An auto-encoder multi-channel engine fault early warning method and system

Abstract

The application discloses a kind of self-encoder multi-channel engine fault early warning method, the method comprises: step S1, data acquisition and pre-processing, step S2, time sequence feature engineering, step S3, self-encoder modeling, step S4, design multi-channel alarm fusion strategy, step S5, daily evaluation and structured output.The multi-warning channel fusion used in the application integrates reconstruction error, residual drift, absolute value gate, extreme spike, direction gate and thermal coupling and other multiple alarm channels, can capture abnormal patterns from different dimensions, cover multiple fault types of sudden failure and progressive failure.

Description

Technical Field

[0001] This invention relates to the field of engine condition monitoring and intelligent fault diagnosis technology, and in particular to an engine multi-channel fault early warning method and system based on a self-encoder. Background Technology

[0002] With the rapid development of the automotive industry, the engine, as the core power component of a vehicle, directly affects the vehicle's safety, reliability, and economy. If engine faults are not detected and warned of in their early stages, they can lead to serious mechanical damage, vehicle breakdowns, or even safety accidents. Traditional engine fault diagnosis methods mainly rely on the fixed threshold alarm mechanism of on-board diagnostic systems (OBD), triggering an alarm when sensor data exceeds a preset threshold. However, this method has the following shortcomings: First, the fixed threshold method cannot adapt to the differences between individual engines. Even engines of the same model exhibit variations in sensor parameter distribution under normal operating conditions due to factors such as manufacturing tolerances, service life, and maintenance conditions. A uniform threshold can easily lead to missed or false alarms. Second, traditional methods focus only on the absolute value of a single sensor indicator, ignoring the coupling relationships between multiple sensors and the temporal trends of parameter changes, thus failing to capture the early, weak signals of progressive faults. Furthermore, the continuous changes in engine operating conditions (such as speed, torque, and temperature) cause the normal range of sensor parameters to dynamically drift, making it difficult for the fixed threshold method to accurately distinguish between changes in operating conditions and actual faults.

[0003] In existing technologies, data-driven fault diagnosis methods are gaining increasing attention, such as traditional machine learning methods like Support Vector Machines (SVM) and Random Forests, as well as deep learning-based methods. However, these methods face the following challenges in practical applications: First, there is far more normal engine operation data than fault data, resulting in a severe imbalance between positive and negative samples, which leads to poor performance of supervised learning methods. Second, engine operating conditions are complex and variable, and the lack of decoupling processing for operating condition factors results in insufficient generalization ability of the model under different operating conditions. Third, a single anomaly detection channel is easily affected by sensor noise and fluctuations in operating conditions, generating a large number of false alarms.

[0004] Therefore, how to automatically learn normal operating mode from engine operating data through unsupervised learning methods and use multi-channel fusion strategies to achieve high-precision, low-false-alarm early fault warning has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a method and system for multi-channel engine fault early warning based on a self-encoder.

[0006] The technical solution adopted to achieve the purpose of this invention is: A self-encoder multi-channel engine fault early warning method includes the following steps: Step S1: Collect multidimensional engine operating data including target parameters, establish the operating condition baseline of the target parameters using ridge regression, and generate residual features; Step S2: Extract the mean, standard deviation, and quantile features of the engine's multidimensional operating data and residual features through a sliding time window to obtain the corresponding window feature vector; Step S3: Standardize the window feature vector of the normally operating engine from step S2 to obtain the input vector. A three-layer fully connected autoencoder neural network is trained using the input vector to obtain the reconstructed vector of the input vector. ; Step S4, based on and Calculate global reconstruction error and reconstruction error drift The target reconstruction error is obtained from the window feature vector based on the target parameters and residual features. ,based on , , The system constructs a base and target reconstruction error alarm channel, an extreme spike alarm channel, a direction-gated absolute value alarm channel, a thermal coupling alarm channel, and a multi-day persistent suppression channel based on the window feature vectors of the target parameters. The base and target reconstruction error alarm channel identifies persistent and gradual anomalies; the extreme spike alarm channel identifies sudden anomalies; the direction-gated absolute value alarm channel distinguishes between normal operating condition fluctuations and anomalies with a clear direction of degradation; the thermal coupling alarm channel identifies composite anomalies of decreased cooling system efficiency; and the multi-day persistent suppression channel summarizes the identification results of each warning channel daily and performs continuous verification, ultimately determining the output and providing a comprehensive observation-level output. Or integrated alarm level output ; Step S5, based on or The alarm status is summarized daily, and daily observation day and alarm day markers are generated. The daily health score is calculated, the warning level is determined, and the structured warning results are output.

