Cold chain equipment fault detection method and system based on sensor

Abstract

The invention provides a sensor-based cold chain equipment fault detection method and system. Belongs to the technical field of cold chain equipment fault detection. The method comprises the following steps: collecting multi-dimensional data of the operation state of the cold chain equipment in real time through a multi-dimensional sensor; performing dimension reduction processing on the standardized multi-dimensional data set by adopting a principal component analysis algorithm, extracting main feature components in the data, and constructing a high-dimensional feature vector set; performing time sequence feature extraction on the multi-dimensional data set by using an LSTM model to obtain a multi-dimensional time sequence feature matrix; fusing the high-dimensional feature vector set with the multi-dimensional time sequence feature matrix to obtain a fused feature matrix; a method of combining principal component analysis and a long short-term memory network is adopted, main features and time sequence features of data are extracted respectively, and the main features and the time sequence features are organically combined through a feature fusion strategy, so that the operation state of the cold chain equipment is described more comprehensively, and the accuracy of fault feature extraction is improved.

Description

TECHNICAL FIELD

[0001] The application provides a sensor-based cold chain equipment fault detection method and system, and belongs to the technical field of cold chain equipment fault detection. BACKGROUND

[0002] Cold chain equipment plays a crucial role in the storage and transportation of food, medicine and other industries, and its stability directly affects the quality and safety of products. However, the running environment of cold chain equipment is complex, and various faults may occur, such as refrigeration system failure, electrical system failure, etc. Traditional cold chain equipment fault detection methods mainly rely on manual inspection and regular maintenance, which is not only inefficient, but also difficult to find potential faults in real time and accurately, which may lead to the expansion of equipment failure and cause serious losses. Therefore, it is of great practical significance to develop an efficient and accurate sensor-based cold chain equipment fault detection method. SUMMARY

[0003] The application provides a sensor-based cold chain equipment fault detection method and system to solve the problems mentioned in the background art.

[0004] The sensor-based cold chain equipment fault detection method provided by the application comprises the following steps:

[0005] S1, a multi-dimensional sensor is used to collect multi-dimensional data of the running state of the cold chain equipment in real time; the collected multi-dimensional data is cleaned and normalized to obtain a standardized multi-dimensional data set;

[0006] S2, a principal component analysis algorithm is used to reduce the dimension of the standardized multi-dimensional data set, extract the main feature components of the data, and construct a high-dimensional feature vector set; an LSTM model is used to extract time sequence features from the multi-dimensional data set to obtain a multi-dimensional time sequence feature matrix; the high-dimensional feature vector set and the multi-dimensional time sequence feature matrix are fused to obtain a fusion feature matrix;

[0007] S3, historical fault case data of the cold chain equipment is obtained, the fusion feature matrix is divided into feature subsets, the correlation between different feature subsets and the historical fault case data is calculated using a correlation analysis method to obtain a feature correlation matrix; according to the feature correlation matrix, the feature subsets related to the fault are screened out, the related feature subsets are further analyzed, the fault feature segments are extracted, and the fault feature segments are reconstructed into a fault feature vector using a feature recombination technology;

[0008] S4, based on historical failure case data, a failure mode feature knowledge base is constructed, each failure mode is described in detail, a support vector machine algorithm is used to classify and identify the failure feature vector, and a candidate failure mode list is obtained; based on a fuzzy comprehensive evaluation method, a failure similarity vector is generated according to the feature matching degree of the failure feature vector and each failure mode in the failure mode feature knowledge base;

[0009] S5, the running environment data of the cold chain equipment is acquired, and a failure correlation factor dataset is constructed; the occurrence probability of the candidate failure mode is preliminarily evaluated by using a Bayesian network model in combination with historical failure case data and the failure correlation factor dataset, and an initial failure mode probability vector is obtained; a dynamic weight adjustment method is used to dynamically correct the initial failure mode probability vector according to the real-time change of the failure correlation factor dataset, and failure probability distribution data is obtained;

[0010] S6, the candidate failure mode list is reordered according to the failure probability distribution data, and a failure mode ordering list based on failure probability is obtained; the internal structure topology data of the cold chain equipment is acquired, the failure mode ordering list and the internal structure topology data of the cold chain equipment are used to perform failure mode loop correlation analysis by using a graph theory algorithm, and a failure mode loop correlation table is obtained; the failure area positioning result data is obtained according to the failure mode loop correlation table and the failure mode ordering list; the failure cause is analyzed in depth in combination with the failure area positioning result data and the failure mode feature knowledge base, and a cold chain equipment failure cause analysis report is generated.

[0011] The sensor-based cold chain equipment failure detection system provided by the application comprises a memory, a processor and a computer program stored on the memory and capable of running on the memory, and the processor executes the program to realize the sensor-based cold chain equipment failure detection method as described in any one of the above.

[0012] The application has the beneficial effects that the selection and arrangement of sensors, data acquisition, cleaning, normalization and other links are planned in detail to ensure the accuracy and reliability of the collected data and provide high-quality data basis for subsequent analysis.

[0013] The method combining principal component analysis and long short-term memory network is used to extract the main features and time sequence features of the data respectively, and the two are organically combined through a feature fusion strategy to more comprehensively describe the running state of the cold chain equipment and improve the accuracy of failure feature extraction.

[0014] A perfect failure mode feature knowledge base is constructed, support vector machine and fuzzy comprehensive evaluation method are used for failure mode recognition and similarity evaluation, which can accurately identify the failure mode and provide a scientific basis for further screening and ordering of the failure mode.

[0015] The method utilizes a Bayesian network and a dynamic weight adjustment method, combines real-time fault correlation factor data, dynamically evaluates and corrects the fault probability, reflects the fault risk of the cold chain equipment in real time, and improves the reliability of fault prediction.

[0016] The method performs fault mode loop correlation analysis and fault area positioning based on internal structure topology data of the cold chain equipment, combines a fault mode feature knowledge base to deeply analyze fault causes, provides accurate maintenance guidance for maintenance personnel, shortens maintenance time, and reduces maintenance cost. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A method step diagram is provided for the application.

[0018] Figure 2 A method step diagram is provided for the application. Figure 1 A S2 step diagram is provided for the application. DETAILED DESCRIPTION

[0019] The preferred embodiments of the application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the application, and are not used to limit the application.

[0020] An embodiment of the application, as shown in Figure 1 The sensor-based cold chain equipment fault detection method comprises the following steps:

[0021] S1, collecting multi-dimensional data of the running state of the cold chain equipment in real time through a multi-dimensional sensor; cleaning and normalizing the collected multi-dimensional data to obtain a standardized multi-dimensional data set;

[0022] S2, performing dimension reduction processing on the standardized multi-dimensional data set by using a principal component analysis algorithm, extracting main feature components in the data, constructing a high-dimensional feature vector set, and using an LSTM model to extract time sequence features of the multi-dimensional data set to obtain a multi-dimensional time sequence feature matrix; fusing the high-dimensional feature vector set and the multi-dimensional time sequence feature matrix to obtain a fusion feature matrix;

[0023] S3, obtaining historical fault case data of the cold chain equipment, dividing the fusion feature matrix into feature subsets, calculating the correlation degree between different feature subsets and the historical fault case data by using a correlation analysis method to obtain a feature correlation degree matrix; according to the feature correlation degree matrix, screening out feature subsets highly related to faults, further analyzing the related feature subsets, extracting fault feature segments, and using feature recombination technology to reconstruct the fault feature segments into a fault feature vector;

[0024] S4, based on historical failure case data, a failure mode feature knowledge base is constructed to describe each failure mode in detail, a support vector machine algorithm is used to classify and identify the failure feature vector, and a candidate failure mode list is obtained; based on a fuzzy comprehensive evaluation method, a failure similarity vector is generated according to the matching degree of the failure feature vector and the characteristics of each failure mode in the failure mode feature knowledge base;

[0025] S5, obtain the running environment data of the cold chain equipment, and construct a failure correlation factor dataset; use a Bayesian network model to combine historical failure case data and the failure correlation factor dataset to preliminarily evaluate the occurrence probability of the candidate failure mode, and obtain an initial failure mode probability vector; use a dynamic weight adjustment method to dynamically correct the initial failure mode probability vector according to the real-time changes of the failure correlation factor dataset, and obtain failure probability distribution data;

[0026] S6, reorder the candidate failure mode list according to the failure probability distribution data to obtain a failure mode ordering list based on failure probability; obtain the internal structure topology data of the cold chain equipment, and based on the failure mode ordering list and the internal structure topology data of the cold chain equipment, use a graph theory algorithm to analyze the failure mode loop correlation, and obtain a failure mode loop correlation table; according to the failure mode loop correlation table and the failure mode ordering list, obtain failure area positioning result data; combine the failure area positioning result data and the failure mode feature knowledge base to analyze the failure cause in depth, and generate a cold chain equipment failure cause analysis report.

[0027] The working principle of the above technical solution is that a plurality of sensors (temperature, humidity, pressure, vibration, and current sensors) are arranged at key positions of the cold chain equipment to collect multi-dimensional data of the equipment running state in real time; the collected raw data is cleaned (noise and outliers are removed) and normalized to ensure the consistency and comparability of the data, forming a standardized multi-dimensional dataset; principal component analysis (PCA) algorithm is used to reduce the dimension of the multi-dimensional dataset, extract the main characteristic components of the data, construct a high-dimensional feature vector set, reduce the data dimension, and retain the key information; a long short-term memory (LSTM) model is used to extract time sequence features of the multi-dimensional dataset to capture the dynamic characteristics of the data over time, and a multi-dimensional time sequence feature matrix is obtained; the high-dimensional feature vector set obtained by dimension reduction processing and the multi-dimensional time sequence feature matrix extracted by LSTM are fused to form a fusion feature matrix, and static and dynamic features are comprehensively utilized.

[0028] The historical failure case data of the cold chain equipment is acquired, including failure occurrence time, type, sensor data, etc. Feature correlation analysis: divide the fusion feature matrix into feature subsets, calculate the correlation degree of different feature subsets with historical failure case data, and screen out feature subsets highly related to failure; further analyze the relevant feature subsets, extract failure feature fragments, and reconstruct them into failure feature vectors using feature reorganization technology, which facilitates subsequent classification and identification.

