Method, device and equipment for predicting risk of failure recurrence of maintenance service and storage medium

Abstract

The application discloses a kind of maintenance service failure recurrence risk pre-judgment method, belong to data processing technical field, comprising: to maintenance process static data, user evaluation semantic data, equipment basic attribute data, equipment runtime sequence data and environmental auxiliary time sequence data are preprocessed;Dynamic time sequence feature is input into AT-LSTM model, and output high-order time sequence characteristic vector and first preliminary recurrence probability;Static cross feature, high-order time sequence characteristic vector and environmental influence coefficient are input into DBN model, and output second preliminary recurrence probability;Based on first preliminary recurrence probability and second preliminary recurrence probability, fusion probability is calculated;Based on expert rule and environmental influence coefficient, fusion probability is calibrated, and output final recurrence probability;Determine the recurrence risk level corresponding to final recurrence probability.The application can improve the accuracy of failure recurrence risk level evaluation.The application also discloses a device, electronic equipment and computer readable storage medium for implementing the above method.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, equipment and storage medium for predicting the risk of recurrence of maintenance service failures. Background Technology

[0002] Electronic equipment is prone to failure and requires repair due to component aging, thermal damage, voltage fluctuations, and other factors. Incomplete repairs or the presence of underlying causes can lead to repeated failures in a short period, resulting in increased downtime losses and decreased customer satisfaction. Therefore, predicting the risk of recurrence allows for verification of repair effectiveness, timely identification of residual risks, and triggering secondary repairs or component replacements to prevent secondary failures. Furthermore, this predictive data can be fed back into maintenance strategy optimization to improve service quality and spare parts management efficiency, thereby reducing overall maintenance costs.

[0003] Currently, a common approach to predict the risk of recurrence of faults after maintenance is the combination of traditional static features and a single model. Specifically, static feature data such as equipment basic information (model, age, number of maintenance histories), fault type codes, and maintenance operation records are used, followed by binary classification prediction (recurrence / non-recurrence) using logistic regression, random forest, or a single neural network. However, static features cannot capture the dynamic degradation trends and real-time fluctuations in equipment operation, and a single model struggles to handle complex nonlinear interactions, exhibits poor generalization ability for rare fault modes, and cannot quantify prediction uncertainty. This results in a significant decrease in prediction accuracy as equipment ages, leading to high false alarm and false negative rates. Summary of the Invention To address the aforementioned problems in the prior art, this invention provides a method, apparatus, equipment, and storage medium for predicting the risk of recurrence of maintenance service failures. The technical problem to be solved by this invention is achieved through the following technical solution: The first aspect of this invention provides a method for predicting the risk of recurrence of maintenance service failures, comprising: Acquire core maintenance data; wherein, the core maintenance data includes static maintenance process data, user evaluation semantic data, basic equipment attribute data, equipment runtime sequence data, and environmental auxiliary time sequence data; The static data of the maintenance process, the semantic data of user evaluation, and the basic attribute data of the equipment are preprocessed to generate static cross features; The device runtime timing data is preprocessed to generate dynamic timing features; The environmental auxiliary time series data is preprocessed to generate environmental impact coefficients; The dynamic temporal features are input into the AT-LSTM model, which outputs a high-order temporal feature vector and a first preliminary recurrence probability. The model input features are input into the DBN model, and the second preliminary relapse probability is output; wherein, the model input features include the static cross features, the higher-order time series feature vector, and the environmental influence coefficient; The fusion probability is calculated based on the first preliminary recurrence probability and the second preliminary recurrence probability; The fusion probability is calibrated based on expert rules and the environmental impact coefficient, and the final recurrence probability is output. Determine the recurrence risk level corresponding to the final recurrence probability.

[0004] The method provided by this invention addresses the core needs of service platform maintenance operations for "accurate prediction of fault recurrence and proactive preventative maintenance." It breaks through the limitations of traditional static features and single models. First, it refines core maintenance data through specialized preprocessing (e.g., time-series completion, smoothing and derivation, and correction of operating parameters by combining environmental data). Then, it uses attention LSTM to capture dynamic precursors of faults, integrates dynamic Bayesian networks to quantify probabilities, and outputs the final risk after adaptive weights and expert rule calibration. This method can capture dynamic degradation trends and real-time operating condition fluctuations in equipment operation, consider performance monitoring data and changes in environmental factors in the short term after maintenance, handle complex nonlinear interaction relationships, has strong generalization ability for rare fault modes, and can quantify prediction uncertainty, thereby improving prediction accuracy.

[0005] In one possible implementation, the preprocessing of the static maintenance process data, user evaluation semantic data, and equipment basic attribute data to generate static cross features includes: The static data of the maintenance process, the basic attribute data of the equipment, and the semantic data of the user evaluation are subjected to data standardization processing to generate the static basic features; wherein, the static basic features include standardized values ​​of the maintenance process, standardized values ​​of the basic attributes of the equipment, and standardized values ​​of the semantic data of the user evaluation. Based on the standardized values ​​of the maintenance process and the standardized values ​​of the equipment's basic attributes, fault association cross terms are constructed, and the static cross features are generated. The preprocessing of the environmental auxiliary time-series data to generate environmental impact coefficients includes: The environmental auxiliary time series data are sequentially completed and smoothed to generate the environmental time series processing features; The environmental time-series processing characteristics are quantified to generate the environmental impact coefficient; The dynamic timing features include device timing processing features and timing derived features; the preprocessing of the device runtime timing data to generate dynamic timing features includes: Based on the aforementioned environmental timing processing characteristics, environmental deviation correction is performed on the device runtime timing data. The corrected device runtime timing data is sequentially completed and smoothed to generate the device timing processing features. The device's timing processing characteristics are dynamically trend-derived, and abnormal parameter precursors are captured to generate the timing-derived characteristics.

[0006] In one possible implementation, the AT-LSTM model includes an LSTM layer, an attention mechanism layer, and a fully connected layer; the step of inputting the dynamic temporal features into the AT-LSTM model and outputting a high-order temporal feature vector and a first preliminary recurrence probability includes: The dynamic temporal features are input into the LSTM layer, and the hidden state sequence of multiple time steps is output. The hidden state sequences of the multiple time steps are input into the attention mechanism layer, the attention score of the hidden state at each time step is calculated, the attention scores of the hidden states at all time steps are normalized to obtain the attention weight of the hidden state at each time step, the hidden states at all time steps are weighted and fused, and the high-order temporal feature vector is output. The higher-order temporal feature vector is input into the fully connected layer, and the first preliminary recurrence probability is output.

[0007] In one possible implementation, the DBN model includes an input layer, a hidden layer, and an output layer; wherein, The input layer includes multiple input feature nodes, including static cross feature nodes, high-order temporal feature vector nodes, and environmental influence coefficient nodes; The hidden layer includes multiple fault factor nodes, each fault factor node corresponds to the core cause of a type of fault, and each fault factor node is connected to multiple input feature nodes as parent feature nodes. The output layer includes an output node, whose parent feature node is all fault factor nodes; The step of inputting the model input features into the DBN model and outputting the second preliminary recurrence probability includes: The joint probability under the fault recurrence state and the joint probability under the fault non-recurrence state are calculated based on the product of the first dependency probability, the dependency probability product term, and the prior probability product term; wherein, the first dependency probability is the probability that the fault recurrence state depends on the fault factor node. The dependency probability product term is the product of the conditional probabilities of all fault factor nodes depending on the parent feature node. The prior probability product term is the product of the prior probabilities of all input feature nodes; The ratio of the joint probability under the fault recurrence state to the total joint probability is calculated to obtain the second preliminary recurrence probability; wherein the total joint probability is the sum of the joint probability under the fault recurrence state and the joint probability under the fault non-recurrence state.

[0008] In one possible implementation, calculating the fusion probability based on the first preliminary recurrence probability and the second preliminary recurrence probability includes: Obtain the test accuracy of the AT-LSTM model; Obtain the test accuracy of the DBN model; Based on the test accuracy of the AT-LSTM model and the test accuracy of the DBN model, the first weight of the first preliminary recurrence probability and the second weight of the second preliminary recurrence probability are calculated respectively. The fusion probability is obtained by weighted summation of the first preliminary recurrence probability and the second preliminary recurrence probability based on the first weight and the second weight.

[0009] In one possible implementation, calibrating the fusion probability based on expert rules and the environmental impact coefficient, and outputting the final recurrence probability, includes: Based on the static cross features, the environmental impact coefficient, the time-series derived features, and the static basic features, the adjustment coefficient corresponding to each expert rule is determined. Calculate the product of the adjustment coefficients corresponding to all expert rules and the fusion probability, and then use... The function is constrained to a probability range to obtain the final recurrence probability.