[0007] In the above technical solution, in step S1, the multi-dimensional operating data of the engine includes engine speed, actual torque percentage, oil temperature, engine coolant temperature, oil pressure, coolant temperature, intake air volume, fuel gas volume, EGR flow rate, catalytic converter exhaust temperature, and fan speed. With oil temperature or engine coolant temperature as the target parameter, ridge regression is used to decouple the target parameter from the operating conditions. With engine speed, actual torque percentage, oil temperature, and engine coolant temperature as independent variables and the target parameter as the dependent variable, the operating condition baseline is fitted. The process of decoupling the target parameters from operating conditions using ridge regression can be represented as follows: ; In the formula, For all candidate ridge regression coefficient vectors In the middle, determine the coefficient vector that minimizes the objective parameter within the curly braces. This is a vector composed of the target parameters at each sampling time. The design matrix consists of engine speed, actual torque percentage, oil temperature, and engine coolant temperature. Let be the vector of ridge regression coefficients to be determined. The regularization coefficient is . Represents the L2 norm, The baseline predicted value for the target parameter under operating conditions; The formula for calculating the residual characteristics is: ; In the formula, resid is the residual characteristic.

[0008] In the above technical solution, in step S3, the mean value of the window feature vector of a normally operating engine is used. Standard deviation The quantile feature is the first quantile feature of the window feature vector of a normally operating engine. 3D features Standardization processing is required; The three-layer fully connected autoencoder neural network includes an encoder and a decoder. The encoder performs dimensionality reduction on the standardized window feature vector, and the decoder restores the input vector after dimensionality reduction by the encoder to obtain the reconstructed vector of the input vector. The dimensionality reduction process of the encoder is represented as follows: ; ; In the formula, and These are the intermediate representation and bottleneck encoding of the encoder, respectively. The weight matrix for the intermediate representation of the encoder. The weight matrix for encoding the bottleneck of the encoder. This is the bias vector for the intermediate representation of the encoder. The bias vector for the bottleneck encoding of the encoder. For activation functions; The The calculation formula is: ; In the formula, Here is the weight matrix of the decoder. This is the bias vector for the decoder; During the training of a three-layer fully connected autoencoder neural network, the objective is to minimize the loss function, and the AdamW optimizer is used to iteratively update the parameters. The formula for calculating the loss function is: ; In the formula, For loss function, The number of samples in a training batch. The sample number. For the first The input vector of each sample; The reconstructed vector of the input vector. This represents the sum of squares of the differences in each dimension.

[0009] In the above technical solution, in step S4, The calculation formula is: ; In the formula, Sampling time, for Global reconstruction error at sampling time This represents the original reconstruction error. For smoothing coefficients; The calculation formula is: ; In the formula, For feature dimension index, For input dimensions, for Sampling time feature dimension The input vector, for Sampling time feature dimension The reconstructed vector of the input vector; The calculation formula is: ; In the formula, for Within a scrolling window of the number of sampling points The median baseline; in, The calculation formula is: ; In the formula, To calculate the median of the reconstruction error sequence within the parentheses, From time Up to the current moment Continuous The global reconstruction error sequence corresponding to each sampling point This represents the number of sampling points; The calculation formula is: ; In the formula, This is the set of window feature vectors corresponding to the target parameters and residual features.

[0010] In the above technical solution, in step S4, or The judgment condition is: it is required that in The sliding window has After the alarm signal appears, it is upgraded to ,in, High-level candidate signals that do not meet the persistence condition remain as Or no alarm will be triggered; Among them, the comprehensive observation level output Represented as: ; In the formula, Based on the reconstruction of abnormal observation signals, To reconstruct the abnormal observation signal for the target, To observe the deterioration trend of the target parameter. For observing signals at extreme peaks, This is a signal for observing thermal coupling anomalies. Representing the logical "OR", when any observation level signal is true, the combined observation signal is... Established.

[0011] Integrated alarm level output Represented as: ; In the formula, This is a continuous basic abnormal alarm signal. To reconstruct abnormal alarm signals for the target, This is an alarm signal for abnormal target parameters. This is an extreme spike warning signal. This is a thermal coupling abnormality alarm signal.

[0012] In the above technical solution, the formula for calculating the daily health score in step S5 is as follows: ; In the formula, Rate your daily health. Where m is the date and m is the alarm channel number. For the first The weight of each channel, For the first The severity coefficient of each channel under the observed signal or alarm signal state. For the first Each channel on date The number of hits within; The criteria for determining the warning level are as follows: At that time, the health level was healthy; At that time, the health level was under a yellow alert. At that time, the health level was under a red alert. The structured early warning results include a list of warning dates, a list of observation dates, daily health scores, the warning level for the last day, and the alarm source channel identifier.

[0013] A second aspect of the present invention is a system for implementing the multi-channel engine fault early warning method of the encoder, the system comprising a data acquisition and preprocessing module, a time-series feature construction module, an autoencoder modeling module, a multi-channel alarm judgment and fusion module, and an early warning evaluation and result output module; The data acquisition and preprocessing module is used to implement step S1; the time series feature construction module is used to implement step S2; the autoencoder modeling module is used to implement step S3; the multi-channel alarm judgment and fusion module is used to implement step S4; and the early warning evaluation and result output module is used to implement step S5.

