Compressor temperature array anomaly prediction method based on twinborn iterative deduction

Abstract

The invention discloses a compressor temperature array anomaly prediction method based on twinborn iterative deduction. The method comprises the following steps: acquiring compressor data; constructing a single temperature parameter twinborn prediction model; full-temperature parameter twinborn iteration deduction is carried out; long-term trend prediction and abnormal accurate early warning of the multi-dimensional temperature array data of the compressor are achieved, the defects of a traditional method in the aspect of multi-temperature parameter collaborative prediction are overcome, and finally the remarkable technical effects of improving the operation reliability of the compressor and guaranteeing safe and efficient operation of industrial production are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of compressor operating state monitoring, and particularly to a method for predicting abnormal temperature arrays of compressors based on twin iterative deduction. Background Art

[0002] Compressors are widely used in petrochemical, refrigeration systems, and various industrial processes. Their operating stability is directly related to the continuity and safety of the overall production process. The temperature parameter is one of the important monitoring indicators of the compressor operating conditions. If the operating temperature is abnormal, it may not only lead to internal lubrication failure, thermal deformation of bearings and rotor materials, or a decrease in sealing performance of the compressor, but also cause phase changes (such as condensation or icing) of the internal gas medium, and in severe cases, result in a decrease in compressor operating efficiency, increased vibration, and even serious mechanical failures.

[0003] Currently, for compressor temperature monitoring, data collection is mostly carried out using single - type sensors such as thermal resistors and thermocouples. However, these single components are vulnerable to electromagnetic interference, working medium pollution, or mechanical vibration, suffering from problems of insufficient data reliability and redundancy. At the same time, traditional monitoring methods usually rely on discrete sensors to collect data, lacking a comprehensive analysis of the overall thermal state evolution trend of the compressor. Especially in the working conditions of multi - stage compressors, variable load, or variable - speed systems, it is impossible to effectively achieve early and accurate warning of abnormal temperature parameter trends, and it is difficult to meet the requirements of intelligent management of compressor operating states in modern industrial production.

[0004] CN115717590A discloses a method for detecting abnormal states of compressors by using PLC data and feature engineering to establish a reconstruction model. Although this method does not require additional sensors, its feature engineering depends on specific domain knowledge and lacks the prediction of multi - step evolution trends of temperature parameters, resulting in limited warning accuracy and lead time.

[0005] CN116181635A discloses a method for detecting anomalies in the exhaust temperature data of subway air - conditioning compressors using the isolation forest algorithm. Although this method has the advantage of unsupervised detection, it is still limited to the static feature analysis of a single temperature parameter, ignoring the co - variation characteristics between the overall temperature arrays of the compressor, and unable to effectively capture the associated abnormal trends of multi - dimensional temperature parameters under dynamic working conditions, affecting the accuracy and stability of fault warning.

[0006] Therefore, the existing methods for predicting compressor temperature anomalies have problems such as single monitoring methods, insufficient data redundancy, and poor multi - dimensional temperature data correlation analysis capabilities. The invention of our side adopts twin iterative deduction and recurrent neural network technologies to solve the problems of long - term accurate prediction of multi - dimensional temperature array data and early warning of abnormal trends of compressors under complex dynamic working conditions. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. In this section, as well as in the abstract and title of the specification of this application, some simplifications or omissions may be made to avoid obscuring the purpose of this section, the abstract, and the title of the specification, and such simplifications or omissions shall not be used to limit the scope of the present invention.

[0008] In view of the above existing problems, the present invention is proposed.