[0010] One possible implementation also includes: Obtain feedback data on fault recurrence results, data on the effectiveness of intervention measures, and supplementary core maintenance data; Based on the fault recurrence result data, the intervention effect data, and the synchronously supplemented core maintenance data, the training set is updated; the AT-LSTM model and the DBN model are retrained using the updated training set, and the control gate weights of the LSTM model and the probability parameters of the DBN model are adjusted to optimize the test accuracy of the AT-LSTM model and the DBN model; and / or, Based on the data on the effectiveness of the intervention measures, new expert rules are added or the adjustment coefficients and triggering conditions corresponding to the expert rules are adjusted; and / or, The probability range of the recurrence risk level is calibrated based on the fault recurrence result data.

[0011] A second aspect of the present invention provides a device for predicting the risk of recurrence of maintenance service failures, comprising: The core data acquisition module is used to acquire core maintenance data; wherein, the core maintenance data includes static maintenance process data, user evaluation semantic data, basic equipment attribute data, equipment operation sequence data, and environmental auxiliary time sequence data; The first preprocessing module is used to preprocess the static data of the maintenance process, the semantic data of user evaluation, and the basic attribute data of the equipment to generate static cross features; The second preprocessing module is used to preprocess the device runtime timing data to generate dynamic timing features; The third preprocessing module is used to preprocess the environmental auxiliary time series data to generate environmental impact coefficients; The first model input-output module is used to input the dynamic temporal features into the AT-LSTM model and output a high-order temporal feature vector and a first preliminary recurrence probability. The second model input / output module is used to input the model input features into the DBN model and output the second preliminary recurrence probability; wherein, the model input features include the static cross features, the higher-order temporal feature vector, and the environmental influence coefficient; The fusion probability calculation module is used to calculate the fusion probability based on the first preliminary recurrence probability and the second preliminary recurrence probability; The fusion probability calibration module is used to calibrate the fusion probability based on expert rules and the environmental influence coefficient, and output the final recurrence probability. The risk level determination module is used to determine the recurrence risk level corresponding to the final recurrence probability.

[0012] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for predicting the risk of recurrence of maintenance service faults provided in the first aspect of the present invention.

[0013] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a method for predicting the risk of recurrence of maintenance service faults provided in the first aspect of the present invention.

[0014] For a detailed description of the second to fourth aspects of the present invention and their various implementations, please refer to the detailed description in the first aspect and its various implementations; and for a detailed description of the beneficial effects of the second to fourth aspects and their various implementations, please refer to the beneficial effect analysis in the first aspect and its various implementations, which will not be repeated here.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for predicting the risk of recurrence of maintenance service failures according to an embodiment of the present invention. Figure 2 This is a structural block diagram of a maintenance service fault recurrence risk prediction device according to an embodiment of the present invention; Figure 3 This is a block diagram of the internal structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0018] This invention provides a method for predicting the risk of recurrence of maintenance service failures. This method is applied to the approving user terminal, which can be a server or a terminal device. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The terminal device can be a smartphone, tablet, desktop computer, wearable device, etc., but is not limited to these.

[0019] Figure 1 This is a flowchart illustrating a method for predicting the risk of recurrence of maintenance service faults provided in this embodiment. Figure 1 As shown, the main process of this method is described below (steps S101 to S109): Step S101: Obtain core maintenance data; wherein, core maintenance data includes static maintenance process data, semantic data of user evaluation, basic equipment attribute data, equipment operation sequence data, and environmental auxiliary time sequence data; Step S102: Preprocess the static data of the maintenance process, the semantic data of user evaluation, and the basic attribute data of the equipment to generate static cross features; Step S103: Preprocess the device runtime timing data to generate dynamic timing features; Step S104: Preprocess the environmental auxiliary time series data to generate environmental impact coefficients; Step S105: Input the dynamic temporal features into the AT-LSTM model and output the high-order temporal feature vector and the first preliminary recurrence probability; Step S106: Input the model input features into the DBN model and output the second preliminary recurrence probability; wherein, the model input features include static cross features, high-order time series feature vectors and environmental influence coefficients; Step S107: Calculate the fusion probability based on the first preliminary recurrence probability and the second preliminary recurrence probability; Step S108: The fusion probability is calibrated based on expert rules and environmental impact coefficients, and the final recurrence probability is output. Step S109: Determine the recurrence risk level corresponding to the final recurrence probability.

[0020] First, the meanings of the five categories of core maintenance data—static maintenance process data, semantic user evaluation data, basic equipment attribute data, equipment runtime sequence data, and environmental auxiliary time-series data—are illustrated with examples as follows: (1) Static data of the maintenance process: on-time arrival rate ( ), service duration compliance rate ( =Actual duration / Standard duration), Repair process standardization score ( (0-10 points, scored by quality inspectors based on key milestones), compliance of parts usage ( 1 = compliant / 0 = non-compliant), maintenance personnel skill level ( (Levels 1-5), maintenance methods () 1 = On-site repair / 2 = Factory return repair / 3 = Remote guidance); (2) User review semantic data: star rating ( (1-5 points), text evaluation sentiment score ( NLP processing is set to -1 to 1), and keyword weights are evaluated ( For example, the TF-IDF value of fault-related keywords such as "abnormal noise" and "water leakage", ranging from 0 to 1. (3) Equipment runtime sequence data: Continuous runtime sequence after maintenance (T={ , ,..., }, where t is the time step, with an interval of 1 hour, n=72 hours, covering the 3-day core observation period after maintenance), and the operating parameters of each time step (voltage U(t), current I(t), temperature T(t), operating power P(t)); (4) Basic attribute data of home appliances: service life ( (year), category () 1 = Refrigerator / 2 = Air Conditioner / 3 = Washing Machine / 4 = Other), Historical Fault Sequence ( (Fault frequency and type coding in the past two years) (5) Environmental auxiliary time series data: environmental temperature series after maintenance ( ), humidity sequence ( ).

[0021] In some optional embodiments, for step S102, it is necessary to first perform data standardization processing on the static data of the maintenance process, the basic attribute data of the equipment, and the semantic data of user evaluation to generate static basic features; wherein, the static basic features include standardized values ​​of the maintenance process, standardized values ​​of the basic attributes of the equipment, and standardized values ​​of the semantic data of user evaluation; then, based on the standardized values ​​of the maintenance process and the standardized values ​​of the basic attributes of the equipment, fault association cross terms are constructed and static cross features are generated.

[0022] Specifically, Z-score standardization can be used to eliminate the difference in units (such as the difference between the units of years of service and skill levels (levels 1-5)) to adapt to model training and avoid a single feature dominating the results. The calculation formula is: Equation (1); In the formula, Let i be the standardized value of feature i. The original value of feature number i. , Let i be the mean and standard deviation of all samples for feature i.

[0023] Equation (1) transforms the original value of each static feature into "the degree of deviation relative to the average value of the feature". The result is normally distributed (mean 0, standard deviation 1), which facilitates horizontal comparison of features with different dimensions.

[0024] For example, the service life of the platform's air conditioning repair. sample mean Year, standard deviation In 2018, a user used an air conditioner for 8 years. =8), after standardization This indicates that the air conditioner's service life is 1.5 standard deviations higher than the average, classifying it as an old appliance with a higher risk of recurring malfunctions. This standardized value will be used later for cross-feature construction and as input to the DBN model.

[0025] For the generation of static cross features, fault-related cross terms are constructed to amplify the impact of core risk factors. The formula and examples are as follows: (1) The fault association cross term is "old equipment + high frequency fault", which will be the standardized service life. frequency of historical failures Multiplication amplifies the combined risk effect of the two factors. Older equipment already has a high failure rate; if this is compounded by high-frequency historical failures, the risk of recurrence will significantly increase. (Constructed static cross-features) The calculation formula is: Equation (2); Example: A certain air conditioner (Used for 8 years) (4 failures in the past 2 years) This static cross feature will be input into the DBN model as a core risk factor.

[0026] (2) The fault-related cross-item is the "core of maintenance quality", that is, the standardization score of the maintenance process. Compliance with the use of accessories The intersection reflects the overall level of maintenance quality. If both standards are met, maintenance reliability is high and the risk of recurrence is low; if either standard is not met, the risk will be amplified. The constructed static intersection feature... The calculation formula is: Equation (3).

[0027] The timing data of the equipment during operation is prone to missing values ​​due to signal interruption or equipment offline, and there is also random noise (such as instantaneous voltage fluctuations). It is necessary to perform completion and smoothing processing to ensure the accuracy of the timing trend and lay the foundation for subsequent extraction of fault precursor features.

[0028] In some optional embodiments, for step S103, the device runtime timing data is preprocessed to generate two types of dynamic timing features: device timing processing features and timing derived features. First, the device runtime timing data is sequentially completed and smoothed to generate device timing processing features; then, the device timing processing features are dynamically trend-derived and parameter anomaly precursors are captured to generate timing derived features. Specifically: Step 1: Time-series data completion (attention-based linear interpolation) For a missing voltage value U(t) at time t, find the set of its neighboring non-missing data points. (e.g., at time t-1, t+1), assign attention weights to each neighboring point. (The closer the time interval, the greater the weight, and the closer it is to the actual operating trend), the weighted calculation yields the completed value. The formula is as follows: .