[0014] A third aspect of the present invention is an electronic device comprising: one or more processors; and a memory for storing one or more programs, wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the self-encoder multi-channel engine fault early warning method.

[0015] A fourth aspect of the present invention is a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the self-encoder multi-channel engine fault early warning method.

[0016] A fifth aspect of the present invention is a computer program product comprising computer-executable instructions which, when executed, implement the self-encoder multi-channel engine fault early warning method.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. Unsupervised learning, no fault samples required: This invention uses an autoencoder that only requires normal operating condition data to train the model, which solves the problems of scarce engine fault data and imbalance of positive and negative samples, and has good engineering applicability.

[0018] 2. Decoupling of operating conditions and reducing false alarms: By fitting the operating condition baseline through ridge regression and calculating residual characteristics, the interference of operating condition changes such as speed and torque on the target parameters is effectively eliminated, allowing the model to focus on the detection of abnormal modes themselves and significantly reducing the false alarm rate.

[0019] 3. Rich temporal features, capturing progressive anomalies: The sliding window statistical features (mean, standard deviation, quantiles) can characterize the temporal dynamics of data from multiple dimensions and have high sensitivity to the early weak signals of progressive faults.

[0020] 4. Multi-channel fusion for comprehensive coverage: It integrates multiple alarm channels such as reconstruction error, residual drift, absolute value gating, extreme spikes, directional gating, and thermal coupling, which can capture abnormal patterns from different dimensions and cover a variety of fault types, including sudden and progressive faults.

[0021] 5. Multi-day continuous suppression to reduce occasional false alarms: Through continuous cross-day testing (at least 2 out of 3 days), it effectively filters out occasional false alarms caused by sensor noise and operating condition fluctuations, thereby improving the reliability of alarms.

[0022] 6. Individualized engine modeling: Each engine is processed independently, and its own historical data is used to establish a baseline, which realizes individualized fault warning and avoids the problem that a uniform threshold model is insensitive to the individual differences of different engines. Attached Figure Description

[0023] Figure 1 The diagram shown is a flowchart of a multi-channel fault early warning method for engines based on a self-encoder according to the present invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0025] Example 1 A method for multi-channel fault early warning of engines based on self-encoder, referring to Figure 1 This includes the following steps: Step S1, Data Acquisition and Preprocessing: Acquire the operating data of the engine multi-dimensional sensor, use ridge regression to decouple the target parameters in the engine multi-dimensional sensor operating data to obtain the fitted operating condition baseline, calculate the difference between the actual value and the predicted value of the fitted operating condition baseline and use it as the residual feature.

[0026] Furthermore, the vehicle network platform acquires multi-dimensional sensor operating data of the engine by engine number and collection time. The multi-dimensional sensor operating data of the engine includes numerical parameters such as engine speed, actual torque percentage, oil temperature, engine water temperature, oil pressure, coolant temperature, intake air volume, fuel gas volume, EGR flow, catalytic converter exhaust temperature, and fan speed.

[0027] Among them, engine speed, actual torque percentage, oil temperature and engine coolant temperature are used in the baseline fitting of the operating conditions; oil pressure or coolant temperature is used as the target parameter; the target parameter, the other numerical sensor parameters and the newly generated feature columns are used together for the sliding window statistics in step S2 and the autoencoder modeling in step S3; coolant temperature, oil temperature and fan speed are also used for absolute value, direction gating and thermal coupling alarm in step S4.

[0028] Ridge regression is used to decouple the target parameters from the operating conditions. The operating condition-related features such as engine speed, actual torque percentage, oil temperature, and engine coolant temperature are used as independent variables, and the target parameters (such as oil pressure or coolant temperature) are used as dependent variables. The operating condition baseline is fitted, and the difference between the actual value of the target parameter and the baseline prediction value is calculated and used as the residual feature to reduce the impact of operating condition changes on the target parameters.

[0029] Furthermore, the process of decoupling the target parameters from operating conditions using ridge regression can be represented as follows: ; In the formula, For all candidate ridge regression coefficient vectors In the middle, determine the coefficient vector that minimizes the objective function within the curly braces. This is a vector composed of the target parameters at each sampling time. The design matrix consists of engine speed, actual torque percentage, oil temperature, and engine coolant temperature. Let be the vector of ridge regression coefficients to be determined. is the regularization coefficient, which is set to 1.0 in this embodiment. Represents the L2 norm, The baseline predicted value of the target parameter under operating conditions.

[0030] The formula for calculating the residual characteristics is: ; In the formula, resid is the difference between the actual value of the target parameter and the predicted value of the baseline of the operating condition.

