A Needle Report Data Analysis System Based on Artificial Intelligence Algorithms

Abstract

This invention discloses a machine needle report data analysis system based on artificial intelligence algorithms, specifically relating to the field of intelligent management and data processing technology in industrial production. It includes modules for data acquisition, edge preprocessing, cloud analysis, and decision feedback: Standardized raw labeled data is acquired through multi-sensor collaborative acquisition and synchronous puncture phase locking; features are refined using hierarchical preprocessing and attention gating algorithms to generate lightweight reports; cloud-based fusion of physical priors and AI algorithms enables lifespan prediction and root cause localization, generating multi-dimensional reports and mining association rules through improved KNN and Apriori algorithms; and process adjustment instructions and maintenance work orders are generated based on scene-adaptive federated incremental learning, with data feedback enabling model iteration. This invention solves problems such as cross-fabric adaptation and ambiguous anomaly recognition, improving needle quality and production efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent management and data processing technology in industrial production, and more specifically, to a needle report data analysis system based on artificial intelligence algorithms. Background Technology

[0002] In the sewing equipment manufacturing industry, the sewing needle, as a core actuator, directly affects stitch quality and production efficiency. Related monitoring and data analysis technologies have been gradually applied to industrial production scenarios. Currently, mainstream needle condition monitoring methods in the industry mostly use a single sensor to collect operational data and determine obvious faults such as needle breakage by setting fixed thresholds. Some advanced solutions combine multiple sensors to collect data on vibration, current, and tension, employing simple time-domain or frequency-domain feature extraction methods in conjunction with traditional machine learning algorithms to achieve basic fault early warning. Simultaneously, statistical reports containing equipment operating parameters, production batches, and quality inspection results are generated during production for post-production review and data recording. Furthermore, lifespan prediction technologies for general industrial equipment are mostly based on training models using single-condition data, and report data analysis focuses on statistical summarization of production data. These technologies have already achieved mature applications in the general machinery manufacturing field.

[0003] However, in practical use, it still has some shortcomings, such as the lack of phase-level synchronization mechanism in multimodal data acquisition, the lack of precise correlation between needle puncture action and sensor signal, the fuzzy extraction of abnormal features due to transmission system interference, and low data signal-to-noise ratio; the life prediction model does not consider the differences in parameters such as fabric elastic modulus and thickness, resulting in poor adaptability across fabric working conditions and large prediction errors. The report data is only used as a statistical record and is not deeply integrated with AI algorithms, lacking abnormal pattern recognition and cross-report association rule mining, and cannot provide effective support for fault root cause localization and process optimization; root cause localization relies heavily on single-type data and does not integrate multimodal evidence chains such as sensors, images, and processes, resulting in low localization accuracy and long time consumption; the technical solution lacks a closed-loop mechanism of "data acquisition-analysis-reporting-decision-iteration", and the model cannot be continuously optimized based on actual production feedback, making it difficult to adapt to dynamically changing production scenarios. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a machine needle report data analysis system based on artificial intelligence algorithms, which solves the problems mentioned in the background art through the following solutions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a machine needle report data analysis system based on artificial intelligence algorithms, comprising:

[0006] Data acquisition module: Based on multi-sensor collaborative acquisition and motor encoder phase-locked synchronization mechanism, the multi-modal data of needle operation is filtered, denoised and labeled into reports to obtain standardized raw labeled data and phase reference signal;

[0007] Edge preprocessing module: Based on hierarchical preprocessing and puncture window periodic sliding attention gating algorithm, feature purification and anomaly detection are performed on the standardized original labeled data and the phase reference signal to obtain preprocessed feature data, anomaly warning level, and generate lightweight preprocessing summary;

[0008] The cloud-based analysis module, based on a physical prior fusion adversarial feature selection network, a cross-fabric transfer learning Gaussian process regression model, and a Bayesian network, performs feature optimization, lifetime prediction, and root cause localization on the preprocessed feature data and the lightweight preprocessed report. This yields optimized fusion features, remaining lifetime predictions, and root cause localization conclusions, and generates a multi-dimensional needle analysis report. Furthermore, an improved KNN algorithm and Apriori algorithm are used to perform anomaly pattern recognition and association rule mining on the structured data of the multi-dimensional needle analysis report, resulting in anomaly entries, pattern labels, and association rule mining results.

[0009] Decision Feedback Module: Based on the scenario-adaptive federated incremental learning mechanism and the quantitative parameter optimization model, the module performs parameter adjustment calculations and generates work orders based on the multi-dimensional needle analysis report and the association rule mining results, thereby obtaining process adjustment instructions and maintenance work orders. After annotating the execution results of the maintenance work orders, the module feeds them back to the cloud analysis module as training samples for model iteration.

[0010] The technical effects and advantages of this invention are as follows:

[0011] 1. Improved accuracy of multimodal data acquisition, overcoming transmission interference and synchronization challenges: Through the deployment of "three-point coordination" sensors, puncture phase locking synchronization mechanism and segmented adaptive gain adjustment algorithm, phase-level accurate acquisition of needle movement data is achieved;

[0012] 2. Improved accuracy of cross-fabric working condition prediction and enhanced adaptability of life prediction: The cloud analysis module integrates the physical prior of the needle-fabric coupling and the adversarial feature selection network, and is combined with the cross-fabric transfer learning Gaussian process regression model. Through working condition clustering and dynamic weight transfer, it adapts to 6 mainstream fabric scenarios, solving the problems of low prediction accuracy and poor working condition adaptability caused by ignoring the differences in fabric parameters.

