A lcp fault early warning method based on time series analysis

Abstract

The application discloses a kind of LCP detection fault early warning method based on time series analysis, it is characterized in that, using LCP material characteristic time sequence and equipment operation time sequence double closed loop driving mode, whole process without artificial selection item, formula parameter is fixed, specifically includes the following steps: S1.multiple-source time series data synchronous acquisition;S2.time series alignment calibration and rigid data cleaning;S3.LCP intrinsic characteristic binding time series enhancement;S4.self-adapting anti-interference processing;S5.time series weak fault feature extraction;S6.four-dimensional time series fusion and crystallization evolution dynamic threshold detection;S7.classified fault early warning and closed loop instruction issue;S8.on-line incremental self-optimization and closed loop iteration;Whole process without artificial selection item, formula parameter is fixed, both ensure the automation of process, standardization, also realize the gradual promotion of data quality-feature clarity-anti-interference ability, provide reliable support for subsequent S5-S8 fault feature extraction, fusion detection, early warning issue and self-optimization.

Description

Technical Field

[0001] This invention relates to the field of fault detection and early warning technology, and in particular to a fault detection and early warning method for LCP based on time series analysis. Background Technology

[0002] The processing of liquid crystal polymers (LCPs) is highly dependent on intrinsic properties such as crystallinity, coefficient of thermal expansion, and mechanical strength. The molding process is easily affected by working conditions such as temperature, vibration, and noise, leading to fluctuations in material properties and equipment failures, which is a recognized technical pain point in this field.

[0003] Currently, relevant technologies have been disclosed in the field of industrial fault early warning. For example, CN120216968B proposes an early warning method based on front-end sensing data. This method enhances acoustic signals, extracts weak features through improved Wavelley transform, integrates change points to divide fault stages, constructs a time-series evolution model based on recurrent neural networks, and adopts chain-based progressive prediction and multi-dimensional early warning decision-making. This method improves detection sensitivity and early warning lead time in the fault monitoring of general mechanical equipment.

[0004] Existing technologies, when combined with LCP processing scenarios, still suffer from insurmountable technical shortcomings. They fail to consider key physical properties such as LCP crystallization evolution, thermal expansion, and mechanical strength, representing a general equipment monitoring solution. Furthermore, LCP failures are essentially caused by the coupling of material properties and equipment status; vibration and acoustics alone cannot identify early-stage faults such as uneven crystallization and performance drift. The lack of a unified reference clock leads to time shifts in multi-source data, failing to meet the high-precision molding requirements of LCP. Moreover, low-computing-power deployment on embedded LCP molding machines is challenging. This invention employs an ultra-lightweight LSTM + time-series distillation transfer learning, requiring only 50 small samples for training and supporting non-stop incremental updates, making it more suitable for industrial deployment. Industrial environments rely on manual parameter selection, resulting in poor reproducibility. This invention features fixed formulas and parameters throughout the process, with no manual selection options, allowing for direct scalable deployment.

[0005] Therefore, there is a need in this field for a fault early warning method that is specifically designed for LCP, driven by a time-series dual closed loop, standardized throughout the entire process, dynamically adaptive, lightweight and deployable, in order to overcome the inherent defects of existing technologies. Summary of the Invention

[0006] This invention provides a method, apparatus, electronic device, and storage medium for fault early warning of LCP detection based on time-series analysis.