[0031] Step S2, Time-series Feature Engineering: Using the engine multidimensional operating data from step S1 and As input features, the mean, standard deviation and quantile features of the engine's multidimensional operating data and residual features are extracted through a sliding time window to obtain the corresponding window feature vector; Define the sliding window size There are 10 time points, corresponding to 10 minutes when the sampling period is 1 minute. The input features of step S2 include the numerical sensor features obtained in step S1.1 and the residual features generated in step S1.2. For each input feature column, the mean, standard deviation, 10th quantile (p10), 50th quantile (p50), and 90th quantile (p90) are calculated within a window to simultaneously characterize the local level, the degree of fluctuation, and the distribution range, thereby enhancing the ability to identify gradual shifts and short-term anomalies. Let the input numerical features be... Dimensions, selected window statistics are If the species is specified, the output will be: Dimensions; This embodiment Therefore, it is expanded to dimension, It can be configured according to the sampling frequency and fault type.

[0032] The input features are a sequence of input features organized by individual engine and ordered by acquisition time. For each sampling time, the input features are defined by that time and its adjacent time units. A window is constructed using historical sampling points, and the window's statistical features, such as mean, standard deviation, and quantiles, are output at that time. In practice, the window results, based solely on historical data, can be obtained by reversing the order of data and restoring the chronological order through rolling statistics, thus avoiding the introduction of data from later times into the current features. Missing values ​​due to insufficient effective sampling points at the beginning of the window are filled using forward and backward padding to ensure consistent feature dimensions across time points; this padding is only used for missing value handling and does not add new samples.

[0033] Step S3, Autoencoder Modeling: Construct a three-layer fully connected autoencoder neural network to standardize the window feature vector of the normally operating engine in Step S2, obtaining the input vector. A three-layer fully connected autoencoder neural network is trained using the input vector to obtain the reconstructed vector of the input vector. .

[0034] The three-layer fully connected autoencoder neural network consists of an encoder and a decoder. The encoder performs dimensionality reduction on the standardized window feature vector. Specifically, the input dimension is set as follows: The dimension of the window feature vector output in step S2 is the dimension of the encoder's final hidden encoding. That is, the dimension of the low-dimensional encoding vector; the output dimension is the same as the input dimension, and The encoder layers use the ReLU activation function to reduce dimensionality, enabling the model to learn low-dimensional essential feature representations of normal data.

[0035] The dimensionality reduction process of the encoder is represented as follows: ; ; In the formula, and These are the intermediate representation and bottleneck encoding of the encoder, respectively. The weight matrix for the intermediate representation of the encoder. The weight matrix for encoding the bottleneck of the encoder. This is the bias vector for the intermediate representation of the encoder. The bias vector for the bottleneck encoding of the encoder. For activation functions; The decoder restores the input vector after dimensionality reduction by the encoder to obtain a reconstructed vector. The formula for calculating the reconstructed vector is as follows: ; In the formula, Here is the weight matrix of the decoder. This is the bias vector for the decoder.

[0036] The autoencoder is trained using only the sliding window statistical feature vector generated in step S2 from a normally operating engine. Before training, the mean of the window feature vector from the normally operating engine is used. and standard deviation For the 3D features After standardization, the autoencoder outputs a reconstructed vector. Mean squared error (MSE) is used as the loss function, and the AdamW optimizer iteratively updates the parameters. The training objective is to minimize the MSE. After training, the three-layer fully connected autoencoder neural network can accurately reconstruct features under normal operating conditions; abnormal operating conditions, due to deviations from the normal pattern, result in larger reconstruction errors.

[0037] The formula for calculating the loss function is as follows: ; In the formula, The number of samples in a training batch; The sample number; For the first One input vector; The reconstructed vector of the input vector; This represents the sum of squares of the differences in each dimension. By minimizing... This enables the autoencoder to learn the joint relationships between features under normal operating conditions.

[0038] Step S4: Design a multi-channel alarm fusion strategy: Calculate the original reconstruction error of each dimension feature based on the input vector and reconstruction vector from step S3. ,right The global reconstruction error sequence is obtained after smoothing with an exponentially weighted moving average. Global reconstruction error and reconstruction error drift The target reconstruction error is obtained from the window feature vector based on the target parameters and residual features. ,based on , , The system constructs a base and target reconstruction error channel, an extreme peak channel, a direction-gated absolute value channel, a thermal coupling channel, and a multi-day persistence suppression channel based on the target parameters and the fan's window feature vector. , and Identifying persistent and progressive anomalies, extreme spike channels are based on , The system identifies sudden anomalies. The directional gating absolute value channel distinguishes between normal operating condition fluctuations and anomalies with a clear direction of degradation based on the window feature vector of the target parameter. The thermal coupling channel identifies composite anomalies of decreased cooling system efficiency based on the window feature vector of the target parameter. The multi-day persistent suppression channel summarizes the identification results from each channel daily and performs continuous verification, ultimately determining the comprehensive observation-level output. Or integrated alarm level output ; Furthermore, The calculation formula is: ; In the formula, Sampling time, for Global reconstruction error at sampling time This represents the original reconstruction error. Let be the smoothing coefficient, where ; The calculation formula is: ; In the formula, For feature dimension index, For input dimensions, for Sampling time feature dimension The input vector, for Sampling time feature dimension The reconstructed vector of the input vector; The calculation formula is: ; In the formula, for Within a scrolling window of the number of sampling points The median baseline.

