A motorcycle helmet production line equipment fault early warning analysis method

Abstract

The application provides a motorcycle helmet production line equipment fault early warning analysis method, through constructing a distributed multi-source sensing data acquisition and cleaning process covering injection molding pressure maintaining, vacuum adsorption, paint spraying leveling and UV curing and the like links, combining a three-layer learnable time convolution working condition semantic fingerprint encoder, realizing multi-modal feature extraction and standardized semantic vector generation; relying on continuous non-alarming working condition history, establishing a health fingerprint-index bimodal basic library with weights, comparing working condition semantic drift in real time, dynamically adjusting an alarm threshold based on a geodesic angle, realizing adaptive calibration of the threshold parameter, the application can significantly improve early identification accuracy and response sensitivity of abnormal trends of the production line equipment, and promote long-term stability and production safety of equipment operation.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent manufacturing and equipment health management technology, and in particular to a method for early warning analysis of equipment failures in a motorcycle helmet production line. Background Technology

[0002] In the field of intelligent manufacturing and industrial equipment health management, technologies for fault early warning, trend assessment, and anomaly detection for complex equipment such as motorcycle helmet production lines are continuously developing. Mainstream solutions generally employ historical statistical distribution thresholds, deep learning time-series models (such as LSTM / GRU), or reinforcement learning strategies for health status monitoring and fault trend prediction, combined with end-to-end acquisition of multi-source sensor data, attempting to achieve early identification and accurate warning of anomalies. Equipment operation data includes highly correlated time-series signals such as hydraulic pressure, temperature, current, environmental parameters, and visual inspection data, driving technological innovations such as multimodal coupling analysis, end-to-end trend modeling, and adaptive threshold calibration. Current industry trends focus on large-scale data-driven intelligent early warning, self-iterative health benchmark maintenance of knowledge bases, online model fine-tuning, and real-time deployment embedded at the edge. Representative technologies typically rely on the following implementation path: establishing initial thresholds for equipment health indicators, forming static or phased intervals through long-term historical data statistical analysis, and then combining this with anomaly detection models to judge real-time data. Some solutions introduce GAN adversarial training, graph neural network propagation, multi-layer residual structures, or meta-reinforcement learning to achieve anomaly trend compensation under dynamic operating conditions. However, these technologies generally face practical deployment obstacles such as high computing power requirements, weak interpretability, and frequent model retraining. Furthermore, in production line scenarios with frequent switching between multiple processes and modes, they often encounter problems such as threshold incompatibility, high false alarm rates, and excessively high maintenance thresholds. In industrial settings such as motorcycle helmet production lines, the current typical application involves collecting multi-source data on equipment operation, configuring fixed threshold criteria, or using historical operating condition clustering to differentiate between equipment health status and fault anomalies. This technology is primarily applicable to batch production lines with simple operating conditions or limited fluctuations, and is ill-suited for dynamic scenarios involving multiple machine models, frequent changes across process stages, and significant differences in operating conditions. Existing industry technologies lack the ability to abstract and respond to the fundamental changes in multiple equipment operating conditions, relying on statistical patterns from past data for threshold setting and alarm strategy formulation. When processes such as injection molding, painting, and curing change rapidly, the thresholds cannot dynamically adjust to the actual operating conditions, often leading to missed hazard reports or frequent false alarms triggered by abnormal disturbances, severely impacting production efficiency and the effectiveness of maintenance decisions. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for early warning analysis of equipment failure in a motorcycle helmet production line.

[0004] The technical solution of this invention is implemented as follows: A method for fault early warning analysis of motorcycle helmet production line equipment, comprising: S1: Acquire multi-source sensor data of motorcycle helmet production line equipment under typical working conditions in injection molding pressure holding section, vacuum adsorption section, paint leveling section and UV curing section. The multi-source sensor data includes hydraulic pressure fluctuation sequence, mold temperature gradient curve, servo motor current harmonic spectrum, ambient temperature and humidity time sequence and visual inspection yield feedback signal. S2: Construct a working condition semantic fingerprint encoder based on the multi-source sensor data, use three learnable temporal convolutional layers to perform feature dimensionality reduction on the input data, force the L2 norm of the output vector to always be 1, and generate a nine-dimensional working condition semantic fingerprint vector with unit spherical constraints. S3: Based on the historical operating data of a continuous 72-hour period without alarms, extract the corresponding operating condition semantic fingerprint vector and its associated equipment health indicators, and construct a dual-modal mapping baseline library of health operating condition fingerprints and health indicators containing confidence weights. S4: Collect sensor data within the current production cycle in real time and generate a real-time working condition semantic fingerprint vector. Search the three health fingerprint samples with the closest Euclidean distance in the dual-modal mapping baseline library of the health working condition fingerprint and health indicators to calculate the semantic drift distance. S5: If the semantic drift distance is greater than the preset semantic drift threshold, it is determined to be a new working condition. A temporary dynamic threshold is generated based on the recent health indicator sliding window. The temporary dynamic threshold parameter is fused with the original static warning threshold to output a dynamic calibration threshold. S6: If the semantic drift distance is less than the preset semantic drift threshold, it is determined to be a stable working condition. The geodesic angle between the real-time working condition semantic fingerprint vector and the matched healthy fingerprint sample on the unit sphere is calculated, and the geodesic angle is mapped to the threshold offset coefficient. The original static warning threshold is fine-tuned based on the threshold offset coefficient to generate the dynamic calibration threshold. S7: Input the dynamic calibration threshold and real-time trend probability into the multi-level fusion criterion module, set a time-sensitive criterion window according to the helmet production cycle, and trigger a first-level warning signal only when the trend probability is continuously higher than the dynamic calibration threshold in multiple consecutive cycle periods. S8: Execute equipment status control actions based on the first-level early warning signal, and at the same time record the mapping relationship between the semantic fingerprint vector of the current working condition and the threshold offset coefficient to update the confidence weight of the dual-modal mapping baseline library of the health working condition fingerprint and health indicators.

[0005] The present invention provides a fault early warning analysis method for motorcycle helmet production line equipment, which has the following beneficial effects: (1) This invention proposes a novel approach based on lightweight semantic fingerprint encoding and geometric threshold control. By constructing a low-dimensional semantic fingerprint encoder with only three layers of temporal convolution and forcing the output vector to lie on a unit sphere, it effectively extracts working condition features that characterize the intrinsic coupling stability of multimodal sensing data, avoiding explicit modeling of specific physical quantities and large-scale parameter training. This design significantly reduces the computational complexity of the model, enabling the system to generate working condition representations in real time at the edge at a frequency of once every 5 seconds, meeting the online monitoring requirements of high-frequency cycles in motorcycle helmet production lines. At the same time, the unit sphere constraint enables different working conditions to form distinguishable clustering structures in geometric space, enhancing the robustness and visualization capability of working condition identification and providing a reliable semantic basis for subsequent dynamic threshold adjustment. (2) This invention introduces an adaptive threshold offset mechanism based on geodesic angle and combines it with a confidence-weighted dual-modal baseline library to achieve dynamic calibration. During online operation, the system calculates the spherical angle between the semantic fingerprint of the current working condition and the nearest healthy sample in the baseline library to generate a nonlinearly variable threshold offset coefficient. While ensuring stability, it achieves gradual relaxation adjustment of the threshold, effectively avoiding false alarms caused by minor fluctuations in the working condition. If no new working condition is found, a cold start process is initiated, and a temporary dynamic threshold is constructed using a short-term sliding window to ensure that the system still has basic discrimination ability even without historical matching items. The entire process does not require retraining the model or updating the parameters, completely eliminating the dependence on complex parameter tuning mechanisms such as meta-learning and reinforcement learning, greatly improving the system's adaptive capability and ease of operation and maintenance. It is particularly suitable for actual scenarios where helmet production line equipment models are diverse and process switching is frequent. (3) To improve the accuracy and anti-interference capability of early warning decisions, this invention designs a multi-level criterion module that integrates trend probability and dynamic threshold, and introduces time-sensitive judgment logic based on production cycle time. This module does not rely on a single instantaneous limit-breaking event, but requires that the health trend be continuously higher than the dynamic threshold for N consecutive complete production cycles before triggering a first-level early warning, thereby effectively filtering false signals caused by short-term disturbances and enhancing the ability to capture the slow degradation process of equipment. All threshold adjustments are applied to the normalized health indicator space, keeping the underlying prediction model structure unchanged, and realizing the decoupling design of "perception-representation-decision", which not only ensures system stability, but also supports flexible expansion and cross-device migration applications. Attached Figure Description

