A data fusion method and system for textile production traceability

Abstract

This invention relates to the field of data fusion technology, and more particularly to a data fusion method and system for textile production traceability. The method includes the following steps: acquiring sensor data and camera data from the textile production workshop, and performing data fusion on the textile production chain data stream to obtain a textile production chain data stream; performing feature cross-calculation based on the textile production chain data stream to generate a processing technology enhancement feature space; fingerprinting the textile production chain data stream to obtain a cross-chain spatiotemporal stamp fingerprint set; evaluating the batch production quality of the textile production line based on the textile production workshop camera data to obtain batch data of the production line to be traced and batch data of the production line to be put into storage; evaluating the feasibility of dyeing optimization for abnormal batches in the batch data of the production line to be traced; and transmitting the batch data of the production line to be put into storage and the batch data of the production line to be traced after dyeing optimization to the workshop warehouse management platform. This invention improves production efficiency and reduces resource waste.

Description

Technical Field

[0001] This invention relates to the field of data fusion technology, and in particular to a data fusion method and system for textile production traceability. Background Technology

[0002] Textile traceability systems, as a key tool to ensure product quality, safety, and sustainability, have become an important technological application in the industry. The introduction of traceability technology helps strengthen the tracking and management of the entire process from raw material procurement, production, to the final product distribution, effectively protecting consumer rights and improving the market competitiveness of enterprise products. Traditional textile production traceability methods mostly rely on a single data source or incomplete information flow. In traditional traceability systems, data from each production stage is managed independently by different enterprises or departments, lacking unified standards and platforms for information sharing and integration. Textile production involves multiple stages, such as raw material procurement, textile processing, dyeing, and garment production. The data generated at each stage is isolated, making cross-departmental and cross-enterprise information connectivity impossible. These data silos make it difficult to obtain traceability information uniformly, failing to form a complete production chain data flow, thus affecting the reliability and efficiency of traceability. Summary of the Invention

[0003] Therefore, it is necessary for the present invention to provide a data fusion method and system for textile production traceability in order to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a data fusion method for textile production traceability includes the following steps:

[0005] Step S1: Acquire sensor data and camera data from the textile production workshop, and perform data fusion of the textile production chain data stream to obtain the textile production chain data stream.

[0006] Step S2: Analyze the yarn dyeing characteristic pattern based on the data flow of the textile production chain, and perform feature cross calculation based on the yarn dyeing characteristic pattern to generate a processing technology enhancement feature space;

[0007] Step S3: Based on the enhanced feature space of the processing technology, perform process fingerprinting on the textile production chain data stream to obtain a cross-chain spatiotemporal stamp fingerprint set;

[0008] Step S4: Based on the video data of the textile production workshop, perform pixel anomaly detection on the finished textile products, and evaluate the production quality of the textile production line batches according to the anomaly detection results, thereby obtaining the production line batch data to be traced and the production line batch data to be put into the warehouse.

[0009] Step S5: Use the cross-chain spatiotemporal stamp fingerprint set to trace the abnormal production line batch data to obtain an abnormal traceability report; evaluate the feasibility of dyeing optimization for abnormal batches based on the abnormal traceability report, and upload it to the workshop production management platform to execute the dyeing optimization task;

[0010] Step S6: Transmit the batch data of the production line to be put into the warehouse and the batch data of the production line to be traced after the dyeing optimization task to the workshop warehouse management platform to execute the warehouse filing task.

[0011] Optionally, step S1 specifically includes:

[0012] Step S11: Acquire sensor data and camera data from the textile production workshop;

[0013] Step S12: Perform data preprocessing and standardization on the sensor data and camera data of the textile production workshop respectively to obtain standardized sensor data and standardized camera data;

[0014] Step S13: Perform dynamic time warping based on standardized sensor data and standardized camera data to obtain a time-aligned dataset;

[0015] Step S14: Extract the joint context features of the temporal alignment dataset, construct a multimodal fusion feature representation, and thus obtain windowed fusion feature data;

[0016] Step S15: Logically reconstruct the windowed feature data, connect the data links of production flow relationship, and thus obtain the textile production chain data flow.

[0017] Optionally, step S14 specifically includes:

[0018] Step S141: Set the sliding time window length to [30s, 120s], and synchronously extract the sensor data and camera data frame sequences of the same time period in the time-aligned dataset within each sliding time window to obtain the time window structure set;

[0019] Step S142: Perform trend gradient encoding on the sensor data in the time window structure set, and extract principal component features from the camera data frame sequence to obtain the preliminary feature vectors in the sensor mode and the preliminary feature vectors in the camera mode, respectively.

