attention weight calculation, and suppresses the redundant dependence between variables; uses the frequency deviation filter fusion device to perform periodic frequency correction on the prediction output and dynamically adjust the prediction result error; designs the joint optimization objective function of disturbance perception. The structure of the structural disturbance perception output fusion module is as follows: Figure 3 As shown, the implementation steps include: S401. Construct a symmetric disturbance mapper. By establishing a symmetric disturbance response channel for each variable, a separable feature representation is constructed under disturbances in different directions. First, a forward disturbance response channel is constructed to simulate the activation effect of the signal on the system state when the variable of the refrigeration unit dataset increases. The mathematical model of the forward disturbance response channel is: ; Where, is the learnable forward perturbation response channel weight, is the forward disturbance response channel bias term, is the response mapping value of the current variable under the disturbance in the rising direction, that is, the rising response, and tanh is the activation function; then the reverse disturbance response channel is constructed to input data Take the negative and map the descending disturbance into the ascending dual signal, so that the positive and negative directions use the same structure. The mathematical model is: ; Where, is the learnable reverse perturbation response channel weight, is the inverse disturbance response channel bias term, is the disturbance response when the variable decreases, that is, the decrease response; after calculating the values ​​of the forward disturbance response channel and the reverse disturbance response channel, a symmetric difference enhancement term is constructed and its structural residual is calculated to capture the directional structural change. The calculation method is to make a difference between the forward disturbance response channel and the reverse disturbance response channel value to obtain , which measures the difference in response of the same variable under two directional perturbations. The absolute value is used to measure the asymmetry of the directional response amplitude. The perturbation expression is fused to construct the final embedding, and its mathematical model is: ; Where, is the disturbance difference weight factor, which is used to dynamically adjust the importance of the asymmetric difference term. The final perturbation embedding result of the variable; finally construct the multivariate perturbation embedding sequence ; S402, using the output of step S401 Construct query Q, key K, and value vector V. Their mathematical models are: ; ; ; Where, 、 、 They are the mapping matrices of query, key, and value, and finally the vector representation corresponding to each variable is obtained. 、 、 ; At the same time, the mutual information matrix between the variable pairs is calculated, and its mathematical model is: ; Where, For variables With variables The joint probability density of For variables With variables The mutual information matrix between them; then the mutual information suppression term is introduced into the constructed attention mechanism, and the mathematical model of the attention mechanism with the mutual information suppression term is: ; Where, is a constant, with an initial value of 64. is a hyperparameter that controls the strength of mutual information suppression and takes a value of 0.5; The mutual information penalty imposed on highly redundant variables is calculated, and finally the variable-level attention weighted aggregation result is output. Its mathematical model is: ; Where n is the number of variables; weighted aggregation is performed on all variables to obtain the aggregate representation , this aggregated representation serves as the main path input to the frequency offset fuser in step S403; S403: Construct a frequency deviation filter fusion device to extract the dominant frequency by analyzing the frequency components of the main path prediction results; construct a frequency modulation function to periodically adjust the prediction output; implement dynamic correction of the prediction frequency deviation error, and improve the adaptability and robustness of the model to periodic disturbances. Specifically, the process includes: The main path output of the prediction model is defined as: ; Where, is the output vector of the self-attention module, is the weight of the linear mapping, is the bias of the linear mapping, is the prediction result without frequency offset adjustment; then the frequency of the signal is estimated using the prediction output window at the most recent moment, and its mathematical model is: ; Where, is the frequency analysis function, k is the frequency estimation window size, which is selected according to the sampling rate and period characteristics, and the initial value is 3; then the modulation function based on the estimated frequency is constructed, and its mathematical model is: ; Where, is the modulation function value, is the phase offset, which is a fixed value of 0.1. Finally, the frequency modulation function is superimposed on the main path prediction as a dynamic adjustment item. Its mathematical model is: ; Where, The frequency compensation amplitude controls the strength of the adjustment item and is usually a learnable parameter. is the frequency offset correction prediction value after final fusion; S404: Construct a joint disturbance perception optimization target. The overall optimization target considers the balance between output prediction accuracy and disturbance compressibility. The optimized mathematical model is: ; Where, is the loss weight coefficient, which is used to balance the relative importance of prediction error and perturbation regularization term, T is the sequence length, is the loss value of the structured perturbation-aware output fusion module.