[0009] To solve the above technical problems, the present invention provides the following technical solutions: Real-time collect the operating condition parameters of the compressor and the first temperature parameter array to form a first state parameter vector at the current moment t; For each target temperature parameter in the first temperature parameter array, using the operating condition parameters at moment t and the first temperature parameter array as inputs, train a single-temperature-parameter twin prediction model for predicting the target temperature at moment t + 1 by using a recurrent neural network; Repeat the above step to establish corresponding single-temperature-parameter twin prediction models for all temperature parameters to form a temperature parameter model set; Using the first state parameter vector at the current moment t as an input, sequentially call the temperature parameter model set to obtain a second temperature parameter array regarding moment t + 1, and circularly iterate the second temperature parameter array and the current operating condition parameters N times to form a second state parameter vector to obtain a multi-step deduction result; Utilize a historical state matrix containing historical state parameter vectors at multiple moments, establish a transformation matrix through principal component analysis, perform dimensionality reduction and inverse transformation on the multi-step deduction result to obtain a deviation baseline for each temperature parameter; Calculate the deviation ratio of each temperature parameter in the multi-step deduction result relative to its deviation baseline. When the deviation ratio of any temperature parameter exceeds the warning threshold, output an abnormal warning signal for the compressor temperature array.

[0010] As a preferred solution of the compressor temperature array abnormal prediction method based on twin iterative deduction according to the present invention, the operating condition parameters at least include current, load percentage, and motor frequency; The first temperature parameter array includes the intake temperature, exhaust temperature, main bearing temperature, inlet bearing temperature of the male rotor, discharge end bearing temperature of the male rotor, thrust bearing temperature of the male rotor, inlet bearing temperature of the female rotor, discharge end bearing temperature of the female rotor, thrust bearing temperature of the female rotor, motor A / B / C phase winding temperatures, and the main motor drive end and non-drive end temperatures at the current moment t.

[0011] As a preferred solution of the compressor temperature array abnormal prediction method based on twin iterative deduction according to the present invention, the formation of the first state parameter vector at the current moment t includes: Normalizing the collected operating condition parameters and the first temperature parameter array to 0~1 respectively; The normalized parameters are synchronized and aligned with a unified timestamp and combined into a vector in chronological order. ; in, is the working condition parameter sub-vector, is the temperature parameter subvector, is the first state parameter vector at the current time t; The first state parameter vector is stored in a state buffer.

[0012] As a preferred solution of the compressor temperature array anomaly prediction method based on twin iterative deduction described in the present invention, the recurrent neural network has six layers, each layer contains 64 neurons, the batch size is 32, the drop rate is 0.1, the learning rate is 0.001, and the number of iterations is not less than 1000 times.

[0013] As a preferred solution of the compressor temperature array anomaly prediction method based on twin iterative deduction described in the present invention, the single temperature parameter twin prediction model for predicting the target temperature at time t+1 obtained by using recurrent neural network training includes: Build with As input, target temperature is the output supervised training sample set; The forward calculation and reverse gradient update are performed using the following formula as the comprehensive loss function:

[0014] Where M is the number of samples, is the normalized time variable, is the time weight attenuation coefficient, , is the regularization weight, , is the mapping matrix, b is the bias vector, and are the operating condition parameter sub-vector and temperature parameter sub-vector of the mth sample respectively, is the attention weight, is the hidden unit activation value, is an exponential linear unit, K is the hidden layer dimension, When it approaches 0, the prediction accuracy of the representation model is optimal. represents the normalized time variable Next The target time corresponding to the sample The actual single temperature parameter value, The predicted The single temperature parameter value of one sample at the normalized time variable Under the condition of

[0015] As a preferred solution of the compressor temperature array anomaly prediction method based on twin iterative deduction described in the present invention, the obtaining of the multi-step deduction result includes: Input the first state parameter vector into the temperature parameter model set by using a sliding time window to predict the second temperature parameter array at the next moment; Fuse the second temperature parameter array with the real-time working condition parameters to form a second state parameter vector as the next round of input; Repeat the above two steps and iterate cyclically N times until a sequence is generated, which is the multi-step deduction result.

[0016] As a preferred solution of the compressor temperature array anomaly prediction method based on twin iterative deduction described in the present invention, the number of times of cyclically iterating N times is 50.