[0006] Figure 1 This is a flowchart of a fault early warning analysis method for motorcycle helmet production line equipment according to the present invention; Figure 2 This is a sub-flowchart of a fault early warning analysis method for motorcycle helmet production line equipment according to the present invention; Figure 3This is another sub-flowchart of the fault early warning analysis method for motorcycle helmet production line equipment of the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. like Figure 1 As shown, this invention provides a method for fault early warning analysis of motorcycle helmet production line equipment, specifically including: S1: Acquire multi-source sensor data of motorcycle helmet production line equipment under typical working conditions in injection molding pressure holding section, vacuum adsorption section, paint leveling section and UV curing section. The multi-source sensor data includes hydraulic pressure fluctuation sequence, mold temperature gradient curve, servo motor current harmonic spectrum, ambient temperature and humidity time sequence and visual inspection yield feedback signal. S2: Construct a working condition semantic fingerprint encoder based on the multi-source sensor data, use three learnable temporal convolutional layers to perform feature dimensionality reduction on the input data, force the L2 norm of the output vector to always be 1, and generate a nine-dimensional working condition semantic fingerprint vector with unit spherical constraints. S3: Based on the historical operating data of a continuous 72-hour period without alarms, extract the corresponding operating condition semantic fingerprint vector and its associated equipment health indicators, and construct a dual-modal mapping baseline library of health operating condition fingerprints and health indicators containing confidence weights. S4: Collect sensor data within the current production cycle in real time and generate a real-time working condition semantic fingerprint vector. Search the three health fingerprint samples with the closest Euclidean distance in the dual-modal mapping baseline library of the health working condition fingerprint and health indicators to calculate the semantic drift distance. S5: If the semantic drift distance is greater than the preset semantic drift threshold, it is determined to be a new working condition. A temporary dynamic threshold is generated based on the recent health indicator sliding window. The temporary dynamic threshold parameter is fused with the original static warning threshold to output a dynamic calibration threshold. S6: If the semantic drift distance is less than the preset semantic drift threshold, it is determined to be a stable working condition. The geodesic angle between the real-time working condition semantic fingerprint vector and the matched healthy fingerprint sample on the unit sphere is calculated, and the geodesic angle is mapped to the threshold offset coefficient. The original static warning threshold is fine-tuned based on the threshold offset coefficient to generate the dynamic calibration threshold. S7: Input the dynamic calibration threshold and real-time trend probability into the multi-level fusion criterion module, set a time-sensitive criterion window according to the helmet production cycle, and trigger a first-level warning signal only when the trend probability is continuously higher than the dynamic calibration threshold in multiple consecutive cycle periods. S8: Execute equipment status control actions based on the first-level early warning signal, and at the same time record the mapping relationship between the semantic fingerprint vector of the current working condition and the threshold offset coefficient to update the confidence weight of the dual-modal mapping baseline library of the health working condition fingerprint and health indicators.

[0009] Step S1: Acquire multi-source sensor data of the motorcycle helmet production line equipment under typical operating conditions in the injection molding and holding pressure section, vacuum adsorption section, paint spraying and leveling section, and UV curing section. The multi-source sensor data includes hydraulic pressure fluctuation sequences, mold temperature gradient curves, servo motor current harmonic spectrum, ambient temperature and humidity time series, and visual inspection yield feedback signals. Specifically, this includes: S1.1: Perform topology identification on the distributed sensor network of the injection molding pressure holding section, vacuum adsorption section, paint leveling section and UV curing section, and synchronously acquire the original sequence of hydraulic pressure fluctuation, the original curve of mold temperature gradient, the original harmonic spectrum of servo motor current, the original time sequence of ambient temperature and humidity and the original feedback signal of visual inspection yield based on the preset sampling frequency, so as to form a multi-source heterogeneous original data acquisition package containing timestamp alignment marks. When performing topology identification on the distributed sensor network in the injection molding and holding section, vacuum adsorption section, paint spraying and leveling section and UV curing section of the motorcycle helmet production line, a sensor node list is independently constructed for each of the four working conditions to clarify the physical location and network connection relationship of the hydraulic pressure sensor, mold temperature sensor, servo motor current sensor, ambient temperature and humidity sensor and vision inspection unit. Based on the topology recognition results, the sampling parameters of each node are uniformly configured, the sampling frequency of the hydraulic pressure sensor is set to match the sampling frequency of the mold temperature sensor, servo motor current sensor, environmental sensor and vision inspection unit, and a global clock source is generated in the data acquisition system as a time synchronization reference. By using time synchronization control logic, trigger acquisition commands are executed on multiple sensor signals, so that the original sequence of hydraulic pressure fluctuation, the original curve of mold temperature gradient, the original harmonic spectrum of servo motor current, the original time sequence of ambient temperature and humidity, and the original feedback signal of visual inspection yield rate are formed into a multi-channel data vector at the same timestamp. To ensure the consistency and traceability of data across different operating conditions, the collected multi-channel data vectors are structured and encapsulated according to production batch and operating condition identifiers, and a timestamp alignment marker field is introduced, which is then mapped to the actual physical timeline of equipment operation through index relationships. Through the above-mentioned topology identification and synchronous acquisition processing method, the operating condition information of the previous step is transformed into a multi-source heterogeneous raw data acquisition package containing timestamp alignment marks, so as to realize the structured and unified input of multi-source data. For example, the distributed sensor network in the A-line injection molding pressure holding section includes four hydraulic pressure sensors (sampling frequency set to 100Hz), three mold temperature sensors (sampling frequency set to 50Hz), two servo motor current sensors (sampling frequency set to 200Hz), one ambient temperature and humidity sensor (sampling frequency set to 10Hz), and one vision inspection unit (sampling frequency set to 1Hz, based on frame capture). After topology recognition, the acquisition system configures the global clock source to millisecond-level resolution and performs interpolation synchronization on the sampling times of each sensor, executing acquisition commands on the signals of each channel at the same millisecond. For example, at 14:00:00.000 on 2024-06-01, the acquired values ​​are: hydraulic pressure 5.23MPa, mold temperature 85.6℃, servo motor current harmonic amplitude sequence length 256, ambient temperature 26.5℃, humidity 58.1%, and vision inspection yield 0.978. The acquisition system encapsulates each signal into a data record according to the operating condition segment identifier "A line_injection molding pressure holding", and adds a timestamp alignment mark 2024-06-01 14:00:00.000, forming a multi-source heterogeneous raw data acquisition package. This acquisition package shows significantly improved time consistency and data traceability in subsequent outlier removal, missing value imputation, and feature standardization processes, thereby enhancing the adaptability and robustness of the model analysis module for operating condition trend prediction and anomaly detection; S1.2: Perform outlier removal and missing value interpolation on the multi-source heterogeneous raw data acquisition package, use the sliding window statistical filtering algorithm to remove noise spikes exceeding three times the standard deviation, and use linear interpolation to fill the data gaps caused by communication packet loss, so as to generate a multi-source sensor cleaning dataset with continuous time axis characteristics. S1.3: Based on the multi-source sensor cleaning dataset, perform dimensional normalization transformation. According to the historical extreme value range of each physical quantity, map the hydraulic pressure fluctuation sequence, mold temperature gradient curve, servo motor current harmonic spectrum, ambient temperature and humidity time series and visual inspection yield feedback signal to the dimensionless range of zero to one, so as to eliminate the numerical deviation caused by the difference in the range of different sensors and obtain a standardized multi-source sensor feature matrix. S1.4: Perform time-frequency domain joint enhancement processing on the standardized multi-source sensing feature matrix, extract the frequency band energy distribution of the servo motor current harmonic spectrum through short-time Fourier transform, and reconstruct the transient change component of the mold temperature gradient curve by wavelet packet decomposition to generate a multi-mode coupled enhancement signal set that integrates time-domain statistics and frequency-domain energy features. S1.5: Based on the multimodal coupling enhancement signal set, perform production cycle slice alignment operation, extract continuous data segments according to the single-piece cycle time window of the helmet production line, and encapsulate multi-dimensional features such as hydraulic pressure fluctuation, mold temperature gradient, servo motor current, ambient temperature and humidity and visual inspection yield into structured time-series data blocks with working condition segment identifiers, so as to output a standardized multi-source sensor data stream for subsequent working condition semantic fingerprint encoders to be directly called.

