A refrigerator fault remote diagnosis method and system

Abstract

This invention belongs to the field of data processing technology, specifically relating to a method and system for remote diagnosis of refrigerator faults. The method includes: acquiring time-series data from multiple sensors within a preset time window of the refrigerator; calculating the time-domain cross-correlation coefficients between pairs of data and the state transition entropy of each data point; combining these to form a diagnostic feature vector and constructing a decision tree; performing post-pruning after construction; obtaining the fault diagnosis types of all leaf nodes in the subtree to be pruned; calculating the weighted fault severity based on a preset fault level knowledge base; calculating a pruning threshold by combining the weak splitting degree of the root node of the subtree to be pruned with the weighted fault severity; if the increase in the validation set error rate caused by replacing the root node with the subtree to be pruned is lower than this threshold, then pruning is performed; for leaf nodes containing multiple fault types, determining the type with the highest fault level as its diagnostic result. This invention improves the reliability and security of remote diagnosis of refrigerator faults.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method and system for remote diagnosis of refrigerator faults. Background Technology

[0002] With the development of smart home technology, modern refrigerators have evolved from simple refrigeration devices into intelligent terminals integrating multiple sensors. Internally deployed temperature, humidity, and current sensors monitor the refrigerator's operating status in real time. These sensors generate massive amounts of multi-channel time-series data during operation, providing crucial data support for remote refrigerator fault diagnosis. Against this backdrop, machine learning methods, especially decision tree models, have become the mainstream technology choice for analyzing sensor data and achieving fault identification due to their strong interpretability and adaptability to structured data.

[0003] However, in practical applications, directly constructed decision tree models are susceptible to interference from data noise and random features, thus affecting the accuracy and reliability of diagnosis. Specifically, unoptimized decision tree models are prone to learning noise and random features in the data during training, resulting in overly complex tree structures and overfitting. To address overfitting, existing technologies typically employ pruning strategies, mainly including pre-pruning and post-pruning. Pre-pruning terminates tree growth early during the decision tree generation process by setting information gain thresholds or node sample number thresholds. Post-pruning, on the other hand, examines non-leaf nodes from the bottom up after generating a complete decision tree. If replacing a non-leaf node with a leaf node improves the model's performance on the validation set, then pruning is performed.

[0004] While pruning strategies can mitigate overfitting to some extent, they still have significant shortcomings in specific scenarios like refrigerator fault diagnosis. In refrigerator fault diagnosis, some faults are far more severe than others. For example, minor refrigerant leaks and fan stalls, while affecting operational efficiency, do not cause immediate equipment failure. However, compressor faults, if not identified promptly, can lead to safety hazards and high repair costs. Existing pre-pruning algorithms rely on manually preset fixed thresholds to prematurely terminate decision tree growth, which cannot match the varying risks of refrigerator faults. Existing post-pruning algorithms, such as error rate reduction pruning or cost complexity pruning, generally lack flexible mechanisms to adjust the pruning intensity. This prevents them from combining the node's partitioning ability with the actual risk level of the faults involved, easily resulting in the incorrect removal of weak branches that distinguish critical faults but have low information gain. Furthermore, pruned leaf nodes may contain samples of multiple fault types. In such cases, existing methods often use a majority rule to determine the diagnostic result. This approach can lead to a large number of low-level faults or normal state samples masking high-level severe faults, causing the diagnostic results to deviate from the actual risk, ultimately leading to safety hazards and economic losses. Summary of the Invention

[0005] Therefore, the purpose of this invention is to propose a remote diagnosis method and system for refrigerator faults, in order to solve the technical problem that the existing decision tree model has poor reliability in remote diagnosis of refrigerator faults, which can easily lead to safety hazards and economic losses.

[0006] To solve the above problems, the technical solution of the remote diagnosis method for refrigerator faults proposed in this invention is as follows:

[0007] Remote diagnostic methods for refrigerator malfunctions include the following steps:

[0008] Acquire multi-channel sensor time-series data of the refrigerator within a preset time window, calculate the time-domain cross-correlation coefficient between each pair of sensor time-series data and the state transition entropy of each sensor time-series data, and combine them into a diagnostic feature vector; construct a decision tree based on the diagnostic feature vector. During the construction process, when the splitting information gain of a non-leaf node is lower than the first threshold, a weak splitting degree inversely proportional to the information gain is recorded for that non-leaf node.