[0007] According to a first aspect of the present invention, a fault early warning method for LCP detection based on time-series analysis is provided. The method includes: S1. Synchronous Acquisition of Multi-Source Time-Series Data: After the equipment is powered on and the operating conditions are stable, the crystallinity of LCP is synchronously acquired using the timestamp collected by the crystallinity sensor as the sole reference clock. Coefficient of thermal expansion Mechanical strength Equipment vibration Equipment acoustics Five-channel timing data, ensuring that the five channels have the same frequency and timestamp, outputting raw timing data with timestamps; S2. Timing Alignment Calibration and Rigid Data Cleaning: Based on the crystallinity in step S1 Using the timestamp as the sole reference, linear interpolation resampling is performed on the other four data streams to unify the data granularity. Dirty data is then removed and missing values ​​are filled in according to fixed rules. After completing the data integrity verification, a regular n×5 time series matrix X(t) is output. S3. LCP Intrinsic Characteristics-Based Time-Series Enhancement: Using LCP crystallinity and thermal expansion coefficient as mandatory constraints, and substituting them into a fixed enhancement formula, the time-series matrix X(t) output in step S2 is enhanced point-by-point. Suspected fault segments are further enhanced. After signal-to-noise ratio verification, the enhancement matrix X_aug(t) is output. The time-series enhancement formula is: Where ω=1.2 and μ=0.5 are fixed parameters. Let λ(t) be the standard crystallinity of LCP, and λ(t) be the real-time coefficient of thermal expansion after cleaning in step S2. , LCP is the standard coefficient of thermal expansion. S4. Adaptive Anti-interference Processing: A unique anti-interference scheme combining db4 wavelet 5-level decomposition and temporal attention weight denoising is adopted to denoise the enhancement matrix X_aug(t) output in step S3, distinguishing between environmental interference and fault segments, and outputting clean data X_pure(t) and interference / fault markers; the temporal attention weight formula is: Where α=1.0 is a fixed parameter, and Δf is the difference between the real-time vibration frequency and 50Hz. The crystallinity data is the result of enhancement in step S3. The crystallinity is the standard for LCP. S5. Extraction of weak time-series fault features: The improved Wigner-Ville transform is used to extract the time-frequency features of the pure fault segment data output in step S4. Redundant features are removed by a fixed screening formula, strong fault features are retained, and the selected feature set F_select(t) and LCP fault time-series feature dictionary are output. S6. Four-dimensional temporal fusion and crystallization evolution dynamic threshold detection: LCP characteristic-related features and equipment operation-related features are fused with fixed weights. The fault determination threshold is updated in real time based on the LCP crystallinity change rate. Fault determination is completed by comparing the fused features with the dynamic threshold. The four-dimensional attention fusion formula is: Where ξ=0.75 is a fixed weight; the dynamic threshold formula is: ,in =0.82 and κ=0.06 are fixed parameters; S7. Graded Fault Early Warning and Closed-Loop Command Issuance: Based on the fault determination result of step S6, execute a three-level early warning according to a fixed threshold, and simultaneously output audible and visual prompts, fault records and equipment control commands, push them to the equipment PLC to execute the corresponding operations, and output early warning records and fault sample data; S8. Online Incremental Self-Optimization and Closed-Loop Iteration: Based on the ultra-lightweight LSTM model, online incremental updates are performed using the fault sample data output in step S7. Model parameters are optimized without system shutdown. The updated model and parameters are then looped back to step S3 to achieve full-process self-optimization iteration. The incremental update formula is: , where η=0.005 is a fixed learning rate.

[0008] Furthermore, in step S1, the sampling frequency is uniformly set to 5ms / group, and a single sensor is selected: crystallinity. The DSC-600 differential scanning calorimeter sensor is used, with a coefficient of thermal expansion of [missing information]. Employing the LDV-1000 laser micro-deformation sensor, mechanical strength Employing the FT-300 tension / compression sensor, vibration Employing PV-200 piezoelectric sensor, acoustic The AS-500 high-precision acoustic sensor was used; during the data acquisition process... by Forced synchronization is triggered based on timestamps, with an average time deviation Δt_avg ≤ 0.5ms; each batch of continuous acquisition lasts for ≥ 10 minutes, and the original data is retained for ≥ 30 days. Acquisition is performed using a DAQ-800 synchronous acquisition card and stored on a 1TB SSD solid-state drive.

[0009] Furthermore, in step S2, linear interpolation resampling uses the interp1 algorithm, with the execution standard being an average time deviation Δt_avg ≤ 0.5ms. If this standard is not met, interpolation is re-executed. The dirty data removal rules are as follows: blank / NaN values ​​are deleted at integer time points, consecutive ≥ 5 constant values ​​are deleted in a whole segment, and data exceeding the range or adjacent differences > 3 times the standard deviation are deleted. The missing value completion rules are as follows: short missing values ​​(≤ 3 points) are completed using the mean of a 5-point sliding window, and long missing values ​​(> 3 points) are marked as 0 and do not participate in subsequent calculations. Data integrity rate ≥ 98% is required to proceed to step S3; otherwise, the process returns to step S1 for re-collection.

[0010] Furthermore, in step S3, when At this time, ω is temporarily adjusted to 1.4 to further amplify the fault characteristics; the signal-to-noise ratio (SNR) is improved by ≥4dB after enhancement. If the standard is not met, ω is finely adjusted with an upper limit of ≤1.5.

[0011] Furthermore, in step S4, a similarity formula is used to distinguish between interference and faulty sections. If S(t) ≥ 0.9, it is determined to be environmental interference, and if S(t) < 0.9, it is determined to be a faulty section. After anti-interference processing, the noise suppression ratio (NSR) ≥ 25dB is considered qualified.