[0039] in, The calculation formula is: ; In the formula, To calculate the median of the reconstruction error sequence within the parentheses, From time Up to the current moment Continuous The global reconstruction error sequence corresponding to each sampling point Let be the number of sampling points, where .

[0040] The calculation formula is: ; In the formula, This is the set of window feature vectors corresponding to the target parameters and residual features.

[0041] The base and target reconstruction error channels were calculated respectively within the first 20 days of the calibration period in the most recent 30 days of data. , and The quantile threshold. The basic and target reconstruction error channels include a basic sub-channel and a target reconstruction sub-channel, which determine whether the engine operating status deviates from the normal data pattern and are used to identify persistent and progressive anomalies; among them, the basic sub-channel requires testing period and Simultaneously exceeding the threshold and meeting the cumulative hit count requirement within the 6-hour window; target reconstruction sub-channel requirements. If the threshold is exceeded, and considering directional gating, secondary confirmation of the basic sub-channel, and multi-day consistency conditions, this channel directly uses the input vector and reconstructed vector from step S3.

[0042] The extreme spike channel (Spike) is detected and The extreme peak values ​​and their cumulative values ​​within a set time window are used to identify sudden anomalies. During the calibration period, the extreme quantile threshold (p99.9) is calculated; during the testing period... and A spike alarm is triggered when the threshold is exceeded and the cumulative number of minutes reaches a specified number within a 12-hour window, in order to capture sudden anomalies.

[0043] The directional gating absolute value channel combines the absolute values ​​of target parameters such as oil pressure and coolant temperature, the direction of change relative to historical baselines, and stable operating conditions represented by speed and torque to distinguish between normal operating condition fluctuations and anomalies with a clear direction of deterioration. It sets thresholds for the absolute values ​​of target parameters and requires that, under stable operating conditions (speed 800-1600 rpm, torque 10-60%), the target parameters must meet directional deviation conditions relative to the 7-day rolling median baseline, i.e., oil pressure drops by more than 15% or coolant temperature rises by more than 5%. After the cumulative number of hits within a 6-hour window reaches the target, an alarm is triggered in conjunction with a multi-day consistency check.

[0044] The thermal coupling channel (dedicated to the water temperature model) integrates the thermal coupling relationship between coolant temperature, engine oil temperature, oil-water temperature difference, and fan speed to identify complex anomalies such as decreased cooling system efficiency. When four conditions are simultaneously met—exceeding the limit for coolant temperature, engine oil temperature, oil-water temperature difference, and low fan speed—and the cumulative hit count reaches a certain threshold within a 12-hour window, a thermal coupling alarm is triggered in conjunction with a multi-day consistency check to capture complex fault modes indicating decreased cooling system efficiency.

[0045] Each alarm channel first generates a candidate alarm signal at the sampling point level and performs preliminary confirmation by accumulating the number of hits within a set time window. Then, a multi-day persistence suppression channel performs cross-day persistence verification on each candidate alarm signal. If anomalies occur for at least two days within a three-day sliding window, its persistence is confirmed to suppress false alarms caused by sensor noise and occasional operating condition fluctuations. The results from each channel after verification by the multi-day persistence suppression channel are fused using a logical OR method, and output when any observation-level condition is met. Output when any alarm level condition is met. ,and The priority is higher than .

[0046] The multi-day persistent suppression channel serves as a post-processing gating mechanism and does not generate raw anomaly scores independently. It is required that... The sliding window has After the alarm signal appears, it is upgraded to ,in, High-level candidate signals that do not meet the persistence condition remain as Alternatively, no alarm may be triggered to filter out occasional fluctuations on a single day.

