A system and method for detecting faults in electronic components

Abstract

The application discloses a kind of electronic component fault detection system and method, it is related to electronic component fault detection technical field, relies on multi-source high-speed data acquisition and online pre-processing, sensor noise and missing value are filtered and packed into micro-batch dataset after, real-time determination fault is carried out again through light weight intelligent fault detection model and incremental update parameter, then with industrial control system is closely integrated, utilizes multi-level risk function to execute millisecond level closed loop protection and returns feedback;Finally, historical record and marked information are uploaded to cloud, and full life cycle management is realized using big data analysis and model iteration;By wavelet filtering, incremental learning and cross-level data integration etc., can effectively shorten fault detection delay, reduce downtime loss, and provide guarantee for predictive maintenance, greatly improve production line safety and efficiency, so as to meet the demand of industrial scene to fault identification and real-time response.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic component fault detection, in particular to an electronic component fault detection system and method. BACKGROUND

[0002] In today's highly automated and compact rhythm industrial production environment, various production lines (such as automated assembly lines, continuous process chemical lines or high-precision machine tool production lines) often acquire multi-source real-time data such as temperature, pressure, vibration and current with millisecond-level sampling frequency. At the same time, in order to ensure efficient production and stable quality, enterprises usually configure multiple sets of monitoring equipment and sensor networks to implement strict detection on key equipment. However, in the face of actual working conditions of high-speed operation of production lines and frequent environmental disturbances, if potential fault signs cannot be quickly identified, small faults often quickly evolve into equipment overheating before they are discovered, and millisecond-level data streams must be analyzed in real time with high fidelity, and quickly linked with the control system to minimize the risk of fault spread and maintain normal production rhythm.

[0003] Currently, the core problem faced by the industrial field in real-time fault detection is how to quickly and accurately analyze massive and noisy multi-source sensor data, and deeply integrate the detection results with the industrial control system without delay or with extremely low delay. Traditional monitoring mostly uses alarm mechanisms based on fixed thresholds, which are difficult to adapt to abnormal patterns caused by equipment aging, environmental changes or working condition fluctuations; when massive data is transmitted to the cloud for complex model analysis, real-time is lost due to network bandwidth and delay limitations. In addition, the standards of sensors and control systems used by different production lines are not the same, and there are great challenges in interface compatibility and data alignment management. If the detection results cannot accurately identify faults and trigger corresponding emergency plans or warnings within milliseconds, local faults can quickly expand into more widespread equipment failures, resulting in a significant decrease in production efficiency and safety. Therefore, an integrated solution that takes into account on-site rapid analysis, noise filtering, model dynamic updating and cloud big data collaborative optimization is urgently needed to fundamentally solve the problem of efficient real-time fault detection of multi-source sensor data in industrial environments, and to achieve true millisecond-level closed-loop response through close integration with the control system.

[0004] Therefore, the present application provides an electronic component fault detection system and method. SUMMARY

[0005] (I) Technical problems solved

[0006] In view of the deficiencies of the prior art, the present application provides a fault detection system and method for electronic components, which relies on multi-source high-speed data acquisition and online preprocessing, filters sensor noise and missing values, and packs them into a micro-batch data set, then uses a lightweight intelligent fault detection model to determine faults in real time and incrementally update parameters, and then integrates closely with an industrial control system, uses a multi-level risk function to perform millisecond-level closed-loop protection and feedback, and finally uploads historical records and label information to the cloud, uses big data analysis and model iteration to realize full life cycle management, provides support for predictive maintenance, and greatly improves production line safety and efficiency, thereby meeting the needs of industrial scenarios for fault identification and real-time response, thereby solving the technical problems described in the background art.

[0007] (Two) Technical solutions

[0008] To achieve the above object, the present application is realized by the following technical solutions:

[0009] A fault detection method for electronic components, comprising,

[0010] When high-frequency sampling triggers multi-source data flow, wavelet filtering and missing value placeholder are used on the original sensor sequence to remove false data, mark transient anomalies, and perform interpolation on delayed data to generate micro-batch packaging to carry sensor ID, timestamp and transient suspicious flag;

[0011] The edge side uses kernel mapping and acceleration transformation to extract feature vectors, and calls an incremental machine learning algorithm to output fault labels and fault confidence in real time, and corrects the model parameters of the fault detection model to adapt to environmental disturbances;

[0012] The on-site PLC or DCS performs high-priority actions or issues moderate warnings based on the multi-level risk index, and records the actual execution and fault labels back to the fault detection model, and optimizes the model parameters based on the real fault confirmation process;

[0013] After collecting micro-batch packaging, fault labels, fault confidence, and multi-level risk index information, the cloud or background system aggregates multi-source data and uses aggregation operations to mine fault patterns, and then issues the optimized fault detection model and maintenance strategy to the edge side, and performs model iteration under the comparison of fault labels.