[0029] Based on historical failure case data, a failure mode feature knowledge base is constructed, which describes in detail the phenomena, causes, features, etc. of each failure mode; Support Vector Machine (SVM) algorithm is used to classify and identify failure feature vectors, and a candidate failure mode list is obtained; based on fuzzy comprehensive evaluation method, according to the feature matching degree of failure feature vector and each failure mode in failure mode feature knowledge base, a failure similarity vector is generated to evaluate the similarity of candidate failure modes.

[0030] The operating environment data of the cold chain equipment (such as environmental temperature, humidity, equipment service life, maintenance records, etc.) are acquired as failure-related factor data to construct a failure-related factor dataset. Using Bayesian network model, combining historical failure case data and failure-related factor dataset, the occurrence probability of candidate failure modes is preliminarily evaluated to obtain an initial failure mode probability vector. According to the real-time changes of the failure-related factor dataset, a dynamic weight adjustment method is used to dynamically correct the initial failure mode probability vector to obtain more accurate failure probability distribution data.

[0031] According to the failure probability distribution data, the candidate failure mode list is reordered to obtain a failure mode ordering list based on failure probability. The internal structure topology data of the cold chain equipment is acquired, based on the failure mode ordering list and the internal structure topology data, the graph theory algorithm is used for failure mode loop correlation analysis to determine the loop position where failure may occur, and a failure mode loop correlation table is obtained. According to the failure mode loop correlation table and the failure mode ordering list, the failure range is further narrowed down for failure area positioning to obtain failure area positioning result data.

[0032] Combined with the failure area positioning result data and the failure mode feature knowledge base, the failure cause is analyzed in depth. According to the analysis result, a cold chain equipment failure cause analysis report is generated to provide accurate maintenance guidance for maintenance personnel, realizing precise detection and diagnosis of cold chain equipment failure.

[0033] The effect of the above technical solution is:

[0034] Through real-time data acquisition of multi-dimensional sensors, combined with the application of principal component analysis, LSTM time series feature extraction and support vector machine classification algorithm, the failure features of cold chain equipment can be accurately captured, and the accuracy of failure detection is improved.

[0035] Through the fault mode circuit correlation analysis and fault area positioning technology, the fault range can be quickly narrowed down, the condition of needing a large amount of time and manual intervention in the traditional fault diagnosis method is reduced, and the efficiency of fault detection is improved.

[0036] Combined with environmental data, equipment service life, maintenance records and other factors, and the dynamic correction method of the Bayesian network model, the fault probability of the cold chain equipment can be evaluated in real time, the early warning of equipment failure is realized, and the accuracy of fault prediction is enhanced.

[0037] Based on the real-time change of dynamic weight adjustment and fault correlation factor data, the system can adaptively adjust the probability distribution of the fault mode, and improve the adaptability of the system to different environments and operating states.

[0038] Through the automatic data cleaning, feature extraction and fault recognition process, the interference of human factors is reduced, so that the system can complete fault diagnosis and repair guidance without too much manual intervention, and the automation level is improved.

[0039] Combined with the fault feature knowledge base, the system can not only provide fault diagnosis, but also in-depth analysis of fault causes, and provide accurate maintenance guidance for maintenance personnel, and enhance the traceability of fault analysis.

[0040] Through accurate fault area positioning and feature recombination technology, the maintenance personnel of the equipment can be provided with targeted fault positioning information, so that the maintenance is more accurate, unnecessary maintenance operations are reduced, and the maintenance cost is reduced.

[0041] Through standardization, dimensionality reduction processing and feature fusion of the collected multi-dimensional data, the efficiency of data processing and analysis in the big data environment is improved, and more efficient support is provided for real-time monitoring and fault detection.

[0042] One embodiment of the present application, the S1, comprises:

[0043] S11, through the multi-dimensional sensors deployed at different key parts of the cold chain equipment, real-time acquisition of the data output by each sensor, setting the data acquisition frequency, determining the acquisition interval according to the running characteristics and fault detection requirements of the cold chain equipment;

[0044] S12 stores the collected raw data in the database, uses a time series database to store, and backs up the stored data;

[0045] S13 adopts a digital filtering algorithm to remove noise in the collected data; based on a combination of statistical methods and machine learning algorithms, the data is detected for outliers; for the detected outliers, interpolation or prediction based on historical data is used for replacement;

[0046] S14 adopts a maximum-minimum normalization method to map the cleaned data to the [0, 1] interval to obtain a standardized multi-dimensional data set.

[0047] The working principle of the above technical solution is: deploy multiple types of sensors at different key positions of the cold chain equipment, including temperature sensors, humidity sensors, pressure sensors, vibration sensors, and current sensors. These sensors can monitor the running state of the equipment in real time and output corresponding data; according to the running characteristics and fault detection requirements of the cold chain equipment, set appropriate data collection frequency. For example, for equipment that needs high-precision monitoring, set to collect data once every second; while for relatively stable running equipment, set to collect data once every minute. Such settings can ensure the real-time nature of the data, and avoid data redundancy and storage pressure.

[0048] Store the collected raw data in a time series database. Time series databases are particularly suitable for storing data that changes over time, can efficiently process large amounts of time series data, and provide fast data query and analysis functions; back up the stored data to prevent data loss or damage. Backup data can be used to restore or further analyze when needed.

[0049] Use a digital filtering algorithm to remove noise in the collected data. Noise may be caused by sensor errors, environmental interference, and other factors, and filtering algorithms can effectively reduce the impact of noise on data quality; based on a combination of statistical methods (such as the 3σ principle) and machine learning algorithms (such as the Isolation Forest algorithm), the data is detected for outliers. The 3σ principle is a normal distribution-based outlier detection method, while the Isolation Forest algorithm is a machine learning-based outlier detection method, and their combination can improve the accuracy of outlier detection; for the detected outliers, interpolation (such as linear interpolation, cubic spline interpolation) or prediction based on historical data is used for replacement. This can ensure the continuity and integrity of the data and avoid the impact of outliers on subsequent analysis.

[0050] Use a maximum-minimum normalization method to map the cleaned data to the [0, 1] interval. Normalization can eliminate the influence of different dimensions and magnitudes on data analysis, making different features comparable. At the same time, normalized data is also easier to process by machine learning algorithms.

[0051] The effect of the above technical solution is:

[0052] By arranging multi-dimensional sensors at key parts of the cold chain equipment and setting appropriate data collection frequency, key information including temperature, humidity, pressure, vibration frequency and current intensity can be accurately and timely collected, thereby ensuring the comprehensiveness and accuracy of data collection;

[0053] Using time series database for data storage and backing up data can effectively avoid data loss or damage, reducing the risk of data storage of the cold chain equipment monitoring system.

[0054] By using digital filtering algorithm to remove noise, combined with statistical method and machine learning algorithm for anomaly detection, the data quality can be significantly improved, the risk of misdiagnosis caused by noise or abnormal data can be reduced, and the stability and reliability of the data can be ensured.

[0055] Using interpolation method and historical data prediction value to replace abnormal data, combined with modern machine learning algorithm (such as isolated forest algorithm) for anomaly detection, the abnormal value processing is more intelligent and automatic, and the need for manual intervention is reduced.

[0056] By using the maximum-minimum normalization method for data standardization, the data scale difference can be effectively eliminated, providing unified and standardized input for subsequent data analysis and modeling, and improving the accuracy of data analysis.

[0057] Using automatic data cleaning, anomaly detection and interpolation processing technology, manual intervention and operation are reduced, the efficiency of data cleaning is improved, and the time cost of data cleaning is reduced.

[0058] Combined with multi-dimensional data collection, digital filtering, anomaly detection and data standardization, the processing capacity of the system for complex data can be effectively improved, and the support capacity for cold chain equipment fault detection and diagnosis is enhanced.

[0059] By optimizing the complete process of data storage, backup, cleaning and standardization, the system can run more stably and efficiently, reducing the maintenance cost and improving the maintainability of the system.

[0060] An embodiment of the present application, as shown in Figure 2 S2, comprising:

[0061] S21, calculating the covariance matrix of the standardized multi-dimensional data set, by calculating the covariance matrix, the linear relationship between different features is obtained;

[0062] S22, solve the eigenvalues and eigenvectors of the covariance matrix; wherein the eigenvalue indicates the information amount contained in the corresponding eigenvector, and the eigenvector indicates the change trend of the data in the direction; according to the size of the eigenvalue, select the first k principal components; project the original data onto the selected principal components to construct a high-dimensional eigenvector set;

[0063] S23, divide the standardized multi-dimensional data set into a training set and a test set according to the time sequence; train the LSTM model using the training set, and adjust the parameters of the model through the back propagation algorithm,

[0064] S24, use the trained LSTM model to extract time sequence features from the multi-dimensional data set; input the data into the model to obtain the output of the model hidden layer, and construct a multi-dimensional time sequence feature matrix;

[0065] S25, according to the characteristics of the high-dimensional eigenvector set and the multi-dimensional time sequence feature matrix, develop a feature fusion strategy; adopt the feature splicing method to splice the two feature matrices in the feature dimension, so that the fused feature matrix contains the main features and time sequence features of the data;

[0066] S26, according to the developed fusion strategy, fuse the high-dimensional eigenvector set and the multi-dimensional time sequence feature matrix to generate a fused feature matrix.

[0067] The working principle of the above technical solution is: by calculating the covariance matrix of the standardized multi-dimensional data set, the linear correlation between the features in the data set can be revealed. Each element in the covariance matrix represents the covariance between two features, that is, how they change together. This step helps to understand the internal relationship between the features and provides a basis for subsequent feature extraction and dimension reduction; solve the eigenvalues and eigenvectors of the covariance matrix, the size of the eigenvalue reflects the information amount contained in the corresponding eigenvector, and the eigenvector indicates the change trend of the data in the direction.