[0013] 3. Efficient multimodal root cause localization for accurate and rapid source tracing: Integrating sensor, image, and process multimodal evidence chains, and through credibility-weighted fusion and 4-level Bayesian network reasoning, hierarchical root cause identification of broken needles and pin defects is achieved;

[0014] 4. Highly efficient edge-cloud collaboration, balancing real-time response and in-depth analysis: The edge preprocessing module rapidly refines core features and provides tiered early warnings through a dedicated attention gating algorithm for the puncture window, with an abnormal command response latency of <10ms. The three-level abnormal linkage mechanism enables precise matching of warnings, speed reduction, and emergency shutdown. The cloud module focuses on in-depth analysis and report generation. Edge-cloud collaboration not only ensures real-time control needs on-site but also enables in-depth data value mining, avoiding the drawbacks of traditional single architectures such as "slow response" or "shallow analysis."

[0015] 5. Intelligent report data analysis to activate the value of data-driven decision-making: Construct a full-process system of "report annotation - multi-dimensional report generation - AI analysis - correlation mining". By improving the KNN algorithm to identify abnormal patterns in reports and the Apriori algorithm to mine cross-report correlation rules, it provides direct basis for root cause location and process optimization, and solves the defects of low utilization rate and lack of in-depth analysis value of existing report data. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0017] Figure 2 This is a schematic diagram of the data acquisition and edge preprocessing structure of the present invention.

[0018] Figure 3 This is a schematic diagram of the cloud analysis module structure of the present invention.

[0019] Figure 4 This is a schematic diagram of the decision feedback and closed-loop optimization structure of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] refer to Figures 1-4 The illustrated needle report data analysis system includes: a data acquisition module, an edge preprocessing module, a cloud analysis module, and a decision feedback module.

[0022] The data acquisition module achieves high-precision acquisition of machine needle operation status data through customized multi-source industrial IoT sensing units and signal synchronization and gain adjustment mechanisms. Specific implementation details are as follows:

[0023] S101. Customized Sensor Deployment Solution: Employing a "three-point collaborative" acquisition layout to achieve accurate signal acquisition.

[0024] The vibration sensor is attached to the upper end of the needle bar, 20mm away from the needle tip, to match the high-frequency vibration transmission path of the needle. A PCB356A16 accelerometer sensor with a range of 0-10g and a sampling frequency of 5kHz is selected.

[0025] The acoustic sensor is deployed below the needle plate, directly facing the needle puncture area. It uses an SGM3770 microphone sensor with a sampling frequency of 1kHz and a pickup range of 20Hz-10kHz.

[0026] The fabric tension sensor is integrated into the core of the feed roller. It is an HBMU9B type tension sensor with a range of 0-50N and a sampling frequency of 1kHz.

[0027] The motor stator current sensor is connected in series in the motor power supply circuit. A LEMLA25-NP Hall effect sensor with a sampling frequency of 2kHz is selected. Each sensor is connected to the edge gateway via an industrial Ethernet PoE interface, with a transmission delay of ≤3ms.

[0028] S102. Puncture Phase Locking Synchronization Mechanism: Based on the physical phase of the needle puncture action, a phase-level synchronization system for multimodal data is constructed.

[0029] Phase reference signal generation: An infrared photoelectric sensor is installed on the side of the needle plate. When the needle tip is flush with the upper edge of the needle plate hole, the sensor outputs a high-level trigger signal with a threshold voltage of 2.5V. This moment is defined as the phase zero point. needle movement cycle The internal structure is divided into pre-puncture stages. to Pre-puncture stage, to The puncture stage to During the recovery phase, a "dual-phase triggered clock" is formed.

[0030] Multimodal data delay calibration: By analyzing the time difference between each sensor signal and the phase reference signal, the acquired data is aligned along the time axis to ensure consistency of different modal data at the same physical action node. Based on adaptive matching of the needle movement cycle to the puncture core window, the core formula for calculating the core window time is as follows:

[0031] ;

[0032] in, For the duration of the core puncture window, The cycle of needle movement ( This mechanism enables multimodal data spatiotemporal synchronization errors to be less than or equal to 5ms, solving the problem of vibration interference in traditional timestamp synchronization.

[0033] S103. Segmented Adaptive Gain Adjustment Algorithm: Designing scenario-based gain adjustment logic for the abnormally sensitive frequency band of 1500-3500Hz for machine needles. Core implementation steps:

[0034] Frequency band division: The vibration spectrum of the needle is divided into three frequency bands: low frequency, sensitive frequency, and high frequency.

[0035] The low-frequency band is 50-1500Hz, dominated by transmission interference;

[0036] The sensitive frequency band is 1500-3500Hz, dominated by abnormal characteristics of the needle.

[0037] The high-frequency band is 3500-5000Hz, dominated by noise.

[0038] Gain coefficient calculation: The gain is dynamically adjusted based on the signal energy proportion of each frequency band. The core calculation formula is as follows:

[0039] ;

[0040] in, Base gain (default is 1). The signal energy in the current frequency band. The total energy of the signal across the entire frequency band. This is the weighting coefficient for sensitive frequency bands (value 3-5). This is the offset coefficient (value 0.2). Low-frequency band. Take 1, high frequency band By setting the value to 0.5, this formula enables targeted gain enhancement in sensitive frequency bands and noise suppression in non-sensitive frequency bands, ultimately improving the signal-to-noise ratio by 40%.