[0017] As a preferred solution of the compressor temperature array anomaly prediction method based on twin iterative deduction described in the present invention, obtaining the deviation baseline of each temperature parameter includes: Normalize the historical state matrix to zero mean, where d is the dimension; Calculate the covariance matrix by using the principal component analysis method: and solve its eigenvalues and the corresponding eigenvectors as the principal components, where T represents the time length of the historical state matrix X, that is, the number of samples in X; Select the smallest integer k that satisfies the cumulative contribution rate of not less than 95% such that and construct a transformation matrix for the first k principal components; Based on the transformation matrix, that is, perform dimensionality reduction on the twin parameter state vector at the t + 1 moment to obtain the dimensionality-reduced principal component information matrix ; Based on the dimensionality-reduced data, perform dimensionality restoration on the dimensionality-reduced data by using the inverse matrix of the transformation matrix. Since information is lost, only the maximum consistent information of all parameters is retained. Therefore, the deviation baseline of each temperature parameter can be obtained through the inverse transformation , where is the inverse matrix of the transformation matrix; Calculate the time mean curve of each temperature parameter in the sequence obtained from the t + 1 to t + N moments in turn, which is defined as the corresponding deviation baseline.

[0018] As a preferred solution of the compressor temperature array anomaly prediction method based on twin iterative deduction described in the present invention, calculate the deviation ratio of each temperature parameter in the multi-step deduction result relative to its deviation baseline, and its calculation method is:

[0019] where is the predicted value of the i-th temperature parameter in the deduction step , is the corresponding deviation baseline value, is the deviation ratio; When , trigger the anomaly warning flag of this temperature parameter.

[0020] Advantages of the present invention: 1. By establishing an accurate and comprehensive basis for equipment operation status data, ensure the accuracy and integrity of the input data of the subsequent prediction model, thereby providing high-quality data guarantee for the prediction of the compressor operation status, and improving the prediction accuracy and reliability; 2. By independently and finely modeling each temperature parameter, improve the prediction accuracy of each individual temperature parameter, avoid the prediction deviation caused by the mutual interference between different parameters, effectively capture the dynamic change trend of a single temperature parameter, and enhance the temperature prediction sensitivity and model refinement; 3. By comprehensively and systematically predicting each temperature parameter, overcome the problem of insufficient data coverage in the traditional single-sensor or small number of sensors monitoring method, realize the in-depth analysis of the overall temperature condition of the compressor, and accurately monitor the overall thermal condition of the compressor; 4. Overcome the problem of prediction lag caused by the simple speculation of the current or short-term state in the prior art, discover potential anomalies in advance, and realize active warning and fault prevention; 5. Eliminate noise and redundant interference information in the data through dimensionality reduction, extract more representative features, thereby accurately capturing the precursors of anomalies, overcoming the false alarms or missed alarms that may be caused by directly using the original prediction data in the traditional method, improving the reliability of anomaly detection, and reducing false alarms and misjudgments; 6. Accurately quantify the degree of anomaly through the deviation ratio, avoid the low efficiency and uncertainty brought by the subjective experience judgment or static threshold method of the traditional monitoring system, improve the accuracy of anomaly determination, predict and prevent the compressor failure risk in advance, and ensure the equipment safety and production continuity. Description of the Drawings

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings. Among them: Figure 1 It is a schematic flow chart of the compressor temperature array anomaly prediction method based on twin iterative deduction shown in the present invention. Specific embodiments

[0022] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific embodiments of the present invention in conjunction with the drawings of the specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all embodiments.