[0068] According to the size of the eigenvalue, select the first k principal components, which can retain the main feature components in the data. Usually, the cumulative contribution rate method is used to determine the value of k, when the cumulative contribution rate reaches a certain threshold (such as 85%), the selected principal components are enough to represent the main information of the original data; project the original data onto the selected principal components to construct a high-dimensional eigenvector set. This step realizes the dimension reduction of the data while retaining the key information in the data.

[0069] The standardized multi-dimensional dataset is divided into a training set and a test set in chronological order to train and validate the LSTM model; the training set is used to train the LSTM model, and the model's parameters are adjusted through the backpropagation algorithm to enable the model to learn the dynamic characteristics of the data over time. LSTM models are particularly suitable for processing time series data because they can capture long-term dependencies in the data. The trained LSTM model is used to extract time series features from the multi-dimensional dataset. The data is input into the model, and the output of the model's hidden layer is obtained, which contains the time series feature information of the data, forming a multi-dimensional time series feature matrix.

[0070] According to the characteristics of the high-dimensional feature vector set and the multi-dimensional time series feature matrix, a feature fusion strategy is developed. The purpose of feature fusion is to combine two different types of features (static features and time series features) to provide a more comprehensive data representation; the two feature matrices are spliced in the feature dimension using feature splicing. In this way, the fused feature matrix contains both the main features and the time series features of the data.

[0071] According to the developed fusion strategy, the high-dimensional feature vector set and the multi-dimensional time series feature matrix are fused to generate a fused feature matrix. This step realizes the organic combination of static features and time series features, providing more rich feature information for subsequent fault mode classification and fault cause analysis.

[0072] The effect of the above technical solution is that by calculating the covariance matrix and extracting the first k principal components, the linear relationships and main information between the features in the dataset can be effectively captured, ensuring the accuracy and comprehensiveness of the feature extraction process.

[0073] The Principal Component Analysis (PCA) method is used to select principal components with an accumulated contribution rate of more than 85%, reducing the dimensionality of the data and reducing the complexity of data processing while retaining important information in the data.

[0074] By training the Long Short-Term Memory (LSTM) model, it can capture the dynamic characteristics of the data over time, enhancing the system's modeling ability for time series data and helping to improve the accuracy and reliability of the prediction.

[0075] By using feature splicing, the high-dimensional feature vector set and the multi-dimensional time series feature matrix are fused, which can retain both the main features and the time series features of the data, thereby improving the effectiveness of feature fusion and ensuring that the fused feature matrix is more comprehensive and accurate.

[0076] By reasonably dividing the training set and test set and using the LSTM model for training, the processing burden of complex data structures during model training can be effectively reduced, improving the efficiency and effectiveness of the training process.

[0077] By using the LSTM model for time series feature extraction, the expression ability of the data in the time dimension can be effectively enhanced, and more accurate and reliable time series information can be provided for subsequent data analysis and prediction.

[0078] Through the design and implementation of the feature fusion strategy, the system can automatically extract and fuse useful features from multi-dimensional data, enhancing the intelligent level of data processing and reducing the dependence on manual intervention.

[0079] By fusing high-dimensional features and time series features, data can be analyzed from multiple angles and levels, improving the system's ability to recognize complex data patterns and enhancing the comprehensiveness of data analysis.

[0080] In one embodiment of the present application, the S22 comprises:

[0081] The eigenvalue decomposition method in linear algebra is used to perform mathematical operations on the covariance matrix to solve its eigenvalues and corresponding eigenvectors.

[0082] The solved eigenvalues are analyzed; by drawing an eigenvalue distribution graph and an information contribution rate curve, the information importance ranking of each eigenvector is intuitively displayed.

[0083] The cumulative contribution rate method is used to dynamically determine the number of principal components k; a cumulative contribution rate threshold is set, and the information contribution rate corresponding to each eigenvalue is sequentially added in descending order of eigenvalues, and when the cumulative contribution rate first reaches or exceeds the set threshold, the number of eigenvectors corresponding to this time is the number of principal components k.

[0084] According to the determined number of principal components k, the eigenvectors corresponding to the first k eigenvalues are selected as the principal components; the original standardized multi-dimensional data set is projected onto the selected principal components for dimensionality reduction processing; the original data matrix is multiplied by the selected principal component matrix through matrix multiplication operation to obtain the dimensionality reduced data matrix, i.e. the high-dimensional feature vector set.

[0085] The working principle of the above technical solution is as follows: the eigenvalue decomposition method in linear algebra is used to perform mathematical operations on the covariance matrix. This process first involves solving the characteristic polynomial of the covariance matrix to obtain the eigenvalues. Eigenvalues are intrinsic properties of matrices, reflecting the stretching degree of the matrix in a certain direction; after obtaining the eigenvalues, the eigenvectors corresponding to each eigenvalue are determined by solving linear equations. Eigenvectors are non-zero vectors that remain unchanged under matrix transformation, representing the trend of data changes in different directions; in order to improve the calculation accuracy and efficiency, efficient numerical calculation algorithms such as QR algorithm are used. These algorithms can reduce errors during the solving process and obtain accurate and reliable eigenvalues and eigenvectors.

[0086] The eigenvalues obtained by solving are analyzed, and the proportion of each eigenvalue in the total sum of all eigenvalues, i.e. the information contribution rate, is calculated. The information contribution rate reflects the amount of information contained in each eigenvector. The larger the eigenvalue, the higher the information contribution rate, indicating that the data direction represented by the eigenvector contains more information. By plotting the eigenvalue distribution graph and the information contribution rate curve, the information importance ranking of each eigenvector can be intuitively displayed. This helps users understand the main features and trends of the data. Considering the data dimensionality reduction requirements and the degree of feature retention, the cumulative contribution rate method is used to dynamically determine the number of principal components k. Set a cumulative contribution rate threshold, such as 85%, which can be adjusted according to actual business needs and data characteristics. According to the order of eigenvalues from large to small, the information contribution rates of each eigenvalue are added in turn. When the cumulative contribution rate first reaches or exceeds the set threshold, the number of eigenvectors corresponding to this time is the number of principal components k. This method can ensure that the selected principal components can retain the main feature information of the data and meet the dimensionality reduction requirements.

[0087] According to the determined number of principal components k, the eigenvectors corresponding to the first k eigenvalues are selected as the principal components. These principal components represent the main features and trends of the data. To verify the rationality and effectiveness of the selected principal components, multiple methods are used for evaluation. On the one hand, the cumulative contribution rate of the selected principal components is calculated to ensure that it reaches or exceeds the set threshold. On the other hand, the distribution of data in the principal component space is observed through visualization methods such as drawing principal component score graphs to check whether there are obvious clustering or abnormal patterns. In addition, cross-validation and other methods can be used to evaluate the performance of the selected principal components in subsequent model training and prediction.

[0088] The original standardized multi-dimensional data set is projected onto the selected principal components for dimensionality reduction processing. Through matrix multiplication, the original data matrix is multiplied by the selected principal component matrix to obtain the dimensionality reduced data matrix, i.e. the high-dimensional eigenvector set. During the construction of the high-dimensional eigenvector set, attention should be paid to maintaining the integrity and consistency of the data. Ensure that the projected data can accurately reflect the characteristics and trends of the original data, providing a reliable data foundation for subsequent analysis and processing.

[0089] The effect of the above technical solution is that by using efficient numerical calculation algorithms such as QR algorithm, the eigenvalues and eigenvectors of the covariance matrix can be quickly and accurately calculated, ensuring the accuracy and efficiency of the calculation, thereby improving the accuracy of subsequent data analysis.

[0090] By calculating the information contribution rate and drawing the eigenvalue distribution graph and the information contribution rate curve, the information importance of each feature vector can be intuitively displayed, which helps to more accurately determine which features dominate in data analysis and enhances the transparency and scientificity of feature evaluation.

[0091] The cumulative contribution rate method is used to dynamically determine the number of principal components k, which can flexibly adjust the feature selection threshold according to the data characteristics, ensure the retention of sufficient information while reducing the dimensionality of data, reduce the subjectivity of the feature selection process, and improve the quality and usability of the reduced data.

[0092] By selecting the first k principal components and projecting the data, the original data is reduced to a lower dimension, significantly reducing the complexity and computational burden of the data, facilitating subsequent analysis and processing.

[0093] By evaluating and visualizing the reduced data matrix, it can be effectively checked whether the selection of principal components is reasonable, and further verified the representativeness and effectiveness of the reduced data, enhancing the reliability of the data reduction result.

[0094] In the data reduction process, the data is projected onto the principal components through matrix multiplication, which can ensure that the projected data accurately reflects the characteristics and trends of the original data, thereby maintaining the integrity and consistency of the data.

[0095] By effectively selecting principal components and using them for subsequent model training, the interference of redundant data can be reduced, the model training efficiency and prediction accuracy can be improved, and the performance of the model can be enhanced.

[0096] Through visualization methods such as drawing principal component score graphs, the distribution of data in the principal component space can be more intuitively observed, the data structure can be checked, and the interpretability of the feature selection process can be enhanced, helping users understand the internal structure and potential patterns of the data.

[0097] An embodiment of the present application, the S26, comprises:

[0098] Before performing feature fusion, the high-dimensional feature vector set and the multi-dimensional time series feature matrix are comprehensively pre-checked; the high-dimensional feature vector set and the multi-dimensional time series feature matrix are processed based on the pre-checking result;

[0099] The processed high-dimensional feature vector set and the multi-dimensional time series feature matrix are dimensionally matched and adjusted; the high-dimensional feature vector set and the multi-dimensional time series feature matrix are spliced in the feature dimension according to the formulated feature splicing strategy; after splicing, a preliminary fusion feature matrix is generated;

[0100] The quality of the generated fusion feature matrix is evaluated, and the fusion feature matrix after quality evaluation is normalized. The quality of the generated fusion feature matrix is evaluated by the following formula:

[0101]

[0102] wherein, represents the quality evaluation value of the fusion feature matrix (the range is [0, 1], and the higher the value, the better the quality); represents the feature dimension number of the feature matrix; represents the Pearson correlation coefficient of feature i and feature j; represents the number of principal components retained by the fusion feature matrix; represents the total number of principal components of the original feature matrix; represents the eigenvalue of the kth principal component; represents the weight coefficient (0≤α≤1, usually set according to task requirements, such as α=0.6).