[0047] Among them, the comprehensive observation level output Represented as: ; In the formula, This represents the overall observation-level output at the current sampling time, and is a Boolean variable; when the output of any observation-level sub-channel is 1... A value of 1 indicates that there are abnormal signs in the engine's operating status that require continuous monitoring, but the alarm-level judgment criteria have not yet been met; when the output of all observation-level sub-channels is 0, The value is 0; This represents the observation-level output of the basic reconstruction anomaly sub-channel; when the global reconstruction error... and residual drift When the number of occurrences of the abnormal sampling points exceeds the corresponding observation-level threshold, and the cumulative number of hits within the preset time window reaches the observation-level requirement, The value is 1 if it is not 0 otherwise, and is used to characterize the overall operating characteristics of the engine as a continuous deviation from the normal data pattern. This represents the observation-level output of the target reconstruction error sub-channel; when the target parameters and their residual characteristics correspond to the target reconstruction error... When the observation-level threshold is exceeded and the preset cumulative hit or auxiliary confirmation conditions are met, The value is 1 if it is not 0 otherwise, and is used to characterize an abnormal deviation between the target parameters such as oil pressure and coolant temperature and the normal correlation mode. This represents the observation-level output of the directional gating absolute value channel; when the target parameter reaches the observation-level absolute value threshold, and under stable operating conditions limited by engine speed and actual torque percentage, it exhibits a preset abnormal change direction relative to the historical rolling baseline, and the cumulative number of hits reaches the observation-level requirement, The value is 1 if it is not 0 otherwise, and it is used to identify abnormal target parameters that have a clear direction of degradation but have not yet reached the alarm level. This represents the observation-level output of the extreme peak channel; when the global reconstruction error... or residual drift If the extreme quantile threshold determined during the calibration period is exceeded, and the cumulative number of occurrences or duration reaches a set value within a preset time window, The value is 1 if it is not 0 otherwise, and it is used to identify sudden anomalies that occur rapidly in a short period of time. This represents the observation-level output of the thermally coupled channel; when a combination of thermally relevant parameters such as coolant temperature, oil temperature, oil-water temperature difference, and fan speed deviates from normal heat dissipation patterns, and reaches the observation-level cumulative condition, The value is 1 if the condition is met, and 0 otherwise, it is used to identify early composite anomalies indicating a decrease in heat dissipation efficiency. For engine fault warning models that do not include thermal coupling judgment, this value is used... Set to 0; This represents the logical "OR", meaning that when any Boolean variable involved in the operation is 1, the result of the logical OR operation is 1. When any observation-level signal is true, Established.

[0048] The comprehensive alarm level output is represented as follows: ; In the formula, This represents the overall alarm level output at the current sampling time; it is a Boolean variable. When the output of any alarm level sub-channel is 1... A value of 1 indicates that the engine operating status has reached the fault warning condition; when the output of all alarm-level sub-channels is 0... The value is 0; when and When both are 1, with The corresponding alarm level status is used as the final output; This indicates the basic alarm output that has been continuously verified and confirmed over several days; it is first based on the global reconstruction error. Residual drift The system generates basic candidate alarms based on the cumulative hit rate within a preset time window, and then summarizes these candidate alarms daily. When the number of abnormal days reaches a preset requirement within a set multi-day sliding window, an alarm is triggered. The value is 1 if the engine malfunction is not present, and 0 otherwise. This is used to confirm that the overall engine malfunction is persistent across days and to suppress false alarms caused by sensor noise or daily operating condition fluctuations. This indicates the alarm level output of the target reconstruction error sub-channel; when the target reconstruction error... When the alarm level threshold is exceeded, and the conditions of cumulative hits, basic anomaly confirmation, or multi-day persistence are met, The value is 1 if the target parameter deviates significantly or continuously from the normal feature relationship otherwise. This indicates the alarm level output of the direction-gated absolute value channel; when the target parameter reaches the alarm level absolute value threshold, and under stable operating conditions, exhibits a preset degradation direction relative to the historical rolling baseline, and meets the requirements of cumulative hits and multi-day persistence, The value is 1 if it is not 0 otherwise. It is used to confirm persistent abnormalities with clear physical significance, such as low oil pressure or high coolant temperature. This indicates the alarm-level output of the extreme spike channel; when the global reconstruction error... or residual drift When the extreme anomaly threshold is exceeded, and the cumulative number of alarms or the duration of alarms reaches the preset time window, the alarm will be triggered. The value is 1 if it is not 0 otherwise, and is used to confirm sudden and large-amplitude anomalies. This indicates the alarm level output of the thermal coupling channel; when a preset abnormal combination of thermal-related conditions such as excessive coolant temperature, excessive engine oil temperature, abnormal oil-water temperature difference, and insufficient fan response is formed, and the cumulative hit and multi-day continuous requirements are met, A value of 1 is used to identify combined faults such as decreased cooling system efficiency or abnormal thermal management; for engine fault warning models that do not include thermal coupling judgment, the value is 0. Set to 0; This indicates a logical "OR", meaning that when any alarm level output is 1, the overall alarm signal WARN is 1.

[0049] Step S5, Daily-level evaluation and structured output: Based on the results obtained in step S4 and It summarizes alarm status daily, marks daily observation days (WATCH_DAY) and alarm days (WARN_DAY), calculates daily health scores, determines warning levels, and outputs structured warning results.

[0050] The formula for calculating the daily health score is as follows: ; In the formula, Rate your daily health. Indicates the date; Indicates the alarm channel number; Indicates the first The weight of each channel; This indicates the severity coefficient of the channel in the observation or alarm level state; Indicates the first Each channel on date The number of hits within.

[0051] Will Limited to a range of 0 to 100, the more times a channel is hit or the higher the severity, the lower the health score.