[0014] Preferably, after collecting the original sensor sequence, align it by timestamp, extract feature coefficients at multiple scales by discrete wavelet transform, and when the energy proportion of the target scale segment is continuously higher than expected, mark it as high-frequency interference and perform suppression operation;

[0015] Transient anomalies are identified by a wavelet energy index for detecting high-energy component anomalies, and a transient anomaly flag is added.

[0016] Preferably, the obtained reconstructed signal sequence is divided into multiple micro-batch data sets, and after recording the metadata at the micro-batch level, the wavelet energy index and the corresponding flag are retained for the corresponding data points within the micro-batch data set, and then the micro-batch package is generated by summarizing.

[0017] Preferably, after simple interpolation or skipping processing of missing values, all sensor channels in the micro-batch data set are obtained and spliced with the wavelet energy index to form an initial feature vector sequence; a self-definable kernel function is introduced to map the initial feature vector sequence to obtain a reduced dimension feature representation, and an acceleration transformation index is introduced to dynamically adjust the dimension reduction strength in the kernel mapping process, and the output feature vector.

[0018] Preferably, on the basis of the feature vector, a lightweight fault detection model is used to quickly predict whether there is a fault and its severity, and the model parameters are updated in small batches or online according to the real feedback;

[0019] A fault confidence is added to the output layer of the model, and when the fault confidence exceeds the preset confidence threshold, an emergency control strategy is quickly executed, and if the fault confidence does not exceed the preset confidence threshold, only a warning or record is selected.

[0020] Preferably, the obtained detection result is transmitted to the on-site PLC or DCS; and a multi-level risk index of the fault confidence and the current system key state is introduced, and according to the landing point of the multi-level risk index in different threshold intervals, the industrial control system executes the corresponding level of protection or alarm, and executes the action instruction directly to the equipment on the production line through the PLC or DCS.

[0021] Preferably, after the PLC or DCS returns the actual action and running state data of the system, the executed action and the fault detection result are paired, and the newly generated system state;

[0022] If the previous fault label has deviation or is insufficient or excessive in judging the fault level, the real label is used to correct the lightweight fault detection model; if the response effect for the same fault type is not good for multiple times, the relationship between the multi-level risk index and the final process result is used to fine-tune the related parameters of the multi-level risk index.

[0023] Preferably, the fault detection records accumulated on the on-site edge device or industrial PC are periodically or real-time transmitted to the cloud or factory background system; a global aggregation operator is constructed to comprehensively measure the global aggregation of a large amount of historical data and on-site key parameters, and after data aggregation, a visual panel is constructed on the cloud or factory background system.

[0024] Preferably, the remaining useful life of the equipment or the fatigue degree of the key components is taken as the target to update the fault detection model; if there is a fault mode or an evolution trend of the existing mode, it is recorded and included in the corresponding maintenance strategy library.

[0025] If there is optimization space for the used fault detection model, an improved version can be packaged and distributed to the field edge device; combined with the remaining service life prediction result of the device and the fault mode analysis, a hierarchical maintenance plan is made.

[0026] A fault detection system of electronic components comprises,

[0027] The data acquisition unit, when high-frequency sampling triggers multi-source data flow, uses wavelet filtering and missing value placeholder to remove false data and mark transient anomalies, and performs interpolation operation on delayed data to generate micro-batch packaging to carry sensor ID, timestamp and transient suspicious flag;

[0028] The feature extraction unit extracts the feature vector by using kernel mapping and acceleration transformation on the edge side, and calls the incremental machine learning algorithm to output the fault label and fault confidence in real time, and corrects the model parameters of the fault detection model to adapt to environmental disturbance;

[0029] The early warning unit executes high-priority action or issues moderate early warning according to the multi-level risk index, and records the actual execution and fault label back to the fault detection model, and optimizes the model parameters according to the real fault confirmation process;

[0030] The feedback iteration unit collects micro-batch packaging, fault label, fault confidence, multi-level risk index information, and the cloud or background system collects multi-source data and mines fault patterns by means of aggregation operation, and then sends the optimized fault detection model and maintenance strategy to the edge side, and executes model iteration under the fault label comparison.