[0020] Step S143: Align the preliminary feature vectors in the sensing mode and the preliminary feature vectors in the camera mode between modes, perform intermodal spatiotemporal correlation modeling, and enhance the signal fusion strength to obtain spatiotemporal aligned embedding features;

[0021] Step S144: Perform upstream and downstream semantic path graph construction and semantic smoothing on the spatiotemporally aligned embedding features of the continuous sliding time window to obtain a semantic fusion vector group;

[0022] Step S145: Reorganize the semantic fusion vector group into time series data according to the time window order of the time window structure set, and perform structured encapsulation to obtain windowed fusion feature data.

[0023] Optionally, step S15 specifically includes:

[0024] Step S151: Perform semantic matching between the semantic features in the windowed fusion feature data and the preset textile production process framework, and construct a process node transmission relationship diagram based on the semantic matching results.

[0025] Step S152: Perform topological sorting on the process node transmission relationship graph, and perform data flow dependency modeling on the sorted node transmission relationship graph to obtain the data flow dependency chain;

[0026] Step S153: Construct a cross-node data flow connector based on the data flow dependency chain;

[0027] Step S154: Use cross-node data flow connectors to connect the data flows of each production node in the process node transmission relationship diagram, and perform spatiotemporal consistency verification to obtain the textile production chain data flow.

[0028] Optionally, the analysis of yarn dyeing characteristic patterns in step S2 specifically includes:

[0029] Extract yarn characteristic data and pH value data of the dyeing process from the data stream of the textile production chain;

[0030] Obtain the textile yarn property library, and extract the corresponding yarn properties from the textile yarn property library based on the yarn characteristic data to obtain yarn property data;

[0031] Spearman rank correlation analysis was performed on yarn characteristic data and pH data during the dyeing process, and feature pairs with Spearman correlation coefficients less than 0.7 were removed to obtain yarn-pH highly correlated feature pairs.

[0032] Based on the high correlation between yarn and pH value, variable relationship pattern recognition is performed. If the yarn density in the feature pair is strongly positively correlated with the dyeing pH value, then a high-density yarn dyeing feature pair is obtained; if the yarn strength in the feature pair is negatively correlated with the dyeing pH value, then a high-strength yarn dyeing feature pair is obtained.

[0033] The high-density yarn dyeing feature pairs and the high-intensity yarn dyeing feature pairs are merged to obtain the yarn dyeing feature pattern.

[0034] Optionally, step S3 specifically includes:

[0035] Step S31: Enhance the feature space according to the processing technology, set the set of process feature quantification parameters, and encode the process feature vectors of each production node in the textile production chain data flow to generate a node feature encoding set.

[0036] Step S32: Perform node comparison learning based on the node feature encoding set, evaluate the similarity of process feature transfer between adjacent nodes, and if the similarity is less than 0.85, mark it as a potential process mutation point, thereby generating a set of potential mutation node pairs;

[0037] Step S33: Perform temporal consistency analysis on the potential variant node pair set. Within a preset time window of 30 minutes, select node pairs that conform to the transmission order of the textile production chain data flow and generate a preliminary spatiotemporally consistent node chain.

[0038] Step S34: Based on the preliminary spatiotemporally consistent node chain, extract the position information and timestamp of each pair of nodes in the spatial layout, and calculate the node transfer delay. If the node transfer delay exceeds 120s, mark the node pair as having a transfer anomaly, thereby generating cross-chain spatiotemporal feature pairs.

[0039] Step S35: Perform feature hash encoding on the cross-chain spatiotemporal feature pairs, integrate all fingerprint identifiers, and construct a cross-chain spatiotemporal stamp fingerprint set.

[0040] Optionally, the anomaly detection of finished textile pixels in step S4 specifically includes:

[0041] Textile production workshop camera data that has been standardized and dynamically time-normalized is subjected to texture feature enhancement processing to extract texture direction, frequency and contrast features, resulting in an enhanced texture image feature matrix.

[0042] A pixel-level image anomaly detection model is constructed based on the enhanced texture image feature matrix. The model detection threshold is set to a significance score greater than 0.6 to identify abnormal pixel regions and output a preliminary abnormal overlay image.

[0043] Morphological structure analysis and pixel clustering are performed on the preliminary anomaly mask image to remove pseudo-anomaly noise areas with an area of ​​less than 50 pixels, and a set of anomaly region bounding boxes is generated based on the pixel cluster center as the annotation basis.

[0044] The set of boundary boxes of abnormal regions is mapped back to the process nodes in the corresponding textile production chain data stream, and the timestamps of the video data from the textile production workshop are combined to generate a set of finished product pixel anomaly detection labels.

[0045] Optionally, the evaluation of batch production quality of textile production line in step S4 specifically includes:

[0046] The finished product pixel anomaly detection tag set is timestamped and batch-aligned according to the node time window in the textile production chain data flow to generate a batch-level anomaly tag matrix.

[0047] Abnormal pixel features are extracted from the batch-level abnormal label matrix to obtain abnormal pixel density, abnormal distribution dispersion and abnormal region coverage.