[0041] S104. Construction of Multidimensional Raw Dataset: The collected data includes machine needle running status data and related attribute data:

[0042] The status data includes amplitude, frequency, and phase characteristics of vibration spectrum sampled at 5kHz; current waveform including RMS value, peak value, and harmonic content; acoustic signal including sound pressure level and spectral peak value; and tension signal including real-time tension value and fluctuation variance. The associated attribute data includes equipment model, needle specifications including diameter and material; fabric parameters including elastic modulus and thickness; and process parameters including stitch density and machine speed.

[0043] By standardizing the data format, all kinds of signals are converted into JSON format. The time axis is anchored by the phase reference signal, and a multi-dimensional needle running raw dataset with 128-dimensional features is constructed to provide high-quality input for subsequent preprocessing modules.

[0044] It should be further explained that the raw data is labeled in a report format: the collected vibration, current, acoustic, tension multimodal data and related attribute data such as equipment model and needle specifications are labeled in a structured manner according to the "report basic fields". The labeled fields include: equipment ID, collection timestamp, working condition label of fabric type and process parameters, sensor type, data amplitude / frequency, phase range, and data quality level determined based on signal-to-noise ratio, forming standardized raw data report entries, providing a unified data entry for subsequent report generation.

[0045] The edge preprocessing module aims to "accurately extract the core features of needle puncture, suppress transmission interference, and quickly link with equipment control." It integrates three mechanisms: adaptive preprocessing, dedicated attention gating detection for the puncture window, and quantitative anomaly linkage. This achieves efficient purification of needle operation data and accurate early anomaly identification. Specific implementation details are as follows:

[0046] S201. Hierarchical Preprocessing of Time Series Data: A hierarchical preprocessing workflow is designed to address trend noise and amplitude differences in multimodal synchronous data, balancing data purity and processing efficiency.

[0047] Detrending adaptive filtering: An adaptive trend extraction method based on local weighted regression is adopted. The trend component of the data is captured by a sliding window. After stripping away the trend noise, a stable signal is obtained. The size of the sliding window is adaptively adjusted according to the needle movement cycle to ensure accurate filtering of trend noise under different vehicle speed conditions.

[0048] Phase anchoring standardization: Based on the puncture phase reference signal, the time series data of each mode are segmented and standardized to avoid the local feature overload caused by traditional global standardization. The data is calibrated according to the statistical characteristics of the current puncture phase interval to ensure that the standardized data retains phase-related local anomalies.

[0049] S202. Puncture Window Dedicated Attention Gated Anomaly Detection Model: Focusing on the core action window of needle puncture, and combining a gating mechanism to suppress transmission interference, the model structure and implementation logic are as follows:

[0050] Puncture core window identification: based on the phase reference signal output by the data acquisition module, i.e. Define the zero point of puncture. to The core puncture window duration is dynamically matched according to the needle rotation speed to ensure accurate coverage of key puncture actions under different working conditions.

[0051] Construction of a periodic sliding attention mask: A sliding attention mask synchronized with the needle movement cycle is designed, assigning high weights only to data within the core window and low weights (weight coefficient ≤ 0.2) to data outside the core window. The core formula for calculating attention weights is:

[0052] ;

[0053] in, For time The corresponding phase attribution coefficients, within the core window Non-core windows , The weighting adjustment factor (value 5) ensures that the attention ratio of the core window data is greater than or equal to 80%.

[0054] Low-frequency vibration suppression gating unit: A gating unit is connected after the attention layer. A low-pass filter core is designed to suppress low-frequency interference (50-200Hz) in the transmission system. The core calculation formula for the gating output is:

[0055] ;

[0056] in, This is the gating adjustment coefficient. For standardized data, It is a low-pass filter function with a cutoff frequency of 250Hz, which achieves a synergistic effect of "core feature preservation + interference suppression" through gating adjustment.

[0057] Anomaly detection inference: The features output by the gate are input into a lightweight fully connected classifier, which outputs anomaly probability values. The core calculation formula is:

[0058] ;

[0059] in, Sigmoid activation function , For output layer parameters, when The abnormal warning process is triggered in a timely manner.

[0060] S203. Quantitative Anomaly Classification and Equipment Interaction: Establish a dual quantitative classification standard based on vibration amplitude and anomaly probability to achieve precise matching between anomaly severity and equipment control commands.

[0061] The definition of the third-level abnormality standard is as follows: Based on 100 sets of machine needle operation data with different degrees of abnormality, the grading threshold is determined through statistical analysis.

[0062] Level 1 abnormality (minor): Furthermore, the vibration amplitude exceeded the baseline by 20%–30%.

[0063] Level 2 abnormality (moderate): Furthermore, the vibration amplitude exceeded the baseline by 30%–50%.

[0064] Level 3 Abnormal (Severe): Furthermore, the vibration amplitude exceeds the baseline by more than 50%, where the baseline vibration amplitude is the average vibration amplitude of a normal needle under the same working conditions.

[0065] Equipment linkage command generation: Based on the anomaly level, corresponding control commands are generated and communicated with the sewing equipment controller via the Modbus-RTU protocol, with a command response latency of less than 10ms.