[0031] (Three) beneficial effects

[0032] The present application provides a kind of fault detection system and method of electronic components, with following beneficial effects:

[0033] By constructing multi-source high-speed data acquisition, lightweight intelligent reasoning, industrial control closed loop and cross-level data integration between industrial field and cloud, fault detection and response are completed within millisecond time sequence, and through large-scale historical data analysis continuous iteration optimization, the following beneficial effects can be significantly brought:

[0034] With the aid of multi-source data synchronization and wavelet filtering method, false data can be effectively removed and transient anomalies are marked, and output cache Method is transmitted to downstream, which greatly reduces noise interference while ensuring time sequence alignment; and by using kernel mapping and acceleration transformation to obtain feature vector And through the incremental machine learning algorithm, real-time fault judgment and confidence construction are realized, so that high detection precision and self-adaptive ability can be maintained under the influence of multiple factors such as equipment aging and working condition change.

[0035] Utilize multi-level risk index and execution action Emergency plan triggering or early warning of the industrial control system is carried out, so that the potential time window of fault diffusion to shutdown is greatly shortened, and feedback flow makes the model parameters Obtain continuous correction opportunity, gradually reduce false alarm and false negative;

[0036] The field collection and control execution record are uploaded to the cloud or factory background, the fault mode and equipment health trend are mined on a large scale dimension combined with the aggregation operator, and then the optimized fault detection model or maintenance strategy is issued to the edge end, so that the cloud-edge collaborative whole life cycle management is realized.

[0037] Overall, the scheme can complete fault identification and closed-loop response in a millisecond period, meet the harsh demand of high real-time of industrial field, and continuously improve the adaptability of the system to new fault modes or long-period equipment wear through cross-level data integration and online incremental learning, so that the safety and stability are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 It is a flowchart of the fault detection method of the electronic component of the present application;

[0039] Figure 2 It is a structural schematic diagram of the fault detection system of the electronic component of the present application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0041] Please refer to Figure 1 The present application provides a fault detection method of electronic component, comprising,

[0042] Step one, when high-frequency sampling triggers multi-source data flow to rush in, the original sensor sequence Wavelet filtering and missing value placeholder are used to remove false data and mark transient anomalies, and interpolation operation is performed on the delayed data to generate micro-batch packaging The sensor ID, timestamp and transient suspicious flag are carried ;

[0043] Step one includes the following:

[0044] Step 101: Multi-source data alignment and wavelet coefficient-based depth filtering:

[0045] The raw sensor sequence was acquired. ,in Indicates sensor number , Indicates the first Each sampling time , and The number of sensors and the length of the sampling sequence are respectively determined; all sensor data are aligned by timestamp, and missing values ​​or severely distorted data points are initially marked.

[0046] For each original sensor sequence To extract the characteristic coefficients at multiple scales, a discrete wavelet transform is performed. Let the selected wavelet basis functions be... Then for In scale With displacement The wavelet coefficients on can be expressed as:

[0047] ;

[0048] In the formula: For sensors In scale and displacement Wavelet coefficients on; For wavelet basis functions used in multi-scale decomposition (such as Morlet or other functions suitable for fault feature extraction); This is a scale parameter, and its numerical range can be selected based on the sampling frequency and the actual fault characteristic frequency band. These are displacement parameters, corresponding to translation on the time axis; For the sampling time, satisfying , The sampling interval;

[0049] By appropriately reconstructing wavelet coefficients and filtering out low-amplitude high-frequency coefficients, high-frequency noise can be reduced while preserving the transient characteristics of the fault as much as possible. The noise components are determined based on the distribution pattern of wavelet coefficients at each scale. When the energy proportion of the target scale segment is consistently higher than expected, it is marked as high-frequency interference and suppression is performed.

[0050] After performing wavelet filtering, a wavelet energy index is defined to detect high-energy component anomalies. Quickly identify transient anomalies:

[0051] ;

[0052] wherein: is the time instant corresponding to the wavelet energy index of the sensor is the selected set of key scales used to extract the scales that best characterize the fault pattern; , in the wavelet transform, and the time instant corresponding to the displacement; is the contribution of the high energy component to the anomaly;

[0053] When the wavelet energy index exhibits a significant spike within a sliding window or a significant difference from the historical baseline, the data at that time instant is preliminarily marked as a transient suspicious value, with the addition of a transient anomaly flag ;

[0054] In use, through the multi-scale decomposition of discrete wavelets and the introduction of the wavelet energy index , short-term fault signs can be effectively retained, and the time alignment of multiple sensors ensures that different physical quantities are comparable at the same time, providing stable and synchronized input conditions for the lightweight reasoning of Step 2. With the help of the transient anomaly flag , suspected fault signs are made explicit at the earliest stage, providing reference labels for the incremental learning of subsequent fault detection models; the introduction of the wavelet energy index into the online filtering link enables it to more sensitively capture fault transient characteristics while maintaining time-frequency localization capabilities.