[0048] A batch quality quantitative evaluation system is constructed based on abnormal pixel density, abnormal distribution dispersion, and abnormal region coverage.

[0049] The batch quality quantitative evaluation system is used to score the batch quality of each batch in the batch-level anomaly label matrix. The scoring range is set to [0,1]. Batches with a quality score lower than 0.6 are marked as batches to be traced. A preliminary quality score result set is generated.

[0050] The preliminary quality score result set is subjected to abnormal fluctuation identification. If the score fluctuation of three consecutive batches exceeds 0.15, it is marked as a batch group with trend quality deviation. The process node identifiers of the textile production chain data flow corresponding to the abnormal pixel area are integrated to output the final production line batch quality assessment report.

[0051] Batches marked as pending traceability in the production line batch quality assessment report are classified as pending traceability production line batch data; batches with a quality score higher than 0.6 and not marked as trending quality deviation batches are classified as pending warehousing production line batch data, and the tag status is synchronized to the production chain data stream.

[0052] Optionally, step S5 specifically includes:

[0053] Step S51: Perform fingerprint matching between the batch data of the production line to be traced and the cross-chain spatiotemporal stamp fingerprint set. Compare the time period of the textile production chain data chain corresponding to the batch data of the production line to be traced with the spatiotemporal stamp distribution range in the fingerprint set, extract the batch fingerprint subset, and generate batch-level fingerprint mapping data.

[0054] Step S52: Based on the node feature hash values ​​in the batch-level fingerprint mapping data, reconstruct the process flow trajectory of the corresponding textile production chain data flow to obtain the reconstructed trajectory node sequence.

[0055] Step S53: Perform anomaly interpolation fitting analysis on the reconstructed trajectory node sequence to identify missing, skipped, or time-reversed node segments in the trajectory and generate anomaly trajectory segment marker set;

[0056] Step S54: Perform cross-source data tracing on the abnormal trajectory segment marker set, and use the textile production chain data chain to retrieve the environmental status, operation records and equipment operating parameters of the corresponding node segment, and extract the abnormal cause feature set;

[0057] Step S55: Perform pattern recognition based on the set of abnormal cause features to identify the abnormal patterns of equipment oscillation deviation and dyeing transition zone drift, and obtain an abnormal source tracing report;

[0058] Step S56: Combine the anomaly traceability report with the yarn dyeing characteristic pattern to establish a batch-level dyeing process reversibility assessment model to evaluate the feasibility of dyeing optimization. If the dyeing optimization feasibility is higher than or equal to 0.6, the anomaly traceability report of the corresponding batch is uploaded to the workshop production management platform, and the corresponding batch status is marked as pending dyeing optimization. The workshop production management platform then schedules the dyeing optimization process. If the dyeing optimization feasibility is lower than 0.6, the anomaly traceability report of the corresponding batch is uploaded to the workshop production management platform, and the corresponding batch status is marked as pending recycling. The workshop production management platform then schedules the batch recycling process.

[0059] This invention achieves comprehensive data fusion across the textile production chain by integrating sensor and camera data from the textile production workshop. This effectively improves the efficiency and accuracy of information integration during the production process, eliminates data silos in traditional production processes, and provides complete data support for traceability systems. By dynamically warping and modeling the spatiotemporal correlation of the textile production chain data flow, it can accurately capture temporal changes during production, ensuring data timeliness and spatial consistency. Data preprocessing and standardization steps ensure that data from different sources can be compared and fused under the same standard, thereby improving overall data quality and information consistency. By setting an appropriate sliding time window (e.g., 30s to 120s) to simultaneously capture and extract features from sensor and camera data, the fusion of temporal and modal features becomes more refined, thus achieving comprehensive and detailed monitoring of the production process. In terms of yarn dyeing characteristic pattern analysis, based on yarn characteristics and pH data during the dyeing process, Spearman rank correlation analysis and pattern recognition were used to eliminate low-correlation feature pairs, focusing on the dyeing characteristics of high-density and high-strength yarns. This optimization step accurately identifies key influencing factors in the dyeing process, further improving the prediction and optimization capabilities of dyeing quality. By setting a reasonable set of process feature quantification parameters and node feature encoding, and implementing node comparative learning and temporal consistency analysis, potential process variation points can be identified and alerted early in production to prevent the spread of quality problems. For the quality assessment of finished textile products, a pixel-level image anomaly detection and batch quality scoring system can effectively identify and mark batches with large quality deviations, enabling timely quality optimization and avoiding waste of defective products due to a lack of accurate quality feedback during production. By introducing a cross-chain spatiotemporal fingerprint set for anomaly tracing of batch data, the accuracy of the tracing report is improved, and a basis for subsequent quality improvement is provided, particularly playing a crucial role in optimizing the dyeing process. By using the dyeing process reversibility assessment model, the feasibility of dyeing optimization can be evaluated, helping the production management platform to accurately schedule dyeing optimization or recycling processes, thereby achieving a reasonable allocation of resources and time in the production process, improving production efficiency, and reducing resource waste.