[0012] Furthermore, in step S5, the improved Wigner-Ville transform uses a Hanning window with a window length of 20ms and an overlap rate of 50%, eliminating interference from cross terms; 11 time-frequency features are extracted, with a feature selection threshold of 0.65. After selection, feature redundancy is reduced by ≥70%, and finally, ≤4 strong fault features are retained.

[0013] Furthermore, in step S6, ( , , These are the characteristic values ​​corresponding to crystallinity, coefficient of thermal expansion, and mechanical strength, respectively. ( , The corresponding eigenvalues ​​are vibration and acoustic, respectively; the correlation coefficient between the fusion feature and the fault label is ≥0.85; the fault determination rule is: if F_fusion(t) > T(t), it is determined as a fault, otherwise it is normal.

[0014] Furthermore, in step S7, the three-level early warning rules are as follows: Level 1 early warning (F_fusion exceeds T(t) by 1.0~1.5 times): yellow light reminder; Level 2 early warning (F_fusion exceeds T(t) by 1.5~2.0 times): orange light shutdown pre-inspection; Level 3 early warning (F_fusion exceeds T(t) by more than 2.0 times): red light emergency shutdown; when an early warning occurs, a screen pop-up window + audio-visual prompt are triggered simultaneously, the fault period, characteristic value, and threshold deviation are automatically saved, and pushed to the equipment PLC to execute deceleration, shutdown, or cooling operations.

[0015] Furthermore, in step S8, the ultra-lightweight LSTM model adopts a single-layer LSTM + temporal distillation transfer learning structure, which can complete the model training with only 50 small samples. It can be deployed on embedded chips with a computing power consumption of ≤30%. The loss is the error between the model's predicted value and the actual fault label. The model is updated online incrementally without interrupting the detection process. The updated parameters are automatically synchronized to steps S3~S6, and a model optimization report is automatically generated every 24 hours.

[0016] A fault early warning system for LCP detection based on time-series analysis is characterized in that it is used to execute the fault early warning method for LCP detection based on time-series analysis as described in any one of claims 1 to 9. The system includes a data acquisition module, a time-series calibration module, a feature enhancement module, an anti-interference module, a feature extraction module, a fusion detection module, an early warning control module, and an online optimization module. The system has no manual selection options throughout the process, the formulas and parameters are fixed, it can be deployed on embedded chips with low computing power, the detection accuracy is stable at ≥98%, and it can be directly used for on-site deployment in LCP molding machines and high-precision testing machines.

[0017] 1. Crystallinity timestamp is the sole criterion, mandatory. Synchronous triggering of data acquisition, with the average time deviation controlled to ≤0.5ms, solves the problems of feature misalignment and fault judgment deviation caused by asynchronous multi-source data, ensuring the accuracy of subsequent data processing. Second, fixed sensor selection and acquisition frequency avoid data source differences caused by different sensors and acquisition frequencies, ensuring data consistency and comparability, and providing a prerequisite for the application of fixed parameter formulas. Third, continuous acquisition time ≥10 minutes and retention of raw data ≥30 days ensure data continuity, facilitate subsequent time series analysis and fault tracing, and provide sufficient historical sample data for online incremental self-optimization in step S8. Fourth, five data channels simultaneously cover LCP material properties and equipment operating status, realizing the linkage monitoring of material properties and equipment operation, avoiding fault omissions caused by single-dimensional data monitoring (such as monitoring only equipment operation without detecting abnormal LCP characteristics, or monitoring only LCP characteristics without detecting equipment operation faults).

[0018] 2. By employing linear interpolation resampling and using crystallinity timestamps as a benchmark, the time deviation of multi-source data was further corrected, ensuring complete uniformity of data granularity. This resolved the minor time deviation issues that might remain in step S1, providing a time-series matching basis for subsequent LCP intrinsic characteristic binding (step S3). Secondly, rigid and fixed dirty data removal and missing value completion rules were adopted to avoid inconsistencies in data cleaning standards caused by manual intervention, ensuring the objectivity and uniformity of data cleaning. Simultaneously, a clear verification standard of ≥98% data integrity was established, preventing low-quality data from entering subsequent processes and reducing false alarms and missed alarms due to data issues. Thirdly, different completion / marking rules were used for short and long missing data, ensuring the usability of short missing data while avoiding interference from long missing data in subsequent analysis, balancing data integrity and accuracy. Fourthly, a closed-loop mechanism was set up to return to S1 for re-collection if the data fails to meet the standards, forming a preliminary closed loop of data collection-cleaning-verification. This ensured that all data entering subsequent steps met quality requirements, improving the reliability of the entire early warning process.