[0055] Step 102, Micro-batch Packaging and Time Sequence Labeling:

[0056] The reconstructed signal sequence obtained after discrete wavelet transform denoising, resampling alignment, missing value placeholder, and transient anomaly labeling in Step 101 is divided into multiple micro-batch data sets with a certain length of time window. Each micro-batch data set can be represented as , where represents the micro-batch index; necessary metadata such as sensor number , time range , data labels, etc. are recorded at the micro-batch level;

[0057] If some time instants are added with transient suspicious flags in Step 101 and corresponding flags; this can be used to identify the true failure likelihood in the inference model of step two by secondary screening and adding a completeness indicator to represent the micro-batch data set ;

[0058] After the above packaging is completed, the micro-batch data is uniformly written into the micro-batch package

[0059] ;

[0060] wherein is the start and end time of the micro-batch; next, these micro-batch packages will be directly read and called by the lightweight fault detection model in step two,

[0061] When used, by packaging the preprocessing results into manageable micro-batch packages , it ensures that the subsequent algorithm can quickly access the feature markers and necessary context information, and in subsequent inference, it can improve the accuracy of fault detection, and with the help of high-dimensional wavelet energy index and transient suspicious flag , it can timely mark when transient anomalies are found, and then improve the adaptability to equipment aging or environmental changes.

[0062] Step two, edge side uses kernel mapping and acceleration transformation to extract feature vectors , and calls incremental machine learning algorithm to output fault markers and confidence in real time, and corrects the model parameters of the fault detection model to adapt to environmental disturbances such as equipment aging and sensor drift;

[0063] The step two includes the following contents:

[0064] Step 201, time series representation based on feature kernel mapping and fast dimensionality reduction:

[0065] Read the micro-batch data set and the corresponding marker information from the micro-batch package , including the transient suspicious flag , the wavelet energy index , and the effective data proportion , etc. For each micro-batch data set , first perform simple interpolation or skip processing on missing values; after obtaining all sensor channels in the micro-batch data set , concatenate them with the wavelet energy index to form an initial feature vector sequence ; ​

[0066] If the current micro-batch dataset contains road sensors and the number of selected critical time points is , then a time-series concatenation vector of length can be obtained after unfolding in the time dimension, and the position index of the transient suspicious flag is retained in it;

[0067] To balance real-time performance and strong discriminability, a self-definable kernel function is introduced to map the initial feature vector sequence to obtain the reduced dimension feature representation , whose basic form can be written as:

[0068] ;

[0069] where represents a nonlinear mapping operation based on the kernel function, which contains multiple orthogonal basis components to highlight fault-sensitive features and suppress redundant noise; for example, if a composite kernel is selected, is the feature mapping corresponding to the composite kernel; is the adjustable weight of the kernel component; wavelet energy information can be combined to give differentiated weighting to sensor time series, avoiding the weakening of transient anomalies by ordinary linear dimension reduction;

[0070] To quickly complete dimension reduction and enhance the reliability of missing data in real-time industrial environments, an acceleration transform index is introduced to dynamically adjust the dimension reduction strength in the kernel mapping process:

[0071] ;

[0072] where: is a model-adjustable hyperparameter used to balance the relationship between the missing rate and the dimension reduction complexity; is the effective data proportion of the micro-batch package ;

[0073] When the effective data proportion is low, there are more missing values, and the acceleration transform index will also decrease, indicating that a more conservative dimension preservation strategy is needed to avoid the loss of key information due to excessive dimension reduction; when the effective data proportion is high, the acceleration transform index correspondingly increases, and a more aggressive projection depth can be used in dimension mapping to obtain a more concise feature representation;

[0074] The final step 201 outputs the feature vector, which can be denoted as: ;​

[0075] wherein, denotes the weighted operation by elements or components; eigenvector will be directly input for real-time fault inference as a subsequent step 202;

[0076] In use, through nonlinear kernel mapping Convert multi-source time series data into a more discriminative feature space, which can fully exploit the wavelet energy index and the potential information of transient suspicious flags ; with the help of the acceleration transform index Dynamically control the depth of dimensionality reduction, which can effectively weaken the impact of missing values on dimensionality reduction, and maximize the compression of redundant dimensions when the data integrity is high, thus balancing real-time and accuracy; introduce the acceleration transform index as a dynamic adjustment coefficient, combined with the proportion of effective data Flexible control of kernel mapping to meet real-time needs in fluctuating industrial environments.