[0010] Step S2: Construct a working condition semantic fingerprint encoder based on the multi-source sensor data, perform feature dimensionality reduction on the input data using three learnable temporal convolutional layers, and enforce the L2 norm of the output vector to always be 1, generating a nine-dimensional working condition semantic fingerprint vector with unit spherical constraints. Specifically, this includes: S2.1: Perform time window slicing on multi-source sensor data such as hydraulic pressure fluctuation sequence, mold temperature gradient curve, servo motor current harmonic spectrum, ambient temperature and humidity time sequence and visual inspection yield feedback signal, and use sliding time window algorithm to generate multi-channel time sequence data blocks containing the complete production cycle, so as to form the original working condition data tensor with spatiotemporal alignment characteristics. S2.2: Perform the first layer of learnable temporal convolution operation based on the original working condition data tensor, extract local high-frequency transient features using a one-dimensional convolution kernel with causal mask, filter noise interference through a nonlinear activation function, and output a primary temporal feature map containing short-term impact response characteristics; S2.3: Perform a second-layer learnable temporal convolution operation on the primary temporal feature map, expand the receptive field using the dilated convolution mechanism to capture mid-frequency periodic fluctuation patterns, and integrate the coupling correlation information between multiple sensors through feature fusion operations to generate an intermediate temporal feature map that characterizes the mid-term stability of the working condition. S2.4: Based on the intermediate temporal feature map, perform the third layer of learnable temporal convolution operation, use the global average pooling strategy to compress the temporal dimension and retain key statistical moment information, project the high-dimensional features to the nine-dimensional latent space through the fully connected mapping layer, and output the unnormalized nine-dimensional working condition semantic candidate vector. Based on the input conditions of the intermediate time series feature map, a high-dimensional time series feature with multi-sensor coupling information and a mid-frequency periodic fluctuation mode is selected as the execution object; Global average pooling is performed on the high-dimensional time series feature data along the time axis. The sum of all time sampling point values ​​of each channel is divided by the number of sampling points to compress the time dimension. Statistical moment information related to operating condition stability is preserved by maintaining the average value of each channel. The global average pooling output is called into the fully connected mapping layer. The mean of each channel is used as the input node of the mapping layer. The matrix multiplication operation is performed using the weight matrix to project the high-dimensional input vector into the nine-dimensional latent space. In matrix multiplication, the coefficients of each projection dimension are set with reference to the uniformity of the historical working conditions in the spherical space, and the output center value of each dimension is adjusted by the bias vector. The unnormalized nine-dimensional working condition semantic candidate vector is output from the mapping layer to ensure that the vector fully reflects the multimodal coupling characteristics and potential working condition modes in the numerical space. By combining global average pooling and fully connected mapping, the intermediate temporal features from the previous step are transformed into unnormalized nine-dimensional working condition semantic candidate vectors, achieving information compression and pattern preservation of high-dimensional temporal features in a low-dimensional latent space. For example, in an embodiment targeting the leveling section of a motorcycle helmet paint spray, the input intermediate-time feature map contains 64 channels, each corresponding to mid-frequency fluctuation information from different sensor combinations, with 1024 time sampling points. The global average pooling operation, for the k-th channel, is calculated using the following formula: in, This represents the feature value of the i-th time sampling point in the k-th channel. This represents the global average value for the k-th channel. The 64-dimensional average value vector is fed into the fully connected mapping layer. The weight matrix of the fully connected layer is 64×9, and the bias vector length is 9. The matrix multiplication operation is as follows: in, This is the weight matrix. It is a 64-dimensional global average vector. For bias vectors, This is a nine-dimensional semantic candidate vector for the working condition. The weight matrix of the mapping layer is optimized during the training phase to ensure that the distribution of each dimension in the nine-dimensional space matches the geometric shape of the historical health condition fingerprint. The output nine-dimensional candidate vector is not normalized and directly enters the L2 norm normalization of the next sub-step to ensure that the original numerical proportions and mode differences are maintained. In this embodiment, the nine-dimensional candidate vector can clearly distinguish the multimodal coupling characteristics of high humidity environment and low humidity environment under the painting leveling condition, and after compressing the time dimension through global average pooling, the processing latency is significantly reduced, meeting the low latency requirements of real-time trend prediction in the production line. S2.5: Perform L2 norm normalization on the nine-dimensional working condition semantic candidate vectors, use vector magnitude calculation and scalar division to force the Euclidean length of the output vector to be always equal to 1, map the candidate vectors onto the unit hyperspherical manifold, and finally generate a nine-dimensional working condition semantic fingerprint vector with unit spherical constraints. When performing norm normalization on the unnormalized nine-dimensional working condition semantic candidate vector output by the preceding step S2.4, the candidate vector is used as the input object, and each of its components is regarded as a coordinate value in nine-dimensional Euclidean space. A module for calculating the modulus is constructed, which obtains the Euclidean length parameter of the vector through the sum of squares operation, and explicitly adds a square root operation to the denominator to eliminate the difference in square dimensions; Using the scalar division operation module, each component in the candidate vector is divided by the above Euclidean length parameter, so that the output vector satisfies the constraint that the Euclidean length is always equal to 1. In the mapping and transformation stage, the normalized vector is projected onto the unit hyperspherical manifold, and its orientation information is kept unchanged while the scale information is eliminated by the mapping rule of the unit spherical coordinate system. Perform unit spherical geometry verification to detect the norm error of the normalized output vector and ensure that the error does not exceed the preset numerical tolerance, thereby verifying the effectiveness of the normalization process; By coordinating the control of vector norm normalization and geometric mapping, the result of the previous step is transformed into a nine-dimensional semantic fingerprint vector that conforms to the unit spherical constraint, thereby achieving the expected technical effect of subsequent semantic drift angle calculation based on spherical geometry. For example, in the leveling stage of a motorcycle helmet paint spraying process, the nine-dimensional candidate vector generated in the previous steps is [0.58, 0.40, 0.71, 0.35, 0.49, 0.61, 0.33, 0.57, 0.46]. The modulus calculation module performs a sum of squares on each component, resulting in a sum of squares of 3.294. The square root operation yields the Euclidean length parameter of 1.815. The scalar division module divides each component by 1.815, resulting in a normalized output vector [0.319, 0.220, 0.392, 0.193, 0.270, 0.336, 0.182, 0.314, 0.254]. The unit spherical geometry verification module performs a norm check on this normalized vector, resulting in a modulus of 1.000, with the error value controlled within 1×10⁻⁶. -5 Within a certain range, the unit spherical constraint condition is satisfied. This normalized vector exhibits stable directional characteristics in the subsequent geodesic angle calculation, effectively reducing angle calculation distortion caused by scale differences when matching healthy fingerprint samples, and significantly improving the accuracy of dynamic threshold self-calibration.