[0019] 3. By using LCP crystallinity and thermal expansion coefficient as mandatory constraints, the binding of material properties and time-series data is achieved, ensuring that time-series enhancement does not deviate from the inherent characteristics of LCP, avoiding the amplification of non-fault signals caused by indiscriminate enhancement, and improving the targeting of fault feature enhancement; secondly, by adopting a fixed-parameter enhancement formula, combined with suspected fault segments ( The ω parameter adjustment mechanism ensures the standardization and lack of manual intervention in the enhancement process, while also specifically amplifying fault features, addressing the pain point of weak and difficult-to-identify features in the early stages of a fault. Thirdly, signal-to-noise ratio (SNR) verification (SNR improvement ≥ 4dB) ensures the enhancement effect; if the standard is not met, parameters are fine-tuned, guaranteeing the quality of the enhanced data and providing clear signal support for anti-interference processing in step S4 and weak fault feature extraction in step S5, reducing the difficulty of subsequent fault feature extraction. Fourthly, the enhancement formula simultaneously correlates with the LCP standard value (…). , The standard-real-time comparison is enhanced by comparing the standard and real-time values, making the enhanced fault characteristics more consistent with the normal characteristic range of LCP and improving the accuracy of subsequent fault judgment; fifth, the output data of this step can be looped back to the model optimization in step S8, providing enhanced high-quality samples for model parameter updates and helping the whole process to self-optimize.

[0020] 4. A unique anti-interference scheme is adopted, using db4 wavelet 5-level decomposition + temporal attention weight denoising. Combined with a fixed-parameter attention weight formula, this avoids the complexity and parameter confusion caused by switching between multiple anti-interference schemes, and also correlates LCP crystallinity data through attention weight. This paper presents three main features: First, it achieves adaptive noise reduction by prioritizing the preservation of fault features and suppressing interference signals, thus solving the problem of difficulty in distinguishing fault features from interference signals in traditional noise reduction methods. Second, it distinguishes between environmental interference and fault segments through a similarity formula, avoiding misjudging environmental interference as faults (false alarms) or fault signals as interference (missed alarms), thereby improving the accuracy of fault identification and providing a labeling basis for subsequent fault feature extraction and judgment. Third, it explicitly requires an NSR ≥ 25dB after anti-interference processing, ensuring the noise reduction effect, filtering out most environmental interference, and enabling subsequent steps to extract weak fault features, making it particularly suitable for high-precision LCP detection fields. The first aspect is the use of a time-series attention weight formula to link the standard value of LCP crystallinity with the real-time value and the difference in vibration frequency. This enables the linkage between anti-interference processing and LCP characteristics and equipment operating status, making the noise reduction process more consistent with the actual scenario of LCP detection and improving the pertinence and reliability of anti-interference. The second aspect is the output of pure data X_pure(t), which provides a high-quality foundation for the extraction of weak fault features in step S5, while the interference / fault labeling provides an auxiliary basis for fault determination in step S6. At the same time, the labeled data can be used as a sample for model optimization in step S8, further improving the model's anti-interference capability.

[0021] With no manual selection options and fixed formula parameters throughout the process, the automation and standardization of the workflow are ensured, while also achieving gradual improvements in data quality, feature clarity, and anti-interference capabilities. This provides reliable support for subsequent fault feature extraction, fusion detection, early warning issuance, and self-optimization in S5-S8. The four-step linkage solves the core pain points of traditional LCP detection, such as data asynchrony, poor quality, weak fault features, and insufficient anti-interference capabilities. This lays the foundation for the high accuracy (stable ≥98%) and low false alarm / missed alarm rate of the entire early warning method, ultimately achieving early warning and accurate judgment of LCP equipment faults, avoiding equipment damage and product quality defects. At the same time, it is compatible with low-computing-power deployment of embedded chips and can be directly applied to industrial sites. Attached Figure Description

[0022] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 A flowchart illustrating the fault warning process for LCP detection according to the present invention is shown; Figure 2 A flowchart of the LCP detection fault early warning method based on time-series analysis according to the present invention is shown. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Furthermore, the terms "and / or" in this article merely describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects are in an "or" relationship.

[0025] Figures 1-2 A flowchart of a fault early warning method for LCP detection based on time-series analysis according to an embodiment of the present invention is shown. The method includes: This experiment selected three LCP-8000 model LCP molding machines of the same type as the test objects, and conducted fault warning detection tests using the method of this invention, the existing single-time detection method, and the existing multi-time loopless detection method, respectively. The test environment was uniformly controlled as follows: temperature 25±1℃, humidity 50±5% RH, power supply voltage 220V±5V, with no additional environmental interference sources; the test material used was standard grade LCP-LX100, with a standard crystallinity of =85%, standard thermal expansion coefficient Standard fault determination benchmark threshold =0.82.