[0103] The working principle of the above technical solution is: before feature fusion, the high-dimensional feature vector set and the multi-dimensional time series feature matrix are comprehensively pre-checked to ensure that there is no missing value in the data. Missing values may affect the subsequent feature fusion and model training effect, so appropriate filling methods (such as mean filling, median filling or prediction filling based on machine learning model) need to be used for processing according to the characteristics and missing proportion of the data; abnormal values in the data are identified by statistical analysis methods (such as Z-score, box plot method), and reasonable methods are used for correction or deletion. Abnormal values may distort the distribution of data and affect the accuracy of feature fusion; ensure that the data types of the two feature matrices are consistent to facilitate subsequent splicing operations. Inconsistent data types may cause errors or inconsistent results in the splicing process.

[0104] The high-dimensional feature vector set and the multi-dimensional time series feature matrix after processing are dimensionally matched and adjusted. If the feature dimensions of the two matrices are inconsistent, the reasons need to be analyzed and corresponding measures are taken. For example, delete the features in the high-dimensional feature vector set that are redundant and not important for subsequent analysis, or integrate the valuable features in the time series feature matrix into the high-dimensional feature vector set through feature expansion; through dimension matching and adjustment, the two matrices have splicable feature dimensions, laying a foundation for subsequent feature splicing operations.

[0105] According to the formulated feature splicing strategy, the high-dimensional feature vector set and the multi-dimensional time series feature matrix are spliced in the feature dimension. During the splicing process, it is necessary to ensure that the splicing order and method meet the strategy requirements, such as splicing after sorting according to the importance of the features; after splicing, a preliminary fusion feature matrix is generated, which contains both the main features and the time series features of the data, providing a more comprehensive feature representation for subsequent data analysis and model training.

[0106] The quality of the preliminary fusion feature matrix is evaluated to check whether there are problems such as data redundancy, high correlation between features, etc. By calculating the correlation coefficient matrix between features, the correlation between features is analyzed, redundant features are identified and removed to improve the generalization ability and computational efficiency of the model; check whether the fusion feature matrix can better reflect the internal rules and characteristics of the data. Through visualization methods (such as scatter plots after principal component analysis dimensionality reduction) or comparative analysis with the original data, the fusion effect is evaluated to ensure that the fused features can more accurately describe the essential characteristics of the data. The fusion feature matrix after quality evaluation is normalized to eliminate the influence of different dimensions and orders of magnitude on data analysis, making different features comparable, and providing more standardized data input for subsequent model training.

[0107] The effect of the above technical solution is that through comprehensive pre-checking before feature fusion, missing values, outliers and inconsistent data types can be effectively identified and processed, ensuring the quality and consistency of the data, thereby avoiding bias or misleading results caused by data problems in the subsequent analysis process.

[0108] By using various methods to handle missing values such as mean filling, median filling, and machine learning model filling, the appropriate processing method can be flexibly selected according to the data characteristics, reducing the risk of processing errors; at the same time, through outlier detection and processing, the data is properly cleaned before feature fusion, avoiding the interference of incorrect data.

[0109] Through dimension matching and adjustment, the consistency of the high-dimensional feature vector set and the multi-dimensional time series feature matrix in the feature dimension is ensured, making the two matrices able to be spliced smoothly, enhancing the operability and flexibility of data processing.

[0110] By splicing according to the importance of features, not only can important features be given priority in the fused matrix, but also the spliced matrix can be more representative and accurate in subsequent analysis, improving the reliability of the model analysis results.

[0111] By calculating the correlation between features and using feature selection methods to remove redundant features, the redundant information in the data is reduced, avoiding the problem of overfitting and low computational efficiency caused by high correlation between features, improving the efficiency of data processing and model training.

[0112] Through quality assessment and visual analysis, it can be verified whether the fusion feature matrix can better reflect the inherent law of data, improve the transparency and accuracy of data analysis, and make the subsequent model better learn and predict.

[0113] By formulating a clear feature splicing strategy and splicing according to feature importance ranking, a flexible and scientific way is provided to handle the fusion problem of high-dimensional data and time series data, improving the effectiveness of subsequent analysis.

[0114] By normalizing the fusion feature matrix, it ensures that features of different dimensions can be compared and analyzed on the same scale, avoiding the impact of dimensional differences, and enhancing the unity and comparability of data.

[0115] The above formula evaluates the quality of the fusion feature matrix through two important factors: the correlation between features (through Pearson correlation coefficient) and the eigenvalues of principal component analysis. By combining these two parts of information, the quality of the feature matrix can be comprehensively evaluated to ensure that not only redundant information is reduced, but also key data features are retained to the greatest extent, thereby improving the expressiveness of the fusion matrix.

[0116] The Pearson correlation coefficient part (by calculating the correlation between feature pairs) aims to identify and reduce the impact of redundant information. High correlation between features may indicate redundant or repetitive information, which can affect the training efficiency and accuracy of the model. Therefore, the formula effectively reduces the impact of redundant features by considering these correlations, enhancing the accuracy and training efficiency of the model.

[0117] By introducing the principal component analysis (PCA) part, the formula ensures that the fusion feature matrix can retain the most important information. The eigenvalues of principal components represent the degree of explanation of each principal component to the original data, and high eigenvalues indicate that the principal component contains more data variability or important information. By retaining the first m principal components, the formula effectively removes inefficient information and enhances the expressiveness of the data.

[0118] The introduction of the weight coefficient makes the formula flexible, allowing for adjustments in importance according to different task requirements. For example, in some tasks, more attention may be paid to reducing the redundancy between features, in which case the weight of the Pearson correlation coefficient can be appropriately increased; in other tasks, more attention may be paid to information retention, and the weight of the principal component analysis part can be appropriately increased. Such design makes the formula have wide applicability and can be flexibly adjusted according to task requirements.

[0119] The final quality evaluation value Q provides a quantitative index (ranging from [0, 1]), which can clearly reflect the quality of the fused feature matrix. The higher the value, the better the quality of the fused matrix. This provides a clear reference for subsequent feature selection, feature engineering or model training, helping developers or researchers to further optimize the feature fusion process according to the quality evaluation results.

[0120] Through pre-checking, dimension matching and adjustment, feature splicing and other processing steps, the fused feature matrix not only effectively combines feature data from different sources, but also further improves the applicability and performance of the matrix through quality evaluation and normalization processing. The finally generated fused feature matrix will have high information quality, which is helpful for subsequent model training or data analysis.

[0121] An embodiment of the present application, the S3, comprises:

[0122] S31, acquire the historical failure case data of cold chain equipment, organize the collected historical failure case data, establish a historical failure case database; classify and label the data, and add a label to each failure case;

[0123] S32, according to the properties and sources of the features in the fused feature matrix, the features are divided into different subsets; through the correlation degree calculation method, the correlation degrees between different feature subsets and historical failure case data are calculated; for each feature subset, the correlation degrees with each failure case data are calculated to obtain a feature correlation degree matrix; the correlation degree is calculated by the following formula:

[0124]

[0125] Among them, represents the correlation degree of feature subset S and failure case f; represents the i-th feature in feature subset S; represents the label vector (or failure feature vector) of failure case f; represents the number of features in feature subset S; represents the weight of feature Si (determined by feature importance, ); represents the correlation measure of feature Si and failure case f; represents the confidence of feature Si (reflecting feature quality, such as missing value rate, abnormal value proportion, etc.); represents the timestamp of feature subset S (if it is time series data); represents the occurrence time of failure case f; represents the time decay factor (controls the decay speed of time correlation);

[0126] S33, analyze the feature correlation matrix, and screen out a feature subset highly related to the fault; set a correlation threshold, and take the feature subset with a correlation greater than the threshold as the feature subset highly related to the fault;

[0127] S34, further analyze the screened feature subset highly related to the fault, and extract a fault feature fragment by using a sliding window and a feature clustering method;

[0128] S35, combine and reconstruct the extracted fault feature fragment by using a feature recombination technology; and combine different fault feature fragments into a fault feature vector by using feature splicing.

[0129] The working principle of the above technical solution is as follows: historical fault case data of cold chain equipment is comprehensively obtained through equipment maintenance records, fault reports, maintenance personnel feedback and other channels. The data includes fault occurrence time, fault type, sensor data at the time of fault occurrence, fault handling process and result and other key information; the collected historical fault case data is sorted, and a structured historical fault case database is established. At the same time, the data is classified and labeled, and is classified in detail according to fault types, fault positions and the like, and corresponding labels are added to each fault case. This step provides a solid foundation for subsequent feature correlation calculation and fault feature extraction.

[0130] According to the properties and sources of the features in the fusion feature matrix, the features are divided into different subsets. For example, temperature-related features and pressure-related features are divided into independent subsets. This step helps to more accurately analyze the correlation between different features and faults; the correlation calculation method is used to calculate the correlation between different feature subsets and historical fault case data. For each feature subset, the correlation between the feature subset and each fault case data is calculated, and a feature correlation matrix is obtained. This matrix directly shows the correlation degree between different feature subsets and faults, and provides a basis for subsequent feature screening.

[0131] A correlation threshold is set, and a feature subset with a correlation greater than the threshold is taken as a feature subset highly related to the fault. This step screens out a feature subset highly related to the fault by setting a reasonable threshold, reduces the data amount of subsequent analysis, and improves the analysis efficiency; the screened feature subset highly related to the fault is further analyzed to understand its characteristics and distribution rules. This step helps to deeply understand the mechanism and influencing factors of fault occurrence, and provides guidance for subsequent fault feature extraction and model construction.

[0132] The sliding window method is used to slide on the feature subset to extract representative feature segments. The size and step of the sliding window are set according to the characteristics of the data and the fault detection requirements to ensure that the extracted feature segments can accurately reflect the characteristics of the fault; similar feature segments are clustered together by using a feature clustering method, which facilitates subsequent analysis and processing. This step helps to reduce data redundancy and improve the accuracy and efficiency of feature extraction.