[0023] Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0024] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0025] According to the embodiments of the present invention, in combination with Figure 1 Shown in the flow chart, a compressor temperature array anomaly prediction method based on twin iterative deduction specifically includes the following steps: S1. Real-time collect the operating condition parameters of the compressor and the first temperature parameter array to form the first state parameter vector at the current moment t; S2. For each target temperature parameter in the first temperature parameter array, using the operating condition parameters at moment t and the first temperature parameter array as inputs, train a single-temperature-parameter twin prediction model for predicting the target temperature at moment t + 1 using a recurrent neural network; S3. Repeat step S2 to establish corresponding single-temperature-parameter twin prediction models for all temperature parameters to form a temperature parameter model set; S4. Using the first state parameter vector at the current moment t as the input, sequentially call the temperature parameter model set to obtain the second temperature parameter array regarding moment t + 1, and cyclically iterate the second temperature parameter array and the current operating condition parameters N times to obtain the multi-step deduction result; S5. Using the historical state matrix containing the historical state parameter vectors at multiple moments, establish a transformation matrix through principal component analysis, perform dimensionality reduction and inverse transformation on the multi-step deduction results, and obtain the deviation baseline of each temperature parameter; S6. Calculate the deviation ratio of each temperature parameter in the multi-step deduction results relative to its deviation baseline. When the deviation ratio of any temperature parameter exceeds the warning threshold, output an abnormal warning signal for the compressor temperature array.

[0026] It should be noted that the collected compressor state parameter data includes operating condition parameters and temperature parameters. Taking the operating condition parameters and full temperature parameters at time t as inputs and the single temperature parameter at time t+1 as the output, a recurrent neural network is used to sequentially construct twin prediction models for single temperature parameters to form a set of temperature parameter twin prediction models; Furthermore, based on the current operating state of the compressor, cyclically call the twin prediction models of each temperature parameter to obtain the single-step deduction results of single temperature parameters, summarize the temperature parameter results of single-step deduction to obtain the temperature array of single-step deduction, and cyclically iterate and deduce according to a fixed number of iteration periods to obtain the temperature array of multi-step twin deduction; During the iterative deduction process, perform state deviation analysis on the deduced temperature array. Specifically, use principal component analysis based on a large amount of historical data to obtain the transformation matrix of the twin distribution law, transform the multi-step twin deduction temperature matrix of the current operating state of the compressor, and perform dimensionality restoration on the data after dimensionality reduction by the inverse matrix of the transformation matrix, retaining the main information to extract the main characteristics of the twin distribution law as the deviation baseline; calculate the difference between the deduced value of the temperature parameter in the twin state parameters and the deviation baseline. If the deviation ratio of this difference exceeds 5%, there is a large deviation in the deduction deviation of this temperature parameter, and there is a potential temperature warning risk.

[0027] Next, in combination with some preferred or optional examples of the present invention, the implementation process and / or effects of certain examples of the present invention will be described more specifically.

[0028] As an example, the operating condition parameters at least include current, load percentage, and motor frequency.

[0029] As an example, the first temperature parameter array includes the intake temperature, exhaust temperature, main bearing temperature, inlet bearing temperature of the male rotor, outlet bearing temperature of the male rotor, thrust bearing temperature of the male rotor, inlet bearing temperature of the female rotor, outlet bearing temperature of the female rotor, thrust bearing temperature of the female rotor, motor A / B / C phase winding temperatures, and the main motor drive end and non-drive end temperatures at the current moment t.

[0030] In an alternative embodiment, form the first state parameter vector at the current moment t, including: Perform 0-1 normalization on the collected operating condition parameters and the first temperature parameter array respectively; Synchronize and align the normalized parameters with a unified timestamp, and combine them into a vector in chronological order ; Wherein, is the sub-vector of operating condition parameters, is the sub-vector of temperature parameters, is the first state parameter vector at the current moment t; Store the first state parameter vector in the state buffer.

[0031] It should be noted that the recurrent neural network adopted in this embodiment has six layers, each layer contains 64 neurons, the batch size is 32, the dropout rate is 0.1, the learning rate is 0.001, and the number of iterations is not less than 1000 times.