[0028] S5. Construct a refrigeration unit fault prediction model. The refrigeration unit fault prediction model runs from data preprocessing, disturbance-driven asymmetric risk modeling, structural disturbance perception output fusion module to the final loss function calculation and optimization. The refrigeration unit fault prediction model runs on the Linux system and uses the PyTorch deep learning framework to build the model. During the training process, the dynamic feedback adjustment mechanism and hyperparameter adjustment ensure that the model adaptively adjusts the weighting coefficient. Finally, the test set is input into the refrigeration unit fault prediction model, and the refrigeration unit failure probability is output to realize the prediction of the refrigeration unit failure. The training structure of the refrigeration unit fault prediction model is as follows: Figure 4 As shown, the specific implementation method is: S501. During the training process, the dynamic feedback adjustment mechanism is the back propagation mechanism, which adjusts the network parameters in reverse by adjusting the loss function value. The hyperparameter settings include: the learning rate is set to 0.0001, the batch size batch_size is set to 64, the optimizer parameter is Adam, and the training round Epochs is 200. Each time the data is loaded from the training set, the forward propagation is performed to calculate the loss and the back propagation is performed to update the weight. The data of each time step is input into the model, and the prediction is carried out in turn through the disturbance-driven asymmetric risk modeling and structural disturbance perception output fusion modules, and the loss function of the disturbance-driven asymmetric risk modeling is fused. Loss function of the structure perturbation perception output fusion module , construct a weighted joint loss function so that the training process takes into account the prediction accuracy of both parts. Its mathematical model is: ; Where, is the weight coefficient used to adjust the loss function With loss function The adjustment weight is 0.6. is the loss function of the refrigeration unit failure prediction model; Figure 5 The figure shows the training diagram of the refrigeration unit fault prediction model. It can be seen from the figure that as the number of iterations Epochs increases, the accuracy of the model continues to improve. The value is constantly decreasing, proving that the accuracy of the prediction model is constantly improving; S502: After using the training set to train the refrigeration unit fault prediction model in step S501, the trained model is verified and the test set is input into the trained refrigeration unit fault prediction model. Figure 6 As shown in the figure, the horizontal axis is the time step, the vertical axis is the failure rate of the refrigeration unit, the black solid point is the actual failure rate value, and the black cross point is the predicted value of the refrigeration unit failure prediction model. It can be seen from the figure that the failure prediction value proposed in this application is relatively close to the true value, the general trend is consistent and the accuracy is high. The experimental results show that the refrigeration unit failure prediction model can predict the failure rate of the refrigeration unit.

Claims

1. A method for predicting failure of a refrigeration unit, characterized in that: The following steps are involved: S1. Collect the dataset related to the refrigeration unit. The dataset consists of multiple data features and constitutes the original dataset. S2. Use trend-weighted interpolation to fill missing values. By introducing a local trend correction term based on the sine function, the changing trend of the original data points before and after is modeled and the missing values ​​are calculated. S3. Construct a disturbance-driven asymmetric risk model and construct a multi-scale disturbance scoring index. Quantify the deviation of variables in the refrigeration unit dataset within different time windows to form a comprehensive score reflecting the local and global disturbance intensity. Use the disturbance score as a dynamic weight to construct an asymmetric risk-guided predictive distribution. By adjusting the weights of high- and low-risk quantiles, directional differentiation of abnormal variable fluctuations and adaptive risk skewness are achieved. Introduce a dynamic structure-driven state update unit to fuse current disturbance characteristics with historical state information to capture time dependence and risk evolution. S4. Construct a structural perturbation perception output fusion module. The structural perturbation perception output fusion module encodes the rising and falling perturbations of each input variable separately through a symmetric perturbation variable mapper to extract asymmetric perturbation features; utilizes a non-mutual information suppression attention mechanism, introduces a mutual information penalty term in the attention weight calculation, and suppresses redundant dependencies between variables; and uses a frequency offset filter fusion to perform periodic frequency correction on the prediction output and dynamically adjust the prediction result error; S5. Construct a refrigeration unit fault prediction model. The refrigeration unit fault prediction model includes data preprocessing, disturbance-driven asymmetric risk modeling, structural disturbance perception output fusion module, and final loss function calculation and optimization. During the training process, a dynamic feedback adjustment mechanism and hyperparameter adjustment ensure that the model adaptively adjusts the weighting coefficients. Finally, the test set is input into the refrigeration unit fault prediction model, and the refrigeration unit failure probability is output to achieve refrigeration unit failure prediction.