[0011] like Figure 2As shown, step S3 involves extracting the corresponding semantic fingerprint vector of the operating condition and its associated equipment health indicators based on historical operating data from a continuous 72-hour period without alarms, and constructing a dual-modal mapping baseline library of health operating condition fingerprints and health indicators containing confidence weights. Specifically, this includes: S3.1: Perform time window slicing on the multi-source sensor data during the 72-hour period without alarms to extract independent working condition data segments corresponding to the injection molding pressure holding section, vacuum adsorption section, spray painting leveling section and UV curing section, thereby obtaining the original data sequence of typical working conditions with timestamps. S3.2: Based on the original data sequence of the typical working conditions, the working condition semantic fingerprint encoder constructed in the previous step is input to perform feature mapping operation to generate a nine-dimensional working condition semantic fingerprint vector with unit spherical constraint, thereby obtaining a set of historical working condition semantic fingerprints that strictly correspond to the historical fault-free period. An input path is established for the original data sequence of the typical working conditions, and each working condition data segment is sequentially sent to the input port of the working condition semantic fingerprint encoder constructed in the previous step to ensure the uniform formatting of the working condition data in the time dimension and physical quantity dimension. The first, second, and third layers of learnable temporal convolution operations and global average pooling operations are performed on the working condition data tensor input to the encoder to preserve the multi-level temporal features after convolution kernel extraction and nonlinear activation filtering, so as to maintain the full spectrum information of short-term impact response, medium-term stability and long-term trend features. The unnormalized nine-dimensional working condition semantic candidate vector output by the fully connected mapping layer is fed into the L2 norm normalization processing module. The vector magnitude is calculated and normalized by division to ensure that the vector falls on the unit hyperspherical manifold and satisfies the spherical geometric constraint. The formula for calculating the vector magnitude is as follows: , where x, y, and z represent the feature components of each dimension, and the modulus is always equal to 1, to achieve geometric normalization and obtain the final nine-dimensional working condition semantic fingerprint vector; The normalized nine-dimensional working condition semantic fingerprint vectors are merged into the historical working condition semantic fingerprint set in the order of timestamps to form a standardized semantic representation dataset that is strictly aligned with the historical fault-free period. Through the above processing method, the original data sequence of typical working conditions in the previous step is transformed into a nine-dimensional semantic fingerprint set of working conditions that satisfies the unit spherical constraint and is suitable for subsequent health indicator correlation calculation, thereby realizing the abstraction and geometric comparability of working condition features. For example, hydraulic pressure fluctuation data, mold temperature gradient curves, and servo motor current harmonic spectra collected in the injection molding and pressure holding section of a motorcycle helmet production line are processed by time window slicing to form a multi-channel time series tensor with a length of 98 seconds, which is then input into the working condition semantic fingerprint encoder. The first layer convolutional kernel is set to a length of 5 and a stride of 1, the second layer convolutional kernel has an expansion coefficient of 2, and the third layer convolutional kernel has a length of 9 and uses global average pooling to compress the time dimension. The candidate vector output by the fully connected mapping layer is [0.35, 0.48, 0.27, 0.15, 0.22, 0.44, 0.31, 0.29, 0.38], which is calculated by L2 norm. The modulus length is 0.947. Each component is divided by this modulus to obtain a normalized vector [0.369, 0.507, 0.285, 0.158, 0.232, 0.465, 0.328, 0.306, 0.401]. This vector is then stored in the corresponding historical no-alarm period semantic set to form the basic representation data for subsequent health indicator matching. In this embodiment, the fingerprint vector output has an angle of less than 0.15 radians with the healthy baseline sample in online testing, verifying the high stability and comparability of the encoder under normal operating conditions. S3.3: Execute a statistical feature extraction algorithm on the original data sequence of the typical working conditions to calculate equipment health indicators such as the mean of vibration kurtosis, the standard deviation of temperature change rate, and the coefficient of variation of coating thickness, thereby obtaining a set of historical equipment health indicators that are precisely aligned with the historical working condition semantic fingerprint set in the time dimension. S3.4: Based on the association between the historical operating condition semantic fingerprint set and the historical equipment health indicator set, a weighted fusion calculation is performed by combining the operating condition duration, data integrity score and cross-sensor consistency score to generate a confidence weight value that characterizes the reliability of the data, thereby obtaining a dual-modal data pair of health operating condition fingerprint and health indicator with confidence weight attribute. The input conditions include a set of historical operating condition semantic fingerprints and a set of historical equipment health indicators that are strictly corresponding to them in time. Both of them are derived from the operating condition coding and statistical feature extraction results completed in the previous steps. Based on the historical operating condition semantic fingerprint set and the historical equipment health indicator set, an association matrix is ​​constructed, and index matching is used to ensure that each operating condition fingerprint vector and the corresponding health indicator are precisely aligned in the same time slice. Calculate the duration of operating conditions on the correlation matrix, calculate the cumulative running time of each type of operating condition during the continuous alarm-free period, and input the duration as a stability reference factor into the weight model. Perform data integrity scoring on the correlation matrix, statistically analyze the effective sampling ratio and missing repair ratio of each health indicator time series, and input the score as a reliability reference factor into the weight model; Cross-sensor consistency score calculation is performed on the correlation matrix. The correlation coefficient matrix between multi-source health indicators is used to evaluate the consistency of signals collected by different sensors under the same operating conditions. This score is then used as a multimodal consistency reference factor and input into the weighting model. A weighted fusion formula is constructed to normalize the duration of the operating condition, the data integrity score, and the cross-sensor consistency score, and then sum them according to a preset weighting ratio. For example: in, These are the confidence weight values. This is a normalized value for duration. This is a data integrity normalized value. The normalized value for consistency score. , , These are the weighting coefficients; The obtained confidence weight values ​​are appended to each pair of health condition fingerprints and health indicators to form a bimodal data record with confidence weight attributes. Through the above weighted fusion processing method, the working condition semantic fingerprint and health indicator time alignment results of the previous step are transformed into a dual-modal data pair with data reliability measurement, thereby realizing the weight initialization preparation of the health baseline knowledge base. For example, in a motorcycle helmet production line, the historical operating condition fingerprint set of the injection molding and pressure holding section contains 120 nine-dimensional unit spherical constraint vectors. The corresponding health indicator set includes three categories of indicators: mean hydraulic vibration kurtosis, standard deviation of mold temperature change rate, and coefficient of variation of coating thickness. Duration statistics show that the cumulative runtime of this condition is 43,200 seconds, normalized to... = The effective sampling ratio of each indicator in the data integrity score calculation is greater than 0.95, and the overall normalized value is... = The cross-sensor consistency score was 0.88, obtained by averaging the correlation coefficient matrix of health indicators, and the normalized value was... = The weighting coefficients are configured as follows: = , = , = Substituting into the formula yields the confidence weights. = × + × + × = The confidence weight is added to the corresponding injection molding pressure holding section health condition fingerprint and health index data pair and then written into the baseline library. The verification results show that this weight can significantly improve the stability and robustness of multi-timescale trend prediction during the dynamic threshold self-calibration process. S3.5: Perform a classification and storage operation on the dual-modal data pairs of health condition fingerprints and health indicators with confidence weight attributes according to the device unit identifier, so as to construct a dual-modal mapping baseline library of health condition fingerprints and health indicators that can be independently retrieved by device, thereby completing the initialization of the benchmark knowledge base for dynamic threshold self-calibration; For the dual-modal data pairs of health condition semantic fingerprints and health indicators with confidence weight attributes, an equipment unit classification index table is established using the equipment unit identifier as the classification index input to map the relationship between the data pairs and the corresponding equipment units. Based on the equipment unit classification index table, the data storage management module is called to group and cache the health data pairs of different equipment units to ensure that the working condition semantic fingerprint and health indicators in each group are consistent in terms of timestamp, working condition segment identifier and confidence weight value. For the data in each group cache, perform index optimization processing, map the nine-dimensional working condition semantic fingerprint vector into a hash key that can be retrieved quickly, and bind the corresponding health indicator set with the confidence weight value to form a bimodal mapping record entry; By using a combined index structure of hash keys and device unit identifiers, a categorized write operation is performed in the storage engine to write each bimodal mapping record entry to the corresponding device-independent data table and attach timestamp metadata tags to support historical data backtracking analysis. The storage verification module is invoked to perform integrity verification on the written bimodal mapping record entries. The verification content includes the correctness of the length of the nine-dimensional working condition semantic fingerprint data, the legality of the health indicator values, and the confidence weight range constraints, to ensure the reliability of the baseline library initialization data. Through the above classification index construction and verification process, the results of the previous step are transformed into a dual-modal mapping baseline library of health condition semantic fingerprint and health indicators that can be independently retrieved by device unit, so as to realize the initialization of the benchmark knowledge base required for dynamic threshold self-calibration. For example, during the baseline library initialization process of the motorcycle helmet injection molding machine unit, the historical nine-dimensional working condition semantic fingerprint vector with no alarms for 72 consecutive hours is used as input, and the equipment unit identifier is set as "A-line injection molding machine". When establishing the classification index table, the set of health indicators corresponding to this unit includes a mean vibration kurtosis of 0.145, a standard deviation of temperature change rate of 0.031, a coating thickness CV value of 0.012, and a confidence weight of 0.92. In the index optimization stage, the nine-dimensional working condition semantic fingerprint vector is mapped to a hash key of length 64, and the above health indicators and confidence weights are bound to form record entries. The composite index structure combines the hash key with the equipment unit identifier, performs classification, writes it to the independent data table of "A-line injection molding machine", and adds a timestamp "2024-05-12 08:00:00". The integrity verification module checks that the nine-dimensional vector dimension is correct, the health indicator values ​​are within the historical extreme range, and the confidence weight falls between 0 and 1. All verifications pass, and the baseline library initialization is completed. The initialization results provide a fast and accurate health benchmark retrieval capability during subsequent dynamic threshold calibration, enabling the dynamic threshold adjustment to maintain significantly improved stability and adaptability in the real-time operating condition changes of the unit.