[0077] Step 202, real-time fault reasoning and incremental update:

[0078] Based on the eigenvector , a lightweight classification or regression model (such as an improved random forest or incremental deep network) is used to quickly predict whether there is a fault and its severity, and the model parameters of the fault detection model are , then the prediction result of the eigenvector in the current micro-batch data set can be represented as:

[0079] ;

[0080] wherein denotes the decision function, and the output can be a fault probability, a risk score, or a specific fault label;

[0081] If only the prediction result is used for binary or multi-classification output, it is often difficult to distinguish different levels of high-risk faults and minor abnormalities, therefore, a fault confidence is added to the model output layer, defined as follows:

[0082] ;

[0083] wherein, is the fault confidence in the fault detection result corresponding to the micro-batch packaging , which is used to quantify the fault risk at the current time (or the current micro-batch packaging);

[0084] : the value range of a smooth monotonic activation function (such as the Logistic function) for mapping the internal weighted sum to the confidence interval; is a compact representation of the feature vector after kernel mapping and acceleration transform of multi-source time series data, time-frequency energy index, missing value placeholder, etc.

[0085] is a square matrix used to describe the coupling relationship between different feature channels and fault modes, usually ; combined with index mapping a nonlinear transformation is applied to the feature vector , so that the fault symptoms are amplified or suppressed in high-dimensional space, thereby improving the accuracy and discrimination of confidence assessment;

[0086] is a control matrix scaling factor when the index is mapped, which can be positive, negative or zero; is a pointer matrix exponent operation, that is:

[0087] ;

[0088] is the weight coefficient when the linear superposition of different index channels (i.e. different transform branches) is taken, which can be any real number; is the number of main activation units selected in the output layer of the model;

[0089] When the fault confidence exceeds the preset confidence threshold, it means that the system state represented by the current micro-batch data set is likely to be at high risk, and emergency control strategies need to be quickly executed; if the fault confidence does not exceed the preset confidence threshold, it can be selected to only give a warning or record;

[0090] To adapt to equipment aging or working condition changes, incremental or online learning needs to be carried out in the industrial field, and the core logic is:

[0091] If a real fault label or further confirmed fault sample is detected, its label information can be returned to this step 202 to correct and incrementally update the model parameters The simplified form of small batch or online correction incremental update can be written as:

[0092] ;

[0093] where, is the learning rate, represents the composite loss function, which is composed of the fault classification focal loss and the confidence regression smoothing loss, is the real label obtained after a certain time delay or in the next stage, which is used to fine-tune or correct the current model;

[0094] If the model adopts a tree structure, the sub-tree can be expanded or the leaf node statistical information can be updated as needed when high-risk fault samples appear. If the model adopts a neural network, the weight can be fine-tuned based on the small batch gradient descent or adaptive optimization algorithm, so as to continuously maintain the sensitivity of the model to system changes without significantly increasing the computational load;

[0095] In use, the lightweight design and the prediction process for micro-batch data sets ensure fast inference, meeting the stringent real-time requirements of industrial scenarios. By introducing a fault confidence to the output layer of the model, the fault risk is quantitatively described, which facilitates the flexible triggering of control strategies at different levels when obtaining the detection results. The fault confidence helps the industrial control system to perform precise scheduling according to different risk levels.

[0096] Step three, the on-site PLC or DCS executes high-priority actions such as emergency shutdown, redundancy switching, or issues a moderate warning based on the multi-level risk index , and records the actual execution and fault label back to the fault detection model to form a closed-loop protection and optimize the model parameters according to the real fault confirmation process;

[0097] The step three includes the following contents:

[0098] Step 301, control instruction generation based on multi-level risk index:

[0099] Receive the detection results from the second step , wherein represents the prediction results of the current micro-batch data set or time slice, i.e., the fault type or fault label, represents the fault confidence of the fault severity; the above results are transmitted to the on-site PLC or DCS through real-time communication protocols such as OPCUA, Modbus, or industrial Ethernet;

[0100] To finely control different fault levels and system states, a multi-level risk index is introduced, which comprehensively considers the fault confidence and the current key state of the system (such as temperature, pressure, flow, and process parameters that are not directly used for fault detection but are very important for control), An exemplary definition is as follows:​

[0101] ;

[0102] indicates the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging

[0103] corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging

[0104] corresponding to the time or time window corresponding to the micro-batch packaging

[0105] ;

[0106] corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging

[0107] corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging

[0108] corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging corresponding to the time or time window corresponding to the micro-batch packaging

[0109] The control instructions are executed by PLC or DCS to directly send action instructions to the equipment on the production line, such as valve opening adjustment, motor speed adjustment, cooling system start, etc., to realize the millisecond-level connection from fault detection to actual action.