[0133] The extracted fault feature segments are combined and reconstructed by using a feature recombination technique. Different fault feature segments are combined into a fault feature vector through feature splicing and other methods. This step integrates the extracted fault feature segments into a higher-level feature representation, providing more comprehensive feature information for subsequent fault pattern classification and fault cause analysis.

[0134] The effect of the above technical solution is that by establishing a historical fault case database and classifying and labeling the data, the fault diagnosis system can obtain accurate fault feature information. By using the correlation analysis of the feature correlation matrix and the fault data, the highly relevant feature subset can be selected, which enhances the accuracy of fault diagnosis.

[0135] Through feature screening and clustering analysis, redundant and irrelevant features can be removed, and only the feature subset closely related to the fault is retained, thereby improving the quality of the data, reducing the influence of redundant information on subsequent analysis, and improving the processing efficiency of the system.

[0136] By using the sliding window and feature clustering method, representative fault feature segments can be effectively extracted from large-scale data. This process enhances the operability of feature extraction, making subsequent analysis more efficient and accurate.

[0137] By recombining and splicing the fault feature segments and combining the analysis of historical data, a more accurate fault prediction model can be constructed. By optimizing the fault feature vector, the state of the equipment can be evaluated more quickly, and possible faults can be predicted in a timely manner, improving the reliability and maintenance efficiency of the equipment.

[0138] By using effective feature screening and clustering methods, unnecessary feature calculations are reduced, and the data processing process is optimized. This can reduce the computational burden and improve the efficiency of the fault detection and analysis process.

[0139] By using the flexible strategy of feature recombination and splicing, the system can adjust the feature combination strategy according to different fault conditions, improving the adaptability of the fault detection system, and maintaining high accuracy and stability when facing variable equipment operating conditions.

[0140] With accurate fault features and correlation analysis, more scientific maintenance recommendations can be provided to equipment maintenance personnel, helping to achieve accurate fault prediction and timely maintenance decisions, thereby reducing equipment failure rate and maintenance cost.

[0141] The above formula can reveal which feature subsets have strong correlation with faults by calculating the correlation degree between feature subsets and fault cases. Through the comprehensive evaluation of weight, correlation, confidence and time decay factor, the formula can effectively capture features highly related to faults. Especially with the introduction of the time decay factor, time-sensitive features can be more accurately evaluated, which helps to improve the accuracy of feature selection.

[0142] By calculating and analyzing the correlation degree of different feature subsets and historical fault cases, it can be identified which features have high prediction value for fault cases in the subset. After screening the highly correlated feature subsets, the accuracy of subsequent fault prediction and diagnosis models can be greatly improved.

[0143] The introduction of the time decay factor τ can effectively control the influence of time correlation on features. Especially in time series data, the correlation degree of features may change over time. By introducing the time factor, the model can more flexibly adapt to time changes and reduce the interference of outdated features.

[0144] The selection of feature subsets is based on the calculated correlation matrix and the set threshold, ensuring that only features highly related to faults are selected. This screening method ensures that the feature input of the model is more refined, effectively avoiding the influence of redundant data and improving the training efficiency and performance of the model.

[0145] By using the sliding window and feature clustering method to extract fault feature fragments, and then combining different fault feature fragments into complete fault feature vectors through feature splicing technology. This method can effectively integrate dispersed information and extract valuable patterns, which helps the training of subsequent models and fault prediction.

[0146] Through comprehensive analysis of historical fault data, feature fusion technology can help extract multi-dimensional and comprehensive feature information, thereby improving the adaptability and diagnostic robustness of the model to different fault types. The combination of different features helps to avoid misjudgment or omission caused by a single feature.

[0147] The confidence of the feature reflects the quality of the data, including the influence of missing values and outliers. By considering the confidence in the correlation calculation, the model error caused by poor data quality can be effectively reduced, and the reliability of the final analysis result can be improved.

[0148] In an embodiment of the present application, the S32 comprises:

[0149] Each feature in the fusion feature matrix is analyzed in detail, and the features are divided into different subsets by multi-dimensional division method according to the nature and source of the features;

[0150] Different correlation calculation methods are selected for different feature subsets and historical fault case data; the selected correlation calculation methods are used to calculate the correlation between each feature subset and each fault case data; for all fault cases, the correlation vectors are combined into a correlation matrix, wherein the rows represent the fault cases and the columns represent the feature subsets;

[0151] The quality of the obtained correlation matrix is evaluated, and corresponding correction measures are taken based on the quality evaluation results.

[0152] The working principle of the above technical solution is as follows: first, each feature in the fusion feature matrix is analyzed in depth to clarify its physical meaning, data source and specific role in the running process of the cold chain equipment. This helps to understand the influence of each feature on the equipment state and provides a basis for subsequent feature subset division; multi-dimensional division method is used to divide the features into different subsets. These dimensions can include physical meaning (such as temperature, pressure, humidity, etc.), data source (such as data collected by different sensors) and correlation between features. Through multi-dimensional division, the division of feature subsets can be both reasonable and practically meaningful, which helps subsequent feature correlation calculation and fault feature extraction.

[0153] Different correlation calculation methods are selected for different feature subsets and historical fault case data. For numerical features, Pearson correlation coefficient, Spearman rank correlation coefficient and other methods can be used; for categorical features, chi-square test, mutual information and other methods can be used. Considering the characteristics of different feature subsets and the complexity of fault case data, the selected correlation calculation methods are optimized. For example, for feature subsets with noise or outliers, more robust correlation calculation methods are used; for high-dimensional feature subsets, dimensionality reduction techniques are used before correlation calculation to improve calculation efficiency and accuracy.

[0154] The selected correlation calculation methods are used to calculate the correlation between each feature subset and each fault case data. In the calculation process, the diversity and complexity of fault case data are fully considered to ensure the accuracy and reliability of the calculation results; for each fault case, the correlation with each feature subset is calculated to obtain a correlation vector. For all fault cases, the correlation vectors are combined into a correlation matrix, wherein the rows represent the fault cases and the columns represent the feature subsets. This matrix directly shows the correlation between different feature subsets and fault cases.

[0155] The obtained correlation matrix is evaluated for quality, and whether there are outliers, missing values, etc. These problems may affect the subsequent feature selection and analysis results; based on the quality evaluation results, take appropriate correction measures. For example, for outliers, median replacement, mean replacement or deletion method can be used for processing; for missing values, interpolation method, regression prediction method or deletion method can be used for filling; the correlation matrix is displayed through visualization methods (such as heat map), which can intuitively observe the correlation degree between feature subsets and fault cases. This helps subsequent feature selection and analysis, and provides intuitive basis for fault feature extraction.

[0156] The effect of the above technical solution is: through the multi-dimensional feature division method, it ensures that the division of feature subsets not only conforms to the physical meaning, but also considers the data source and correlation. This division method improves the rationality of the feature subsets, ensures that each feature subset has practical significance, and helps subsequent correlation analysis and feature extraction.

[0157] By dividing the features with high correlation into the same subset and using appropriate dimension reduction techniques, the computational complexity caused by high-dimensional feature subsets is reduced. Combined with the nature of the features and the source of the data, unnecessary feature redundancy is avoided, and the computational efficiency is improved.

[0158] A variety of correlation calculation methods are used for different feature subsets and fault cases, and the calculation methods are optimized. For example, more robust calculation methods are used to handle outliers and noise, ensuring the accuracy and reliability of the calculation results.

[0159] By accurately calculating the correlation between feature subsets and fault cases, more powerful evidence can be provided for subsequent fault prediction. Through fine feature analysis and optimized calculation methods, the accuracy of fault prediction is improved, and the warning ability of the system is enhanced.

[0160] Through the correction measures for outliers and missing values, the interference of data noise is effectively reduced, the accuracy of correlation calculation is ensured, and the quality of feature data is improved, providing more reliable data support for subsequent analysis.

[0161] By using visualization methods such as heat map, the correlation between feature subsets and fault cases can be intuitively displayed, providing more intuitive basis for subsequent feature selection and analysis. This visualization method helps to quickly identify key features, improve the efficiency and accuracy of data analysis.

[0162] According to the characteristics of different feature subsets, targeted correlation degree calculation methods and optimization strategies are adopted to ensure that the system can adapt to different types of feature data and complex fault cases. This improves the adaptability of the system in various operating environments, enabling it to maintain high accuracy when facing various complex fault situations.

[0163] In one embodiment of the present application, the S4 comprises:

[0164] S41, based on historical fault case data, a fault mode feature knowledge base is constructed; each fault mode is described in detail, and a mapping relationship between the fault mode and the feature vector is established; the fault mode feature knowledge base is continuously improved and updated;

[0165] S42, support vector machine (SVM) algorithm is used for classification and identification of fault feature vectors; the fault feature vectors in the historical fault case data are used as training samples, and the corresponding fault modes are used as labels to train the support vector machine model;

[0166] S43, the trained support vector machine model is used for classification and identification of new fault feature vectors; the fault feature vector is input into the model, and the model outputs the corresponding fault mode to obtain a candidate fault mode list;

[0167] S44, a fault similarity evaluation model is constructed based on a fuzzy comprehensive evaluation method; according to the feature matching degree of the fault feature vector and each fault mode in the fault mode feature knowledge base, the fuzzy comprehensive evaluation model is used to calculate the similarity between the fault feature vector and each fault mode; the similarity results are combined to form a fault similarity vector.

[0168] The working principle of the above technical solution is: based on historical fault case data, a fault mode feature knowledge base is constructed. Each fault mode is described in detail, covering key information such as fault phenomenon, fault cause, fault feature, and fault impact. This step provides a rich knowledge base for subsequent fault classification and identification; a mapping relationship between the fault mode and the feature vector is established, so that each fault mode can correspond to a specific feature vector. This helps to convert abstract fault modes into quantifiable feature vectors, facilitating subsequent machine learning processing; with the emergence of new fault cases and the continuous development of fault detection technology, the fault mode feature knowledge base is continuously improved and updated. Regularly review and revise the knowledge base, add new fault modes and feature information, delete outdated information, and ensure the accuracy and integrity of the knowledge base.