[0012] like Figure 3 As shown, step S4 involves: real-time acquisition of sensor data within the current production cycle and generation of a real-time operating condition semantic fingerprint vector; retrieving the three health fingerprint samples with the closest Euclidean distance from the dual-modal mapping baseline library of health condition fingerprints and health indicators to calculate the semantic drift distance. Specifically, this includes: S4.1: Acquire real-time multi-source sensor data consisting of hydraulic pressure fluctuation sequence, mold temperature gradient curve, servo motor current harmonic spectrum, ambient temperature and humidity time sequence and visual inspection yield feedback signal within the current production cycle. Use time sliding window mechanism to perform synchronization alignment and noise suppression processing on the real-time multi-source sensor data to generate a standardized real-time multi-source sensor data matrix with a unified timestamp reference. S4.2: Based on the standardized real-time multi-source sensor data matrix, input it into the working condition semantic fingerprint encoder that has been constructed in the previous step, and use the three learnable temporal convolutional layers inside the working condition semantic fingerprint encoder to perform multi-level feature extraction and nonlinear transformation processing on the standardized real-time multi-source sensor data matrix to output an intermediate layer feature tensor containing high-dimensional abstract features. Based on the standardized real-time multi-source sensor data matrix, the working condition semantic fingerprint encoder that has been constructed and optimized by weight parameters in the previous steps is called. The multi-channel data corresponding to hydraulic pressure fluctuation, mold temperature gradient, servo motor current harmonic spectrum, ambient temperature and humidity and visual inspection yield are arranged in time axis and input into the first layer of learnable temporal convolutional network of the encoder. The convolutional kernel length and causal mask mode are set to ensure that subsequent sampling in the sampling sequence will not affect the previous feature extraction results. Based on the local high-frequency transient feature map output from the first layer, it is input into the second layer of the temporal convolutional network with dilation coefficient. By setting the dilation coefficient, the receptive field coverage is expanded to retrieve the mid-frequency periodic pattern, and cross-channel feature fusion operation is performed to preserve the coupling response between hydraulic temperature and current harmonics, forming a mid-level temporal feature map containing mid-term stable features. The intermediate temporal feature map is input into the global average pooling and fully connected mapping unit of the third-layer temporal convolutional network to compress the temporal dimension and retain global statistical features, while realizing the projection transformation from the high-dimensional original space to the nine-dimensional latent space. The feature vectors output by the convolutional layer in the above process are sparsified and enhanced by nonlinear activation functions such as ReLU, so that the noise response in the high-dimensional feature tensor is suppressed below the threshold and the equipment operation status mode is highlighted. Combining the feature dimension constraint mechanism, an intermediate layer feature tensor containing high-dimensional abstract features is generated in the third convolution output stage. This tensor has complete spatiotemporal dependency relationship and multimodal fusion characteristics, which satisfies the input conditions of subsequent principal component projection processing. By leveraging the synergistic effect of convolution kernel parameters, dilation coefficients, pooling windows, and nonlinear functions, the standardized real-time multi-source sensor data matrix from the previous step is transformed into a high-dimensional intermediate layer feature tensor with pattern separation capabilities in both the time and frequency domains, thereby achieving a precise mapping of operational features from original sensor measurements to semantic representation vectors. For example, data from each channel of the standardized real-time multi-source sensor data matrix—hydraulic pressure fluctuation, mold temperature gradient, servo motor current harmonic spectrum, ambient temperature and humidity, and visual inspection yield—are arranged at a sampling frequency of 100Hz and input into the first-layer convolutional network. The convolutional kernel length is set to 5, the stride to 1, and a causal mask pattern is applied to ensure temporal order, extracting short-term high-frequency peaks and transient fluctuation features. The features output from the first-layer convolution are mapped to the second-layer dilated convolutional network, with a dilation coefficient of 4 and a kernel length of 7, fusing cross-channel features and capturing mid-cycle patterns such as the stable fluctuation pattern of hydraulic pressure during the injection molding holding phase. The output of the second-layer convolution is input into the third-layer global average pooling unit, with the pooling window covering the entire time dimension. This is then projected onto a nine-dimensional latent space through a fully connected mapping layer and sparsified using ReLU to obtain a nine-dimensional high-dimensional feature tensor. For example, for the servo motor harmonic spectrum channel, its average energy value at the output of the third-layer convolution is... After ReLU transformation, it is still Noise energy value such as The value is then set to zero, thereby improving the purity of the feature signal. This nine-dimensional feature tensor is further compressed into a real-time working condition semantic fingerprint vector that satisfies the unit spherical constraint in the subsequent principal component projection. The verification results show that when there is a sudden temperature fluctuation in the UV curing section, this process can significantly improve the model's ability to capture non-stationary modes, and the output high-dimensional abstract features show high robustness and high adaptability in semantic drift detection. S4.3: Perform principal component dimensionality reduction projection operation on the intermediate layer feature tensor, extract the first three-dimensional principal component components and splice them to form a nine-dimensional original feature vector, and then perform L2 norm normalization constraint operation on the nine-dimensional original feature vector to generate a real-time working condition semantic fingerprint vector that strictly satisfies the unit spherical geometric constraint condition. When performing feature projection processing on the intermediate layer feature tensor, the input object is a standardized intermediate layer feature tensor output by the preceding temporal convolutional layer and containing multi-dimensional abstract features. Its dimension is identified as M×N, where M represents the number of feature channels and N represents the number of temporal sampling points. Based on this feature tensor, a covariance matrix is ​​constructed. By solving its eigenvalues ​​and eigenvector array, a principal component vector group sorted by contribution rate is obtained. The original high-dimensional feature space is mapped to the principal component subspace using a projection operator matrix. The first three principal component components are truncated to capture the signal energy and operating condition change trend in the direction of maximum variance. Arrange the corresponding values ​​of each principal component in different feature domains by channel and perform a concatenation operation to obtain an original feature vector of length nine. Euclidean modulus calculation is performed on the original nine-dimensional eigenvector, and L2 norm normalization constraints are constructed using the following formula: in, The original nine-dimensional eigenvector is represented by the sum of squares symbol, which indicates the summation of the squares of each component, and the square root operation is used to obtain the vector magnitude. Divide each component by the modulus to ensure that the length of the output vector is always equal to one, thereby mapping it onto the unit spherical manifold to form a real-time working condition semantic fingerprint vector that strictly satisfies the geometric constraints of the unit spherical surface. By combining principal component dimensionality reduction projection with L2 norm constraints, the high-dimensional feature tensor of the intermediate layer in the previous step is transformed into low-dimensional, normalized real-time working condition semantic fingerprint data that can be used for semantic drift measurement and baseline library retrieval, thus achieving the expected technical effect of significantly improving online matching accuracy and computational efficiency. For example, in the real-time monitoring of the injection molding and pressure holding section of a motorcycle helmet production line, assuming the intermediate layer feature tensor dimension is 12×100 and the covariance matrix size is 12×12, the first three terms in the eigenvalue sequence are 4.2, 3.8, and 3.1, respectively. The corresponding eigenvectors are used to construct the projection matrix, projecting the original 12-dimensional features into 3-dimensional principal component components. The values ​​of each principal component are taken from three different sensor domains (hydraulic pressure, mold temperature, and current harmonics) and concatenated to generate a nine-dimensional original feature vector [0.52, 0.48, 0.50, 0.60, 0.55, 0.58, 0.49, 0.47, 0.51]. Its L2 norm is calculated. The result is 1.57. Dividing each component of the original nine-dimensional feature vector by 1.57 yields the normalized fingerprint vector [0.331, 0.306, 0.318, 0.382, 0.350, 0.369, 0.312, 0.299, 0.325], whose magnitude is always equal to 1, satisfying the unit sphere constraint. In this embodiment, the normalized fingerprint vector can quickly locate the nearest neighbor sample matching the working condition in the baseline library, significantly improving the accuracy of semantic drift distance measurement and ensuring the stable execution of the dynamic threshold self-calibration mechanism under different working conditions. S4.4: Call the bimodal mapping baseline library of health condition fingerprint and health index constructed in the previous steps, use the real-time working condition semantic fingerprint vector as the query index, perform Euclidean distance metric calculation in the bimodal mapping baseline library of health condition fingerprint and health index to filter out the three health fingerprint sample sets that are closest to the real-time working condition semantic fingerprint vector in space. S4.5: Based on each healthy fingerprint sample in the three sets of healthy fingerprint samples and the real-time operating condition semantic fingerprint vector, calculate the Euclidean space straight-line distance between each pair, and take the minimum value of the Euclidean space straight-line distance to generate a semantic drift distance scalar that represents the degree of deviation of the current operating state from the health benchmark. Based on the pairwise matching pairs formed by the three health fingerprint sample sets selected from the previous sub-steps and the real-time working condition semantic fingerprint vectors, the input objects are set to be nine-dimensional vector data that strictly satisfy the unit spherical constraint. For each matching pair, perform Euclidean distance measurement in multidimensional space, sum the squares of each component of the nine-dimensional vector difference and take the square root to obtain the straight-line distance value; A pairwise calculation mode is adopted to ensure that the distance between each healthy fingerprint sample and the real-time fingerprint vector is calculated independently with double-precision floating-point accuracy. In the process of distance calculation, the Euclidean distance formula is defined as follows: in This represents the i-th dimension component of the real-time operating condition fingerprint. This represents the i-th dimension component of a fingerprint sample under healthy operating conditions; This formula maps the sum of squares of vector differences to a distance scalar. Performing this operation on three healthy fingerprint samples generates three sets of distance value sequences. The minimum value operation is performed on the obtained distance value sequence, and the minimum distance value is used as the semantic drift distance scalar to represent the degree of deviation between the current operating state and the health benchmark; By using the above Euclidean distance measurement and minimum value filtering process, the search results of the previous step are transformed into core indicators that can quantitatively describe the degree of deviation of the current working condition, thereby achieving a benchmark deviation measurement that is sensitive to and robust to changes in working conditions. For example, in the paint leveling section of a motorcycle helmet production line, the nine-dimensional components of the real-time working condition semantic fingerprint vector are set as [0.12, 0.35, 0.47, 0.51, 0.26, 0.19, 0.33, 0.44, 0.29]. The three healthy fingerprint sample components are H1: [0.15, 0.30, 0.50, 0.49, 0.28, 0.22, 0.31, 0.46, 0.27], H2: [0.10, 0.37, 0.44, 0.53, 0.25, 0.18, 0.34, 0.42, 0.30], and H3: [0.13, 0.36, 0.45, 0.50, 0.27, 0.20, 0.33, ...]. [0.45, 0.28]. For H1, the sum of squared differences is calculated dimension by dimension and the square root is taken to obtain a distance value of 0.065; for H2, the distance value is 0.057; and for H3, the distance value is 0.060. The minimum value of the three distance values ​​is selected to obtain a semantic drift distance scalar of 0.057, which serves as the core quantitative indicator of the deviation of the working condition in this scenario. In performance verification, when the semantic drift distance is less than the preset threshold of 0.08, the system is determined to be in a stable working condition, and the subsequent calculation of the geodesic angle is used for threshold fine-tuning. When the distance value increases significantly and exceeds the threshold, the new working condition judgment and temporary dynamic threshold generation process are triggered, thereby significantly improving the robustness and adaptability of the fault warning system under dynamic working conditions.