[0032] In an alternative embodiment, a single-temperature-parameter twin prediction model for predicting the target temperature at time t + 1 is trained by using a recurrent neural network, including: Construct a supervised training sample set with as the input and the target temperature as the output; Perform forward calculation and backward gradient update through the following formula as the comprehensive loss function:

[0033] Wherein, M is the number of samples, is the normalized time variable, is the time weight decay coefficient, , is the regularization weight, , is the mapping matrix, b is the bias vector, and are respectively the sub-vector of operating condition parameters and the sub-vector of temperature parameters of the m-th sample, is the attention weight, is the hidden unit activation value, is the exponential linear unit, K is the hidden layer dimension, and when it approaches 0, it indicates that the model prediction accuracy is optimal, represents the normalized time variable at the -th actual single-temperature-parameter value corresponding to the target time represents the predicted single-temperature-parameter value of the -th sample at the normalized time variable ; And introduce an interval constraint mapping to the original output of the network:[[]]

[0034] Among them, σ is the Sigmoid function, and are respectively the minimum and maximum values of the j-th temperature parameter in the historical data; And introduce interval constraint mapping to update the comprehensive loss function, so that the model converges within the physically feasible temperature range.

[0035] It should be noted that the comprehensive loss function in this embodiment is the single-temperature-parameter twin prediction model, and the physically feasible temperature range is the industry standard range.

[0036] Furthermore, obtaining the multi-step deduction results includes: Input the first state parameter vector into the temperature parameter model set by using a sliding time window to predict the second temperature parameter array at the next moment; Fuse the second temperature parameter array with the real-time working condition parameters to form the second state parameter vector as the next-round input; Repeat the above two steps, iterate N times in a loop until the sequence is generated, which is the multi-step deduction result.

[0037] As an example, the number of iterations N is 50.

[0038] It should be noted that for the deduction results obtained from the single-temperature-parameter twin prediction model, the state parameters including each temperature should follow the distribution law of the state parameters. Therefore, it is necessary to perform deviation analysis on the deduced and predicted state vectors to identify whether there is a deterioration trend in the compressor temperature; For example, use the principal component analysis method to construct a deviation identification model based on the distribution law of state parameters. By extracting the main change directions in the data, project the high-dimensional data into a low-dimensional space to complete the extraction of the trend characteristics of the twin state parameters. After back-projecting each parameter, eliminate the fluctuation deviation; Based on the data after dimensionality reduction, perform dimensionality restoration on the data after dimensionality reduction by using the inverse matrix of the transformation matrix. Since information is lost, only retain the maximum consistent information of all parameters; For the results predicted at each time step, the difference between the predicted value and the deviation baseline can be analyzed; in this embodiment, taking the twin deduction results at the moment as an example, the deviation baseline of each twin temperature parameter can be obtained through inverse transformation.

[0039] Exemplarily, obtaining the deviation baseline of each temperature parameter includes: Normalize the historical state matrix to zero mean to eliminate the influence of dimension, where d is the dimension; Calculate the covariance matrix by the principal component analysis method: , and solve its eigenvalues and the corresponding eigenvectors as the principal components, where T represents the time length of the historical state matrix X, that is, the number of samples in X; Select the smallest integer k that satisfies the cumulative contribution rate of not less than 95% such that , and construct a transformation matrix for the first k principal components ; Based on the transformation matrix, that is, perform dimensionality reduction on the twin parameter state vector at time t + 1 to obtain the dimensionality-reduced principal component information matrix ; Based on the dimensionality-reduced data, perform dimensionality recovery on the dimensionality-reduced data by the inverse matrix of the transformation matrix. Since information is lost, only the maximum consistent information of all parameters is retained. Therefore, the deviation baseline of each temperature parameter can be obtained through inverse transformation , where is the inverse matrix of the transformation matrix; Calculate the time mean curve of each temperature parameter in the sequence obtained from time t + 1 to time t + N in turn, and define it as the corresponding deviation baseline.

[0040] It is not difficult to understand that principal component analysis is an unsupervised dimensionality reduction technique used to simplify high-dimensional data sets while retaining as much key information (variance) in the original data as possible. Its core goal is to transform the original features into a new set of independent (uncorrelated) features (principal components), and these new features are sorted according to the defined data variance size.

[0041] In this embodiment, the essence of the matrix transformation applying principal component analysis is: extracting core features by projecting onto the principal component space; using inverse transformation to restore dimensions to generate a low-noise baseline; establishing a dynamic and robust deviation baseline in a data-driven manner, which solves the bottleneck of traditional methods in multi-parameter collaborative analysis and dynamic working condition adaptability, thereby improving the accuracy and reliability of anomaly warning.