[0052] Based on the severity and frequency of alarms from each channel on the same day, a health score of 0-100 is generated, and the warning level is determined. The warning level is determined according to the following rules: when the health score is greater than 80 ( When the time is ), the health level for that day is healthy ( When the health score is between 70 and 80 ( When the health level is yellow (as indicated by the warning), the health alert level for that day is yellow. When the health score is less than 70 ( When the health level is red alert, the health level for that day is red alert. ).

[0053] The structured early warning results include: a list of alarm dates, a list of observation dates, daily health scores, the warning level of the last day (healthy / yellow warning / red warning), and an alarm source channel identifier, which are used for vehicle maintenance decision support and fault diagnosis assistance.

[0054] Example 2 A system for implementing the self-encoder multi-channel engine fault early warning method in Embodiment 1, the system comprising a data acquisition and preprocessing module, a time-series feature construction module, a self-encoder modeling module, a multi-channel alarm judgment and fusion module, and an early warning evaluation and result output module.

[0055] The data acquisition and preprocessing module is used to implement step S1; the time series feature construction module is used to implement step S2; the autoencoder modeling module is used to implement step S3; the multi-channel alarm judgment and fusion module is used to implement step S4; and the early warning evaluation and result output module is used to implement step S5.

[0056] Furthermore, the data acquisition and preprocessing module, the time series feature construction module, the autoencoder modeling module, the multi-channel alarm judgment and fusion module, and the early warning assessment and result output module can be deployed on a server. The modules can transmit engine operating data, time series features, autoencoder reconstruction results, multi-channel alarm results, and early warning assessment results sequentially through data interfaces.

[0057] Example 3 This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the self-encoder multi-channel engine fault early warning method described in Embodiment 1 above.

[0058] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0059] Example 4 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the self-encoder multi-channel engine fault early warning method described in Embodiment 1 above.

[0060] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0061] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0062] Example 5 This embodiment provides a computer program product including computer-executable instructions stored on a machine-readable storage medium (such as a disk, flash memory, or optical disk). When executed by at least one data processing device (such as a microprocessor or digital signal processor), the instructions cause the data processing device to implement the steps in the self-encoder multi-channel engine fault early warning method described in Embodiment 1.

[0063] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0064] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0065] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0066] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A self-encoder multi-channel engine fault early warning method, characterized in that, Includes the following steps: Step S1: Collect multidimensional engine operating data including target parameters, establish the operating condition baseline of the target parameters using ridge regression, and generate residual features; Step S2: Extract the mean, standard deviation, and quantile features of the engine's multidimensional operating data and residual features through a sliding time window to obtain the corresponding window feature vector; Step S3: Standardize the window feature vector of the normally operating engine from step S2 to obtain the input vector. A three-layer fully connected autoencoder neural network is trained using the input vector to obtain the reconstructed vector of the input vector. ; Step S4, based on and Calculate global reconstruction error and reconstruction error drift The target reconstruction error is obtained from the window feature vector based on the target parameters and residual features. ,based on , , The system constructs a base and target reconstruction error alarm channel, an extreme spike alarm channel, a direction-gated absolute value alarm channel, a thermal coupling alarm channel, and a multi-day persistent suppression channel based on the window feature vectors of the target parameters. The base and target reconstruction error alarm channel identifies persistent and gradual anomalies; the extreme spike alarm channel identifies sudden anomalies; the direction-gated absolute value alarm channel distinguishes between normal operating condition fluctuations and anomalies with a clear direction of degradation; the thermal coupling alarm channel identifies composite anomalies of decreased cooling system efficiency; and the multi-day persistent suppression channel summarizes the identification results of each warning channel daily and performs continuous verification, ultimately determining the output and providing a comprehensive observation-level output. Or integrated alarm level output ; Step S5, based on or The alarm status is summarized daily, and daily observation day and alarm day markers are generated. The daily health score is calculated, the warning level is determined, and the structured warning results are output.

2. The self-encoder multi-channel engine fault early warning method as described in claim 1, characterized in that, In step S1, the multi-dimensional engine operating data includes engine speed, actual torque percentage, oil temperature, engine coolant temperature, oil pressure, coolant temperature, intake air volume, fuel gas volume, EGR flow rate, catalytic converter exhaust temperature, and fan speed. With oil temperature or engine coolant temperature as the target parameter, ridge regression is used to decouple the target parameter from the operating conditions. With engine speed, actual torque percentage, oil temperature, and engine coolant temperature as independent variables and the target parameter as the dependent variable, the operating condition baseline is fitted. The process of decoupling the target parameters from operating conditions using ridge regression can be represented as follows: ； In the formula, For all candidate ridge regression coefficient vectors In the middle, determine the coefficient vector that minimizes the objective parameter within the curly braces. This is a vector composed of the target parameters at each sampling time. The design matrix consists of engine speed, actual torque percentage, oil temperature, and engine coolant temperature. Let be the vector of ridge regression coefficients to be determined. The regularization coefficient is . Represents the L2 norm, The baseline predicted value for the target parameter under operating conditions; The formula for calculating the residual characteristics is: ； In the formula, resid is the residual characteristic.