[0029] Example: Voltage data is missing at t=20 hours after an air conditioner repair. Adjacent non-missing data points are t=19 hours (U=220V) and t=21 hours (U=222V), each with a 1-hour time interval. Weights are... After completion, U(20) = 0.5 × 220 + 0.5 × 222 = 221V. The completed sequence will then undergo a smoothing step to avoid missing values ​​causing the AT-LSTM to miss trends.

[0030] Step 2: Time-series data smoothing (Savitzky-Golay filtering) Centered on time t, a sliding window is formed by taking k time steps before and after (a total of m = 2k + 1 points), and filtering coefficients are used. (Determined through least squares fitting to mitigate noise effects) Weighted average is used to obtain the smoothed voltage value. This preserves the core trend while eliminating random interference. The formula is: , j=-k,…,k.

[0031] Example: With k=2 (window size m=5), filter coefficients... At a certain time t, the voltages at adjacent points are [220, 221, 225, 222, 221] (including the instantaneous fluctuation of 225V). After smoothing... This eliminates instantaneous noise. The smoothed sequence, as a temporal processing feature, will be directly used for temporal derived feature construction and as input to the AT-LSTM model.

[0032] Step 3: Time-series derivation (serving AT-LSTM model) Based on the smoothed time series, dynamic trend features are derived to capture precursors of parameter anomalies. The formulas and examples are as follows: (1) Calculate the smoothed voltage at time t Smoothed voltage at time t-1 The difference This reflects the voltage change trend; a positive value indicates a voltage increase, and a negative value indicates a voltage decrease. The larger the absolute value, the more obvious the trend and the more prominent the precursor to an anomaly. The parameter trend formula is: .

[0033] Example: After smoothing, U=220.5V at t=19 hours, U=221V at t=20 hours, and U=223V at t=21 hours. , The voltage rise trend intensifies, indicating potential problems such as poor contact in the circuit.

[0034] (2) Calculate the voltage moving variance for the first 5 time steps before time t. This reflects voltage stability; the larger the variance, the more drastic the parameter fluctuations, and the higher the risk of fault recurrence (e.g., large voltage fluctuations in air conditioning circuits make them prone to short circuits again). The parameter stability formula is... In the formula, Let j be the smoothed voltage. This is the average value of all smoothed voltages within the sliding window.

[0035] Example: The 5 smoothed voltage values ​​t=22 hours ago are [220.5, 221, 223, 222.8, 224], mean ,but A large variance indicates poor voltage stability.

[0036] It should be noted that the above example only illustrates the feature derivation for the completed and smoothed voltage. Further feature derivation is needed for operating parameters such as the completed and smoothed current, temperature, and operating power. By concatenating the time-derived features with the smoothed original time-series sequence, and using the result as input to the AT-LSTM model, dynamic fault precursors can be accurately captured.

[0037] By combining and deriving features, the correlation features of fault recurrence are strengthened. All derived features serve as the core inputs of subsequent models, thus solving the problem of "low information density of raw data".

[0038] In some optional embodiments, for step S104, the environmental auxiliary time series data are first completed and smoothed sequentially to generate environmental time series processing features; then the environmental time series processing features are quantified to generate environmental impact coefficients.

[0039] Environmental auxiliary time-series data typically consists of environmental temperature and humidity sequences for the 72 hours following maintenance. Like equipment runtime sequence data, this type of data is prone to missing values ​​due to signal acquisition interruptions or equipment offline issues, and also contains random noise from environmental monitoring (such as instantaneous fluctuations in temperature and humidity). The calculation of the environmental impact coefficient E relies on the average environmental temperature and humidity values ​​for the 72 hours following maintenance. If the data is missing or noisy, it will directly lead to deviations in the mean calculation, thus affecting the accuracy of the environmental impact coefficient E and ultimately reducing the accuracy of fault recurrence risk prediction. Therefore, it is necessary to first complete and smooth the environmental auxiliary time-series data sequentially, following the same preprocessing logic as the equipment runtime sequence data (attention-based linear interpolation completion + Savitzky-Golay filtering smoothing) to ensure data integrity and accuracy. This will not be elaborated further here.

[0040] Ambient temperature and humidity can affect equipment operating parameters (e.g., high temperatures accelerate circuit aging, high humidity causes component corrosion). It is necessary to quantify the environmental impact factor E to correct operating parameters and the final risk value, avoiding prediction errors caused by environmental interference. The formula for calculating the environmental impact factor E is as follows: In the formula, , These are the average ambient temperature and humidity values ​​72 hours after the maintenance. , This refers to the standard ambient temperature and humidity for the corresponding equipment category (e.g., standard temperature of 25℃ and humidity of 50% for air conditioners). The value of E ranges from 0.8 to 1.2. The closer it is to 1.2, the greater the impact of the environment on the operation of the home appliance.

[0041] Example: Average ambient temperature after an air conditioner repair (standard ), average humidity (standard ),but The environmental impact coefficient is relatively high, indicating that high temperature and humidity will accelerate circuit aging, and this factor needs to be amplified in the risk assessment. E will ultimately be incorporated into the comprehensive risk value calculation to improve the accuracy of predictions under complex environments.

[0042] Furthermore, since high temperature and humidity environments directly affect the operating parameters of equipment circuits such as voltage and current, if uncorrected operating parameters are used directly to analyze fault precursors, feature extraction will be biased due to environmental interference. Therefore, it is necessary to use environmental temperature and humidity time series data to correct the environmental deviation of the equipment operating parameters at the corresponding time steps, and then perform completion, smoothing and derivative feature construction based on the corrected operating parameters to ensure that the time series features can truly reflect the operating status of the home appliance itself, rather than a false state under environmental interference.

[0043] In this embodiment, modeling is implemented in three steps based on the preprocessed feature data: 1) using AT-LSTM to extract dynamic fault precursors from time-series derived features; 2) using DBN to fuse time-series features and static cross features to quantify the recurrence probability; 3) using adaptive weights + expert rules for calibration to output the final probability.

[0044] In this embodiment, the AT-LSTM model includes an LSTM layer, an attention mechanism layer, and a fully connected layer. The input is a smoothed time series sequence (e.g., Time-series derived features , By using LSTM to capture long-term and short-term dependencies and introducing an attention mechanism to strengthen the weights of key time steps corresponding to fault precursors, the problem of insufficient sensitivity of traditional LSTM to core nodes is solved, and a high-order temporal feature vector is output. .

[0045] Specifically, the dynamic temporal features are input into the LSTM layer, which outputs a sequence of hidden states at multiple time steps. The sequence of hidden states at multiple time steps is then input into the attention mechanism layer to calculate the attention score of the hidden state at each time step. The attention scores of the hidden states at all time steps are normalized to obtain the attention weight of the hidden state at each time step. The hidden states at all time steps are then weighted and fused to output a high-order temporal feature vector. The high-order temporal feature vector is then input into the fully connected layer to output the first preliminary recurrence probability.

[0046] The core advantage of LSTM (Long Short-Term Memory) networks is their ability to remember long-term trends in time-series data while forgetting irrelevant information. Through three control gates—the forget gate, the input gate, and the output gate—it simulates the human thought process of "memory-update-output," accurately capturing long-term changes in time-series data (such as the complete evolution of air conditioner voltage from stable to slightly fluctuating to drastic fluctuations), thus avoiding the short-term memory limitations of traditional neural networks. The three control gates work collaboratively; the specific formulas and in-depth explanations are as follows: (1) Gate of Oblivion Its core function is to filter and discard invalid historical information while retaining useful memories; its formula is: In the formula, The Sigmoid activation function maps input values ​​to a range of 0-1, where 0 represents discarding completely and 1 represents retaining completely, essentially acting as a switch. This is the hidden state of the previous time step (t-1), which can be understood as the core information remembered in the previous time step (such as the stable state of the air conditioner voltage at time t-1). The input features at the current time (time t) specifically refer to "smoothed time series + time-derived features" (such as voltage value and voltage trend at time t). Voltage variance ); The weight matrix and bias terms are continuously optimized during model training to adjust the influence of input information and ensure that the forget gate can accurately filter information. The output of the forget gate is (0-1). The closer it is to 1, the more historical information it contains from the previous moment. The more useful it is, the more it needs to be kept. The closer the value is to 0, the more invalid the information from the previous moment is, and it needs to be discarded to avoid interfering with the current judgment.

[0047] Example: After an air conditioner was repaired, the voltage remained stable for t=1-19 hours. (The forgetting gate at t=20) (Close to 1), indicating "retaining the stable memory at time t=19"; when the voltage fluctuation intensifies at time t=21 ( Forget Gate at t=21 (Close to 0) indicates that "the slight fluctuation information at time t=20 is weakened and the focus is on the current violent fluctuation", so as to avoid the interference of the previous stable information on the judgment of the current anomaly.

[0048] (2) Input gate Its core function is to update the cell state and record current information. This involves the following formulas: Equation (4); Equation (5); Equation (4) determines which information is updated; in Equation (4), Similar to the forget gate, it also outputs 0-1 through the Sigmoid function. 1 means "a certain type of information at the current moment is very important and needs to be completely updated into long-term memory", and 0 means "the current information is useless and does not need to be updated".