[0066] Level 1 Anomaly: Outputs a warning command, triggering an audible and visual alarm on the equipment's control panel, but does not affect operation;

[0067] Level 2 Anomaly: Output a speed reduction command to control the vehicle speed to decrease by 10%, and record the abnormal data at the same time;

[0068] Level 3 anomaly: Output emergency stop command, cut off motor drive power, and save multimodal data snapshot at the moment the anomaly occurs.

[0069] S204. Lightweight Report Generation of Preprocessing Results: After data detrending, standardization, gating filtering, and anomaly detection are completed at the edge, a report is automatically generated. The report includes core fields: Device ID, processing time window, operating condition information, key preprocessing indicators, local anomaly warning level, and data cache index. The report format uses a lightweight JSON structure, is stored locally through the edge gateway for 72 hours, and synchronized to the cloud for local operators to quickly view and integrate with cloud reports. Key indicators include signal-to-noise ratio improvement, anomaly probability value, and core window feature statistics.

[0070] S205. Preprocessed data output and caching: The time series feature data after detrending, standardization, and gating filtering is encapsulated in the format of "phase interval + feature type" and uploaded to the cloud analysis module through the MQTT protocol of the edge gateway.

[0071] Meanwhile, a circular cache pool is built at the edge with a cache capacity of 10 needle movement cycles. The cached data is used for backtracking analysis when an anomaly occurs, and when the cached data reaches a threshold, the old data is automatically overwritten to ensure efficient utilization of edge storage resources.

[0072] The cloud-based analysis module integrates three major mechanisms: multi-domain feature physical prior fusion, cross-condition transfer learning prediction, and multi-modal evidence chain root cause localization, to achieve in-depth analysis and decision support for the machine needle's operating status. Specific implementation details are as follows:

[0073] S301. Multi-domain Feature Hierarchical Extraction and Physical Prior Modeling: For time-series feature data uploaded from the edge, a two-layer feature system of "data-driven features + physical prior constraints" is constructed.

[0074] Multi-domain feature hierarchical extraction: Core features of each modality data are extracted from the time domain, frequency domain, and time-frequency domain respectively to form an initial feature set (128 dimensions):

[0075] Time-domain characteristics: Calculate the peak value, kurtosis, waveform factor, and fluctuation variance of vibration, current, and tension signals;

[0076] Frequency domain features: The time-domain signal is converted to the frequency domain through Fast Fourier Transform (FFT) to extract the spectral peak, spectral entropy, and center frequency;

[0077] Time-frequency domain features: Wavelet transform is used to extract the time-frequency entropy and wavelet energy of the signal, and to characterize the abrupt change features of the signal in different time-frequency windows.

[0078] Needle-Fabric Coupled Physical Prior Modeling: Constructing a quantitative coupled physical model to transform the interaction laws between the needle and the fabric into characteristic constraints: Core coupling equations:

[0079] ;

[0080] in, The force of the needle puncture. This is the coupling coefficient (values ​​range from 0.8 to 1.2). The elastic modulus of the fabric. The diameter of the needle. For fabric thickness;

[0081] Constructing a mapping matrix of "process parameters - physical characteristics" based on coupling equations The 16-dimensional physical prior parameters are mapped to the initial feature space to form a physical constraint feature set; the parameters contain... , , wait.

[0082] S302. Adversarial Feature Selection Network with Physics Prior Fusion;

[0083] To address redundant and noisy features in the initial feature set, physical prior constraints are introduced to optimize the adversarial feature selection process, thereby improving feature quality and model generalization ability.

[0084] Network structure design: The architecture adopts a three-module structure of "generator-discriminator-prior constraint". The generator is responsible for feature reconstruction, the discriminator distinguishes between high-quality and low-quality features, and the prior constraint introduces physical rules to constrain the feature distribution.

[0085] Prior fusion and loss function design: The physical constraint feature set and the initial feature set are weighted and fused before being input into the generator. The core fusion formula is:

[0086] ;

[0087] in, For the initial feature set, For physical constraint feature set, The fusion weight (initial value 0.6);

[0088] The adversarial training loss function combines adversarial loss and physical constraint loss to ensure that features are both discriminative and conform to physical laws.

[0089] Optimize feature output: Through network training convergence, with 100 training epochs and a learning rate of 0.001, the output dimension is reduced to 64-dimensional optimized fusion feature set, reducing feature redundancy by 60% and improving physical consistency by 45%.

[0090] S303. Cross-Fabric Working Condition Transfer Learning Gaussian Process Regression Life Prediction: To address the issue of poor model adaptability under different fabric working conditions, a fusion prediction model of "working condition clustering + weight transfer" is designed to improve the prediction accuracy of needle remaining service life (RUL).

[0091] Cross-fabric working condition clustering: The K-Means algorithm is used to cluster historical data by fabric type. The clustering index is "puncture force variance - vibration spectrum similarity". The clustering is determined by calculating the similarity of vibration spectrum under different fabric working conditions. The number of clusters K=6, covering mainstream fabrics such as cotton, chemical fiber, and leather.

[0092] Transfer learning Gaussian process regression model:

[0093] Baseline model training: The Gaussian process regression baseline model is trained under the working conditions (typical fabrics) of each cluster center. The squared exponential kernel is selected as the kernel function to ensure the model's fitting accuracy to typical working conditions.