[0042] Furthermore, it should be noted that principal component analysis filters high-frequency noise and minor fluctuations by retaining the principal components with the largest variance (95% cumulative contribution rate), making the baseline focus on the stable laws of historical data and reducing false alarms; the baseline after inverse transformation is essentially the "average mode" of the historical normal state, which can reflect the inherent co-variation law of temperature parameters (such as the correlation between bearing temperature and load), rather than an isolated threshold; in the scenario of variable load / speed, traditional static thresholds are prone to failure, while the baseline obtained by principal component analysis can dynamically capture the joint distribution of multi-parameters and is more adaptable to complex working conditions.

[0043] It should be noted that the calculation method for calculating the deviation ratio of each temperature parameter in the multi-step deduction result relative to its deviation reference line is as follows:

[0044] Wherein, is the predicted value of the i-th temperature parameter in the deduction step , is the corresponding deviation reference line value, is the deviation ratio; When , an abnormal warning flag for this temperature parameter is triggered.

[0045] Furthermore, a dynamic warning threshold is set for the i-th temperature parameter:

[0046] Wherein, is the global minimum safety threshold, is the deviation amplification factor, is the rolling standard deviation within the most recent S deduction steps (S≥N) ; For each deduction step , a Boolean decision flag is generated:

[0047] Wherein, is the deviation ratio of the i-th temperature parameter in the deduction step , is the warning threshold dynamically calculated for the i-th temperature parameter; 1 indicates that the warning for this temperature parameter is triggered in this deduction step; 0 indicates that the warning is not triggered; Exemplarily, when the warning is triggered, it is immediately written into the alarm queue and a temperature array abnormal warning signal is sent to the compressor control system through the MQTT protocol. At the same time, the trigger timestamp, the overlimit parameter number and its value are recorded.

[0048] The foregoing principal component analysis and matrix conversion methods for the historical state matrix can be carried out by means and methods in the prior art and will not be elaborated herein.