[0049] Equation (5) generates the new information to be updated; in Equation (5), It is the hyperbolic tangent activation function, which is used to convert the input information at the current time step ( The purpose of mapping the values ​​to the range of -1 to 1 is to compress the information range and avoid unstable model training due to excessively large values. It refers to the "new information to be updated into long-term memory" at the current moment, which includes the core content of the temporal characteristics at the current moment (such as the trend information of the drastic voltage fluctuation at t=21).

[0050] also, These are all parameters optimized during model training, used to accurately filter and generate new information.

[0051] Example: At time t=21, the air conditioner voltage fluctuates drastically. , ), at this time the input gate (Close to 1), indicating that "the voltage fluctuation information at the current moment is very important and needs to be fully updated into long-term memory"; at the same time, information to be updated is generated. (Positive values ​​indicate an upward trend in fluctuations, and the magnitude of the value indicates the strength of the trend), used for subsequent updates to long-term memory.

[0052] (3) Cell state renewal Its core function is to integrate "historical memories filtered through the forgetting gate" and "new information generated by the input gate" to form long-term memory of the current moment. Its formula is: In the formula, It is an element-wise product (point-to-point multiplication), that is, the corresponding elements of two vectors are multiplied, not matrix multiplication; This refers to the "cell state" of the previous moment, which is the long-term memory of the previous moment (such as the voltage steady state memory at time t-1). The historical memory filtered by the forgetting gate—the long-term memory of the previous moment multiplied by the forgetting gate weight—is equivalent to "retaining useful historical memories and discarding invalid ones." The new information filtered by the input gate—the new information to be updated—is multiplied by the input gate weight, which is equivalent to "keeping only the new information that is useful at the current moment"; It is the long-term memory (cell state) of the current moment, which is a fusion of historical useful memory and current useful new information, and is also the core carrier for LSTM to remember long-term trends.

[0053] Example: Time t=21: Long-term memory of the previous time step (Representing slight fluctuations in memory at time t-20), Forget Gate Output ,so (Only a small amount of slight fluctuation memory at time t-20 is retained); Input gate output To be updated with new information ,so (Focusing on the dramatic fluctuations at the current time t-21); ultimately, the current time will be stored in long-term memory. It clearly records the trend change "from slight fluctuations to sharp fluctuations".

[0054] (4) Output gate Its core function is to filter useful information from long-term memory at the current moment and generate the hidden state for use by subsequent attention mechanisms. This involves the following formula: Equation (6); Equation (7); Equation (6) controls the output information; in Equation (6), It is the control signal for the output gate, which outputs 0-1 through the Sigmoid function. Its function is to filter the current long-term memory. The information in the code is 1, which means "a certain type of information needs to be output", and 0, which means "a certain type of information will not be output for the time being, and will continue to be retained in long-term memory".

[0055] Equation (7) generates the hidden state In equation (7), It is to store the long-term memory of the current moment. Mapping to the range of -1 to 1 compresses the information range; The hidden state at the current moment is the long-term memory after output gate filtering, containing core temporal information from both the current and past moments, essentially a "summary memory of the current moment"; the hidden states for all time steps (t=1 to t=72) are also included. This will form a hidden state sequence. This sequence will be directly input into the subsequent attention mechanism to extract features of key temporal nodes.

[0056] Example: Output gate at time t=21 (Close to 1) indicates that "the memory of the drastic fluctuations at the current moment needs to be fully output"; Therefore, the current state is hidden. This value will be used as the core summary memory at time t=21 and added to the hidden state sequence. middle.

[0057] Hidden state sequence output by LSTM layer The model contains a summary of memories from 72 time steps. However, not all time step memories are equally important. Memories from time steps with clear pre-fault indicators (such as drastic voltage fluctuations between t=20 and 22 hours) are more valuable for predicting fault recurrence. Therefore, an attention mechanism is introduced, assigning attention weights to the hidden states of each time step. This allows the model to focus on core time steps, amplifying the impact of pre-fault indicator information and ultimately generating a high-order temporal feature vector. .

[0058] Optionally, for the attention mechanism layer, the importance of each time step needs to be evaluated first, i.e., the attention score for each time step is calculated using the following formula: Equation (8); In equation (8), The hidden state at time t Attention score These are the parameters used in model training. The function maps the score to a range of -1 to 1. The higher the score, the more important the information at that time step is for predicting fault recurrence.

[0059] Then, weight normalization is performed, that is, the sum of the weights of all time steps is equal to 1, which facilitates weighted fusion. The calculation formula is as follows: Equation (9); In equation (9), These are normalized attention weights, which, through an exponential function and summation, convert the importance scores of all time steps into values ​​between 0 and 1, with the sum of the weights of all time steps being 1. The core logic is that the higher the importance score of a time step, the greater its weight after normalization. For example, for time steps with obvious early signs of failure, t=20-22 hours. It will be significantly higher than other time steps.

[0060] Then, temporal feature vector fusion is performed, which involves weighted fusion of the hidden states at all time steps to focus on core information. The calculation formula is as follows: Equation (10); In equation (10), The final extracted high-order temporal feature vector has a dimension of d (d is determined by model training optimization). Essentially, it is the "weighted average of the hidden states of all time steps", but it highlights the information of high-weight time steps (fault precursor nodes). It is equivalent to "extracting the most core fault precursor features from the memory of 72 time steps". It will be directly input into the DBN model to participate in the recurrence probability calculation.

[0061] Example: An air conditioner experiences severe voltage fluctuations between t=20 and 22 hours (derived characteristic) The hidden states at these three time steps (all ≥ 1.71) Their importance scores are calculated. It is significantly higher than other time steps (average) Attention weights are obtained after normalization. (The three factors combined account for 0.6, representing 60% of the total weight), while the weights of the other 70 time steps combined are only 0.4; the final weighted fusion yields the temporal feature vector. The initial recurrence probability of the AT-LSTM is output after passing through a fully connected layer (mapping the vector to between 0 and 1). This probability primarily reflects the risk of recurrence caused by drastic voltage fluctuations.

[0062] This embodiment introduces a DBN (Dynamic Bayesian Network) to fuse the static cross features constructed earlier. Timing characteristics of AT-LSTM output (etc.) A multi-node dependency model (e.g., voltage fluctuation → circuit stability → recurrence risk) is constructed to quantify the impact of each feature on the probability of fault recurrence and output the preliminary recurrence probability. It balances interpretability with the accuracy of probabilistic reasoning.

[0063] The DBN model consists of an input layer, hidden layers, and an output layer. The input layer includes multiple input feature nodes, which in turn include static cross feature nodes and higher-order temporal feature vector nodes (i.e., the output of the AT-LSTM model). It is decomposed into multiple nodes, such as voltage fluctuation intensity, voltage trend slope, etc. (corresponding to each dimension of the time series feature vector) and environmental impact coefficient nodes.

[0064] The hidden layer comprises multiple fault factor nodes, each corresponding to a core cause of a type of fault. Each fault factor node is connected to multiple input feature nodes as parent feature nodes. The hidden layer serves as the intermediate node connecting the input and output layers, directly inducing fault recurrence. Each fault factor node corresponds to a core cause of a type of fault, and each fault factor node has a clearly defined parent feature node (i.e., a feature node in the input layer that influences that fault factor). A specific example of this correspondence is shown below (using an air conditioning circuit fault scenario): Failure factors Circuit stability (direct causes: circuit interface oxidation, voltage fluctuations) → parent feature node (input layer): (Repair quality affects interface oxidation) (Voltage fluctuation intensity, timing characteristic nodes) (Voltage trend slope, time series characteristic nodes); Failure factors Component aging degree (direct causes: service life, environmental impact) → parent feature node (input layer): (Old home appliances + frequent malfunctions) (Environmental coefficients, indirectly used as input layer nodes); Failure factors : Reliability of maintenance process (direct causes: maintenance personnel skills, process standardization) → Parent feature node (input layer): (Maintenance personnel skills) (Repair quality); In simple terms, the parent feature node is all the input layer feature nodes that affect a certain fault factor. Each fault factor has a clear parent feature to avoid ambiguity in the parent feature.

[0065] The output layer includes an output node, representing the fault recurrence state, with values ​​{1, 0} – 1 indicates the fault will recur, and 0 indicates the fault will not recur. The parent feature node of the output node is all fault factor nodes (i.e., fault factors). , , All of these can affect the final relapse status.

[0066] Based on the above structure, the core logic of DBN is: input layer feature nodes → influence hidden layer fault factor nodes → fault factor nodes comprehensively influence the recurrence state of the output layer. By quantifying this causal relationship, the second preliminary recurrence probability is calculated. .