[0094] Dynamic weight transfer: For new fabric conditions, calculate the cluster similarity between them and each benchmark condition, dynamically assign weights to the benchmark models and fuse predictions. Core weight calculation and fusion formulas:

[0095] ;

[0096] ;

[0097] in, For the first The weights of the baseline model, For new working conditions and the first Similarity between the baseline operating conditions For the first The predicted values ​​of the baseline model;

[0098] Prediction results output: The model outputs the predicted value and confidence interval of the remaining service life of the needle (confidence level 95%). The prediction error is less than or equal to ±8% under cross-fabric conditions, which is 25% lower than the error of the traditional Gaussian process regression model.

[0099] S304. Multimodal Evidence Chain Weighted Fusion Bayesian Network Root Cause Localization: When a broken pin or pin quality defect event occurs, the root cause localization process is initiated, integrating multimodal evidence to achieve hierarchical root cause identification.

[0100] Multimodal evidence chain construction:

[0101] Sensor evidence: Extracting the characteristics of sudden vibration amplitude changes, current fluctuations, and sudden tension changes at the moment of anomaly occurrence to form a 16-dimensional sensor evidence vector. ;

[0102] Image evidence: Semantic features such as skipped stitches, wrinkles, and pinhole offsets are extracted from finished product stitch defect images using a convolutional neural network (CNN), outputting an 8-dimensional image evidence vector. ;

[0103] Process evidence: Deviations in process parameters such as stitch density, suture tension, and machine speed are extracted to form a 12-dimensional process evidence vector. ;

[0104] Evidence chain credibility weighting: Weights are dynamically assigned based on the identification accuracy of each evidence chain. The core weight calculation formula is as follows:

[0105] ;

[0106] in, , , The recognition accuracy rates for sensors, images, and process evidence were respectively set at 0.92, 0.94, and 0.88 in the experiment, and the final weight values ​​were determined accordingly. , , Weighted fusion of evidence vectors forms a unified multimodal evidence input.

[0107] Bayesian network root cause reasoning: Construct a Bayesian network with four levels: "equipment failure - process deviation - material problem - environmental impact". Input the fused evidence vector into the network and calculate the posterior probability of each root cause through Bayesian reasoning. The core reasoning logic is based on the relationship between evidence and root cause to achieve accurate quantification of root cause probability.

[0108] Hierarchical root cause output: Outputs the top-3 root causes with posterior probabilities and their confidence levels, displayed in a hierarchy of "core root cause - secondary root cause - related influencing factors". The root cause localization accuracy is greater than or equal to 92%, and the localization time is less than or equal to 9 minutes, which is 40% more efficient than traditional Bayesian network localization.

[0109] S305. Multi-dimensional Needle Analysis Report Generation: Based on cloud-based analysis results, four core reports are generated. The reports use a standardized XML format, include a unified relational index, and support cross-report linked queries.

[0110] Needle health status report: Fields include device ID, statistical period (day / week / batch), current health score, which is based on RUL prediction value and abnormal probability quantification, with a full score of 100 points, core feature trend, including the time series changes of the top 5 key features such as vibration peak and current harmonics, health level (excellent / good / medium / poor), and warning prompts.

[0111] Stitch Quality Trend Report: Fields include production batch, fabric type, process parameter combination, stitch pass rate (related to image detection results), defect type distribution (e.g., skipped stitches, wrinkles, pinhole offset), correlation between defects and machine needle status (based on cloud-based root cause localization results), and quality trend slope (determined as rising / falling / stable).

[0112] Maintenance warning report: Fields include device ID, current RUL prediction value of the needle, confidence interval, maintenance priority, which is dynamically determined according to RUL and production plan, high / medium / low, suggested maintenance time window, list of spare parts to be replaced, and historical maintenance record association index;

[0113] Process optimization report: Fields include working condition label, current process parameters, optimization suggestion parameters are quantitative adjustment values, optimization basis is based on root cause localization conclusions combined with historical best data, expected optimization effect such as the increase in pin pass rate, the percentage increase in pin life, and optimization execution status (pending execution / executed / executed invalid).

[0114] S306. AI Analysis and Correlation Mining of Report Data:

[0115] Structured storage of report data: The four core reports are synchronized to the cloud report database, adopting a "master table + slave table" architecture: The master table stores basic report information, including report ID, generation time, and associated device / batch. The slave table stores specific field data for each report. A two-way association index is established through the report ID and device ID to support joint queries across reports, time periods, and devices.

[0116] AI-driven report data parsing:

[0117] Report anomaly pattern recognition: The report anomaly detection algorithm based on the improved KNN is used to perform time series analysis on the core indicators of the four major reports, namely health score, pass rate, RUL value and process parameter deviation, to identify anomaly patterns such as a sudden drop in health score, pass rate below the threshold for three consecutive batches, and RUL predicted value fluctuation exceeding 20%, and output anomaly report items and pattern labels such as sudden failure type, gradual decline type, poor process adaptation type, etc.

[0118] Cross-report association rule mining: The Apriori algorithm is used to mine the association relationships between different reports. The core mining logic is to use the RUL value of the maintenance warning report, the parameter deviation of the process optimization report, and the defect type of the stitch quality trend report as association items. The mining rule is such as "when the stitch tension deviation is >0.5N and the vibration peak exceeds the baseline by 30%, the probability of stitch skipping defects increases by 45%". A report on the mining of report association rules is generated to provide supplementary basis for root cause localization.