[0049] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting abnormal conditions of a compressor temperature array based on twin iterative deduction, characterized in that Including: Collecting the operating condition parameters of the compressor and the first temperature parameter array in real time to form a first state parameter vector at the current moment t; For each target temperature parameter in the first temperature parameter array, using the operating condition parameters at moment t and the first temperature parameter array as inputs, training with a recurrent neural network to obtain a single-temperature-parameter twin prediction model for predicting the target temperature at moment t + 1; Repeating the above step to establish corresponding single-temperature-parameter twin prediction models for all temperature parameters to form a temperature parameter model set; Using the first state parameter vector at the current moment t as an input, sequentially calling the temperature parameter model set to obtain a second temperature parameter array regarding moment t + 1, and circularly iterating the second temperature parameter array and the current operating condition parameters to form a second state parameter vector N times to obtain a multi-step deduction result; Using a historical state matrix containing historical state parameter vectors at multiple moments, establishing a transformation matrix through principal component analysis, performing dimensionality reduction and inverse transformation on the multi-step deduction result to obtain a deviation baseline for each temperature parameter; Calculating the deviation ratio of each temperature parameter in the multi-step deduction result relative to its deviation baseline, and when the deviation ratio of any temperature parameter exceeds the warning threshold, outputting an abnormal warning signal for the compressor temperature array; 2. The compressor temperature array anomaly prediction method based on twin iterative deduction according to claim 1, characterized in that, The operating condition parameters at least include current, load percentage, and motor frequency; The first temperature parameter array includes the intake temperature, exhaust temperature, main bearing temperature, inlet bearing temperature of the male rotor, outlet bearing temperature of the male rotor, thrust bearing temperature of the male rotor, inlet bearing temperature of the female rotor, outlet bearing temperature of the female rotor, thrust bearing temperature of the female rotor, motor A / B / C phase winding temperatures, and main motor drive end and non-drive end temperatures at the current moment t; 3. The compressor temperature array anomaly prediction method based on twin iterative deduction according to claim 1 or 2, characterized in that The forming of the first state parameter vector at the current moment t includes: Performing 0~1 normalization on the collected operating condition parameters and the first temperature parameter array respectively; Synchronize and align the normalized parameters with a unified timestamp, and combine them into a vector in chronological order ; Among them, is the sub-vector of operating condition parameters, is the sub-vector of temperature parameters, is the first state parameter vector at the current moment t; Storing the first state parameter vector into a state buffer; 4. The compressor temperature array anomaly prediction method based on twin iterative deduction according to claim 1, characterized in that The recurrent neural network has six layers, each layer contains 64 neurons, the batch size is 32, the dropout rate is 0.1, the learning rate is 0.001, and the number of iterations is not less than 1000 times; 5. The method for predicting abnormal compressor temperature array based on twin iterative deduction according to claim 3, characterized in that, The training with a recurrent neural network to obtain a single-temperature-parameter twin prediction model for predicting the target temperature at moment t + 1 includes: Construct a supervised training sample set with as the input and the target temperature as the output; Performing forward calculation and backward gradient update using the following formula as the comprehensive loss function: ； where M is the number of samples, is the normalized time variable, is the time weight decay coefficient, , is the regularization weight, , is the mapping matrix, b is the bias vector, and are the sub-vector of operating condition parameters and the sub-vector of temperature parameters of the m-th sample respectively, is the attention weight, is the activation value of the hidden unit, is the exponential linear unit, K is the dimension of the hidden layer, and when it approaches 0, it indicates that the model prediction accuracy is optimal, represents the normalized time variable at the target time corresponding to the -th sample, represents the actual single temperature parameter value of the -th sample at the normalized time variable ​ 6. The compressor temperature array anomaly prediction method based on twin iterative deduction according to claim 5, wherein The obtaining of the multi-step deduction result includes: Using a sliding time window to input the first state parameter vector into the temperature parameter model set to predict the second temperature parameter array at the next moment; Fusing the second temperature parameter array and the real-time operating condition parameters to form a second state parameter vector as the next round of input; Repeat the above two steps and iterate N times in a loop until the sequence is generated, which is the result of the multi-step deduction.

7. The compressor temperature array anomaly prediction method based on twin iterative deduction according to claim 6, wherein, The number of times of circularly iterating N times is 50; 8. The method for predicting abnormal compressor temperature array based on twin iterative deduction according to claim 1, characterized in that Obtaining the deviation baseline for each temperature parameter includes: Normalize the historical state matrix to zero mean, where d is the dimension; Calculate the covariance matrix by the principal component analysis method: , and solve its eigenvalues and the corresponding eigenvectors are the principal components, where T represents the time length of the historical state matrix X, that is, the number of samples in X; Select the smallest integer k that satisfies the cumulative contribution rate of not less than 95% such that , construct a transformation matrix for the first k principal components ; Based on the transformation matrix, that is, for the twin parameter state vector at time t + 1 perform dimensionality reduction to obtain the principal component information matrix after dimensionality reduction ; Based on the data after dimensionality reduction, dimensionality restoration is performed on the data after dimensionality reduction by the inverse matrix of the transformation matrix. Since information is lost, only the maximum consistent information of all parameters is retained. Therefore, the deviation baseline of each temperature parameter can be obtained through inverse transformation , where is the inverse matrix of the transformation matrix; Sequentially calculating the time mean curve of each temperature parameter in the sequence obtained from moment t + 1 to moment t + N, which is defined as the corresponding deviation baseline; 9. The compressor temperature array anomaly prediction method based on twin iterative deduction according to claim 8, wherein Calculating the deviation ratio of each temperature parameter in the multi-step deduction result relative to its deviation baseline, and its calculation method is: ； in, is the temperature parameter in the deduction step The predicted value in is the corresponding deviation baseline value, is the deviation ratio; When an abnormal warning flag for this temperature parameter is triggered.

Images