[0067] Specifically, the joint probability under the fault recurrence state and the joint probability under the fault non-recurrence state are calculated first based on the product of the first dependency probability, the dependency probability product term, and the prior probability product term. Here, the first dependency probability is the probability that the fault recurrence state depends on the fault factor node; the dependency probability product term is the product of the conditional probabilities of all fault factor nodes depending on their parent feature nodes; and the prior probability product term is the product of the prior probabilities of all input feature nodes. Then, the ratio of the joint probability under the fault recurrence state to the total joint probability is calculated to obtain the second preliminary recurrence probability. Here, the total joint probability is the sum of the joint probability under the fault recurrence state and the joint probability under the fault non-recurrence state.

[0068] For the joint probability distribution, the causal relationship of "input layer features → hidden layer failure factors → output layer recurrence state" is transformed into a computable probability formula. The overall probability is decomposed into the product of the probabilities of each node, simplifying subsequent probabilistic reasoning. The core logic of the formula is: Overall joint probability = Probability of recurrence state depending on failure factor × Probability of each failure factor depending on its parent feature × Prior probability of each input feature. This decomposition transforms the complex causal relationship into a computable product of probabilities, reducing the difficulty of reasoning.

[0069] The formula for calculating the joint probability is: Equation (11); In equation (11), The joint probability represents the probability that three events occur simultaneously: the input layer feature node F takes the current value, the hidden layer fault factor node H takes the current value, and the output layer recurrence state node Y takes the current value. It is the basis for subsequent calculation of the recurrence probability. This is the conditional probability, also known as the first dependency probability, representing the probability that the recurrence state Y in the output layer takes the current value given that all fault factor nodes H in the hidden layer take the current value; for example, in cases of poor circuit stability ( =Unstable), components are severely aged ( =Severe), Repair process is average ( Under normal conditions, the probability of fault recurrence (Y=1); The term "dependency product" represents the computation of the product of the corresponding parent feature nodes for all hidden layer fault factor nodes h. Given the current value, the probability that the fault factor h takes the current value is multiplied together. This indicates traversing each fault factor node in the hidden layer (e.g.) , , ); This represents the set of parent feature nodes corresponding to a certain fault factor h, that is, all feature nodes in the input layer that affect the fault factor h; for example... (Circuit stability) parent feature node , (Component aging) parent feature node ; For example, in the parent feature node =0.65 (Medium repair quality) =1.71 (large voltage fluctuation intensity) Under the condition that =2 (large voltage trend slope), the fault factor The probability of instability; For the prior probability product term, it means that for all input layer feature nodes f, the prior probability of the input layer feature node f taking the current value is calculated, and then all these probabilities are multiplied together. This represents traversing each feature node of the input layer, i.e., f in P(f) of the user question, where f is a feature node of the input layer. For example, f could be... (Intersection features) (Intersection features) (Time-series characteristic nodes), E (Environmental impact coefficient nodes), etc.; This represents the prior probability of the input layer feature node f, i.e., the probability that the feature node takes the current value among all historical samples. For example, among historical samples... =1.8 The sample proportion of (old + high frequency failure) is 15%, then P( =1.8)=0.15.

[0070] Based on joint probabilistic inference, the current input feature F and time-series features are derived. Under the given conditions, the probability of fault recurrence (Y=1), i.e., the second preliminary recurrence probability. This is the core output of the DBN model; the core logic of the formula is that the initial recurrence probability = the joint probability of (current feature + current failure factor + recurrence) / the total probability of (current feature + current failure factor), which is essentially the conditional probability of failure recurrence given that all current features and failure factors are known; its formula is: Equation (12); In equation (12), The initial recurrence probability ranges from 0 to 1. The closer the probability is to 1, the higher the risk of recurrence. For the specific form of joint probability, where It is the set of static cross-feature nodes in the input layer. It is the set of temporal feature nodes of the input layer (the two combined are the above F); the numerator represents the probability that "current input feature (static + temporal), current fault factor, fault recurrence (Y=1)" will occur simultaneously, and the calculation method is Equation (11). The summation term represents the sum of the joint probabilities of "fault recurrence (Y=1)" and "fault non-recurrence (Y=0)" under the conditions of "current input feature and current fault factor". Since Y only has two values, 0 and 1, this summation term is essentially the total probability of the current input feature and the current fault factor occurring simultaneously, used to normalize the numerator to ensure... The value ranges from 0 to 1.

[0071] Since DBN has a large number of nodes, direct calculation is complex. Therefore, the Gibbs sampling algorithm (a commonly used probabilistic inference algorithm) is adopted to approximate the true probability value through multiple samplings, eliminating the need to manually calculate the probability of all nodes and reducing computational complexity.

[0072] Example: Referring to the previous case: After an air conditioner is repaired, the current values ​​of the input layer feature node (F) are as follows (all are preprocessed values): Static cross feature nodes ( ): =1.8 (old + high frequency of failures) =0.65 (Medium repair quality) =0.8 (Standardized value of maintenance personnel skill level, upper-middle); Time-series characteristic nodes ( ): =1.71 (voltage fluctuation intensity) =2 (voltage trend slope); environmental factor node (E=1.068).

[0073] The current value of the hidden layer fault factor node (H) (derived based on inference from the parent feature node): (Circuit stability): Unstable (Parent feature node value:) =0.65、 =1.71、 =2, both pointing to poor circuit stability); (Component aging level): Severe (Parent feature node value:) =1.8 and E=1.068, both indicating severe component aging. (Reliability of maintenance process): Generally (Parent feature node value:) =0.8、 =0.65, indicating a generally average process.

[0074] Step 1: Calculate the numerator (joint probability when Y=1). Based on the joint probability formula, break it down and calculate: ① P(Y=1|H): The conditional probability of recurrence when the fault factor is {unstable, severe, moderate}, derived from training on historical samples. =Unstable, =Serious, =General)=0.8; ② Each failure factor depends on the probability product of its parent features (derived from training on historical samples): P( =Unstable| ): Parent characteristics =0.65、 =1.71、 =2, probability =0.85; P( =Severe| ): Parent characteristics =1.8, E=1.068, probability=0.9; P( =General| ): Parent characteristics =0.8、 =0.65, probability =0.7; Product = 0.85 × 0.9 × 0.7 ≈ 0.5355; ③ The product of prior probabilities for each input feature node (derived from historical sample statistics): P( =1.8)=0.15、P( =0.65)=0.2、P( =0.8)=0.25、P( =1.71)=0.18、P( =2)=0.12, P(E=1.068)=0.2; Product = 0.15 × 0.2 × 0.25 × 0.18 × 0.12 × 0.2 ≈ 0.000324; Numerator = 0.8 × 0.5355 × 0.000324 ≈ 0.0001388; Step 2: Calculate the denominator (the sum of the joint probabilities of Y=1 and Y=0). The calculated joint probability for Y=1 is approximately 0.0001388. Calculate the joint probability when Y=0 (using the same method as Y=1): P(Y=0|H)=0.2 (derived from historical training), the other two terms are the same as when Y=1 (because the values ​​of H and F remain unchanged), so the joint probability of Y=0 = 0.2 × 0.5355 × 0.000324 ≈0.0000347; Denominator = 0.0001388 + 0.0000347 ≈ 0.0001735; Step 3: Calculate the preliminary recurrence probability = Numerator / Denominator ≈ 0.0001388 / 0.0001735 ≈ 0.8. This is a simplified calculation; in practice, after approximation using Gibbs sampling and optimization of model parameters, the final value will be... =0.68; It should be noted that the probability values ​​obtained from historical training and historical sample statistics in the example are all based on training and statistics of a large amount of historical maintenance data (such as 20,000 records) on the platform. In actual applications, there is no need to calculate manually. The algorithm automatically calls historical data to complete the reasoning. The breakdown here is for the purpose of understanding the formula.

[0075] In this embodiment, the advantages of AT-LSTM (strong generalization, fitting complex time-series patterns) and DBN (strong interpretability, quantization feature dependence) are combined. The weights of the two are dynamically allocated using adaptive weights, and then calibrated using expert rules (connected with previous static and time-series features) to output the final recurrence probability. It balances accuracy and interpretability.

[0076] For step S107, first obtain the test accuracy of the AT-LSTM model and the DBN model; then, based on the test accuracy of the AT-LSTM model and the test accuracy of the DBN model, calculate the first weight of the first preliminary recurrence probability and the second weight of the second preliminary recurrence probability respectively; finally, perform a weighted summation of the first preliminary recurrence probability and the second preliminary recurrence probability based on the first weight and the second weight to obtain the fusion probability.

[0077] In some embodiments, the first weight The calculation formula is Second weight The calculation formula is , Fusion probability The calculation formula is .

[0078] In the formula, The test accuracy of the AT-LSTM model (based on historical maintenance data verification, such as 89% accuracy on the test set). The test accuracy of the DBN model (e.g., 87% accuracy on the test set); weights , The model is dynamically adjusted according to its accuracy. The core logic is that the higher the accuracy of the model, the greater its weight, and the stronger its impact on the final fusion probability. This avoids fusion bias caused by fixed weights and ensures that the fusion result takes into account both the advantages of time series fitting and probability quantization.

[0079] Example: AT-LSTM model test accuracy verified by historical data. DBN model test accuracy The weights are calculated as follows: , Based on the previous case, the initial recurrence probability of AT-LSTM Preliminary recurrence probability of DBN Then the fusion probability (Retain four decimal places for easy calibration later).