[0110] When in use, the fault model output can be quickly converted into specific action instructions by means of real-time communication protocol and the hardware characteristics of the field PLC / DCS, thereby shortening the potential time window of fault diffusion and equipment damage, and placing the fault confidence level In the embedded matrix exponential operation and trace operation, a more functional multi-level risk index is formed According to the system state, the high-risk, medium-risk and low-risk can be distinguished.

[0111] Step 302, feedback of execution results and incremental strategy iteration:

[0112] After executing the control instructions, the PLC / DCS will return the actual action of the system (such as adjusting the valve opening, switching the standby equipment, etc.) and the subsequent collected running state data to this step 302;

[0113] Through the timestamped record, the execution action is paired with the fault detection result , and the newly generated system state is saved, wherein reflects the key process parameters in several sampling periods after executing the control;

[0114] If it is found through a period of observation or artificial maintenance that the previous fault mark has deviation or is insufficient or excessive in judging the fault level, the real label needs to be fed back to the second step (especially step 202) to update the parameters, so as to correct the judgment ability of the lightweight fault detection model.

[0115] If the response effect is not good for multiple times for the same fault type (such as multiple false shutdowns or delayed shutdowns), the relationship between the multi-level risk index and the final process result can be recorded locally or in the cloud to further fine-tune the weight matrix , the benchmark matrix B or the corresponding risk threshold described in step 301 to reduce the failure risk in similar scenarios later; this optimization process can be completed on a local edge device (industrial PC), or it can be transmitted to the cloud or factory background system for large-scale data analysis (corresponding to the fourth step), and then the adjusted configuration is issued to the field to realize the iterative optimization of cloud and edge integration; in order to ensure the unity of the marking information with the previous steps, any new label and system state Add same batch index micro batch Or timestamp tag, avoid data isolation.

[0116] By comparing multi-level risk index The calculated value and the actual execution result , determine whether the control action is timely and effective, and improve the subsequent response to realize the rapid evaluation of real-time control effect; If the actual fault label is confirmed or the response failure condition is found, the fault detection algorithm in the second step can be updated, and the threshold or weight matrix in step 301 in the third step can be corrected, forming a closed loop of detection-decision-relearning; Industrial sites often have variable working conditions. Through the incremental feedback of this step 302, the system can adapt to the access of new equipment, the replacement of production formula or the aging of equipment, and adapt to different industrial processes, so as to maintain efficient and accurate fault detection and control capability.

[0117] Step four, collect micro batch packaging, fault label , fault confidence , multi-level risk index information, cloud or background system collects multi-source data and mines complex fault mode by means of aggregation operation, and then downloads the optimized fault detection model and maintenance strategy to the edge side, and executes model evolution under fault label Comparison;

[0118] The step four includes the following contents:

[0119] Step 401, multi-source data collection and visual analysis:

[0120] The fault detection records (such as , fault determination , fault confidence , multi-level risk index , and control system actual execution information , etc.) accumulated on the field edge device or industrial PC are transmitted to the cloud or factory background system regularly or in real time, so as to complete the big data collection;

[0121] Store the uploaded data in a distributed database, and set retrieval index according to micro batch packaging And process equipment ID (such as sensor ID or PLC number); In addition to fault diagnosis information and control response data, device maintenance history, operator annotation, device working condition change record and other external data sources can also be integrated;

[0122] Construct global aggregation operator for comprehensive measurement of massive historical data and key parameters on site , which can be written as:

[0123] ;

[0124] wherein: is a set of indices of historical micro-batch packages or time slices; is a set of indices of current micro-batch packages is a multi-dimensional matrix representation after merging in the cloud (e.g. combined with and external device status, etc.);

[0125] is a weight factor, related to fault severity or data freshness; is a global coupling matrix, used to measure the collaborative pattern of multi-device, cross-period information when analyzing across batches; is a control matrix for large-scale historical data weighting, which can be optimized offline in the cloud or specified manually; is a matrix trace operation, used to compress the complex matrix result into a scalar;

[0126] Through such high-order aggregation, the commonality of fault patterns can be amplified and noise can be suppressed in large-scale samples, and a global data basis is provided for the predictive maintenance strategy of step 402.