[0009] After pruning the constructed decision tree, for any subtree to be pruned, obtain the fault diagnosis type corresponding to all its leaf nodes, and calculate the weighted fault severity of the subtree according to the preset fault level knowledge base.

[0010] A pruning threshold is calculated based on the weak split degree of the root node of the subtree and the weighted fault severity. The pruning threshold is directly proportional to the weak split degree and inversely proportional to the weighted fault severity. When the increase in the validation set error rate caused by replacing the root node with the subtree is lower than the pruning threshold, the subtree is pruned.

[0011] For any leaf node of the pruned decision tree, if the leaf node contains training samples belonging to multiple fault types, then according to the fault level knowledge base, the type with the highest fault level among the multiple fault types is determined as the diagnosis result of the leaf node.

[0012] Furthermore, the method for obtaining the diagnostic feature vector is as follows:

[0013] The timing data of four sensors—compressor temperature sensor, condenser temperature sensor, evaporator temperature sensor, and internal temperature sensor—are collected within a preset time period at a preset sampling frequency.

[0014] The value range of time-series data from each sensor is discretized into multiple intervals, the data sequence is converted into a state sequence, a state transition probability matrix is ​​constructed, and the state transition entropy is calculated based on the state transition probability matrix.

[0015] The calculated values ​​of all time-domain cross-relationships are sequentially concatenated with all state transition entropy values ​​to form the diagnostic feature vector.

[0016] Furthermore, the method for calculating the weak splitting degree of the non-leaf nodes is as follows:

[0017] Weak splitting degree is defined as The splitting information gain of non-leaf nodes is ;

[0018] For any non-leaf node, if the split information gain of the non-leaf node is... If it is below the first threshold, then ;in, To prevent the smoothing constant from having a denominator of zero;

[0019] If the non-leaf node splits information gain If it is not lower than the first threshold, then It is 0.

[0020] Furthermore, the method for calculating the weighted fault severity is as follows:

[0021] Establish the aforementioned fault level knowledge base, and assign a unique, numerical fault level to each fault type and normal state;

[0022] By statistically analyzing all training samples covered by the subtree to be pruned, calculating the proportion of each fault type in the subtree, and multiplying it by the corresponding fault level in the fault level knowledge base, the weighted average of the fault levels of the subtree is calculated as the weighted fault severity.

[0023] Furthermore, the pruning threshold is calculated as follows:

[0024] The weighted fault severity is defined as follows: The pruning threshold is The calculation formula is: ;in, For balance coefficient, Used to avoid a denominator of zero.

[0025] Further, the step of pruning the subtree when the increment of the validation set error rate caused by replacing the subtree with the root node is lower than the pruning threshold includes:

[0026] Calculate the number of misclassified samples of the subtree on the validation set. ;

[0027] Calculate the number of misclassified samples of the root node on the validation set after replacing the subtree with the root node. ;

[0028] Calculate the error rate increment , , The total number of samples in the validation set;

[0029] when < When the time comes, a pruning operation is performed, replacing the root node with the subtree.

[0030] Furthermore, the method for determining the diagnostic result is as follows:

[0031] If a leaf node contains at least two different types of training samples, then query the fault level knowledge base to obtain the fault level corresponding to each type.

[0032] Compare the various fault levels and select the fault type with the highest level value as the diagnosis result of the leaf node.

[0033] The technical solution of the refrigerator fault remote diagnosis system proposed in this invention is as follows:

[0034] The refrigerator remote fault diagnosis system includes the following modules:

[0035] The module is used to acquire time-series data from multiple sensors of the refrigerator within a preset time window, calculate the time-domain cross-correlation coefficient between any two pairs of sensor time-series data and the state transition entropy of each sensor time-series data, and combine them into a diagnostic feature vector; a decision tree is constructed based on the diagnostic feature vector. During the construction process, when the splitting information gain of a non-leaf node is lower than the first threshold, a weak splitting degree that is inversely proportional to the information gain is recorded for that non-leaf node.

[0036] The calculation module is used to perform post-pruning on the constructed decision tree. For any subtree to be pruned, it obtains the fault diagnosis type corresponding to all its leaf nodes and calculates the weighted fault severity of the subtree according to the preset fault level knowledge base.