[0026] During the experiment, three types of minor faults commonly found in LCP molding machines were simulated: abnormal crystallization, excessive equipment vibration, and abnormal acoustic signals. Fault simulation employed a signal injection method to ensure the fault signal amplitude was 1.2 times the normal signal amplitude, guaranteeing consistent fault intensity across all test groups. Before the experiment, all equipment was debugged to ensure stable operation. Sensors were uniformly installed: a differential scanning calorimeter (DSC-600) was installed at the LCP melt outlet, a laser micro-deformation sensor (LDV-1000) was installed on the molding die surface, a tensile / compression sensor (FT-300) was installed on the material conveying mechanism, a piezoelectric sensor (PV-200) was installed in the middle of the equipment body, and a high-precision acoustic sensor (AS-500) was installed on the top of the equipment. All sensors were connected to an embedded detection terminal via a synchronous acquisition card, and the terminal was equipped with a 1TB solid-state drive for data storage.

[0027] The LCP detection fault early warning method based on time-series analysis described in this embodiment specifically includes the following steps: S1. Synchronous acquisition of multi-source time-series data; After the three experimental devices have been running for 30 minutes and the operating conditions have stabilized, using the crystallinity timestamp collected by the differential scanning calorimeter as the sole reference clock, five channels of time-series data are synchronously collected at a collection frequency of 5ms / group: LCP crystallinity. Coefficient of thermal expansion Mechanical strength Equipment vibration Equipment acoustics ;in by Synchronization is triggered by a timestamp-based system, with the average time deviation Δt_avg ≤ 0.5ms. Each batch of data is collected continuously for 15 minutes, and the raw data is retained for 30 days. Data transmission status is monitored in real time during the collection process to ensure no data loss.

[0028] S2, Timing Alignment Calibration and Rigid Data Cleaning; Based on the crystallinity in step S1 Using the timestamp as the sole reference, a one-dimensional linear interpolation algorithm is used to perform linear interpolation resampling on the other four data streams, unifying the data granularity to 5ms / group, ensuring that the average time deviation Δt_avg ≤ 0.5ms, and re-interpolating if the standard is not met.

[0029] Dirty data is removed according to fixed rules: blank values / non-numerical values ​​corresponding to whole time points are deleted; whole data segments corresponding to no less than 5 consecutive constant values ​​are deleted; data that exceeds the range or whose difference between adjacent values ​​is greater than 3 times the standard deviation are deleted.

[0030] Missing value imputation rules: Short missing values ​​(no more than 3) are imputed using the mean of a 5-point sliding window; long missing values ​​(more than 3) are marked as 0. Complete data integrity verification to ensure a data integrity rate of no less than 98%. If the rate is not met, return to step S1 for re-collection. The final output is a well-organized 18000×5 time series matrix X(t) (15 minutes × 120 sets / minute = 18000 sets of data).

[0031] To eliminate the impact of differences in physical dimensions and numerical ranges on subsequent fusion calculations, the five time-series data were normalized separately, mapping all data to the [0,1] numerical range. The normalization formula is as follows: x_{i,norm}(t)=\frac{x_i(t)-x_{i,min}}{x_{i,max}-x_{i,min}} In the formula: x_{i,norm}(t) is the normalized value of the i-th data; x_i(t) is the original cleaned data of the i-th data; x_{i,min} and x_{i,max} are the minimum and maximum values ​​of the i-th data range, respectively. The normalized data enters the subsequent enhancement and noise reduction steps.

[0032] S3, LCP intrinsic characteristics are bundled with timing enhancements; Using LCP crystallinity and thermal expansion coefficient as mandatory constraints, the time series matrix X(t) output in step S2 is enhanced point-by-point by substituting them into the time series enhancement formula:

[0033] In the formula: the enhancement weight coefficient ω = 1.2, and the offset correction coefficient μ = 0.5; =85% is the standard crystallinity of LCP; λ(t) is the real-time thermal expansion coefficient after cleaning and normalization, x2(t); The standard thermal expansion coefficient is LCP.

[0034] When detected At this point, ω is temporarily adjusted to 1.4 to further amplify the fault characteristics. After enhancement, the signal-to-noise ratio (SNR) is checked to ensure that the SNR improvement is not less than 4dB. If the standard is not met, ω is finely adjusted with an upper limit not exceeding 1.5. The final output is the enhancement matrix X_aug(t).

[0035] S4. Adaptive anti-interference processing; A unique anti-interference scheme combining db4 wavelet five-level decomposition and temporal attention weight denoising is adopted to denoise the enhancement matrix X_aug(t) output in step S3.