[0169] The support vector machine (SVM) algorithm is used to classify and identify the fault feature vectors. The fault feature vectors in the historical fault case data are used as training samples, and the corresponding fault modes are used as labels to train the support vector machine model. This step enables the model to learn the mapping relationship between the fault feature vectors and the fault modes; the classification performance of the model is optimized by adjusting the parameters of the model, such as the kernel function type, the penalty factor, etc. Different parameter settings may affect the classification accuracy and generalization ability of the model, so detailed parameter tuning is required.

[0170] The trained support vector machine model is used to classify and identify new fault feature vectors. The fault feature vectors are input into the model, and the model outputs the corresponding fault modes to obtain a list of candidate fault modes. This step realizes the rapid classification and identification of new fault cases; based on the fuzzy comprehensive evaluation method, a fault similarity evaluation model is constructed. The evaluation factor set (such as the dimensions of the fault feature vector), the evaluation set (such as "high similarity", "medium similarity", "low similarity", etc.), and the weight of each evaluation factor are determined. This step provides an evaluation framework and standard for subsequent fault similarity calculation.

[0171] According to the matching degree of the fault feature vectors and the characteristics of each fault mode in the fault mode feature knowledge base, the fuzzy comprehensive evaluation model is used to calculate the similarity between the fault feature vectors and each fault mode. The similarity results are combined into a fault similarity vector, providing a quantitative basis for further analysis and processing of the fault mode.

[0172] The effect of the above technical solution is: by constructing a detailed fault mode feature knowledge base and continuously updating it, the mapping relationship between the fault mode and the feature is ensured to be timely. Combined with the classification ability of the support vector machine algorithm, different fault modes can be accurately identified, thereby improving the accuracy of the fault diagnosis system.

[0173] Through the fuzzy comprehensive evaluation method, the fault similarity is carefully evaluated, which can reasonably distinguish between multiple similar fault modes and reduce the possibility of misjudgment. This similarity evaluation mechanism enhances the processing ability of complex fault conditions, thereby reducing diagnostic errors.

[0174] Since the technical solution clearly states that the fault mode feature knowledge base should be continuously updated based on new fault cases and detection techniques, the adaptability of the system to new fault modes is enhanced. As the system evolves, it can identify more diverse and complex fault types, enhancing the long-term usability of the system.

[0175] The support vector machine algorithm uses historical fault case data as training samples, not only improves the efficiency of training, but also improves the performance of fault feature vector classification recognition by optimizing model parameters. This enables the system to quickly adapt to different fault situations and reduces resource waste during the training process.

[0176] By constructing a similarity calculation method based on fuzzy comprehensive evaluation, redundant fault feature vector data can be effectively processed, and irrelevant information can be avoided when identifying fault patterns, thereby improving the system's computing efficiency and data processing quality.

[0177] The introduction of the fuzzy comprehensive evaluation method enables the system to comprehensively evaluate multi-dimensional and multi-level fault features, further enhancing the intelligence and flexibility of the diagnosis process. With the support of fuzzy logic systems, fuzzy and uncertain data can be processed to make more accurate fault pattern matching.

[0178] Regularly reviewing and revising the design of the knowledge base ensures that the system can adapt to changing fault patterns and technical requirements. This updating mechanism not only ensures the real-time and reliability of the system, but also facilitates the maintenance and optimization of the fault diagnosis system by technical personnel.

[0179] Through the technical solution of combining support vector machines with fuzzy comprehensive evaluation, the system can consider different dimensions of fault features, enabling it to maintain high recognition ability when facing complex and variable fault situations.

[0180] In one embodiment of the present application, the S44 comprises:

[0181] Each dimension of the fault feature vector is analyzed in depth, and the comment set is designed in detail according to the actual fault diagnosis requirements and precision requirements;

[0182] The weights of each evaluation factor are determined based on the combination of multiple methods; a quantitative matching model is constructed for the matching degree of the fault feature vector and the characteristics of each fault mode in the fault mode feature knowledge base;

[0183] According to the determined evaluation factor set, comment set, weight distribution, and feature matching degree quantification model, the fuzzy comprehensive evaluation method is used for calculation;

[0184] For each evaluation factor, the membership value is calculated according to the membership function of each comment level to obtain the fuzzy evaluation matrix; and the fuzzy evaluation matrix and the weight vector are subjected to fuzzy synthesis operation to obtain the similarity evaluation results of the fault feature vector and each fault mode;

[0185] The evaluation results of multiple evaluation factors are integrated, and a synthesis operator is used to obtain the final fault similarity vector.

[0186] The working principle of the above technical solution is: the dimensions of the fault feature vector are analyzed in depth, not only focusing on traditional numerical characteristics, but also mining potential semantic characteristics and correlation characteristics. For example, for the temperature feature, in addition to the numerical size, the temperature change trend and its correlation with other features (such as pressure, humidity) are also analyzed; combined with the information in the fault mode feature knowledge base, a comprehensive and representative evaluation factor set is determined. These factors should be able to fully reflect the characteristics of the fault feature vector and be closely related to the fault mode; the evaluation factors are classified and coded for subsequent processing and calculation. Classification helps to organize and manage evaluation factors, and coding facilitates computer processing and storage.

[0187] According to the actual fault diagnosis requirements and accuracy requirements, the evaluation set is designed in detail. In addition to the common "high similarity" "moderate similarity" "low similarity", more detailed evaluation levels such as "extremely high similarity" "weak similarity" can also be added; clear definitions and judgment criteria are established for each evaluation level to ensure the consistency and accuracy of the evaluation. This helps different evaluators reach a consensus and improves the reliability of the evaluation results.

[0188] The weights of each evaluation factor are determined based on a combination of expert experience method, analytic hierarchy process (AHP), and entropy weight method; the weights obtained by the three methods are integrated, and the final weight distribution is determined by weighted average or other combination methods. This method fully considers expert experience, data characteristics, and objective information, making the weight distribution more reasonable and accurate.

[0189] A quantitative matching model is constructed for the matching degree of the fault feature vector and the features of each fault mode in the fault mode feature knowledge base; a variety of similarity calculation methods are used, such as Euclidean distance, cosine similarity, dynamic time warping (DTW), and the appropriate calculation method is selected according to the properties and characteristics of different features; considering the importance difference of the features, the similarity calculation results of different features are weighted to obtain more accurate feature matching degree quantitative values.

[0190] For each evaluation factor, the membership value is calculated according to the membership function of each evaluation level, and a fuzzy evaluation matrix is obtained; the fuzzy evaluation matrix and the weight vector are subjected to fuzzy synthesis operation to obtain the similarity evaluation results of the fault feature vector and each fault mode. This step converts the fuzzy relationship between the evaluation factors and the evaluation levels into specific similarity values.

[0191] The evaluation results of multiple evaluation factors are integrated and subjected to comprehensive processing using synthesis operators (such as weighted average type operator, maximum and minimum type operator, etc.); the integrated evaluation results form a fault similarity vector, which intuitively shows the similarity between the fault feature vector and each fault mode.

[0192] The effect of the above technical solution is that by deeply analyzing each dimension of the fault feature vector and combining the temperature change trend and the correlation between temperature and other features, the potential information of the fault can be more comprehensively captured, and the recognition accuracy of the fault mode is improved.

[0193] By fine-grained design of the evaluation set, the evaluation levels such as "extremely high similarity" and "weak similarity" are added, so that the similarity determination between the fault and the mode is more detailed and accurate, and the misjudgment rate is reduced.

[0194] A variety of similarity calculation methods are adopted, and the expert experience method, the analytic hierarchy process (AHP) and the entropy weight method are combined to comprehensively distribute the evaluation factor weight, so that the adaptability and robustness of the system to various fault features under different environments and conditions are improved.

[0195] Combined with the Euclidean distance, the cosine similarity and the dynamic time warping (DTW), the most suitable similarity calculation method can be selected according to the properties of different features, so that the accuracy of the fault feature matching is improved.

[0196] By weighting the similarity calculation results of different features, the contribution of each feature to the fault mode matching can be accurately reflected, and the influence of the feature importance on the final diagnosis result is enhanced.

[0197] By using the fuzzy comprehensive evaluation method, the membership degree of each evaluation factor is calculated, and the similarity evaluation result is obtained through fuzzy synthesis operation, so that the flexibility and intelligence of the diagnosis process are enhanced, and the fuzzy and uncertain fault data can be processed.

[0198] By comprehensively operating the multiple similarity calculation methods and the fuzzy evaluation matrix, the calculation process is simplified, and the calculation complexity in the fault diagnosis process is reduced.

[0199] By combining the expert knowledge and the data-driven method to comprehensively distribute the evaluation factor weight, the system can be continuously updated and adjusted, and the recognition ability of new fault modes and the long-term adaptability are enhanced.

[0200] By clearly designing the evaluation level, the weight distribution and the similarity calculation, the entire fault diagnosis process is more transparent and interpretable, which is helpful for the understanding and optimization of the system by the technical personnel.

[0201] By clearly defining the standards and calculation process of the evaluation level, the consistency of the fault diagnosis result is ensured, the influence of human factors on the diagnosis result is reduced, and the accuracy of the overall system is improved.

[0202] One embodiment of the present application, the S5, comprises:

[0203] S51, obtain the running environment data of the cold chain equipment as the fault correlation factor data; pre-process the collected fault correlation factor data;

[0204] S52, use the Bayesian network model, combine the historical fault case data and the fault correlation factor data set, and construct a fault probability evaluation model;

[0205] S53, according to the constructed Bayesian network model, calculate the occurrence probability of the candidate fault mode under the condition of the current fault correlation factor, and obtain an initial fault mode probability vector; through the inference algorithm of the Bayesian network, update the probability distribution of each fault mode node according to the known fault correlation factor data;

[0206] S54, develop a dynamic weight adjustment method, determine the initial weight of each factor according to historical data and expert experience, and establish a weight adjustment rule;

[0207] S55, adopt the dynamic weight adjustment method, dynamically correct the initial fault mode probability vector according to the real-time change of the fault correlation factor data set; periodically update the weight of each factor, recalculate the occurrence probability of the candidate fault mode, and obtain more accurate fault probability distribution data, which reflects the fault risk of the cold chain equipment in real time.