[0119] Report data feature extraction: For time-series data in reports, such as health scores and pass rate trends, a temporal convolutional network (TCN) is used to extract trend features, which are then integrated into the optimized feature set of cloud analytics to improve the accuracy of lifespan prediction and root cause localization.

[0120] S307. Analysis Results and Report Encapsulation and Feedback Interface: Optimized fusion features, remaining lifetime predictions, root cause analysis results, four core reports, and report association rule mining reports are encapsulated in the format of "Equipment ID-Timestamp-Operating Condition Tag" and transmitted to the decision feedback module via a RESTful API interface. Simultaneously, a model iteration data pool is constructed to store analysis results, actual lifetime data, root cause verification results, and report parsing data for each operating condition.

[0121] The decision feedback module integrates three mechanisms: scenario-adaptive federated incremental learning, quantitative process parameter optimization, and standardized work order annotation. This enables the precise conversion of analysis results into production execution and the continuous iteration of the model. Specific implementation details are as follows:

[0122] S401. Scenario-Adaptive Federated Incremental Learning Mechanism: Addressing the model update interference problem caused by the lack of scenario differentiation in traditional federated learning, a hierarchical learning system of "scenario partitioning - node deployment - dynamic updating" is constructed to ensure that model iteration accurately adapts to different production conditions.

[0123] Scene segmentation and node deployment: Production scenarios are divided into six core categories based on "fabric type + equipment model": cotton-flatbed sewing machine, synthetic fiber-flatbed sewing machine, leather-flatbed sewing machine, cotton-overlock sewing machine, synthetic fiber-overlock sewing machine, and leather-overlock sewing machine. Each scenario deploys an independent edge learning node, while a central coordination node is deployed in the cloud. Edge nodes are responsible for local data training, while the central node is responsible for parameter aggregation. Parameters are transmitted between nodes via an encrypted communication protocol to ensure data privacy.

[0124] Scene similarity and data filtering: After new work order data is generated, its affiliation is determined by calculating its feature matching degree with each core scene. Only data with the highest matching degree is assigned to the scene for training, avoiding cross-scene data interference. Data is assigned to the corresponding scene when the matching degree reaches a set threshold. Data from niche scenes must have a sample size of 50 or more to participate in training; if the sample size is insufficient, it is temporarily stored in the cloud data pool.

[0125] Dynamic update triggering and parameter aggregation: The update trigger condition is set to "the average prediction error of three consecutive work orders in the same scene is greater than 5%", after which the incremental learning process is started. Edge nodes use stochastic gradient descent (SGD) to optimize local model parameters, and center nodes aggregate the parameters of each edge node through a weighted average. The core aggregation formula is:

[0126] ;

[0127] in, These are the aggregated global model parameters. For the first Local parameters of each edge node For the first Local sample size of each node This represents the total number of samples across all scenarios. The aggregated parameters are synchronized to each edge node and the cloud analysis module to achieve iterative adaptation of the model to different scenarios.

[0128] S402. Quantitative Dynamic Optimization and Decision Generation of Process Parameters: Based on root cause analysis and capacity prioritization, a quantitative process parameter optimization model is constructed, outputting directly actionable parameter adjustment suggestions.

[0129] Process parameter weight allocation: The influence weight is calculated based on the information entropy of each process parameter such as suture tension, machine speed, and stitch density. The parameters with the highest weights, such as suture tension, machine speed, and stitch density, are prioritized for optimization, and their weights are determined according to the proportion of information entropy.

[0130] Quantitative adjustment calculation: Combining root cause analysis results with the historical best process parameter database, the parameter adjustment amount is calculated. Core formula:

[0131] ;

[0132] in, For the first The adjustment amount of each parameter, These are the historically optimal parameter values. The current parameter value. To adjust the coefficient, a dynamic value is assigned based on capacity priority: higher capacity is prioritized. Medium priority low priority .

[0133] The final output is a quantitative adjustment suggestion, such as "reduce the stitch tension by 0.4N and increase the vehicle speed by 8rpm".

[0134] S403. Standardized Work Order Generation and Multi-Dimensional Annotation: Based on quantitative optimization suggestions and maintenance requirements, structured executable work orders are generated, and a standardized annotation system is established to provide high-quality feedback data for model iteration.

[0135] Work orders are generated in a structured manner: Each work order comprises four core modules: basic information, process adjustment instructions, preventative maintenance schedule, and spare parts replacement list. Basic information includes equipment ID, production batch, and operating condition tag; process adjustment instructions include quantitative parameter adjustment values ​​and adjustment steps; the preventative maintenance schedule includes maintenance time windows and maintenance items; and the spare parts replacement list includes the model, quantity, and replacement cycle of the replacement needles. Work orders are encapsulated in XML format and pushed to the equipment controller and maintenance management system via an industrial internet platform interface.

[0136] Work order execution results are annotated in multiple dimensions: They are categorized into three scenarios: "Effective parameter optimization," "Effective maintenance," and "Ineffective," with annotations based on quantitative indicators.

[0137] Effective parameter optimization: After adjustment, the pin pass rate increased by greater than or equal to 5%, and the production efficiency decreased by less than or equal to 3%;

[0138] Effective maintenance: After maintenance, the probability of needle malfunction is reduced by more than or equal to 40%, and the remaining lifespan is extended by more than or equal to 10%;

[0139] Invalid: The above indicators were not met.