[0013] Step S5: If the semantic drift distance is greater than a preset semantic drift threshold, it is determined to be a new operating condition. A temporary dynamic threshold is generated based on a recent health indicator sliding window. A dynamic calibration threshold is output by fusing the temporary dynamic threshold parameter with the original static warning threshold. Specifically, this includes: S5.1: Perform Euclidean distance calculation on the real-time operating condition semantic fingerprint vector and the three health fingerprint samples retrieved from the dual-modal mapping baseline library of the health operating condition fingerprint and health indicators to obtain a semantic drift distance value that characterizes the degree of deviation of the current operating state from the health benchmark. S5.2: Based on the semantic drift distance value and the preset semantic drift threshold, a size comparison and judgment process is performed to generate a condition type determination flag indicating whether the current production scenario belongs to a known stable condition or an unknown new condition; S5.3: If the working condition type determination flag indicates an unknown new working condition, then perform rolling quantile statistical analysis on the equipment health index sliding window data within the last thirty minutes to extract temporary dynamic threshold parameters that can reflect the characteristics of short-term sudden fluctuations. The sliding window data of equipment health indicators within the last 30 minutes are time-synchronized and calibrated, and the health indicators from vibration sensors, temperature sensors and coating thickness detection units are arranged into a continuous data sequence of 30 minutes in length according to a unified timestamp reference. A window segmentation operation is performed on the calibrated health indicator data sequence, dividing the 30-minute data into several fixed-length sub-windows. The length of the sub-window is set as an integer multiple of the helmet production cycle to ensure that the data within each sub-window has the integrity of the production cycle. A rolling update mechanism is implemented for the health indicator data in each sub-window, adding the health indicator samples newly added at the most recent sampling time to the window, while removing the oldest sample, so as to realize the dynamic translation of the window content over time. Perform quantile statistical analysis within the scrolling updated child window, based on preset quantile level parameters. Calculating the quantile thresholds for health indicators involves the mathematical calculation of quantiles, assuming the window sample set is... The sample size is The quantile position index is The formula for calculating the position index is: ,in The function is called the floor function, and the sample value corresponding to the resulting index is the quantile threshold parameter. The quantile threshold parameters of each sub-window are aggregated in chronological order to form a quantile time curve. The latest value of this curve is used as a temporary dynamic threshold output, and a time variability index is attached to characterize the strength of short-term fluctuations. Through the above rolling quantile statistical analysis, the unknown new working condition judgment result of the previous step is transformed into a temporary dynamic threshold parameter that can adapt to short-term sudden fluctuations, so as to realize the threshold adaptive capability without retraining. For example, when a motorcycle helmet production line encounters new operating conditions, the vibration kurtosis mean sensor output is recorded at a sampling frequency of 1Hz, the standard deviation of the temperature change rate is sampled every 2 seconds, and the coefficient of variation of the coating thickness is sampled every 5 seconds. Within the last thirty minutes, the window length is set to an integer multiple of the single-piece cycle time of 98 seconds, i.e., a window length of 294 seconds. Each window contains 294 vibration kurtosis samples, 147 temperature change rate samples, and 59 coating thickness samples. A rolling update mechanism is used; newly sampled vibration kurtosis values ​​are added to the end of the window, and the oldest kurtosis values ​​are discarded. Within each window, the vibration kurtosis mean sample set is arranged in ascending order, with the number of samples... =294, the quantile position index is calculated as follows: = The 265th sample value was used as the temporary threshold for vibration kurtosis. Within the same window, the same calculations were performed on the temperature change rate and coating thickness indicators to obtain their respective quantile thresholds. The quantile thresholds of the three indicators were then weighted and aggregated with weights of 0.5, 0.3, and 0.2 to form a temporary dynamic threshold parameter. Verification showed that this temporary dynamic threshold can respond to sudden fluctuations within the first two minutes of a new operating condition, avoiding missed detections caused by fixed thresholds, and preventing false alarms due to short-term anomalies, significantly improving the stability and adaptability of early warning judgment. S5.4: Based on the temporary dynamic threshold parameter and the original static early warning threshold, perform fusion calculation to output the final dynamic calibration threshold signal that adapts to the current multi-timescale trend prediction requirements; Based on the temporary dynamic threshold parameters output in the previous step, combined with the pre-set original static warning threshold, the threshold fusion operation module is called to perform multi-parameter consistency verification, verifying the computability of the three in the numerical domain, and ensuring that there are no null or illegal values ​​that cause fusion failure. The verified temporary dynamic threshold parameters and the original static early warning threshold are input into the weighted fusion calculation unit. Linear or nonlinear combination processing is performed according to the preset weight coefficient matrix. The generation of the weight coefficient matrix refers to the working condition type judgment flag and the fluctuation characteristics of health indicators to ensure that the fusion result can take into account both short-term abnormal sensitivity and long-term trend stability. A multi-scale adapter is used to map the threshold of the fused output to a time scale. For the case where the trend prediction module has both short-term and long-term windows, the short-term adaptation threshold and the long-term adaptation threshold are calculated separately, and the threshold of the corresponding window is selected for comparison in the trend probability evaluation stage. The dynamic threshold signal generator is called to encapsulate the multi-scale adaptation threshold, add timestamp, working condition segment identifier and fusion method identifier, and output the final dynamic calibration threshold signal as the input of the subsequent multi-level fusion criterion module. By using multi-parameter weighted fusion and time-scale mapping, the temporary dynamic threshold parameter or threshold offset coefficient variable from the previous step is transformed into a dynamic calibration threshold signal that meets the adaptation requirements of multi-time-scale trend prediction, thereby achieving the expected technical effect of threshold adaptation under dynamic working conditions. For example, in the leveling stage of motorcycle helmet painting, the original static warning threshold is set as follows: The temporary dynamic threshold parameter is During the fusion process, the weights of short-term anomalies in the weight coefficient matrix are set to... Long-term stable weights are set to Execute the short-term fusion formula: ,in The short-term fusion threshold is calculated as follows: 2. Execute the long-term fusion formula: ,in The long-term fusion threshold is calculated as follows: V and V' are mapped to a short-term trend prediction window (5 production cycles) and a long-term trend prediction window (20 production cycles), respectively, and encapsulated as a dynamically calibrated threshold signal. Verification results show that, under this operating condition, the trend probability within the short-term window can be calibrated to the threshold. The system accurately identifies sudden fluctuations, while the long-term window is at the threshold. This allows for stable tracking of gradual equipment degradation, significantly improving the adaptability and robustness of the early warning system.

[0014] Step S6: If the semantic drift distance is less than a preset semantic drift threshold, it is determined to be a known stable operating condition. The geodesic angle between the real-time operating condition semantic fingerprint vector and the matched healthy fingerprint sample on a unit sphere is calculated, and the geodesic angle is mapped to a threshold offset coefficient of a nonlinear slowly varying function to drive the original threshold to be finely adjusted in a more relaxed direction to generate a dynamic calibration threshold. Specifically, this includes: S6.1: If the semantic drift distance is less than the preset semantic drift threshold, it is determined to be a stable working condition. The real-time working condition semantic fingerprint vector and the matched health fingerprint sample are subjected to unit spherical projection verification processing to confirm that both vectors satisfy the constraint that the L2 norm is always 1, and a standardized fingerprint vector pair with unit spherical geometric characteristics is obtained, which provides a compliant input basis for subsequent distance measurement based on spherical geometry. S6.2: Perform vector dot product operation on the standardized fingerprint vector pair to obtain the cosine similarity value that represents the degree of linear correlation between the two vectors in the multidimensional space. Transform the Euclidean distance relationship in the high-dimensional space into the cosine value index in the angle domain as an intermediate transition parameter for calculating the spherical angle. Based on the standardized fingerprint vector pair, the high-precision linear algebra operation module is called to perform component-level product operation, and the i-th component of the real-time working condition semantic fingerprint vector is multiplied with the i-th component of the healthy fingerprint sample to form a component product sequence. The component product sequence is input into a summation operator, and all component product values ​​are summed to obtain the numerical representation of the vector dot product. A normalization verification mechanism is used to verify the consistency between the dot product value and the unit sphere constraint condition, ensuring that the dot product value is within the compliant range of [-1, 1] under the premise that the magnitude of both is always equal to 1; Calculate cosine similarity using the following formula: in, This represents the i-th component of the real-time operating condition semantic fingerprint vector. This represents the i-th dimension component of the matched healthy fingerprint sample vector. Represents the dimension of a vector. Cosine similarity index; This step transforms the Euclidean distance relationship in high-dimensional space into a cosine value index in the angle domain, providing a transition parameter for the subsequent inverse cosine mapping of geodesic angles, and realizing a differentiable transformation from linear correlation to angle quantization. By using vector dot product and cosine value solving, the result of the previous step is transformed into a quantitative index that represents the degree of linear correlation between the semantic vectors of two working conditions, thereby achieving a standardized assessment of the similarity of working conditions. For example, in the condition monitoring of the painting and leveling section of a motorcycle helmet production line, the parameters of the real-time condition semantic fingerprint vector v are configured as [0.32, 0.41, 0.28, 0.12, 0.15, 0.22, 0.13, 0.19, 0.26], and the parameters of the matched health fingerprint sample w are configured as [0.30, 0.39, 0.27, 0.14, 0.16, 0.21, 0.11, 0.20, 0.25]. Both satisfy the unit spherical constraint condition with an L2 norm of 1. Performing component product operation yields the product sequence [0.096, 0.1599, 0.0756, 0.0168, 0.024, 0.0462, 0.0143, 0.038, 0.065], and summing them yields the dot product result of 0.5368. In the formula, n=9. Substitute this into the above sequence to calculate the cosine similarity. The cosine similarity falls within the [-1,1] compliant range, indicating that the real-time operating conditions and the health baseline have a moderately high correlation in the multidimensional spherical space. The application effect is that in subsequent geodesic angle calculations, the angle value is small, the threshold offset coefficient is low, and thus the dynamic calibration threshold only undergoes a slight, lenient adjustment, effectively avoiding false alarms caused by transient interference. S6.3: The cosine similarity value is nonlinearly mapped using the inverse cosine function to calculate the geodesic angle between the real-time working condition semantic fingerprint vector and the matched healthy fingerprint sample on a unit sphere, thereby obtaining an angular quantification index that accurately represents the amplitude of the working condition semantic drift and eliminating the distortion error of linear distance measurement in spherical space. S6.4: Perform nonlinear slowly varying function transformation processing based on the geodesic angle to construct a threshold offset coefficient that increases smoothly with the increase of the angle, obtain a weight adjustment factor that can reflect the degree of deviation of the working condition and avoid sudden interference, and realize a continuous differentiable mapping from the geometric angle space to the threshold adjustment range. S6.5: Perform gradient fine-tuning on the original static early warning threshold based on the threshold offset coefficient to generate a dynamic calibration threshold that adapts to the current real-time operating conditions, obtain a final judgment benchmark that combines long-term trend stability and short-term anomaly sensitivity, and complete the online self-calibration closed loop of the threshold parameter. The input data includes the threshold offset coefficient variable output by step S6.4 and the original static warning threshold preset by the system, and both are within the standardized health indicator space, satisfying the requirement of numerical domain consistency. The threshold offset coefficient is used as a control factor for the adjustment range. Gradient fine-tuning is performed based on the current numerical benchmark of the original static early warning threshold to smoothly change the judgment benchmark and avoid false alarms caused by sudden offset. A gradient-based fine-tuning computational model is constructed, in which the adjustment step size parameter is set. This parameter is jointly determined by the degree of semantic drift of real-time operating conditions, the confidence weight of healthy operating conditions in the baseline library, and the prediction stability target of multiple time scales, forming a weighting factor for threshold update. A control law combining linear superposition and exponential gradual variation is adopted to transform the threshold offset coefficient into gradient weight through a gradual variation function, and to perform stepwise increase and decrease on the original static threshold to ensure that the change of the threshold curve has continuous differentiability. The transition from the current static value to the dynamic value is achieved through a threshold update iterative formula: in, The original static early warning threshold, This is the threshold offset coefficient. To adjust the step size, This is a smoothing factor. Adjusted by... The steepness of the threshold change curve is controlled to meet the smoothness requirements under different production cycles; After the iteration is completed, the output dynamic calibration threshold is passed to the subsequent early warning judgment layer, and together with the model trend probability, it participates in the state consistency verification of the multi-level fusion criterion. Through dynamic adaptation of the threshold, the immediate capture of short-term anomalies and the stable tracking of long-term trends are achieved. By using gradient fine-tuning and smooth control, the geodesic angle mapping result from the previous step is transformed into a dynamic threshold adjustment factor, realizing online self-calibration of the threshold parameter and obtaining a final judgment benchmark that combines long-term trend stability and short-term anomaly sensitivity. For example, a health baseline threshold for injection molding holding pressure section equipment. At the angle between the geodesics is When the value is in radians, the threshold offset coefficient is obtained by mapping through a slowly varying function. Set adjustment step size Smoothing factor According to the formula, the fine-tuning increment is... The dynamic calibration threshold output is In a real-time production environment, this threshold is used to determine the trend probability of the spray leveling stage. When the trend probability of three consecutive cycle periods is higher than this dynamic threshold, the system successfully triggers a first-level warning. No false alarms caused by instantaneous disturbances occurred in the measured data, which verifies the robustness improvement effect of the threshold self-calibration closed loop in multi-timescale prediction.