2. A refrigeration unit fault prediction method according to claim 1, characterized in that: In step S1, the data features cover multi-dimensional data features during the operation of the refrigeration unit, mainly including: temperature, pressure, flow, liquid level and oil quality characteristics of the refrigerant and lubricating oil systems, which are used to monitor the working fluid status inside the unit; vibration, noise, bearing temperature and speed characteristics of the core components of the compressor, pump and fan, which are used to reflect the mechanical health status of the equipment; current, voltage, power, power factor, harmonic content, voltage fluctuation and power quality characteristics, which are used to identify electrical anomalies; ambient temperature, humidity and atmospheric pressure characteristics, which are used to evaluate the impact of the external environment on the performance of the unit; refrigerant leakage and unit air tightness characteristics, which are used for fault safety detection.

3. The method for predicting failure of a refrigeration unit according to claim 2, characterized in that: In step S2, the missing values ​​in the data are processed and the trend weighted interpolation method is used to fill the missing data in the original data set. Missing values ​​at all times, using The data point before time and the subsequent data points Calculate the missing values, and the mathematical model of the trend weighted interpolation method is: ； Where, for The difference data of time, for Time data, for Time data, is the local trend adjustment factor, is a local trend correction term based on sinusoidal function modulation, and its mathematical model is: ; The weight of the impact of the previous time point on the current missing value is calculated based on the ratio of the time interval between the current missing time point and the next time point to the total time interval. The weight of the impact of the next time point on the current missing value is calculated based on the ratio of the time interval between the current missing time point and the previous time point to the total time interval.

4. A refrigeration unit fault prediction method according to claim 3, characterized in that: In step S3, a multi-scale perturbation scoring index is constructed to capture the relative degree of change of the target variable in multiple time windows and define the perturbation scale set , each is the span of the time window, defined at each scale The mathematical model of the local sliding deviation under φ is: ； Where, is the target variable value at the current moment, is the target variable value at the jth moment in the past, ranging from ts to t-1, s is the time step, is the arithmetic mean of the s observations before the current time t, is the absolute deviation of the current value relative to the average value of the past s steps, reflecting the short-term, medium-term and long-term disturbance intensity, and then combined with the weighting factors of each scale Aggregate multi-scale perturbation scoring metric , which indicates the relative historical disturbance intensity at the current moment.

5. The method for predicting failure of a refrigeration unit according to claim 4, characterized in that: In step S3, an asymmetric risk-guided prediction distribution is constructed, and the perturbation score index reveals the variable trend as a regulating factor for risk distribution modeling. The method includes: The risk guidance factor is calculated by Sigmoid mapping , which is used to control the offset direction of the asymmetric distribution and is represented by the state Construct low-high risk quantiles; achieve low-risk quantiles The mathematical model is: ； Where, is the state representation vector at time t, representing the comprehensive information of historical disturbances, and the state representation vector is calculated by the dynamic structure driven memory unit in claim 6, is a symmetric weight matrix, capturing The covariation relationship between the dimensions in is the state representation vector The 2-norm of is the norm weight coefficient, which adjusts the influence of the norm term on the quantile. is the quantile bias term, is the low-risk quantile, indicating the lower bound of the predicted distribution; Achieving high-risk quantiles The mathematical model is: ； Where, is a symmetric weight matrix, is the state representation vector The 1-norm of is the norm weight coefficient, is the quantile bias term, is the high-risk quantile, indicating the upper bound of the predicted distribution; when When the state is too concentrated, Suppressed by the two-norm, tends to be conservative, when Indicates that multiple variables are active. Pulled up by the 1 norm, focus on the upper tail prediction; finally predict the distribution model, its mathematical model is: ； Finally, a joint optimization objective function is constructed. The first part is the log-likelihood loss term, and the second part is the perturbation score response term. The two parts are weighted and balanced by the risk adjustment coefficient. The mathematical model is: ； Where, is the risk adjustment factor, is the overall loss function, is the set of all trainable parameters of the model, including the state update module, quantile prediction module, and perturbation score weight. Forecast distribution at the current moment in the true value The logarithmic probability density at ; Where, is the prediction target at time t The conditional probability density function of For is the lower tail truncated probability density of the benchmark, For is the upper tail truncated probability density of the benchmark.