[0080] Due to model fusion probability While balancing accuracy and generalization, there may be "bias in extreme scenario scenarios" (such as old equipment + high-frequency failure + harsh environment, but the model fusion probability does not fully amplify the risk). Therefore, expert rules (based on more than 10 years of experience in the maintenance field, interpretable and adjustable) are introduced to adjust the fusion probability. Perform calibration and output the final recurrence probability. Furthermore, the environmental impact coefficient E calculated earlier is incorporated to further improve accuracy.

[0081] In some optional embodiments, for step S108, firstly, based on static cross features, environmental influence coefficients, temporal derived features, and static basic features, the adjustment coefficient corresponding to each expert rule is determined; then, the product of the adjustment coefficients corresponding to all expert rules and the fusion probability is calculated, and then... The function is restricted to a probability range to obtain the final recurrence probability.

[0082] In this embodiment, expert rules are linked to the preceding features to enhance practicality and enable direct implementation. Specifically: (1) If static cross features (Old equipment + high frequency of failures) and environmental impact factor (With significant environmental impact), then (Amplified risk, with a maximum of 0.95); (2) If static cross features (Excellent repair quality) and time-series derived characteristics (If the operating parameters are stable), then (To reduce risk, the lower limit is 0.05); (3) If the user rates the sentiment value (Strongly negative reviews) and contains keywords related to malfunctions ( ),but (Amplifying risks); (4) If the maintenance personnel's skill level (Highest level) and the repair method is factory return repair. ),but (Reduce risk); (5) If the appliance has been used for a certain number of years Equipment that is over 5 years old and has a history of failures ≥ 5 times ( ),but (Forcibly raising the risk floor); (6) If the time series trend characteristics If the absolute value of three consecutive time steps is ≥1.5 (parameter mutation), then... (Amplifying the risk of dynamic anomalies).

[0083] It is important to note that the core expert rules all use preprocessed / derived feature data, and do not directly use the original data before preprocessing. For example: User review semantic data: Expert rule (3) using preprocessed data (The sentiment score of the text evaluation is -1 to 1 in NLP processing.) (Evaluation keyword weights, TF-IDF processed to 0-1), non-original star rating text, unprocessed evaluation content; Static data during maintenance: Expert rule (4) using pre-processed data (Maintenance personnel skill levels, levels 1-5, original dimensions but participating feature standardization) (Repair method, 1 / 2 / 3 coding, which is the original coding but without additional preprocessing) (The core cross-characteristics of maintenance quality are standardized) × Derivation), and the expert rule (2) also uses the derived version. Non-original process standardization scores and original values ​​for component compliance; Equipment basic attribute data: Expert rules (1), (5) using preprocessed / derived data (Years of use, original values ​​but included in standardization and cross-features) (Construction) (The cross-characteristics of old equipment and high-frequency faults are represented by standardized...) × derivative), (Historical fault sequence, coded / standardized values), not the original service life, uncoded historical fault records.

[0084] In addition, some data is raw encoded data (such as...) Repair methods 1 / 2 / 3 For skill levels 1-5, this type of data does not require standardization / completeness (dimensionless issues, non-missing data), and falls under the category of "preprocessing without additional operations," rather than "directly using the raw unstructured / uncoded data before preprocessing," which aligns with the overall characteristics and usage logic of the algorithm.

[0085] Based on features such as static cross features, environmental impact coefficient, time-series derived features, and static basic features, we query the expert rules that meet the criteria and obtain the adjustment coefficients corresponding to the expert rules that meet the criteria. If the expert rules do not meet the criteria, the adjustment coefficient is set to 1.

[0086] Then, substituting all adjustment coefficients into the calibration formula, we can calculate the final recurrence probability P. The calibration formula is: In the formula, The adjustment coefficient is set for each expert rule (if the rule is met, the corresponding coefficient is used; otherwise, 1 is used). This is the product of the coefficients of all rules that meet the criteria. The function is used to limit the final probability range to between 0.05 and 0.95, avoiding prediction distortion caused by extreme values ​​(such as P=1 or P=0), and ensuring the rationality and practicality of the probability.

[0087] Example: Fusion probability ,feature: (Complies with Rule 1) (Does not comply with rule 1) (Does not comply with rule 2) (Does not comply with rule 3) (Does not comply with rule 4) and (Does not comply with rule 5) However, it does not have three consecutive time steps (does not meet rule 6); it only meets one rule (rule 1 does not satisfy condition E, therefore, no rule is met). Ultimate recurrence probability (Retain two decimal places for subsequent risk classification).

[0088] In some optional embodiments, for step S109, based on the final recurrence probability Based on the actual needs of maintenance operations, risks are divided into four levels (thoroughly correcting contradictions in risk classification, clearly defining the scope of each level, with no overlap or omissions), corresponding to different pre-emptive intervention measures, achieving a closed loop of "prediction-intervention-risk reduction," and providing a basis for risk level assessment in subsequent settlement processes. A risk value R linked to the final recurrence probability P can be used for classification (R=P×10, for easy and intuitive level division), clearly defining the scope of each level and thoroughly resolving the "contradiction between medium and high risk ranges." The specific classification is as follows: Low risk: probability of eventual recurrence Corresponding risk value Key features: Stable operation of home appliances, high repair quality, no obvious signs of failure, and extremely low probability of recurrence; Low to medium risk: probability of eventual recurrence Corresponding risk value Key characteristics: Slight abnormalities exist (such as minor parameter fluctuations), but there is no superposition of core risk factors, and the probability of recurrence is low. Medium to high risk: probability of eventual recurrence Corresponding risk value Key characteristics: There is an overlap of core risk factors (such as old equipment + parameter fluctuations), obvious early signs of failure, and a high probability of recurrence. High risk: probability of eventual recurrence Corresponding risk value Key characteristics: Multiple risk factors are superimposed (old + high frequency failure + harsh environment + parameter mutation), resulting in an extremely high recurrence probability and requiring urgent intervention.

[0089] Furthermore, differentiated pre-intervention measures are formulated for different recurrence risk levels. All measures incorporate the features extracted earlier (such as temporal anomalies and static cross features) to ensure the targeted nature of the intervention. Simultaneously, the intervention results are fed back into the model to optimize the accuracy of subsequent predictions. Specifically: (1) Low risk (R∈[0.5,3)): Routine follow-up, no additional intervention required; send an operation reminder once every 7 days after repair, push home appliance maintenance suggestions once a month, and continuously monitor based on home appliance operation sequence data. If there are no abnormalities, no further action is required. (2) Low to medium risk (R∈[3,6)): Mild intervention to reduce the probability of recurrence; arrange a remote parameter inspection once every 3 days after maintenance (focusing on monitoring time-series derived features). , ), and simultaneously push targeted maintenance guides (such as component inspection suggestions for old equipment); (3) Medium to high risk (R∈[6,8)): Moderate intervention, enhanced monitoring and maintenance; remote inspection within 24 hours after repair, time series parameters monitored once every 2 days, and on-site re-inspection arranged once 7 days after repair (focusing on checking the quality of repair). , Circuit stability h1), replace easily worn parts (such as aging interfaces); (4) High risk (R∈[8,9.5]): Emergency intervention to prevent recurrence; arrange on-site re-inspection within 12 hours after repair, comprehensively investigate fault factors (circuit stability, component aging, repair process), replace core vulnerable components, arrange inspections once on the 1st, 3rd and 7th day after repair, and adjust the operating parameters of home appliances simultaneously (combined with the environmental impact coefficient E correction) to reduce operating load.

[0090] Furthermore, this method is not a static model, but rather constructs a closed-loop mechanism of "prediction-intervention-feedback-optimization". By combining actual maintenance results (feedback data on fault recurrence, data on the effectiveness of intervention measures, and core maintenance data supplemented in sync), it continuously optimizes model parameters, expert rules, and risk classification standards to ensure that the algorithm adapts to the dynamic changes in maintenance business in the long term, thereby improving prediction accuracy and practicality.

[0091] Record the final result of each repair (whether it recurs) and the effectiveness of intervention measures (e.g., whether the recurrence rate decreases after intervention for medium- and high-risk cases), and simultaneously supplement with new repair data (static + time-series) to update the training set. Periodically (e.g., quarterly), retrain the AT-LSTM model and DBN model using the updated training set, and adjust the control gate weights of the LSTM model. etc.) and the probability parameters of the DBN model ( (etc.) to optimize the test accuracy of AT-LSTM and DBN models. .

[0092] It should be noted that the training of the DBN model is divided into two stages: unsupervised pre-training adjusts the weight matrix and the biases of the visible / hidden layers, and supervised fine-tuning fine-tunes the weights / biases while fitting conditional probability parameters such as P(Y|H) and P(h|F_h). The weights / biases are the underlying training parameters of DBN, and P(Y|H) and P(h|F_h) are the business layer probability parameters after the underlying parameters are transformed. The relationship between the two is "deep learning training → probabilistic graphical model transformation". In the fault recurrence scenario, the transformed probability parameters (P(Y|H) and P(h|F_h)) are finally used, and the weights / biases are only used as intermediate parameters in the training process.