[0037] The pruning module is used to calculate a pruning threshold based on the weak split degree of the root node of the subtree and the weighted fault severity. The pruning threshold is directly proportional to the weak split degree and inversely proportional to the weighted fault severity. When the increase in the validation set error rate caused by replacing the root node with the subtree is lower than the pruning threshold, the subtree is pruned.

[0038] The determination module is used to determine the highest fault level among the multiple fault types for any leaf node of the pruned decision tree if the leaf node contains training samples belonging to multiple fault types, based on the fault level knowledge base, and use this as the diagnosis result of the leaf node.

[0039] Furthermore, the method for obtaining the diagnostic feature vector is as follows:

[0040] The timing data of four sensors—compressor temperature sensor, condenser temperature sensor, evaporator temperature sensor, and internal temperature sensor—are collected within a preset time period at a preset sampling frequency.

[0041] The value range of time-series data from each sensor is discretized into multiple intervals, the data sequence is converted into a state sequence, a state transition probability matrix is ​​constructed, and the state transition entropy is calculated based on the state transition probability matrix.

[0042] The calculated values ​​of all time-domain cross-relationships are sequentially concatenated with all state transition entropy values ​​to form the diagnostic feature vector.

[0043] Furthermore, the method for calculating the weak splitting degree of the non-leaf nodes is as follows:

[0044] Weak splitting degree is defined as The splitting information gain of non-leaf nodes is ;

[0045] For any non-leaf node, if the split information gain of the non-leaf node is... If it is below the first threshold, then ;in, To prevent the smoothing constant from having a denominator of zero;

[0046] If the non-leaf node splits information gain If it is not lower than the first threshold, then It is 0.

[0047] Furthermore, the method for calculating the weighted fault severity is as follows:

[0048] Establish the aforementioned fault level knowledge base, and assign a unique, numerical fault level to each fault type and normal state;

[0049] By statistically analyzing all training samples covered by the subtree to be pruned, calculating the proportion of each fault type in the subtree, and multiplying it by the corresponding fault level in the fault level knowledge base, the weighted average of the fault levels of the subtree is calculated as the weighted fault severity.

[0050] Furthermore, the pruning threshold is calculated as follows:

[0051] The weighted fault severity is defined as follows: The pruning threshold is The calculation formula is: ;in, For balance coefficient, Used to avoid a denominator of zero.

[0052] Further, the step of pruning the subtree when the increment of the validation set error rate caused by replacing the subtree with the root node is lower than the pruning threshold includes:

[0053] Calculate the number of misclassified samples of the subtree on the validation set. ;

[0054] Calculate the number of misclassified samples of the root node on the validation set after replacing the subtree with the root node. ;

[0055] Calculate the error rate increment , , The total number of samples in the validation set;

[0056] when < When the time comes, a pruning operation is performed, replacing the root node with the subtree.

[0057] Furthermore, the method for determining the diagnostic result is as follows:

[0058] If a leaf node contains at least two different types of training samples, then query the fault level knowledge base to obtain the fault level corresponding to each type.

[0059] Compare the various fault levels and select the fault type with the highest level value as the diagnosis result of the leaf node.

[0060] The beneficial effects of this invention are as follows: First, by collecting time-series data from multiple sensors in the refrigerator, calculating the time-domain cross-correlation coefficient and state transition entropy, and constructing a multi-dimensional diagnostic feature vector, this invention comprehensively captures the correlation characteristics and change patterns of the operating status of the refrigerator's core components, providing an accurate data foundation for fault identification. Second, by recording weak splitting degrees during the decision tree construction process and combining weak splitting degrees with weighted fault severity to calculate dynamic pruning thresholds during the pruning stage, this invention avoids the problem of key branch deletion or redundant branch retention caused by the rigidity of fixed thresholds in traditional pre-pruning. It also solves the defect of traditional post-pruning that ignores fault risk levels based solely on error rate, effectively protecting the diagnostic path for distinguishing high-severity faults. Simultaneously, it prunes low-value redundant branches to suppress model overfitting, improving the model's generalization ability to refrigerator operating data under different working conditions and environments. Finally, by adopting a decision rule that prioritizes the highest fault level for leaf nodes after pruning, this invention solves the drawback of the traditional majority rule masking high-risk faults, ensuring that serious faults can be identified and warned in a timely manner, thus enhancing the reliability and security of remote refrigerator fault diagnosis. Attached Figure Description

[0061] Figure 1 This is a flowchart of the steps in the remote diagnosis method for refrigerator faults according to the present invention. Detailed Implementation