6. A refrigeration unit fault prediction method according to claim 5, characterized in that: In step S3, a state aggregation mechanism based on structural gating is proposed. By introducing a structural selection mechanism and a dynamic state fusion method, the optimal combination is automatically selected from multiple state update paths at each moment, thereby enhancing the model's ability to represent complex patterns. The method includes: The input features and perturbation score indicators of the current time t , historical observations Joint encoding to form a structure-aware tensor To represent the structural response characteristics of the current state; through the gating function softmax, the structure perception tensor is mapped to the weights of M structural subunits to calculate the generated structure selection weight , It can be expressed as: , and satisfies , each For the The contribution of each structural unit to the current state; then construct the structural sub-state update unit and introduce M independent state update fast response units , each used to model different types of state transitions, the output of each substructure is: ; Where, For the A new rapid response unit, is the transformation matrix of the input variables, is the transformation matrix of the historical target variable, is the bias term of the fast response unit, is a fast response rate factor that controls the driving strength of the new input to the current state. The larger the factor, the more sensitive the response. The final state is determined by the weighted output of each substructure. Weighted fusion calculation .

7. A refrigeration unit fault prediction method according to claim 6, characterized in that: In step S4, a symmetric disturbance mapper is constructed, and a symmetric disturbance response channel is established for each variable to construct a separable feature representation under disturbances in different directions. First, a forward disturbance response channel is constructed to input data. After adding bias through weighted linear mapping, the hyperbolic tangent function tanh is used for nonlinear activation, and the output reflects the response intensity corresponding to the variable during the increase process, which is used to simulate when the variable of the refrigeration unit dataset increases; Then build the reverse disturbance response channel to input data Take the negative value to map the disturbance in the descending direction to the dual signal in the ascending direction, and use the same structure as the positive and negative directions; after calculating the values ​​of the forward disturbance response channel and the reverse disturbance response channel, construct a symmetric difference enhancement term and calculate its structural residual to capture the directional structural change. The calculation method is to make a difference between the values ​​of the forward disturbance response channel and the reverse disturbance response channel to obtain , which measures the difference in response of the same variable under two directional perturbations. The absolute value is used to measure the asymmetry of the directional response amplitude. The perturbation expression is fused to construct the final embedding, and its mathematical model is: ； Where, is the disturbance difference weight factor, which is used to dynamically adjust the importance of the asymmetric difference term. The final perturbation embedding result of the variable; Finally, construct a multivariate perturbation embedding sequence .

8. The method for predicting failure of a refrigeration unit according to claim 7, characterized in that: In step S4, the output of claim 7 is used With the mapping matrix 、 、 Perform dot product operations to construct query Q, key K, and value vector V, and calculate the vector representation corresponding to each variable 、 、 ; At the same time, the mutual information matrix between the variable pairs is calculated, and its mathematical model is: ； Where, For variables With variables The joint probability density of For variables With variables The mutual information matrix between Then the mutual information suppression term is introduced into the constructed attention mechanism , the mathematical model of the attention mechanism that introduces the mutual information suppression term is: ； Where, is a constant, is a hyperparameter that controls the strength of mutual information suppression. is the output of the attention mechanism with the mutual information suppression term added, The mutual information penalty imposed by the operation on highly redundant variables; the adjusted attention weights are applied to the corresponding value vectors, and the weighted representations of all variables are aggregated to form a fused variable-level feature representation , aggregate n variables and get the aggregate representation , the aggregated representation serves as the main path input to the frequency offset fuser.

9. The method for predicting failure of a refrigeration unit according to claim 8, characterized in that: In step S4, a frequency deviation filter fusion device is constructed to extract the dominant frequency by analyzing the frequency components of the main path prediction result; Construct a frequency modulation function to periodically adjust the predicted output and dynamically correct the predicted frequency deviation error. This includes: By extracting the comprehensive representation vector from the attention module Perform a linear transformation on the learnable weight vector and add a bias term to obtain the prediction result without frequency bias adjustment ; Then, the frequency of the signal is estimated using the predicted output window at the most recent moment. The mathematical model is: ； Where, is the frequency analysis function, k is the frequency estimation window size, which is selected according to the sampling rate and period characteristics; then the modulation function based on the estimated frequency is constructed, and its mathematical model is: ； Where, is the modulation function value, is the phase offset; the frequency modulation function is then superimposed on the main path prediction as a dynamic adjustment item, where the amplitude of the adjustment item is dynamically determined by the learnable control factor, and finally the fused frequency offset correction prediction value is obtained ; Construct a joint disturbance perception optimization objective. The overall optimization objective considers the balance between output prediction accuracy and disturbance compressibility. The mathematical model for optimization is: ； Where, is the loss weight coefficient, which is used to balance the relative importance of prediction error and perturbation regularization term, T is the sequence length, is the loss value of the structured perturbation-aware output fusion module.

Images