[0093] Based on data on the effectiveness of intervention measures, adjust the coefficients and triggering conditions of expert rules (e.g., if the relapse rate does not decrease after intervention of a certain rule, reduce its adjustment coefficient); add special scenario rules (e.g., risk rules for new types of home appliances) to improve the rule system.

[0094] Based on the failure recurrence data, calibrate the probability range of risk classification (if the actual recurrence rate of medium- and high-risk cases is too high, the range should be appropriately narrowed and intervention strengthened); adjust the linkage between risk values ​​R and P to ensure that the classification matches actual business needs.

[0095] Through closed-loop optimization, the core metrics of the algorithm will continue to improve. The specific verification metrics are as follows (based on iterative verification of historical data): the initial accuracy rate is 88%, which increases by 1-2% after each quarterly optimization, and eventually stabilizes at over 92%; the recurrence rate of medium and high risk before optimization is 65%, which drops to below 40% after optimization.

[0096] This method constructs a four-layer innovative architecture of "deep mining of temporal features - adaptive fusion of multiple models - dynamic risk iterative correction - closed-loop intervention feedback". Its core logic is as follows: first, the core maintenance data is refined through special preprocessing (including temporal completion, smoothing and derivation, and correction by combining environmental data, etc.), then attention LSTM is used to capture dynamic precursors of faults, dynamic Bayesian network is fused to quantify the probability, the final risk is output after adaptive weight and rule calibration, and finally the intervention mechanism is linked and the model is optimized in reverse.

[0097] Based on the same inventive concept, embodiments of the present invention provide a device for predicting the risk of recurrence of maintenance service failures. Figure 2 This is a structural block diagram of a maintenance service fault recurrence risk prediction device 200 provided in an embodiment of the present invention. Figure 2 As shown, the maintenance service fault recurrence risk prediction device 200 mainly includes: The core data acquisition module 201 is used to acquire core maintenance data; among which, the core maintenance data includes static maintenance process data, semantic data of user evaluation, basic equipment attribute data, equipment operation sequence data, and environmental auxiliary time sequence data; The first preprocessing module 202 is used to preprocess the static data of the maintenance process, the semantic data of user evaluation, and the basic attribute data of the equipment to generate static cross features. The second preprocessing module 203 is used to preprocess the device runtime timing data to generate dynamic timing features; The third preprocessing module 204 is used to preprocess the environmental auxiliary time series data to generate environmental impact coefficients; The first model input-output module 205 is used to input dynamic temporal features into the AT-LSTM model and output a high-order temporal feature vector and a first preliminary recurrence probability. The second model input / output module 206 is used to input the model input features into the DBN model and output the second preliminary recurrence probability; wherein, the model input features include static cross features, high-order time series feature vectors and environmental influence coefficients; The fusion probability calculation module 207 is used to calculate the fusion probability based on the first preliminary recurrence probability and the second preliminary recurrence probability; The fusion probability calibration module 208 is used to calibrate the fusion probability based on expert rules and environmental influence coefficients, and output the final recurrence probability. The risk level determination module 209 is used to determine the recurrence risk level corresponding to the final recurrence probability.

[0098] In some optional embodiments, the first preprocessing module 202 is specifically used to perform data standardization processing on the static data of the maintenance process, the basic attribute data of the equipment, and the semantic data of user evaluation to generate static basic features; wherein, the static basic features include standardized values ​​of the maintenance process, standardized values ​​of the basic attributes of the equipment, and standardized values ​​of the semantic data of user evaluation; and to construct fault association cross terms and generate static cross features based on the standardized values ​​of the maintenance process and the standardized values ​​of the basic attributes of the equipment. The third preprocessing module 204 is specifically used to complete and smooth the environmental auxiliary time series data in sequence to generate environmental time series processing features; and to quantify the environmental time series processing features to generate environmental impact coefficients. The dynamic timing features include equipment timing processing features and timing derived features; the second preprocessing module 203 is specifically used to correct environmental deviations in the equipment runtime timing data based on the environmental timing processing features; to perform completion and smoothing processing on the corrected equipment runtime timing data in sequence to generate equipment timing processing features; and to perform dynamic trend derivation and capture of parameter anomalies in the equipment timing processing features to generate timing derived features.

[0099] In some optional embodiments, the AT-LSTM model includes an LSTM layer, an attention mechanism layer, and a fully connected layer; the first model input / output module 205 is specifically used to input dynamic temporal features into the LSTM layer and output a hidden state sequence at multiple time steps; input the hidden state sequence at multiple time steps into the attention mechanism layer, calculate the attention score of the hidden state at each time step, normalize the attention scores of the hidden states at all time steps to obtain the attention weight of the hidden state at each time step, perform weighted fusion on the hidden states at all time steps, and output a high-order temporal feature vector; input the high-order temporal feature vector into the fully connected layer and output the first preliminary recurrence probability.

[0100] In some optional embodiments, the DBN model includes an input layer, a hidden layer, and an output layer; wherein, the input layer includes multiple input feature nodes, which include static cross feature nodes, high-order temporal feature vector nodes, and environmental influence coefficient nodes; the hidden layer includes multiple fault factor nodes, each fault factor node corresponding to the core cause of a type of fault, and each fault factor node is connected to multiple input feature nodes as parent feature nodes; the output layer includes an output node, whose parent feature node is all fault factor nodes.

[0101] The second model input / output module 206 is specifically used to calculate the joint probability in the fault recurrence state and the joint probability in the fault non-recurrence state based on the product of the first dependency probability, the dependency probability product term, and the prior probability product term; wherein, the first dependency probability is the probability that the fault recurrence state depends on the fault factor node; the dependency probability product term is the product of the conditional probabilities of all fault factor nodes depending on the parent feature node; the prior probability product term is the product of the prior probabilities of all input feature nodes; and the ratio of the joint probability in the fault recurrence state to the total joint probability is calculated to obtain the second preliminary recurrence probability; wherein, the total joint probability is the sum of the joint probability in the fault recurrence state and the joint probability in the fault non-recurrence state.

[0102] In some optional embodiments, the fusion probability calculation module 207 is specifically used to obtain the test accuracy of the AT-LSTM model; obtain the test accuracy of the DBN model; calculate the first weight of the first preliminary recurrence probability and the second weight of the second preliminary recurrence probability based on the test accuracy of the AT-LSTM model and the test accuracy of the DBN model, respectively; and perform a weighted summation of the first preliminary recurrence probability and the second preliminary recurrence probability based on the first weight and the second weight to obtain the fusion probability.

[0103] In some optional embodiments, the fusion probability calibration module 208 is specifically used to determine the adjustment coefficient corresponding to each expert rule based on static cross features, environmental influence coefficients, time-series derived features, and static basic features; calculate the product of the adjustment coefficients corresponding to all expert rules and the fusion probability, and then apply... The function is restricted to a probability range to obtain the final recurrence probability.

[0104] In some optional embodiments, the maintenance service failure recurrence risk prediction device 200 further includes: The feedback data acquisition module is used to acquire feedback data on fault recurrence results, data on the effectiveness of intervention measures, and synchronously supplemented core maintenance data. The training optimization module is used to update the training set based on fault recurrence result data, intervention effect data, and synchronously supplemented core maintenance data; retrain the AT-LSTM model and DBN model using the updated training set, adjust the control gate weights of the LSTM model and the probability parameters of the DBN model, to optimize the test accuracy of the AT-LSTM model and DBN model; and / or, The expert rule adjustment module is used to add new expert rules or adjust the adjustment coefficients and trigger conditions of expert rules based on intervention effect data; and / or, The probability range fusion probability calibration module is used to calibrate the probability range of recurrence risk level based on fault recurrence result data.

[0105] The functional modules in the embodiments of this invention can be integrated together to form an independent unit, such as integrated into a processing unit, or each module can exist physically separately, or two or more modules can be integrated to form an independent unit. The integrated unit can be implemented in hardware or as a software functional unit. If the function is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0106] Various variations and specific examples of the methods provided in the embodiments of the present invention are also applicable to the maintenance service fault recurrence risk prediction device provided in this embodiment. Through the foregoing detailed description of the maintenance service fault recurrence risk prediction method, those skilled in the art can clearly understand the implementation method of the maintenance service fault recurrence risk prediction device in this embodiment. For the sake of brevity, it will not be described in detail here.

[0107] Figure 3 This is a structural block diagram of an electronic device 300 provided in an embodiment of the present invention. Figure 3 As shown, the electronic device 300 includes a memory 301, a processor 302, and a communication bus 303; the memory 301 and the processor 302 are connected through the communication bus 303.

[0108] The memory 301 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 301 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for at least one function, and instructions for implementing the maintenance service fault recurrence risk prediction method provided in the above embodiments, etc. The data storage area may store data involved in the maintenance service fault recurrence risk prediction method provided in the above embodiments, etc.