[0015] Step S7: The dynamic calibration threshold and real-time trend probability are input to the multi-level fusion criterion module. A time-sensitive criterion window is set according to the helmet production cycle. A first-level warning signal is triggered only when the trend probability is continuously higher than the dynamic calibration threshold for multiple consecutive cycle periods. Specifically, this includes: S7.1: Obtain the input data of the multi-level fusion criterion module, including the dynamic calibration threshold and real-time trend probability generated in the previous step. The real-time trend probability is specifically calculated as follows: the ratio of semantic drift distance to the preset semantic drift threshold is taken as the first component, and the ratio of geodesic angle to π / 2 is taken as the second component; then, the first component and the second component are assigned preset weight coefficients (e.g., 0.5 each) and weighted and summed to obtain a fusion value between 0 and 1. This fusion value is the real-time health trend probability. The closer the value is to 1, the higher the degree of deviation of the current working condition from the health benchmark and the more significant the fault trend. Based on the single-piece production cycle parameters of the helmet production line, the time axis is discretized and sliced ​​to generate a time-sensitive criterion window sequence that is strictly aligned with the physical production rhythm, so as to establish the basic time unit for subsequent state assessment. S7.2: Based on the time-sensitive criterion window sequence, a sliding window mapping operation is performed on the real-time trend probability within each independent time-sensitive criterion window to transform the continuous probability stream data into a discrete probability sample set arranged according to the production cycle. By comparing the magnitude relationship between the values ​​in each discrete probability sample set and the dynamic calibration threshold, a cycle-level over-limit judgment result sequence containing Boolean flag states is generated. The real-time trend probability data of each independent criterion window in the time-sensitive criterion window sequence is subjected to continuous sampling index mapping processing. The original trend probability flow in the criterion window is discretized and sliced ​​according to the time interval set by the production cycle parameter to form a probability sampling point sequence that strictly corresponds to the physical production cycle. A sliding window transformation is performed on the probability sampling point sequence within the criterion window. Based on the preset sliding step size and window length, multiple local probability subsequences are generated on the time axis. Each subsequence covers the trend probability record within a complete production cycle. For each probability sample value in a local probability subsequence, the dynamic calibration threshold parameter is called to perform a numerical comparison operation. Samples greater than the threshold are marked as out of limit and samples less than or equal to the threshold are marked as not out of limit, and Boolean flag bits corresponding to the sample positions are obtained. The Boolean flag bits in each local probability subsequence are concatenated in chronological order to generate a beat-level overlimit judgment vector covering the entire criterion window period, ensuring that each beat position corresponds to a unique overlimit judgment result; The beat-level overlimit judgment vectors of each criterion window period are summarized in sequence to form a complete serialized beat-level overlimit judgment result sequence, which provides the basic input for subsequent time series coherence analysis and continuous period counting logic; By using the above sliding window mapping and sample-by-sample threshold comparison processing method, the time-sensitive criterion window sequence result of the previous step is transformed into a beat-level judgment Boolean sequence that covers the entire cycle and can be directly used for judging the persistence of exceeding limits, thereby realizing the accurate quantification of the trend probability of exceeding limits at the production beat scale. For example, in a motorcycle helmet production line, the single-piece production cycle time is set to 98 seconds, the time-sensitive criterion window length is set to 3 cycle times, the sliding step is 10 seconds, and the window length is 98 seconds. The real-time trend probability flow within a certain criterion window is discretized to obtain 9 sampling points per cycle. The predicted probability value of each sampling point is compared with a dynamic calibration threshold of 0.72. Within one criterion window, the sampling point values ​​are 0.75, 0.78, 0.70, 0.73, 0.76, 0.68, 0.80, 0.77, and 0.69. The formula is used... in, These are trend probability sample values. To dynamically calibrate the threshold, output a Boolean flag. The above sampling points were compared to obtain the Boolean label sequence 1, 1, 0, 1, 1, 0, 1, 1, 0, corresponding to the beat-level over-limit judgment results. The Boolean sequences within multiple criterion windows were concatenated in chronological order to form a beat-level over-limit judgment result sequence with a total length of 27. This sequence clearly quantifies whether the trend probability within each beat cycle is in an over-limit state, providing direct input for subsequent continuous cycle counting analysis. In testing, the Boolean judgment of this method showed significantly improved consistency with manual acceptance judgment, avoiding false alarms caused by short-term probability fluctuations. S7.3: Using the preset continuous cycle counting logic, perform temporal coherence analysis on the beat-level over-limit judgment result sequence, count the number of beat-level over-limit judgment results continuously marked as over-limit state, and compare the counted number with the preset minimum number of continuous triggering cycles of the system to generate a time dimension conformity verification flag characterizing the persistence of the fault, so as to distinguish between occasional fluctuations and real degradation trends. The sequence of over-limit judgment results at the beat level is used as the input object, and a continuous period counting buffer is established to store the Boolean flag state in each beat cycle. A step-by-step indexing scan is performed on the contents of the buffer. When a period marked as out of bounds is encountered, the continuous counter value is incremented by one unit, and when a non-out-of-bounds state is encountered, the counter value is reset to zero, thus forming a mapping relationship between continuous count values ​​and periodic sequences. Apply the maximum value extraction operation to each consecutive count value in the above mapping relationship to obtain the longest consecutive over-limit period length index within the current analysis window; The longest consecutive over-limit period length index is compared with the system's preset minimum consecutive trigger period number, and a time dimension compliance verification flag is generated using the following formula: in For compliance verification, The longest consecutive over-limit period length. This is the preset minimum number of consecutive trigger cycles; When the comparison result meets the threshold condition, the compliance check flag is set to the valid state; otherwise, it is set to the invalid state. By using time-series coherence analysis and cycle length comparison processing, the sequence of beat-level overlimit judgment results from the previous step is transformed into a time-dimensional conformity verification flag that can be used to distinguish between occasional fluctuations and real degradation trends, thereby achieving continuous verification before the high-confidence warning is triggered. For example, in a motorcycle helmet production line, the sampling period is set to 98 seconds, and the analysis window length is set to 600 seconds. The sequence of beat-level over-limit judgment results {1,1,0,1,1,1,0,1} within a certain operating segment is subjected to continuous counting buffering processing to obtain the continuous count value sequence {1,2,0,1,2,3,0,1}. The longest continuous over-limit cycle length L is 3. The preset minimum number of continuous trigger cycles T is 4, as shown in the formula. If the calculated value F is invalid, the system determines that the operating state of this segment is an occasional fluctuation. If the continuous over-limit marker sequence of another operating segment is {1,1,1,1,1}, the continuous count value sequence is {1,2,3,4,5}, and the longest continuous over-limit cycle length L is 5, the calculation formula outputs F as a valid state. The system determines that the operating state of this segment is a real degradation trend and meets the high confidence triggering condition. The final output time dimension compliance verification flag enters the early warning triggering process after verification, realizing the accurate capture of production line anomalies. S7.4: Based on the logical state of the time dimension compliance verification flag, when the number of consecutively marked over-limit judgment results reaches or exceeds the minimum number of consecutive trigger cycles, the early warning triggering mechanism is activated to generate a first-level early warning signal with high confidence, and the first-level early warning signal is output to the equipment status control execution unit to complete the final decision conversion from probability assessment to deterministic alarm.