[0140] The annotation results synchronously record the process parameters, equipment status, and production index data before and after the adjustment, forming a complete feedback data chain.

[0141] S404. Report-Driven Closed-Loop Optimization: Deep Integration of Report Data and Decision Making: The decision feedback module receives four core reports and a report association rule mining report from the cloud analysis output, and uses the report analysis results as the core basis for decision optimization.

[0142] Process parameter optimization and refinement: When the process optimization report shows tension deviations across batches of the same type of fabric, and the association rule mining report confirms a strong correlation between this deviation and skipped stitch defects, an adjustment coefficient is calculated. An additional 0.1 weight is added to ensure that parameter adjustments are more targeted;

[0143] Dynamic adjustment of maintenance schedule: Combining the maintenance priority of the maintenance early warning report with the external production plan report, when high-priority maintenance needs conflict with peak production periods, based on the health score trend of the needle health status report, the maintenance time that can be delayed is predicted, with a maximum of no more than 2 production batches, and temporary process adjustment suggestions such as reducing the machine speed by 5% are generated to reduce the risk of failure.

[0144] Emergency Decision Making: When a sudden failure item is identified in the report's abnormal mode, an emergency decision-making process is automatically triggered, skipping the routine parameter adjustment steps and directly outputting an emergency shutdown or speed reduction command of 30%, while simultaneously pushing a maintenance work order to maintenance personnel.

[0145] S405. Report Data Feedback and Model Iteration: After organizing the annotated work order execution results, equipment operation data, and report data changes according to standardized fields, the data is fed back to the model iteration data pool and report database of the cloud analysis module.

[0146] Data weight allocation: In the returned data, the weight of samples marked as "effective optimization" or "effective maintenance" is increased by 20%, samples with "clear abnormal patterns" are included in the key training set, and samples with "report association rules verified" are given an additional weight of 0.1, which improves the model's ability to identify complex patterns in report data and the accuracy of decision-making.

[0147] Model iteration trigger: When the cumulative amount of backflow data reaches 100 records / scenario, or when the proportion of "invalid" labels in a certain scenario report is ≥30% for two consecutive periods, the local update process of scenario adaptive federated incremental learning will be automatically triggered.

[0148] S406. Layered Visualization and Closed-Loop Feedback Interface: Design a layered visualization interface that adapts to different roles, enabling accurate delivery of analysis results, and simultaneously construct a standardized feedback interface:

[0149] Layered visual interface:

[0150] Operator interface: Displays real-time process parameter adjustment suggestions, abnormal warning information, work order execution steps, and presents the current pin quality and needle status trend in chart form;

[0151] Maintenance personnel interface: Displays the predicted remaining life of the needle, the preventive maintenance schedule, and spare parts replacement reminders; supports maintenance record entry.

[0152] Management interface: Displays statistics on the health status of all equipment needles in the workshop, work order execution efficiency, and production quality trends, providing support for decision-making on capacity and maintenance cost optimization.

[0153] Closed-loop feedback interface: The labeled work order execution results and equipment operation data are fed back to the model iteration data pool of the cloud analysis module via a RESTful API. The data pool stores data according to scenarios, providing training data for scenario-adaptive federated incremental learning, and recording performance metrics before and after model iteration.

[0154] It needs to be further explained that,

[0155] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0156] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A needle report data analysis system based on artificial intelligence algorithms, characterized in that, include: Data acquisition module: Based on multi-sensor collaborative acquisition and motor encoder phase-locked synchronization mechanism, the multi-modal data of needle operation is filtered, denoised and labeled into reports to obtain standardized raw labeled data and phase reference signal; Edge preprocessing module: Based on hierarchical preprocessing and puncture window periodic sliding attention gating algorithm, feature purification and anomaly detection are performed on the standardized original labeled data and the phase reference signal to obtain preprocessed feature data, anomaly warning level, and generate lightweight preprocessing summary; The cloud-based analysis module, based on a physical prior fusion adversarial feature selection network, a cross-fabric transfer learning Gaussian process regression model, and a Bayesian network, performs feature optimization, lifetime prediction, and root cause localization on the preprocessed feature data and the lightweight preprocessed report. This yields optimized fusion features, remaining lifetime predictions, and root cause localization conclusions, and generates a multi-dimensional needle analysis report. Furthermore, an improved KNN algorithm and Apriori algorithm are used to perform anomaly pattern recognition and association rule mining on the structured data of the multi-dimensional needle analysis report, resulting in anomaly entries, pattern labels, and association rule mining results. The cross-fabric transfer learning Gaussian process regression model is described below: Cross-fabric working condition clustering: The K-Means algorithm is used to cluster historical data by fabric type. The clustering index is "puncture force variance - vibration spectrum similarity". The clustering is determined by calculating the similarity of vibration spectrum under different fabric working conditions. The number of clusters K=6, covering six mainstream fabrics. Baseline model training: The Gaussian process regression baseline model is trained under the working conditions of each cluster center. The kernel function is selected as the squared exponential kernel to ensure the model's fitting accuracy to typical working conditions. Dynamic weight transfer: For new fabric conditions, calculate the cluster similarity between them and each benchmark condition, dynamically assign weights to the benchmark models and fuse predictions. Core weight calculation and fusion formulas: ； ； in, For the first The weights of the baseline model, For new working conditions and the first Similarity between the baseline operating conditions For the first The predicted values ​​of the baseline model; Prediction results output: The model outputs the predicted value and confidence interval of the remaining service life of the needle. The prediction error is less than or equal to ±8% under different fabric conditions, which is 25% lower than the error of the traditional Gaussian process regression model. The improved KNN algorithm described above uses an improved KNN-based report anomaly detection algorithm to perform time-series analysis on the core indicators of multi-dimensional needle analysis reports, such as health score, pass rate, RUL value, and process parameter deviation, identify anomaly patterns, and output anomaly report entries and pattern labels. Decision Feedback Module: Based on the scenario-adaptive federated incremental learning mechanism and the quantitative parameter optimization model, the module performs parameter adjustment calculations and generates work orders based on the multi-dimensional needle analysis report and the association rule mining results, obtaining process adjustment instructions and maintenance work orders. After annotating the execution results of the maintenance work orders, the module feeds them back to the cloud analysis module as training samples for the physical prior fusion adversarial feature selection network, cross-fabric transfer learning Gaussian process regression model, Bayesian network, improved KNN algorithm, and Apriori algorithm iteration.