[0062] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0063] Specific embodiments of the remote refrigerator fault diagnosis method proposed in this invention:

[0064] like Figure 1 As shown, the remote diagnostic method for refrigerator faults includes the following steps:

[0065] S1. Obtain multi-channel sensor time-series data of the refrigerator within a preset time window, calculate the time-domain cross-correlation coefficient between each pair of sensor time-series data and the state transition entropy of each sensor time-series data, and combine them into a diagnostic feature vector; construct a decision tree based on the diagnostic feature vector. During the construction process, when the splitting information gain of a non-leaf node is lower than the first threshold, a weak splitting degree inversely proportional to the information gain is recorded for that non-leaf node.

[0066] In this step, the time-domain cross-correlation coefficient is used to describe the temporal correlation of time-series data from different sensors, and the state transition entropy is used to describe the degree of disorder or uncertainty in the changes of time-series data from a single sensor between different sensor states.

[0067] In this embodiment, a compressor temperature sensor, a condenser temperature sensor, an evaporator temperature sensor, and an internal temperature sensor are deployed inside the refrigerator. Data is continuously collected at a frequency of 1Hz and captured within a preset time window of 30 minutes, forming four channels of sensor time-series data. For any two channels of sensor time-series data, the Pearson correlation coefficients of these two channels under different time delays are calculated to obtain a time-domain cross-correlation sequence. The value range of each channel of sensor time-series data is divided into 5 equal parts, i.e., multi-interval discretization is performed, discretizing the data sequence into a state sequence, constructing a 5×5 state transition probability matrix, and calculating the state transition entropy based on the state transition probability matrix. The 6 time-domain cross-correlation values ​​and 4 state transition entropy values ​​are sequentially concatenated to form a 10-dimensional diagnostic feature vector.

[0068] Specifically, the time-series data from the four sensors collected yields a time series containing 1800 data points for each sensor. The data range of each time series is divided into five equally wide intervals. For example, the compressor temperature range of 20℃ to 70℃ is divided into five states: 20℃~30℃, 30℃~40℃, 40℃~50℃, 50℃~60℃, and 60℃~70℃, thus converting the original data sequence into a state sequence. The frequency of transitions from one state to another within the aforementioned time window is statistically analyzed, constructing a 5×5 state transition probability matrix. Information entropy, i.e., state transition entropy, is calculated based on this matrix using the Shannon entropy formula. Six time-domain cross-correlation values ​​are calculated between each pair of the four sensor time-series data points. For example, the cross-correlation coefficient between compressor temperature and condenser temperature is 0.85. These six time-domain cross-correlation values ​​are then sequentially concatenated with the calculated four state transition entropy values ​​to form a diagnostic feature vector containing 10 elements.

[0069] Of course, in other embodiments, the type of sensor is not limited to the four types mentioned above. Other types of sensors, such as compressor operating current sensors, can also be added. In this case, repeating the above steps can form a diagnostic feature vector under a specific time window.

[0070] The C4.5 decision tree generation algorithm is used, employing diagnostic feature vectors with fault labels as the training set to construct the decision tree. At each non-leaf node, all features in the diagnostic feature vector are traversed, and the information gain after data partitioning using each feature is calculated. The information gain is calculated by subtracting the weighted average of the information entropies of all child nodes after feature partitioning from the information entropy of the parent node. The feature with the largest information gain is selected as the splitting criterion for the non-leaf node. A first threshold for information gain is set during this process. If the maximum information gain calculated for the current non-leaf node is less than the first threshold, it is considered a weak splitting node, and a weak split degree is calculated for it. This weak split degree value is a preset constant divided by the information gain of the non-leaf node, and this weak split degree value is associated with and stored for the non-leaf node.

[0071] In an optional embodiment, the weak splitting degree of non-leaf nodes is calculated as follows:

[0072] Weak splitting degree is defined as The splitting information gain of non-leaf nodes is ;

[0073] For any non-leaf node, if the split information gain of the non-leaf node is... If it is below the first threshold, then ;in, To prevent the smoothing constant from having a denominator of zero;

[0074] If the non-leaf node splits information gain If it is not lower than the first threshold, then It is 0.

[0075] S2, perform post-pruning on the constructed decision tree. For any subtree to be pruned, obtain the fault diagnosis type corresponding to all its leaf nodes, and calculate the weighted fault severity of the subtree according to the preset fault level knowledge base.