[0127] In the cloud or factory background system, a visualization panel is constructed to display the real-time and historical trends of each production line or device, including fault occurrence frequency, downtime length, risk level distribution, etc.; some significant indicators (such as , ) can also be analyzed by heat map or time series superposition; an interactive exploration function is provided for operators or managers, which can be used for deep mining and comparison of specific periods or fault types, laying a cognitive foundation for subsequent maintenance and scheduling decisions;

[0128] When used, the detection results and control logs scattered in the edge side and local control system are upgraded to large-scale queryable data assets to avoid information silos and support long-term retention. Through the fusion of external data and fault detection results, a richer context can be formed in the cloud, which helps to accurately locate the root cause of the fault and identify potential risks;

[0129] Step 402, life cycle predictive maintenance and model iteration:

[0130] After completing the big data collection in step 401, a comprehensive method of machine learning, signal processing or domain knowledge (such as device health modeling) is used to deeply mine historical fault cases, pre-fault feature and control response effect; with the target of device remaining useful life or key component fatigue degree, a long-time series fault detection model can be trained or updated in the cloud, such as based on time series recurrent network, graph model or other advanced algorithms suitable for engineering field;

[0131] When a new failure mode is found (such as a combination of sensor waveforms that the edge has never seen before) or an evolutionary trend of an existing mode (such as a gradual increase in the frequency of failure at a certain temperature range), it can be recorded and incorporated into the corresponding maintenance strategy library;

[0132] If the cloud analysis result shows that there is a significant optimization space for the fault detection model (including feature extraction and inference algorithm) used in the second step, an improved version can be packaged and distributed to the on-site edge device; combined with the remaining service life prediction result and the fault mode analysis, a hierarchical maintenance plan is developed, such as replacing the vulnerable parts in advance, arranging the maintenance period, optimizing the spare parts inventory, etc.;

[0133] Unlike the real-time closed-loop control in the third step, which focuses on emergency protection, the predictive maintenance here is biased towards medium and long-term planning, which can avoid blind shutdown or excessive maintenance, and improve the utilization rate of the device,

[0134] When used, with the help of full life cycle data deposition, not only can the risk be better predicted before failure, but also the device life can be extended and the downtime loss can be reduced. Cloud big data analysis combined with edge incremental update makes the fault detection model continuously adapt to changes in the industrial environment, reduces false positives, false negatives, and late reports, reduces unnecessary expenses caused by regular large-scale maintenance through predictive maintenance, and avoids repeated emergency shutdowns caused by relying solely on real-time control; The deep fault mode analysis of the cloud is directly linked to the update of the edge fault detection model, which realizes the seamless connection of macro big data and local micro-batch detection, and breaks through the limitations of traditional local or cloud processing.

[0135] Please refer to Figure 2 The present application provides a kind of fault detection system of electronic components, comprising,

[0136] Data acquisition unit, when high-frequency sampling triggers multi-source data flow to rush in, wavelet filtering and missing value placeholder are used to original sensor sequence to remove false data, mark transient anomaly, and interpolation operation is carried out on delayed late data, to generate micro-batch package to carry sensor ID, time stamp and transient suspicious flag;

[0137] Feature extraction unit, edge side uses kernel mapping and acceleration transformation to extract feature vector, and calls incremental machine learning algorithm to output fault label and fault confidence in real time, and corrects the model parameters of fault detection model to adapt to environmental disturbance;

[0138] Early warning unit, on-site PLC or DCS executes high-priority action or issues moderate early warning according to multi-level risk index, and records actual execution and fault label backflow to fault detection model, and optimizes model parameters according to real fault confirmation process;

[0139] The feedback iteration unit collects the micro-batch packaging, fault label, fault confidence, and multi-level risk index information, and the cloud or background system collects multi-source data and mines fault patterns by means of aggregation operation, and then the optimized fault detection model and maintenance strategy are sent to the edge side, and model iteration is performed under the fault label comparison.