2. The needle report data analysis system based on artificial intelligence algorithm according to claim 1, characterized in that, The multi-sensor collaborative acquisition adopts a three-point collaborative acquisition layout, including: a vibration sensor attached to the upper end of the needle bar, an acoustic sensor deployed below the needle plate facing the needle puncture area, a fabric tension sensor integrated at the center of the feed roller shaft, and a motor stator current sensor connected in series with the motor power supply circuit.

3. The needle report data analysis system based on artificial intelligence algorithm according to claim 1, characterized in that, The hierarchical preprocessing includes a two-stage processing flow of detrending adaptive filtering and phase anchoring standardization; wherein, the detrending adaptive filtering uses an adaptive sliding window method to remove trend noise from the data; the phase anchoring standardization is performed based on the puncture phase interval divided by the phase reference signal to retain phase-related local features.

4. The needle report data analysis system based on artificial intelligence algorithm according to claim 1, characterized in that, The abnormal warning levels include: Level 1 anomaly is defined as an anomaly probability greater than or equal to 0.6 and less than 0.7, and the vibration amplitude exceeds the normal baseline under the same working conditions by 20% to 30%. Level II anomaly is defined as an anomaly probability greater than or equal to 0.7 and less than 0.85, and the vibration amplitude exceeds the normal baseline under the same working conditions by 30% to 50%. Level 3 anomaly is defined as an anomaly probability greater than 0.85 and a vibration amplitude exceeding the normal baseline under the same working conditions by more than 50%.

5. The needle report data analysis system based on artificial intelligence algorithm according to claim 1, characterized in that, The remaining life prediction value includes: the estimated remaining life of the needle in hours, output by the cross-fabric transfer learning Gaussian process regression model; the remaining life prediction value is accompanied by a 95% confidence interval, and the Gaussian process regression model is adapted to at least six fabric working conditions with different combinations of elastic modulus and thickness parameters through a cross-fabric transfer learning mechanism, the fabric working conditions include at least three corresponding working conditions of cotton, chemical fiber and leather, and the prediction error under the at least three corresponding working conditions is no greater than ±8%.

6. The needle report data analysis system based on artificial intelligence algorithm according to claim 5, characterized in that, The multi-dimensional needle analysis reports include: needle health status report, needle quality trend report, maintenance early warning report, and process optimization report; The needle health status report includes a health score and core characteristic trends based on the predicted remaining lifespan and the probability of abnormality. The pin quality trend report includes pin pass rate, defect type distribution and its correlation with pin status; The maintenance early warning report includes the predicted remaining lifespan of the needle, maintenance priority, and recommended maintenance time window; The process optimization report includes process parameters, optimization suggestion parameters generated based on root cause analysis and historical data, and expected optimization results.

7. A needle report data analysis system based on artificial intelligence algorithms according to claim 6, characterized in that, The process adjustment instructions include: assigning weights based on the degree of influence of process parameters on the output results; determining adjustment coefficients based on root cause localization conclusions and the historical best process parameter library, combined with the current capacity priority; and calculating the quantitative adjustment amount of each process parameter based on the historical best parameter values ​​in the historical best process parameter library, the current parameter values ​​in the process parameters, and the adjustment coefficients.

8. The needle report data analysis system based on artificial intelligence algorithm according to claim 1, characterized in that, The maintenance work order includes: a structured work order generated based on the process adjustment instructions and maintenance requirements; The structured process order shall include at least: basic information for identifying equipment and production batch, process adjustment instructions including quantitative parameter adjustment values ​​and operation steps, preventive maintenance schedule including maintenance time windows and maintenance items, and a spare parts replacement list including the model and quantity of the needles to be replaced. The execution result of the maintenance work order will be marked as "valid" or "invalid" according to the preset quantitative indicators, and after being marked, it will be fed back to the cloud analysis module as a training sample for model iteration.

9. A needle report data analysis system based on artificial intelligence algorithms according to claim 1, characterized in that, The scenario-adaptive federated incremental learning mechanism includes: dividing production scenarios according to fabric type and equipment model and deploying independent edge learning nodes; filtering data and assigning scenarios based on the feature matching degree between new work order data and each scenario; and when the update trigger condition is met, the cloud central node aggregates the local model parameters of each edge node by weighted average to realize scenario-based incremental update of model parameters.

Images