[0016] Step S8: Based on the first-level early warning signal, execute equipment status control actions, and simultaneously record the mapping relationship between the current operating condition semantic fingerprint vector and the threshold offset coefficient to update the confidence weight of the dual-modal mapping baseline library of the health operating condition fingerprint and health indicators. Specifically, this includes: S8.1: Obtain the first-level early warning signal and the corresponding real-time equipment operating status parameters output by the multi-level fusion criterion module, and execute status control actions such as shutdown protection, speed reduction operation or process parameter rollback based on the preset equipment safety control strategy library, so as to generate equipment status control execution records containing action type identifiers and execution timestamps; S8.2: Extract the real-time operating condition semantic fingerprint vector generated at the time of this warning trigger and the calculated threshold offset coefficient. Use the real-time operating condition semantic fingerprint vector as the index key and the threshold offset coefficient as the association value to construct a single mapping relationship data pair to form a current operating condition feature mapping unit for baseline library update. S8.3: Call the dual-modal mapping baseline library of health condition fingerprint and health indicators, retrieve the historical health fingerprint entry with the smallest Euclidean distance to the real-time condition semantic fingerprint vector in the current condition feature mapping unit, so as to locate the target baseline data record to be updated and obtain its currently stored original confidence weight value. S8.4: Based on the threshold offset coefficient in the current working condition feature mapping unit and the severity level of the action in the equipment status control execution record, a weight decay factor calculation model is constructed, and the original confidence weight value of the target baseline data record is weighted and corrected to generate an updated confidence weight value that reflects the latest working condition deviation. S8.5: Replace the original confidence weight value in the target baseline data record with the updated confidence weight value, and write the current working condition feature mapping unit as a new sample into the dual-modal mapping baseline library of health working condition fingerprint and health indicators according to the equipment unit identifier, so as to complete the online iteration and adaptive evolution of the baseline library knowledge. The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0017] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for fault early warning analysis of motorcycle helmet production line equipment, characterized in that, Includes the following steps: S1: Acquire multi-source sensor data of motorcycle helmet production line equipment under different production conditions; S2: Construct a working condition semantic fingerprint encoder based on the multi-source sensor data, and use the working condition semantic fingerprint encoder to perform feature dimensionality reduction processing on the multi-source sensor data to generate a working condition semantic fingerprint vector. S3: Based on the historical operating data of the preset alarm-free period, extract the corresponding operating condition semantic fingerprint vector and its associated equipment health indicators, and construct a dual-modal mapping baseline of health operating condition fingerprint and health indicators. S4: Real-time acquisition of sensor data within the current production cycle and generation of real-time working condition semantic fingerprint vector; retrieval of the nearest health fingerprint sample in the dual-modal mapping baseline library of health condition fingerprint and health indicators to calculate semantic drift distance. S5: If the semantic drift distance is greater than the preset semantic drift threshold, it is determined to be a new working condition. A temporary dynamic threshold is generated based on the recent health indicator sliding window. The temporary dynamic threshold parameter is fused with the original static warning threshold to output a dynamic calibration threshold. S6: If the semantic drift distance is less than the preset semantic drift threshold, it is determined to be a stable working condition. The geodesic angle between the real-time working condition semantic fingerprint vector and the matched healthy fingerprint sample on the unit sphere is calculated, and the geodesic angle is mapped to a threshold offset coefficient. The original static warning threshold is fine-tuned based on the threshold offset coefficient to generate the dynamic calibration threshold.

2. The method for fault early warning analysis of motorcycle helmet production line equipment according to claim 1, characterized in that, Following step S6, the following is also included: S7: Based on the dynamic calibration threshold and the real-time trend probability, a time-sensitive criterion window is set according to the helmet production cycle. When the real-time trend probability is continuously higher than the dynamic calibration threshold in multiple consecutive cycle periods, an early warning signal is triggered. S8: Execute equipment status control actions based on the warning signal, and at the same time record the mapping relationship between the semantic fingerprint vector of the current working condition and the threshold offset coefficient to update the confidence weight of the dual-modal mapping baseline library of the health working condition fingerprint and health indicators.

3. The method for fault early warning analysis of motorcycle helmet production line equipment according to claim 1, characterized in that, The multi-source sensor data includes hydraulic pressure fluctuation sequence, mold temperature gradient curve, servo motor current harmonic spectrum, ambient temperature and humidity time sequence, and visual inspection yield feedback signal.

4. The method for fault early warning analysis of motorcycle helmet production line equipment according to claim 1, characterized in that, The feature dimensionality reduction process is performed using three learnable temporal convolutional layers. These three layers extract high-frequency transient, mid-frequency periodic, and global trend features, respectively. The convolutional layers are followed by global average pooling and parameter matrix mapping. Finally, the nine-dimensional working condition semantic fingerprint is normalized to fall on a unit spherical manifold through L2 norm normalization.

5. The method for fault early warning analysis of motorcycle helmet production line equipment according to claim 1, characterized in that, Step S3 specifically includes: Time window slicing was performed on the multi-source sensor data during the 72-hour period without alarms to extract independent working condition data segments corresponding to the injection molding pressure holding section, vacuum adsorption section, paint spraying leveling section and UV curing section, and obtain the original data sequence of typical working conditions. Based on the original data sequence of the typical working conditions, the feature mapping operation is performed on the constructed working condition semantic fingerprint encoder to generate the working condition semantic fingerprint vector, thereby obtaining the historical working condition semantic fingerprint set. Statistical feature extraction is performed on the original data sequence of the typical operating conditions to calculate equipment health indicators, thereby obtaining a set of historical equipment health indicators; Based on the association between the historical operating condition semantic fingerprint set and the historical equipment health indicator set, a weighted fusion calculation is performed by combining the operating condition duration, data integrity score and cross-sensor consistency score to generate a confidence weight value, thereby obtaining a dual-modal data pair of health operating condition fingerprint and health indicator with confidence weight attribute. The dual-modal data pairs of health condition fingerprints and health indicators with confidence weight attributes are classified and stored according to the device unit identifier to construct a dual-modal mapping baseline library of health condition fingerprints and health indicators.

6. The method for fault early warning analysis of motorcycle helmet production line equipment according to claim 5, characterized in that, The dual-modal mapping baseline library of health condition fingerprints and health indicators is stored according to the device unit identifier. Multi-sample nine-dimensional fingerprints can be quickly retrieved through hash index and support historical data time-series backtracking and incremental writing of new data.

7. The method for fault early warning analysis of motorcycle helmet production line equipment according to claim 1, characterized in that, Step S4 specifically includes: Acquire real-time multi-source sensor data within the current production cycle, and use a time sliding window mechanism to perform synchronization alignment and noise suppression processing on the real-time multi-source sensor data to generate a standardized real-time multi-source sensor data matrix. Based on the standardized real-time multi-source sensor data matrix, it is input into the constructed working condition semantic fingerprint encoder. The three learnable temporal convolutional layers inside the working condition semantic fingerprint encoder are used to perform multi-level feature extraction and nonlinear transformation processing on the standardized real-time multi-source sensor data matrix, and output the intermediate layer feature tensor. Principal component dimensionality reduction projection operation is performed on the intermediate layer feature tensor, the first three-dimensional principal component components are extracted and spliced ​​to form a nine-dimensional original feature vector, and L2 norm normalization constraint operation is performed on the nine-dimensional original feature vector to generate a real-time working condition semantic fingerprint vector. The constructed dual-modal mapping baseline library of health condition fingerprints and health indicators is invoked. The real-time working condition semantic fingerprint vector is used as the query index. Euclidean distance metric calculation is performed in the dual-modal mapping baseline library of health condition fingerprints and health indicators to filter out the three sets of health fingerprint samples that are closest to the real-time working condition semantic fingerprint vector in space. Based on each healthy fingerprint sample in the three healthy fingerprint sample sets and the real-time working condition semantic fingerprint vector, the Euclidean distance between each pair is calculated, and the minimum value of the Euclidean distance is calculated to generate a semantic drift distance scalar.

8. The method for fault early warning analysis of motorcycle helmet production line equipment according to claim 7, characterized in that, The semantic drift distance scalar represents the degree to which the current operating state deviates from the health benchmark.

9. The method for fault early warning analysis of motorcycle helmet production line equipment according to claim 1, characterized in that, Step S5 specifically includes: The semantic drift distance is calculated by performing Euclidean distance calculation on the real-time working condition semantic fingerprint vector and three health fingerprint samples retrieved from the dual-modal mapping baseline library of health working condition fingerprint and health indicators to obtain the semantic drift distance value. Based on the semantic drift distance value and the preset semantic drift threshold, a size comparison and judgment process is performed to generate a condition type determination flag indicating whether the current production scenario belongs to a stable condition or a new condition. If the condition type determination flag indicates an unknown new condition, then perform rolling quantile statistical analysis on the equipment health index sliding window data within the last 30 minutes to extract temporary dynamic threshold parameters. The temporary dynamic threshold parameter is fused with the original static early warning threshold to output the final dynamic calibration threshold that adapts to the current multi-timescale trend prediction requirements.

10. The method for fault early warning analysis of motorcycle helmet production line equipment according to claim 1, characterized in that, Step S6 also includes: outputting a threshold offset coefficient based on a nonlinear slowly varying function of the geodesic angle, updating the original static early warning threshold in a gradient manner, and finally generating a dynamic calibration threshold that adapts to the current real-time operating conditions.

Images