[0208] The working principle of the above technical solution is as follows: first, the running environment data (such as environmental temperature, humidity, equipment service life, maintenance record, etc.) and running parameters (such as refrigerant pressure, current fluctuation, etc.) of the cold chain equipment are obtained through sensors, equipment monitoring systems, etc. These data are used as fault correlation factor data, which provides a basis for subsequent fault probability evaluation; the collected fault correlation factor data is pre-processed, including data cleaning, missing value processing, anomaly detection and correction, etc., to ensure the accuracy and reliability of the data. The pre-processed data will be used to construct the fault probability evaluation model.

[0209] Using the Bayesian network model, combining historical fault case data and fault correlation factor data set, determine the nodes in the network. The nodes include fault mode nodes and fault correlation factor nodes, each node represents a variable or state; analyze the causal relationship between nodes to determine which factors are direct or indirect causes of fault occurrence. At the same time, according to historical data and expert experience, determine the conditional probability distribution between nodes, i.e. the probability distribution of a node given the state of its parent nodes.

[0210] According to the constructed Bayesian network model, the occurrence probability of the candidate failure mode under the condition of the current failure correlation factors is calculated. This is achieved through the inference algorithm of the Bayesian network, and the probability distribution of each failure mode node is updated according to the known failure correlation factor data; the calculated occurrence probability of each failure mode forms an initial failure mode probability vector, which intuitively shows the possibility of the occurrence of each failure mode under the condition of the current failure correlation factors.

[0211] According to historical data and expert experience, the initial weights of each failure correlation factor on the failure probability are determined. These weights reflect the relative importance of different factors in the failure process; weight adjustment rules are developed to consider the degree of influence of different failure correlation factors on the failure probability over time. For example, the influence of equipment service life on failure probability will gradually increase with the increase of service time.

[0212] A dynamic weight adjustment method is adopted to dynamically correct the initial failure mode probability vector according to the real-time changes of the failure correlation factor data set. This is achieved by periodically updating the weights of each factor and recalculating the occurrence probability of the candidate failure mode; through dynamic correction, more accurate failure probability distribution data are obtained, which can reflect the failure risk of cold chain equipment in real time. This is of great significance for timely discovery of potential failures, taking preventive measures and optimizing equipment maintenance plans.

[0213] The effect of the above technical solution is that by constructing a failure probability evaluation model based on Bayesian network, combining historical failure cases and failure correlation factor data, the failure risk of cold chain equipment can be more accurately evaluated, thereby improving the accuracy of failure prediction.

[0214] By dynamically adjusting the weights of failure correlation factors, the influence of each factor on the failure probability can be adjusted in real time, reducing the failure probability caused by factors such as equipment aging or changes in operating environment.

[0215] By dynamically correcting the failure mode probability vector and periodically updating the weights of each factor, the system can reflect the failure risk of cold chain equipment in real time, enhancing the real-time and flexibility of the failure diagnosis process.

[0216] According to the dynamically updated failure probability distribution data, the failure risk of each part of the equipment can be more accurately evaluated, thereby improving the accuracy of equipment maintenance and avoiding unnecessary maintenance or neglecting potential failure points.

[0217] Through the dynamic weight adjustment method, especially considering the influence of equipment service life on failure probability, the aging of the equipment can be monitored in real time, enhancing the recognition and monitoring ability of the system to the aging of the equipment.

[0218] By integrating environmental data, operating parameters, and historical failure cases, the system can more accurately predict equipment failures, reducing the probability of failure and improving the overall safety and reliability of cold chain equipment.

[0219] By identifying potential failures and assessing their probability of occurrence in advance, preventive measures can be taken in time to avoid serious equipment failures, thereby reducing the cost of failure repair and equipment downtime.

[0220] By introducing a dynamic weight adjustment mechanism, the system can automatically adjust the evaluation of failure modes according to changes in the operating environment (such as temperature, humidity, etc.), enhancing the system's adaptability to different environmental changes.

[0221] By combining historical data, expert experience, and real-time data, the system uses a Bayesian network model for failure mode reasoning, improving its intelligent level and enabling autonomous dynamic decision-making and failure warning.

[0222] By combining failure correlation factor data with a Bayesian network model, more accurate decisions can be made based on data and algorithms, enhancing the accuracy of data-driven decisions.

[0223] One embodiment of the present invention, the S6, includes:

[0224] S61, reorder the candidate failure mode list according to the failure probability distribution data, arrange the candidate failure modes in order of failure probability from high to low; set a failure probability threshold, determine the failure modes with failure probability greater than the threshold as key failure modes for key analysis and processing;

[0225] S62, obtain the internal structure topology data of the cold chain equipment; based on the obtained internal structure topology data, establish a topology model of the cold chain equipment; represent the components in the equipment as nodes and the connection relationship as edges to construct a topology graph;

[0226] S63, based on the failure mode sorting list and the internal structure topology model of the cold chain equipment, use a graph theory algorithm to perform failure mode loop correlation analysis;

[0227] S64, according to the correlation analysis result, determine the loop where the failure may occur, and obtain the failure mode loop correlation table; for each failure mode, list the loops that may be affected;

[0228] S65, according to the failure mode loop correlation table and the failure mode sorting list, further narrow down the failure range and locate the failure area;

[0229] S66, according to the failure cause analysis result, generate a cold chain equipment failure cause analysis report; at the same time, provide corresponding treatment suggestions and repair schemes; provide the failure cause analysis report to the maintenance personnel.

[0230] The working principle of the above technical solution is: according to the fault probability distribution data, the candidate fault mode list is reordered, and arranged from high to low according to the fault probability. This helps to focus on fault modes with higher fault probability, thereby improving the efficiency of fault detection and diagnosis; set a fault probability threshold, and determine the fault modes with fault probability greater than the threshold as key fault modes. Focusing on the analysis and processing of these key fault modes helps to quickly locate and solve potential high-risk faults.

[0231] The internal structure topology data of the cold chain equipment is obtained, including the connection relationship between components, position information, etc. These data are the basis for establishing the topology model; based on the obtained topology data, the topology model of the cold chain equipment is established. The components in the equipment are represented as nodes, and the connection relationship is represented as edges, to construct a topology graph. This helps to visually show the structure and connection relationship inside the equipment.

[0232] Based on the fault mode ranking list and the internal structure topology model of the cold chain equipment, graph theory algorithm is used for fault mode loop correlation analysis. Analyze the association between each key fault mode and the nodes and edges in the topology graph to determine the loop position where the fault may occur; through the analysis of the graph theory algorithm, the loop position where the fault may occur can be accurately determined, providing an important basis for subsequent fault area positioning.

[0233] According to the correlation analysis results, the loop where the fault may occur is determined, and the fault mode loop correlation table is generated. For each fault mode, list the loops that may be affected, providing detailed information for fault area positioning; combining the fault mode loop correlation table and the fault mode ranking list, further narrow down the fault range. By analyzing the association between fault modes and loops, the specific area where the fault occurs can be determined; when determining the fault area, the actual layout and operation of the equipment also need to be considered to ensure the accuracy of the positioning.

[0234] According to the fault cause analysis results, a cold chain equipment fault cause analysis report is generated. The report content includes fault mode, fault area, fault cause, fault influence, etc. information, and provides corresponding treatment suggestions and maintenance schemes; the fault cause analysis report is provided to the maintenance personnel, to provide accurate maintenance guidance for the maintenance personnel. Maintenance personnel can quickly locate the fault position according to the information in the report, take effective maintenance measures, and improve the efficiency and quality of maintenance.

[0235] The effect of the above technical solution is: by sorting fault modes according to fault probability and focusing on high-probability fault modes, the efficiency of fault detection and diagnosis can be improved, and the waste of resources on low-probability faults is reduced.

[0236] By combining the fault mode circuit correlation table and the device topology graph, the fault area can be accurately located, reducing the uncertainty in locating the fault and improving the accuracy of fault location.

[0237] By combining fault mode sorting and device topology model for circuit correlation analysis, the correlation between fault modes and internal structure of the device can be comprehensively analyzed, ensuring the completeness and comprehensiveness of fault cause analysis.

[0238] By generating detailed fault cause analysis report and providing clear fault area and cause information, maintenance personnel can quickly locate and handle faults, significantly reducing maintenance time.

[0239] By sorting fault modes and setting fault probability threshold, high-probability fault modes can be controlled and focused on, effectively reducing the risk of potential fault modes in complex systems.

[0240] Based on device topology data and fault mode sorting, accurate fault prediction can be made for each part of the internal structure of the device, reducing the impact of potential faults on the overall performance of the device.

[0241] By providing accurate and comprehensive fault cause analysis report, it helps maintenance personnel quickly understand the root cause of the fault, reduces the possibility of misjudgment and error repair, and improves maintenance quality and efficiency.

[0242] By accurately sorting and analyzing the fault modes of cold chain equipment, potential problems can be discovered and preventive maintenance can be performed in a timely manner, thereby enhancing the reliability of the equipment and reducing downtime.

[0243] Using graph algorithm and topology model for fault circuit analysis and correlation, the intelligent level of the system in fault detection and diagnosis is improved, which can autonomously identify and infer the possibility of fault occurrence.

[0244] By generating detailed fault cause report, maintenance personnel can make more accurate decisions based on data-driven analysis, improving the effectiveness of maintenance work.

[0245] One embodiment of the present application is a sensor-based cold chain equipment fault detection system, which includes a memory, a processor and a computer program stored in the memory and executable on the memory, the processor executes the program to realize the sensor-based cold chain equipment fault detection method as described in any one of the above.