[0076] Specifically, a fault level knowledge base is established, assigning a unique, numerical fault level to each possible fault type and normal state of the refrigerator. For example, normal operation is defined as level 0, slight frost as level 1, poor door seal as level 2, compressor starting difficulty as level 4, and refrigerant leakage as level 5. The pruning process traverses each non-leaf node of the decision tree from bottom to top. For any non-leaf node, it is considered the root node of a subtree to be pruned. All leaf nodes under this subtree are traversed, and the number and types of faults covered by the training samples in each leaf node are counted. The statistical results of all leaf nodes are summarized to obtain all fault types that the subtree can diagnose and the total number of samples for each type. The weighted fault severity of this subtree is calculated as follows: by statistically analyzing all training samples covered by the subtree, the proportion of each fault type in the subtree is calculated, and multiplied by the corresponding fault level in the fault level knowledge base, the weighted average of the fault levels of the subtree is calculated as the weighted fault severity.

[0077] In an optional embodiment, “normal” is defined as level 0, “minor refrigerant leak” is defined as level 1, “fan stall” is defined as level 2, and “compressor failure” is defined as level 3;

[0078] Define weighted fault severity as The calculation formula is as follows: ;in, This is an index of the fault types appearing in the samples covered by the subtree. Fault type The corresponding level This belongs to the fault type The number of training samples, The total number of training samples covered by the subtree.

[0079] Specifically, the established fault level knowledge base will classify the levels of normal states. Set to 0, the level of minor refrigerant leakage. Set to 1, fan stall level Let's set it to 2. Assume a subtree to be pruned covers 100 training samples, including 50 normal samples, 30 samples with minor refrigerant leaks, and 20 samples with stalled fans. Equals 50, Equals 30, The value is 20. Based on the weighted fault severity calculation formula, the weighted fault severity of the subtree to be pruned is calculated. It is 0.7.

[0080] S3. Calculate a pruning threshold based on the weak split degree of the root node of the subtree and the weighted fault severity. The pruning threshold is directly proportional to the weak split degree and inversely proportional to the weighted fault severity. When the increase in the validation set error rate caused by replacing the subtree with the root node is lower than the pruning threshold, the subtree is pruned.

[0081] In this embodiment, the pruning threshold is defined as follows: The calculation formula is as follows: ;in, For balance coefficient, To weight the severity of the fault, This is used to avoid the denominator being zero. The weak splitting degree of the root node of the subtree is calculated and stored in step S1. When performing the pruning step, it is directly read and used to calculate the above-mentioned pruning threshold.

[0082] The specific pruning process is as follows:

[0083] Calculate the number of misclassified samples of the subtree on the validation set. ;

[0084] Calculate the number of misclassified samples of the root node on the validation set after replacing the subtree with the root node. ;

[0085] Calculate the error rate increment , , The total number of samples in the validation set;

[0086] when < When the time comes, a pruning operation is performed, replacing the root node with the subtree to be pruned.

[0087] Among them, the number of misclassified samples of the subtree on the validation set The calculation method is as follows: Testing is performed on an independent validation set. The sample data in this validation set is classified through the decision paths of the subtrees. The total number of samples whose diagnostic results from the subtree to be pruned do not match the true labels of the validation set samples is counted. For example, in a validation set containing 1000 samples, the number of misclassified samples from the subtree to be pruned is recorded as 50. Equals 50. The number of misclassified samples by the root node on the validation set. The calculation method is the same as above, and will not be repeated here.

[0088] S4. For any leaf node of the pruned decision tree, if the leaf node contains training samples belonging to multiple fault types, then according to the fault level knowledge base, determine the type with the highest fault level among the multiple fault types as the diagnosis result of the leaf node.

[0089] Specifically, after all pruning is complete, each leaf node of the decision tree is examined. If a leaf node contains training samples of at least two different types, the fault level knowledge base is queried to obtain the fault level corresponding to each type. The fault levels are compared, and the fault type with the highest level value is selected as the diagnosis result for the leaf node.

[0090] In an optional embodiment, the leaf node contains training samples of two types: minor refrigerant leakage and fan stall. The fault level knowledge base is queried to obtain level 1 for minor refrigerant leakage and level 2 for fan stall. Since the level of fan stall is higher than that of minor refrigerant leakage, fan stall is determined as the diagnosis result of the leaf node.