[0036] The formula for temporal attention weights is:

[0037] In the formula: adjustment coefficient α = 1.0; Δf is the difference between the real-time vibration frequency and the 50Hz power frequency.

[0038] The similarity formula S(t) is used to distinguish between interference signals and fault signal segments: S(t) ≥ 0.9 is judged as environmental interference, and S(t) < 0.9 is judged as a fault segment. After anti-interference processing, the noise suppression ratio is ensured to be no less than 25dB, and the clean data X_pure(t) and interference / fault labeling results are output.

[0039] S5. Extraction of features of weak temporal faults; Extract time-frequency features from the clean data of the fault segment output in step S4, and extract a total of 11 time-frequency features. Eliminate redundant features by using a fixed filtering rule (filtering threshold is 0.65) to ensure that the feature redundancy is reduced by no less than 70%. Finally, retain 3 strong fault features and output the selected feature set F_select(t) and LCP fault time series feature library.

[0040] S6, Four-dimensional temporal fusion and dynamic threshold detection of crystallization evolution; The fusion formula, which integrates LCP material properties with equipment operation characteristics using fixed weights, is as follows:

[0041] In the formula: the fusion allocation coefficient ξ = 0.75.

[0042] Characteristic components of LCP material: F_LCP=0.55f_x1+0.225f_x2+0.225f_x3, where f_x1, f_x2, and f_x3 are the characteristic values ​​corresponding to crystallinity, coefficient of thermal expansion, and mechanical strength, respectively, and the sum of the weighting coefficients is 1; Device operating characteristic components: F_device=0.65f_x4+0.35f_x5, where f_x4 and f_x5 are the characteristic values ​​corresponding to vibration and acoustics, respectively, and the sum of the weighting coefficients is 1.

[0043] The fault determination threshold is updated in real time based on the LCP crystallinity change rate. The dynamic threshold formula is:

[0044] Where: benchmark threshold =0.82; threshold correction coefficient κ=0.06; v(t) is the real-time change rate of crystallinity; T_std is the threshold baseline correction amount.

[0045] Fault determination is achieved by comparing fused features with dynamic thresholds, ensuring that the correlation coefficient between the fused features and the fault label is not less than 0.85. The determination rule is: if F_fusion(t) > T(t), it is determined as a fault; otherwise, it is normal.

[0046] S7. Graded fault early warning and closed-loop command issuance; Based on the fault determination result of step S6, a three-level early warning is executed according to a fixed threshold: Level 1 Warning (F_fusion exceeds T(t) by 1.0~1.5 times): Yellow light alert; Level 2 warning (F_fusion exceeds T(t) by 1.5~2.0 times): Orange light indicates shutdown and pre-inspection; Level 3 warning (F_fusion exceeds T(t) by more than 2.0 times): Red light, emergency shutdown.

[0047] When an early warning is issued, a pop-up window on the screen and an audio-visual prompt are triggered simultaneously. The fault segment time, characteristic value, and threshold deviation are automatically saved and pushed to the device's programmable logic controller to execute the corresponding shutdown or pre-inspection operation.

[0048] S8, Online Incremental Self-Optimization and Closed-Loop Iteration; Based on the ultra-lightweight long short-term memory network model (single-layer long short-term memory network, temporal distillation transfer learning structure), online incremental updates are performed using the fault sample data output in step S7.

[0049] The incremental update formula is:

[0050] In the formula: learning rate η = 0.005; θ_old is the model parameters before update; θ_new is the model parameters after update; Loss is the model loss function.

[0051] The model can be trained with only 50 small samples, and the computing power consumption is no more than 30%. The model is updated online incrementally without interrupting the detection process. The updated parameters are automatically synchronized to steps S3 to S6, and a model optimization report is automatically generated every 24 hours.

[0052] Method of the present invention 99.2 18 6.8 32.5 78.3 25.6 99.1 98.7 Existing single time-series detection methods 78.5 45 1.2 15.3 32.1 22.3 97.8 72.4 Existing multi-time-series loop-free detection methods 86.7 32 2.5 20.1 45.8 42.7 98.2 83.6 Blank control group 0 No response 0 0 0 0 97.5 0 The experimental data show that: in terms of detection accuracy, the method of this invention reaches 99.2%, which is 20.7 percentage points higher than the single time-series method and 12.5 percentage points higher than the multi-time-series loop-free method; the early warning response speed is only 18ms, which is 60% and 43.75% higher than the two existing methods, respectively; in terms of data processing performance, the signal-to-noise ratio is improved by 6.8dB, the noise suppression ratio is 32.5dB, and the feature redundancy is reduced by 78.3%, all of which are significantly better than existing technologies; in terms of computing power and integrity, the computing power utilization is 25.6%, balancing low computing power and high data integrity (99.1%); the early warning classification accuracy reaches 98.7%, which is a significant improvement over existing methods.