[0140] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solutions. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0141] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0142] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be realized by other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0143] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. can be located in one place or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0144] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A fault detection method for electronic components, characterized in that: include, When high-frequency sampling triggers a surge of multi-source data streams, wavelet filtering and missing value placeholders are applied to the original sensor sequence to remove pseudo data and mark transient anomalies. Interpolation is then performed on the aligned and delayed data to generate micro-batch packages to carry sensor IDs, timestamps, and transient suspicious flags. The edge side uses kernel mapping and accelerated transformation to extract feature vectors, and calls incremental machine learning algorithms to output fault labels and fault confidence in real time, and corrects the model parameters of the fault detection model to adapt to environmental disturbances. The on-site PLC or DCS executes high-priority actions or issues medium-level warnings based on multi-level risk indices, and feeds back the actual execution records and fault labels to the fault detection model, and optimizes the model parameters based on the actual fault confirmation process; After collecting information on micro-batch packaging, fault marking, fault confidence, and multi-level risk index, the cloud or back-end system collects multi-source data and uses aggregation operations to mine fault patterns. The optimized fault detection model and maintenance strategy are then distributed to the edge side, and the model is continuously iterated based on fault label comparison. After acquiring the original sensor sequence, it is aligned by timestamp. The feature coefficients at multiple scales are extracted by discrete wavelet transform. When the energy proportion of the target scale segment is consistently higher than expected, it is marked as high-frequency interference and suppression is performed. Transient anomalies are identified by wavelet energy index, which is used to detect anomalies in high-energy components, and transient anomaly flags are attached. The obtained reconstructed signal sequence is divided into multiple micro-batch datasets. After recording metadata at the micro-batch level, wavelet energy index and corresponding flags are retained for the corresponding data points within the micro-batch dataset, and then the micro-batch encapsulation is generated. After simple interpolation or skipping of missing values, all sensor channels in the micro-batch dataset are obtained and concatenated with the wavelet energy index to form an initial feature vector sequence. A customizable kernel function is introduced to map the initial feature vector sequence to obtain the dimensionality-reduced feature representation. An accelerated transformation index is introduced to dynamically adjust the dimensionality reduction intensity in the kernel mapping process and output the feature vector.

2. The fault detection method for electronic components according to claim 1, characterized in that: Based on the feature vector, a lightweight fault detection model is used to quickly predict whether a fault exists and its severity, and the model parameters are incrementally updated in small batches or online based on real feedback. Add a fault confidence level to the model output layer. When the fault confidence level exceeds the preset confidence threshold, an emergency control strategy is executed quickly. If the fault confidence level does not exceed the preset confidence threshold, only a warning or recording is selected.

3. The method for fault detection of electronic components according to claim 2, characterized in that: The acquired test results are transmitted to the on-site PLC or DCS; It also introduces a multi-level risk index based on fault confidence and the current critical state of the system. Based on the landing points of the multi-level risk index in different threshold ranges, the industrial control system executes corresponding levels of protection or alarms and issues action commands directly to the production line equipment through PLC or DCS.

4. The fault detection method for electronic components according to claim 3, characterized in that: After the PLC or DCS transmits back the actual actions and operating status data of the system, it matches the executed actions with the fault detection results and saves the newly generated system status. If the previous fault labeling is biased or the fault level is judged insufficiently or excessively, the real label will be used to correct the lightweight fault detection model; if multiple responses for the same fault type are not effective, the relevant parameters of the multi-level risk index will be fine-tuned by the relationship between the multi-level risk index and the final process result.

5. The fault detection method for electronic components according to claim 4, characterized in that: Fault detection records accumulated on field edge devices or industrial PCs will be transmitted to the cloud or factory back-end system on a regular or real-time basis; a global aggregation operator will be built to summarize massive historical data and key field parameters, and a visualization panel will be built in the cloud or factory back-end system after data aggregation is completed.

6. The fault detection method for electronic components according to claim 5, characterized in that: Update the fault detection model with the remaining service life of the equipment or the fatigue of key components as the target; if there are fault modes or evolution trends of existing modes, record them and incorporate them into the corresponding maintenance strategy library. If there is room for optimization in the fault detection model used, the improved version can be packaged and distributed to the field edge devices; Based on the predicted remaining service life of the equipment and failure mode analysis, a tiered maintenance plan is developed.

7. A fault detection system for electronic components, employing the method described in any one of claims 1 to 6, characterized in that: include, The data acquisition unit, when high-frequency sampling triggers the influx of multi-source data streams, applies wavelet filtering and missing value placeholders to the original sensor sequence to remove pseudo data and mark transient anomalies, and performs interpolation operations on the aligned and delayed data to generate micro-batch packages to carry sensor ID, timestamp and transient suspicious flags; The feature extraction unit extracts feature vectors at the edge using kernel mapping and accelerated transformation, and calls an incremental machine learning algorithm to output fault labels and fault confidence in real time, and corrects the model parameters of the fault detection model to adapt to environmental disturbances. The early warning unit, on-site PLC or DCS, executes high-priority actions or issues medium-level early warnings based on multi-level risk indices, feeds back actual execution records and fault labels to the fault detection model, and optimizes model parameters based on the actual fault confirmation process; The feedback iteration unit collects micro-batch packaging, fault marking, fault confidence, and multi-level risk index information. The cloud or back-end system aggregates multi-source data and uses aggregation operations to mine fault patterns. The optimized fault detection model and maintenance strategy are then distributed to the edge side, and the model is continuously iterated based on fault label comparison.

Images