[0091] This invention protects critical branches that distinguish severe faults from being mistakenly deleted by using a dynamic pruning strategy that combines fault severity levels. It also incorporates a leaf node determination method that prioritizes the diagnosis of the highest-level faults, thus achieving more reliable and safer remote fault diagnosis for refrigerators.

[0092] Specific embodiments of the refrigerator fault remote diagnosis system proposed in this invention:

[0093] The refrigerator remote fault diagnosis system includes the following modules:

[0094] The module is used to acquire time-series data from multiple sensors of the refrigerator within a preset time window, calculate the time-domain cross-correlation coefficient between any two pairs of sensor time-series data and the state transition entropy of each sensor time-series data, and combine them into a diagnostic feature vector; a decision tree is constructed based on the diagnostic feature vector. During the construction process, when the splitting information gain of a non-leaf node is lower than the first threshold, a weak splitting degree that is inversely proportional to the information gain is recorded for that non-leaf node.

[0095] The calculation module is used to perform post-pruning on the constructed decision tree. For any subtree to be pruned, it obtains the fault diagnosis type corresponding to all its leaf nodes and calculates the weighted fault severity of the subtree according to the preset fault level knowledge base.

[0096] The pruning module is used to calculate a pruning threshold based on the weak split degree of the root node of the subtree and the weighted fault severity. The pruning threshold is directly proportional to the weak split degree and inversely proportional to the weighted fault severity. When the increase in the validation set error rate caused by replacing the root node with the subtree is lower than the pruning threshold, the subtree is pruned.

[0097] The determination module is used to determine the highest fault level among the multiple fault types for any leaf node of the pruned decision tree if the leaf node contains training samples belonging to multiple fault types, based on the fault level knowledge base, and use this as the diagnosis result of the leaf node.

[0098] It should be noted that the processing procedures of the construction module, calculation module, pruning module, and determination module are the same as those of steps S1, S2, S3, and S4 in the above embodiment of the remote diagnosis method for refrigerator faults, and will not be described in detail here.

[0099] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0100] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A remote diagnostic method for refrigerator faults, characterized in that, Includes the following steps: Acquire multi-channel sensor time-series data of the refrigerator within a preset time window, calculate the time-domain cross-correlation coefficient between pairs of sensor time-series data and the state transition entropy of each sensor time-series data, and combine them into a diagnostic feature vector. The method for obtaining state transition entropy is as follows: divide the data range of each time series into multiple equally wide intervals to convert the original data sequence into a state sequence; The frequency of transitions from one state to another within a time window is counted, a state transition probability matrix is ​​constructed, and the information entropy, i.e., the state transition entropy, is calculated based on the state transition probability matrix. A decision tree is constructed based on the diagnostic feature vector. During the construction process, when the splitting information gain of a non-leaf node is lower than the first threshold, a weak splitting degree inversely proportional to the information gain is recorded for that non-leaf node. The calculation method is as follows: Weak splitting degree is defined as The splitting information gain of non-leaf nodes is ; For any non-leaf node, if the splitting information gain of the non-leaf node is... If it is below the first threshold, then ;in, To prevent the smoothing constant from having a denominator of zero; if the information gain of a non-leaf node is split. If it is not lower than the first threshold, then =0; Post-pruning is performed on the constructed decision tree. For any non-leaf node, it is considered the root node of a subtree to be pruned. All leaf nodes under this subtree are traversed, and the number and types of faults covered by the training samples in each leaf node are counted. For any subtree to be pruned, the fault diagnosis types corresponding to all its leaf nodes are obtained, and the weighted fault severity of the subtree is calculated according to the preset fault level knowledge base. The calculation method is as follows: Establish a fault level knowledge base and assign a unique, numerical fault level to each fault type and normal state; by statistically analyzing all training samples covered by the subtree to be pruned, calculate the proportion of each fault type in the subtree, multiply it by the corresponding fault level in the fault level knowledge base, and calculate the weighted average of the fault levels of the subtree as the weighted fault severity. The pruning threshold is calculated based on the weak split degree of the root node of the subtree and the weighted fault severity. The pruning threshold is directly proportional to the weak split degree and inversely proportional to the weighted fault severity. When the increase in the validation set error rate caused by replacing the root node with the subtree is lower than the pruning threshold, the subtree is pruned. For any leaf node of the pruned decision tree, if the leaf node contains training samples belonging to multiple fault types, then based on the fault level knowledge base, the type with the highest fault level among the multiple fault types is determined as the diagnosis result of the leaf node.