[0053] This invention deeply binds LCP material properties with equipment operating status through a closed-loop process of multi-source time-series synchronous acquisition, LCP intrinsic enhancement, adaptive anti-interference, four-dimensional feature fusion, dynamic threshold determination, hierarchical early warning, and online incremental optimization. This enables high-precision, fast-response, and low-computing-power early warning of minor faults, effectively solving the defects of existing technologies and possessing outstanding substantive features and significant progress.

[0054] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A fault early warning method for LCP detection based on time-series analysis, characterized in that, The LCP material property timing sequence and equipment operation timing dual closed-loop drive mode is adopted, with no manual selection options and fixed formula parameters throughout the process. The specific steps are as follows: S1. Synchronous Acquisition of Multi-Source Time-Series Data: After the equipment is powered on and the operating conditions are stable, the crystallinity of LCP is synchronously acquired using the timestamp collected by the crystallinity sensor as the sole reference clock. Coefficient of thermal expansion Mechanical strength Equipment vibration Equipment acoustics Five-channel timing data, ensuring that the five channels have the same frequency and timestamp, outputting raw timing data with timestamps; S2. Timing Alignment Calibration and Rigid Data Cleaning: Based on the crystallinity in step S1 Using the timestamp as the sole reference, linear interpolation resampling is performed on the other four data streams to unify the data granularity. Dirty data is then removed and missing values ​​are filled in according to fixed rules. After completing the data integrity verification, a regular n×5 time series matrix X(t) is output. S3. LCP Intrinsic Characteristics-Based Time-Series Enhancement: Using LCP crystallinity and thermal expansion coefficient as mandatory constraints, and substituting them into a fixed enhancement formula, the time-series matrix X(t) output in step S2 is enhanced point-by-point. Suspected fault segments are further enhanced. After signal-to-noise ratio verification, the enhancement matrix X_aug(t) is output. The time-series enhancement formula is: Where ω=1.2 and μ=0.5 are fixed parameters. Let λ(t) be the standard crystallinity of LCP, and λ(t) be the real-time coefficient of thermal expansion after cleaning in step S2. , LCP is the standard coefficient of thermal expansion. S4. Adaptive Anti-interference Processing: A unique anti-interference scheme combining db4 wavelet 5-level decomposition and temporal attention weight denoising is adopted to denoise the enhancement matrix X_aug(t) output in step S3, distinguishing between environmental interference and fault segments, and outputting clean data X_pure(t) and interference / fault markers; the temporal attention weight formula is: Where α=1.0 is a fixed parameter, and Δf is the difference between the real-time vibration frequency and 50Hz. The crystallinity data is the result of enhancement in step S3. The crystallinity is the standard for LCP. S5. Extraction of weak time-series fault features: The improved Wigner-Ville transform is used to extract the time-frequency features of the pure fault segment data output in step S4. Redundant features are removed by a fixed screening formula, strong fault features are retained, and the selected feature set F_select(t) and LCP fault time-series feature dictionary are output. S6. Four-dimensional temporal fusion and crystallization evolution dynamic threshold detection: LCP characteristic-related features and equipment operation-related features are fused with fixed weights. The fault determination threshold is updated in real time based on the LCP crystallinity change rate. Fault determination is completed by comparing the fused features with the dynamic threshold. The four-dimensional attention fusion formula is: Where ξ=0.75 is a fixed weight; the dynamic threshold formula is: ,in =0.82 and κ=0.06 are fixed parameters; S7. Graded Fault Early Warning and Closed-Loop Command Issuance: Based on the fault determination result of step S6, execute a three-level early warning according to a fixed threshold, and simultaneously output audible and visual prompts, fault records and equipment control commands, push them to the equipment PLC to execute the corresponding operations, and output early warning records and fault sample data; S8. Online Incremental Self-Optimization and Closed-Loop Iteration: Based on the ultra-lightweight LSTM model, online incremental updates are performed using the fault sample data output in step S7. Model parameters are optimized without system shutdown. The updated model and parameters are then looped back to step S3 to achieve full-process self-optimization iteration. The incremental update formula is: , where η=0.005 is a fixed learning rate.