3. The self-encoder multi-channel engine fault early warning method as described in claim 1, characterized in that, In step S3, the mean of the window feature vector of a normally operating engine is used. Standard deviation The quantile feature is the first quantile feature of the window feature vector of a normally operating engine. 3D features Standardization processing is required; The three-layer fully connected autoencoder neural network includes an encoder and a decoder. The encoder performs dimensionality reduction on the standardized window feature vector, and the decoder restores the input vector after dimensionality reduction by the encoder to obtain the reconstructed vector of the input vector. The dimensionality reduction process of the encoder is represented as follows: ； ； In the formula, and These are the intermediate representation and bottleneck encoding of the encoder, respectively. The weight matrix for the intermediate representation of the encoder. The weight matrix for encoding the bottleneck of the encoder. This is the bias vector for the intermediate representation of the encoder. The bias vector for the bottleneck encoding of the encoder. For activation functions; The The calculation formula is: ； In the formula, Here is the weight matrix of the decoder. This is the bias vector for the decoder; During the training of a three-layer fully connected autoencoder neural network, the objective is to minimize the loss function, and the AdamW optimizer is used to iteratively update the parameters. The formula for calculating the loss function is: ； In the formula, For loss function, The number of samples in a training batch. The sample number. For the first The input vector of each sample; The reconstructed vector of the input vector. This represents the sum of squares of the differences in each dimension.

4. The self-encoder multi-channel engine fault early warning method as described in claim 1, characterized in that, In step S4 The calculation formula is: ； In the formula, Sampling time, for Global reconstruction error at sampling time This represents the original reconstruction error. For smoothing coefficients; The calculation formula is: ； In the formula, For feature dimension index, For input dimensions, for Sampling time feature dimension The input vector, for Sampling time feature dimension The reconstructed vector of the input vector; The calculation formula is: ； In the formula, for Within a scrolling window of the number of sampling points The median baseline; in, The calculation formula is: ； In the formula, To calculate the median of the reconstruction error sequence within the parentheses, From time Up to the current moment Continuous The global reconstruction error sequence corresponding to each sampling point This represents the number of sampling points; The calculation formula is: ； In the formula, This is the set of window feature vectors corresponding to the target parameters and residual features.

5. The self-encoder multi-channel engine fault early warning method as described in claim 1, characterized in that, In step S4 or The judgment condition is: it is required that in The sliding window has After the alarm signal appears, it is upgraded to ,in, High-level candidate signals that do not meet the persistence condition remain as Or no alarm will be triggered; Among them, the comprehensive observation level output Represented as: ； In the formula, Based on the reconstruction of abnormal observation signals, To reconstruct the abnormal observation signal for the target, To observe the deterioration trend of the target parameter. For observing signals at extreme peaks, This is a signal for observing thermal coupling anomalies. Representing the logical "OR", when any observation level signal is true, the combined observation signal is... Established; Integrated alarm level output Represented as: ； In the formula, This is a continuous basic abnormal alarm signal. To reconstruct abnormal alarm signals for the target, This is an alarm signal for abnormal target parameters. This is an extreme spike warning signal. This is a thermal coupling abnormality alarm signal.

6. The self-encoder multi-channel engine fault early warning method as described in claim 1, characterized in that, In step S5, the formula for calculating the daily health score is: ； In the formula, Rate your daily health. Where m is the date and m is the alarm channel number. For the first The weight of each channel, For the first The severity coefficient of each channel under the observed signal or alarm signal state. For the first Each channel on date The number of hits within; The criteria for determining the warning level are as follows: At that time, the health level was healthy; At that time, the health level was under a yellow alert. At that time, the health level was under a red alert. The structured early warning results include a list of warning dates, a list of observation dates, daily health scores, the warning level for the last day, and the alarm source channel identifier.

7. A system for implementing the self-encoder multi-channel engine fault early warning method as described in claim 1, characterized in that, The system includes a data acquisition and preprocessing module, a time-series feature construction module, an autoencoder modeling module, a multi-channel alarm judgment and fusion module, and an early warning assessment and result output module. The data acquisition and preprocessing module is used to implement step S1; the time series feature construction module is used to implement step S2; the autoencoder modeling module is used to implement step S3; the multi-channel alarm judgment and fusion module is used to implement step S4; and the early warning evaluation and result output module is used to implement step S5.

8. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the self-encoder multi-channel engine fault early warning method as described in claim 1.

9. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed, are used to implement the self-encoder multi-channel engine fault early warning method as described in claim 1.

10. A computer program product, characterized in that, The aforementioned computer program product includes computer-executable instructions that, when executed, implement the self-encoder multi-channel engine fault early warning method as described in claim 1.

Images