2. The remote fault diagnosis method for refrigerators according to claim 1, characterized in that, The method for obtaining the diagnostic feature vector is as follows: The timing data of four sensors—compressor temperature sensor, condenser temperature sensor, evaporator temperature sensor, and internal temperature sensor—are collected within a preset time period at a preset sampling frequency.

3. The remote fault diagnosis method for refrigerators according to claim 1, characterized in that, The pruning threshold is calculated as follows: The weighted fault severity is defined as follows: The pruning threshold is The calculation formula is: ;in, For balance coefficient, Used to avoid a denominator of zero.

4. The remote fault diagnosis method for refrigerators according to claim 3, characterized in that, The step of pruning the subtree when the increment of the validation set error rate caused by replacing the subtree with the root node is lower than the pruning threshold includes: Calculate the number of misclassified samples of the subtree on the validation set. ; Calculate the number of misclassified samples of the root node on the validation set after replacing the subtree with the root node. ; Calculate the error rate increment , , The total number of samples in the validation set; when < When the time comes, a pruning operation is performed, replacing the root node with the subtree.

5. The remote fault diagnosis method for refrigerators according to claim 1, characterized in that, The method for determining the diagnostic result is as follows: If a leaf node contains at least two different types of training samples, then query the fault level knowledge base to obtain the fault level corresponding to each type. Compare the various fault levels and select the fault type with the highest level value as the diagnosis result of the leaf node.

6. A remote diagnostic system for refrigerator faults, characterized in that, Includes the following modules: The module is used to acquire multi-channel sensor time-series data of the refrigerator within a preset time window, calculate the time-domain cross-correlation coefficient between pairs of sensor time-series data and the state transition entropy of each sensor time-series data, and combine them into a diagnostic feature vector. The method for obtaining state transition entropy is as follows: divide the data range of each time series into multiple equally wide intervals to convert the original data sequence into a state sequence; The frequency of transitions from one state to another within a time window is counted, a state transition probability matrix is ​​constructed, and the information entropy, i.e., the state transition entropy, is calculated based on the state transition probability matrix. A decision tree is constructed based on the diagnostic feature vector. During the construction process, when the splitting information gain of a non-leaf node is lower than the first threshold, a weak splitting degree inversely proportional to the information gain is recorded for that non-leaf node. The calculation method is as follows: Weak splitting degree is defined as The splitting information gain of non-leaf nodes is ; For any non-leaf node, if the splitting information gain of the non-leaf node is... If it is below the first threshold, then ;in, To prevent the smoothing constant from having a denominator of zero; if the information gain of a non-leaf node is split. If it is not lower than the first threshold, then =0; The calculation module is used to perform post-pruning on the constructed decision tree. For any non-leaf node, it is treated as the root node of a subtree to be pruned. All leaf nodes under this subtree are traversed, and the number and types of faults covered by the training samples in each leaf node are counted. For any subtree to be pruned, the fault diagnosis types corresponding to all its leaf nodes are obtained, and the weighted fault severity of the subtree is calculated based on a preset fault level knowledge base. The calculation method is as follows: Establish a fault level knowledge base and assign a unique, numerical fault level to each fault type and normal state; by statistically analyzing all training samples covered by the subtree to be pruned, calculate the proportion of each fault type in the subtree, multiply it by the corresponding fault level in the fault level knowledge base, and calculate the weighted average of the fault levels of the subtree as the weighted fault severity. The pruning module is used to calculate the pruning threshold based on the weak split degree of the root node of the subtree and the weighted fault severity. The pruning threshold is directly proportional to the weak split degree and inversely proportional to the weighted fault severity. When the increase in the validation set error rate caused by replacing the root node with the subtree is lower than the pruning threshold, the subtree is pruned. The determination module is used to determine the highest fault level among the multiple fault types for any leaf node of the pruned decision tree if the leaf node contains training samples belonging to multiple fault types, based on the fault level knowledge base, and use this as the diagnosis result of the leaf node.

7. The remote fault diagnosis system for refrigerators according to claim 6, characterized in that, The method for obtaining the diagnostic feature vector is as follows: The timing data of four sensors—compressor temperature sensor, condenser temperature sensor, evaporator temperature sensor, and internal temperature sensor—are collected within a preset time period at a preset sampling frequency.

Images