2. The method according to claim 1, characterized in that, In step S1, the sampling frequency is uniformly set to 5ms / group, and a single sensor is selected: crystallinity. The DSC-600 differential scanning calorimeter sensor is used, with a coefficient of thermal expansion of [missing information]. Employing the LDV-1000 laser micro-deformation sensor, mechanical strength Employing the FT-300 tension / compression sensor, vibration Employing PV-200 piezoelectric sensor, acoustic The AS-500 high-precision acoustic sensor was used; during the data acquisition process... by Forced synchronization is triggered based on timestamps, with an average time deviation Δt_avg ≤ 0.5ms; each batch of continuous acquisition lasts for ≥ 10 minutes, and the original data is retained for ≥ 30 days. Acquisition is performed using a DAQ-800 synchronous acquisition card and stored on a 1TB SSD solid-state drive.

3. The method according to claim 1, characterized in that, In step S2, linear interpolation resampling uses the interp1 algorithm, with the standard being an average time deviation Δt_avg ≤ 0.5ms. If this standard is not met, interpolation is re-implemented. The dirty data removal rules are: delete blank / NaN values ​​at integer time points, delete segments with ≥ 5 consecutive constant values, and delete data that exceeds the range or whose adjacent differences are > 3 times the standard deviation. The missing value completion rules are: short missing values ​​(≤ 3 points) are completed using the mean of a 5-point sliding window, and long missing values ​​(> 3 points) are marked as 0 and do not participate in subsequent calculations. Data integrity rate ≥ 98% is required to proceed to step S3; otherwise, return to step S1 for re-collection.

4. The method according to claim 1, characterized in that, In step S3, when At this time, ω is temporarily adjusted to 1.4 to further amplify the fault characteristics; the signal-to-noise ratio (SNR) is improved by ≥4dB after enhancement. If the standard is not met, ω is finely adjusted with an upper limit of ≤1.

5.

5. The method according to claim 1, characterized in that, In step S4, a similarity formula is used to distinguish between interference and faulty sections. If S(t) ≥ 0.9, it is determined to be environmental interference, and if S(t) < 0.9, it is determined to be a faulty section. After anti-interference processing, the noise suppression ratio (NSR) ≥ 25dB is considered qualified.

6. The method according to claim 1, characterized in that, In step S5, the improved Wigner-Ville transform uses a Hanning window with a window length of 20ms, an overlap rate of 50%, and no cross-term interference. Eleven time-frequency features were extracted, with a feature selection threshold of 0.

65. After selection, feature redundancy was reduced by ≥70%, and ≤4 strong fault features were ultimately retained.

7. The method according to claim 1, characterized in that, In step S6, ( , , These are the characteristic values ​​corresponding to crystallinity, coefficient of thermal expansion, and mechanical strength, respectively. ( , These are the eigenvalues ​​corresponding to vibration and acoustics, respectively. The correlation coefficient between the fusion feature and the fault label is ≥0.85; the fault determination rule is: if F_fusion(t) > T(t), it is determined to be a fault, otherwise it is normal.

8. The method according to claim 1, characterized in that, In step S7, the three-level early warning rules are as follows: Level 1 early warning (F_fusion exceeds T(t) by 1.0~1.5 times): yellow light reminder; Level 2 early warning (F_fusion exceeds T(t) by 1.5~2.0 times): orange light shutdown pre-inspection; Level 3 early warning (F_fusion exceeds T(t) by more than 2.0 times): red light emergency shutdown; when an early warning occurs, a screen pop-up window and audio-visual prompts are triggered simultaneously, and the fault period, characteristic value, and threshold deviation are automatically saved and pushed to the equipment PLC to execute deceleration, shutdown, or cooling operations.

9. The method according to claim 1, characterized in that, In step S8, the ultra-lightweight LSTM model adopts a single-layer LSTM + temporal distillation transfer learning structure, and the model training can be completed with 50 small samples. It can be deployed on embedded chips with a computing power consumption of ≤30%. The loss is the error between the model's predicted value and the actual fault label. The model is updated online incrementally without interrupting the detection process. The updated parameters are automatically synchronized to steps S3~S6, and a model optimization report is automatically generated every 24 hours.

10. A fault early warning system for LCP detection based on time-series analysis, characterized in that, The system is used to execute the LCP detection fault early warning method based on time series analysis according to any one of claims 1 to 9, the system comprising a data acquisition module, a time series calibration module, a feature enhancement module, an anti-interference module, a feature extraction module, a fusion detection module, an early warning control module, and an online optimization module; The system requires no manual selection throughout the process, and the formulas and parameters are fixed. It can be deployed with low computing power on embedded chips, and the detection accuracy is stable at ≥98%. It can be directly used for on-site deployment of LCP molding machines and high-precision testing machines.

Images