[0109] Processor 302 may include one or more processing cores. Processor 302 executes instructions, programs, code sets, or instruction sets stored in memory 301, and calls data stored in memory 301 to perform various functions and process data as described in this application. Processor 302 may be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), controller, microcontroller, and microprocessor. It is understood that for different devices, the electronic devices used to implement the functions of processor 302 may also be other types, and this embodiment of the invention does not specifically limit the specific devices used.

[0110] The communication bus 303 may include a path for transmitting information between the aforementioned components. The communication bus 303 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The communication bus 303 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 The symbol is represented by only one double arrow, but this does not indicate that there is only one bus or one type of bus. Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0111] This invention also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described in the above embodiments for predicting the risk of recurrence of maintenance service failures.

[0112] In this embodiment, the computer-readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. The computer-readable storage medium can be, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any combination thereof. Specifically, the computer-readable storage medium can be a portable computer disk, a hard disk, a USB flash drive, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), staging random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory stick, floppy disk, optical disk, magnetic disk, mechanical encoding device, or any combination thereof.

[0113] The computer program in this embodiment includes functions for executing... Figure 1 The program code for the method shown may include instructions corresponding to the execution of the method steps provided in the above embodiments. The computer program may be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded via a network (e.g., the Internet, local area network, wide area network, and / or wireless network) to an external computer or external storage device. The computer program may be executed entirely on the user's computer as a standalone software package.

[0114] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0115] Additionally, it should be understood that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0116] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting the risk of recurrence of maintenance service failures, characterized in that, include: Acquire core maintenance data; wherein, the core maintenance data includes static maintenance process data, user evaluation semantic data, basic equipment attribute data, equipment runtime sequence data, and environmental auxiliary time sequence data; The static data of the maintenance process, the semantic data of user evaluation, and the basic attribute data of the equipment are preprocessed to generate static cross features; The device runtime timing data is preprocessed to generate dynamic timing features; The environmental auxiliary time series data is preprocessed to generate environmental impact coefficients; The dynamic temporal features are input into the AT-LSTM model, which outputs a high-order temporal feature vector and a first preliminary recurrence probability. The model input features are input into the DBN model, and the second preliminary relapse probability is output; wherein, the model input features include the static cross features, the higher-order time series feature vector, and the environmental influence coefficient; The fusion probability is calculated based on the first preliminary recurrence probability and the second preliminary recurrence probability; The fusion probability is calibrated based on expert rules and the environmental impact coefficient, and the final recurrence probability is output. Determine the recurrence risk level corresponding to the final recurrence probability.

2. The method as described in claim 1, characterized in that, The preprocessing of the static data of the maintenance process, the semantic data of user evaluation, and the basic attribute data of the equipment to generate static cross features includes: The static data of the maintenance process, the basic attribute data of the equipment, and the semantic data of the user evaluation are subjected to data standardization processing to generate the static basic features; wherein, the static basic features include standardized values ​​of the maintenance process, standardized values ​​of the basic attributes of the equipment, and standardized values ​​of the semantic data of the user evaluation. Based on the standardized values ​​of the maintenance process and the standardized values ​​of the equipment's basic attributes, fault association cross terms are constructed, and the static cross features are generated. The preprocessing of the environmental auxiliary time-series data to generate environmental impact coefficients includes: The environmental auxiliary time series data are sequentially completed and smoothed to generate the environmental time series processing features; The environmental time-series processing characteristics are quantified to generate the environmental impact coefficient; The dynamic timing features include device timing processing features and timing derived features; the preprocessing of the device runtime timing data to generate dynamic timing features includes: Based on the aforementioned environmental timing processing characteristics, environmental deviation correction is performed on the device runtime timing data. The corrected device runtime timing data is sequentially completed and smoothed to generate the device timing processing features. The device's timing processing characteristics are dynamically trend-derived, and abnormal parameter precursors are captured to generate the timing-derived characteristics.

3. The method as described in claim 1 or 2, characterized in that, The AT-LSTM model includes an LSTM layer, an attention mechanism layer, and a fully connected layer; the step of inputting the dynamic temporal features into the AT-LSTM model and outputting a high-order temporal feature vector and a first preliminary recurrence probability includes: The dynamic temporal features are input into the LSTM layer, and the hidden state sequence of multiple time steps is output. The hidden state sequences of the multiple time steps are input into the attention mechanism layer, the attention score of the hidden state at each time step is calculated, the attention scores of the hidden states at all time steps are normalized to obtain the attention weight of the hidden state at each time step, the hidden states at all time steps are weighted and fused, and the high-order temporal feature vector is output. The higher-order temporal feature vector is input into the fully connected layer, and the first preliminary recurrence probability is output.

4. The method as described in claim 1 or 2, characterized in that, The DBN model includes an input layer, a hidden layer, and an output layer; wherein, The input layer includes multiple input feature nodes, including static cross feature nodes, high-order temporal feature vector nodes, and environmental influence coefficient nodes; The hidden layer includes multiple fault factor nodes, each fault factor node corresponds to the core cause of a type of fault, and each fault factor node is connected to multiple input feature nodes as parent feature nodes. The output layer includes an output node, whose parent feature node is all fault factor nodes; The step of inputting the model input features into the DBN model and outputting the second preliminary recurrence probability includes: The joint probability under the fault recurrence state and the joint probability under the fault non-recurrence state are calculated based on the product of the first dependency probability, the dependency probability product term, and the prior probability product term; wherein, the first dependency probability is the probability that the fault recurrence state depends on the fault factor node. The dependency probability product term is the product of the conditional probabilities of all fault factor nodes depending on the parent feature node. The prior probability product term is the product of the prior probabilities of all input feature nodes; The ratio of the joint probability under the fault recurrence state to the total joint probability is calculated to obtain the second preliminary recurrence probability; wherein the total joint probability is the sum of the joint probability under the fault recurrence state and the joint probability under the fault non-recurrence state.

5. The method as described in claim 1 or 2, characterized in that, The calculation of the fusion probability based on the first preliminary recurrence probability and the second preliminary recurrence probability includes: Obtain the test accuracy of the AT-LSTM model; Obtain the test accuracy of the DBN model; Based on the test accuracy of the AT-LSTM model and the test accuracy of the DBN model, the first weight of the first preliminary recurrence probability and the second weight of the second preliminary recurrence probability are calculated respectively. The fusion probability is obtained by weighted summation of the first preliminary recurrence probability and the second preliminary recurrence probability based on the first weight and the second weight.

6. The method as described in claim 2, characterized in that, The calibration of the fusion probability based on expert rules and the environmental impact coefficient, and the output of the final recurrence probability, include: Based on the static cross features, the environmental impact coefficient, the time-series derived features, and the static basic features, the adjustment coefficient corresponding to each expert rule is determined. Calculate the product of the adjustment coefficients corresponding to all expert rules and the fusion probability, and then use... The function is constrained to a probability range to obtain the final recurrence probability.

7. The method as described in any one of claims 1, 2, or 6, characterized in that, Also includes: Obtain feedback data on fault recurrence results, data on the effectiveness of intervention measures, and supplementary core maintenance data; Based on the fault recurrence result data, the intervention effect data, and the synchronously supplemented maintenance core data, the training set is updated; the AT-LSTM model and the DBN model are retrained using the updated training set, and the control gate weights of the LSTM model and the probability parameters of the DBN model are adjusted to optimize the test accuracy of the AT-LSTM model and the DBN model. And / or, Based on the data on the effectiveness of the intervention measures, new expert rules are added or the adjustment coefficients and triggering conditions corresponding to the expert rules are adjusted. And / or, The probability range of the recurrence risk level is calibrated based on the fault recurrence result data.

8. A device for predicting the risk of recurrence of maintenance service failures, characterized in that, include: The core data acquisition module is used to acquire core maintenance data; wherein, the core maintenance data includes static maintenance process data, user evaluation semantic data, basic equipment attribute data, equipment operation sequence data, and environmental auxiliary time sequence data; The first preprocessing module is used to preprocess the static data of the maintenance process, the semantic data of user evaluation, and the basic attribute data of the equipment to generate static cross features; The second preprocessing module is used to preprocess the device runtime timing data to generate dynamic timing features; The third preprocessing module is used to preprocess the environmental auxiliary time series data to generate environmental impact coefficients; The first model input-output module is used to input the dynamic temporal features into the AT-LSTM model and output a high-order temporal feature vector and a first preliminary recurrence probability. The second model input / output module is used to input the model input features into the DBN model and output the second preliminary recurrence probability; wherein, the model input features include the static cross features, the higher-order temporal feature vector, and the environmental influence coefficient; The fusion probability calculation module is used to calculate the fusion probability based on the first preliminary recurrence probability and the second preliminary recurrence probability; The fusion probability calibration module is used to calibrate the fusion probability based on expert rules and the environmental influence coefficient, and output the final recurrence probability. The risk level determination module is used to determine the recurrence risk level corresponding to the final recurrence probability.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the maintenance service fault recurrence risk prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the maintenance service fault recurrence risk prediction method according to any one of claims 1 to 7.

Images