[0246] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A sensor-based fault detection method for cold chain equipment, characterized in that, The method includes: S1. Real-time collection of multi-dimensional data on the operating status of cold chain equipment is achieved through multi-dimensional sensors; the collected multi-dimensional data is cleaned and normalized to obtain a standardized multidimensional dataset. S2. Principal component analysis (PCA) is used to reduce the dimensionality of the standardized multidimensional dataset, extract the main feature components in the data, and construct a high-dimensional feature vector set. Then, the LSTM model is used to extract the temporal features of the multidimensional dataset to obtain a multidimensional temporal feature matrix. The high-dimensional feature vector set and the multidimensional temporal feature matrix are fused to obtain a fused feature matrix. S3. Obtain historical failure case data of cold chain equipment, divide the fused feature matrix into feature subsets, use correlation analysis to calculate the correlation degree between different feature subsets and historical failure case data, and obtain the feature correlation degree matrix; based on the feature correlation degree matrix, select feature subsets related to failure, further analyze the relevant feature subsets, extract failure feature fragments, and use feature recombination technology to reconstruct the failure feature fragments into failure feature vectors. S4. Based on historical failure case data, construct a failure mode feature knowledge base, describe each failure mode in detail, and use the support vector machine algorithm to classify and identify the failure feature vectors to obtain a candidate failure mode list; based on the fuzzy comprehensive evaluation method, generate a failure similarity vector according to the feature matching degree between the failure feature vector and each failure mode in the failure mode feature knowledge base. S5. Obtain operating environment data of cold chain equipment and construct a fault correlation factor dataset; use a Bayesian network model, combined with historical fault case data and the fault correlation factor dataset, to conduct a preliminary assessment of the occurrence probability of candidate fault modes and obtain an initial fault mode probability vector; adopt a dynamic weight adjustment method to dynamically correct the initial fault mode probability vector according to the real-time changes of the fault correlation factor dataset and obtain fault probability distribution data. S6. Reorder the candidate fault mode list based on the fault probability distribution data to obtain a fault mode ranking list based on fault probability; obtain the internal structural topology data of the cold chain equipment; based on the fault mode ranking list and the internal structural topology data of the cold chain equipment, use graph theory algorithm to perform fault mode loop association analysis to obtain a fault mode loop association table; obtain fault area location result data based on the fault mode loop association table and the fault mode ranking list; combine the fault area location result data and the fault mode feature knowledge base to conduct in-depth analysis of the fault causes and generate a cold chain equipment fault cause analysis report.

2. The sensor-based cold chain equipment fault detection method according to claim 1, characterized in that, S1 includes: S11. Collect data output from each sensor in real time through multi-dimensional sensors, set the data acquisition frequency, and determine the acquisition interval based on the operating characteristics and fault detection requirements of the cold chain equipment. S12 stores the collected raw data in a database, using a time-series database, and backs up the stored data. S13 uses a digital filtering algorithm to remove noise from the collected data; it uses a combination of statistical methods and machine learning algorithms to detect outliers; and it replaces detected outliers with interpolation or predicted values ​​based on historical data. S14 uses the max-min normalization method to map the cleaned data to the [0,1] interval, resulting in a standardized multidimensional dataset.

3. The sensor-based cold chain equipment fault detection method according to claim 1, characterized in that, The S2 includes: S21. Calculate the covariance matrix of the standardized multidimensional dataset, and obtain the linear relationship between different features by calculating the covariance matrix; S22. Solve for the eigenvalues ​​and eigenvectors of the covariance matrix; where the eigenvalues ​​represent the amount of information contained in the corresponding eigenvectors, and the eigenvectors represent the trend of data change in that direction; select the first k principal components based on the magnitude of the eigenvalues; project the original data onto the selected principal components to construct a high-dimensional eigenvector set; S23. Divide the standardized cube into training and test sets in chronological order; use the training set to train the LSTM model and adjust the model parameters using the backpropagation algorithm; S24. Use the trained LSTM model to extract temporal features from the multidimensional dataset; input the data into the model to obtain the output of the hidden layer of the model, and form a multidimensional temporal feature matrix; S25. Based on the characteristics of high-dimensional feature vector sets and multi-dimensional time-series feature matrices, a feature fusion strategy is formulated. The two feature matrices are concatenated in the feature dimension using a feature concatenation method, so that the fused feature matrix contains both the main features and time-series features of the data. S26. According to the established fusion strategy, the high-dimensional feature vector set and the multi-dimensional time series feature matrix are fused to generate a fused feature matrix.

4. The sensor-based cold chain equipment fault detection method according to claim 3, characterized in that, S22 includes: Using the eigenvalue decomposition method in linear algebra, mathematical operations are performed on the covariance matrix to solve for its eigenvalues ​​and corresponding eigenvectors. The obtained eigenvalues ​​are analyzed; by plotting eigenvalue distribution maps and information contribution rate curves, the information importance ranking of each eigenvector is displayed intuitively. The number of principal components k is dynamically determined using the cumulative contribution rate method. A cumulative contribution rate threshold is set, and the information contribution rates corresponding to each feature value are accumulated sequentially in descending order of feature value. When the cumulative contribution rate reaches or exceeds the set threshold for the first time, the number of principal components k is obtained. Based on the determined number of principal components k, the eigenvectors corresponding to the first k eigenvalues ​​are selected as principal components; the original standardized multidimensional dataset is projected onto the selected principal components to perform dimensionality reduction; the original data matrix is ​​multiplied by the selected principal component matrix through matrix multiplication to obtain the dimensionality-reduced data matrix, i.e., the high-dimensional eigenvector set.

5. The sensor-based cold chain equipment fault detection method according to claim 3, characterized in that, S26 includes: Before feature fusion, the high-dimensional feature vector set and the multi-dimensional temporal feature matrix are pre-checked; the high-dimensional feature vector set and the multi-dimensional temporal feature matrix are then processed based on the pre-check results. The processed high-dimensional feature vector set and multi-dimensional temporal feature matrix are subjected to dimensional matching and adjustment; according to the established feature splicing strategy, the high-dimensional feature vector set and multi-dimensional temporal feature matrix are spliced ​​in the feature dimension; after splicing, a preliminary fusion feature matrix is ​​generated. The quality of the initially generated fusion feature matrix is ​​evaluated, and the fusion feature matrix after quality evaluation is normalized.

6. The sensor-based cold chain equipment fault detection method according to claim 1, characterized in that, The S3 includes: S31. Obtain historical failure case data of cold chain equipment, organize the collected historical failure case data, and establish a historical failure case database; classify and label the data, and add tags to each failure case; S32. Based on the properties and sources of the features in the fusion feature matrix, the features are divided into different subsets; the correlation degree is calculated between different feature subsets and historical fault case data using the correlation degree calculation method; for each feature subset, the correlation degree between it and each fault case data is calculated to obtain the feature correlation degree matrix; S33. Analyze the feature correlation matrix and select the feature subset that is highly correlated with the fault; set a correlation threshold and take the feature subset with a correlation greater than the threshold as the feature subset that is highly correlated with the fault. S34. Further analyze the selected fault-related feature subset and extract fault feature fragments using sliding window and feature clustering methods. S35. Using feature recombination technology, the extracted fault feature fragments are combined and reconstructed; feature splicing is used to combine different fault feature fragments into a fault feature vector.

7. The sensor-based cold chain equipment fault detection method according to claim 1, characterized in that, The S4 includes: S41. Based on historical failure case data, construct a failure mode feature knowledge base; provide a detailed description of each failure mode and establish a mapping relationship between failure modes and feature vectors; continuously improve and update the failure mode feature knowledge base. S42. Use the support vector machine algorithm to classify and identify fault feature vectors; use the fault feature vectors in historical fault case data as training samples and the corresponding fault modes as labels to train the support vector machine model. S43. Use the trained support vector machine model to classify and identify the new fault feature vectors; input the fault feature vectors into the model, and the model outputs the corresponding fault modes to obtain a list of candidate fault modes. S44. Based on the fuzzy comprehensive evaluation method, construct a fault similarity evaluation model; according to the degree of feature matching between the fault feature vector and each fault mode in the fault mode feature knowledge base, use the fuzzy comprehensive evaluation model to calculate the similarity between the fault feature vector and each fault mode; combine the similarity results into a fault similarity vector.

8. The sensor-based cold chain equipment fault detection method according to claim 1, characterized in that, The S5 includes: S51. Obtain operating environment data of cold chain equipment as fault-related factor data; preprocess the collected fault-related factor data; S52. Using a Bayesian network model, combined with historical failure case data and failure correlation factor dataset, construct a failure probability assessment model. S53. Based on the constructed Bayesian network model, calculate the probability of occurrence of candidate fault modes under the current fault association factors to obtain the initial fault mode probability vector; and update the probability distribution of each fault mode node based on the known fault association factor data using the Bayesian network inference algorithm. S54. Develop a dynamic weight adjustment method, determine the initial weight of each factor based on historical data and expert experience, and establish weight adjustment rules. S55. A dynamic weight adjustment method is adopted to dynamically correct the initial failure mode probability vector based on the real-time changes of the failure-related factor dataset; the weights of each factor are updated periodically, and the occurrence probability of candidate failure modes is recalculated to obtain more accurate failure probability distribution data, which reflects the failure risk of cold chain equipment in real time.

9. The sensor-based cold chain equipment fault detection method according to claim 1, characterized in that, The S6 includes: S61. Reorder the candidate fault mode list according to the fault probability distribution data, and arrange the candidate fault modes in order of fault probability from high to low; set a fault probability threshold, and identify fault modes with a fault probability greater than the threshold as key fault modes for key analysis and processing. S62. Obtain the internal structure topology data of the cold chain equipment, and establish a topology model of the cold chain equipment based on the obtained internal structure topology data; represent the components in the equipment as nodes, the connection relationship as edges, and construct a topology graph; S63. Based on the fault mode ranking list and the internal structural topology model of cold chain equipment, graph theory algorithm is used to perform fault mode loop association analysis. S64. Based on the correlation analysis results, determine the loops where the fault may occur and obtain the fault mode loop correlation table; for each fault mode, list the loops it affects. S65. Based on the fault mode loop association table and the fault mode sorting list, further narrow down the fault range and locate the fault area. S66. Based on the results of the fault cause analysis, generate a fault cause analysis report for the cold chain equipment; at the same time, provide corresponding handling suggestions and maintenance plans; and provide the fault cause analysis report to maintenance personnel.

10. A sensor-based fault detection system for cold chain equipment, characterized in that, It includes a memory, a processor, and a computer program stored on and executable on the memory, wherein the processor executes the program to perform the functions as claimed in the claim. The sensor-based cold chain equipment fault detection method described in any of the above.

Images