[0060] Optionally, this specification also provides a data fusion system for textile production traceability, used to execute the data fusion method for textile production traceability as described above. The data fusion system for textile production traceability includes:

[0061] The data fusion module is used to acquire sensor data and camera data from the textile production workshop, and to perform data fusion of the textile production chain data stream to obtain the textile production chain data stream.

[0062] The yarn feature pattern recognition module is used to analyze yarn dyeing feature patterns based on the data flow of the textile production chain, and to perform feature cross calculations based on the yarn dyeing feature patterns to generate a processing technology enhancement feature space.

[0063] The fingerprinting module is used to perform process-based fingerprinting of textile production chain data streams based on the enhanced feature space of the processing technology, resulting in a cross-chain spatiotemporal stamp fingerprint set.

[0064] The anomaly detection module is used to detect anomalies in the pixels of finished textile products based on camera data from the textile production workshop, and to evaluate the production quality of textile production line batches based on the anomaly detection results, thereby obtaining batch data of production lines to be traced and batch data of production lines to be put into storage.

[0065] The abnormal production traceability module is used to trace abnormal production data of production line batches by using cross-chain spatiotemporal stamp fingerprint set to obtain an abnormal traceability report; evaluate the feasibility of dyeing optimization for abnormal batches based on the abnormal traceability report, and upload it to the workshop production management platform to execute the dyeing optimization task;

[0066] The warehouse transmission module is used to transmit batch data of production lines to be put into storage and batch data of production lines to be traced after dyeing optimization to the workshop warehouse management platform to perform warehouse filing tasks.

[0067] The present invention relates to a data fusion system for textile production traceability. This system can implement any of the data fusion methods for textile production traceability of the present invention. It is used to combine the operation and signal transmission medium between various modules to complete the data fusion method for textile production traceability. The modules within the system cooperate with each other, thereby improving production efficiency and reducing resource waste. Attached Figure Description

[0068] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0069] Figure 1 This is a schematic diagram of the steps in the data fusion method for textile production traceability of the present invention;

[0070] Figure 2 This is a detailed flowchart of step S1 in the present invention;

[0071] Figure 3 This is a detailed flowchart of step S3 in the present invention;

[0072] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0073] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0074] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0075] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0076] To achieve the above objectives, please refer to Figures 1 to 3 This invention provides a data fusion method for textile production traceability, the method comprising the following steps:

[0077] Step S1: Acquire sensor data and camera data from the textile production workshop, and perform data fusion of the textile production chain data stream to obtain the textile production chain data stream.

[0078] In this embodiment, sensors for temperature, humidity, tension, pH, and conductivity, deployed at key process nodes in the textile production workshop, collect environmental and equipment status data at 1-second intervals. Simultaneously, high-definition industrial cameras installed at the fabric output sections of each production line capture color video streams with a resolution of 1920×1080 and a frame rate of 10fps. The sensor data and image data are synchronized with a unified timestamp and integrated using a three-layer fusion mechanism: time alignment, spatial mapping, and signal strength weighting processing, forming a textile production chain data stream with node identifiers and location information. An error correction module is introduced during the fusion process to limit the maximum time drift to within ±2 seconds, ensuring data synchronization accuracy of over 95%.

[0079] Step S2: Analyze the yarn dyeing characteristic pattern based on the data flow of the textile production chain, and perform feature cross calculation based on the yarn dyeing characteristic pattern to generate a processing technology enhancement feature space;

[0080] In this embodiment, raw characteristic data of yarn are extracted from the textile production chain data stream, including yarn fineness (unit: te [30s, 120s]), moisture regain, color adsorption coefficient, and real-time pH curve during the dyeing process. By matching yarn characteristic values ​​with pH values, Spearman rank correlation analysis is used to assess their correlation, retaining only characteristic pairs with a correlation coefficient not less than 0.7. Relationships such as a positive correlation between yarn density (150–180 te [30s, 120s]) and pH value, and a negative correlation between high-strength yarn (breaking strength ≥ 28 cN / te [30s, 120s]) and pH value are identified. These characteristic pairs are then categorized into two groups: high-density alkaline-suitable characteristic category and high-strength alkaline-avoidant characteristic category, forming identifiable yarn dyeing pattern labels for subsequent processing optimization. The yarn dyeing characteristic patterns in the textile production chain data stream are analyzed to extract relevant raw feature data, such as dyeing temperature (60–90℃), dye concentration (unit: g / L), pH value (range: 4–10), yarn density (unit: te[30s, 120s]), and color difference value (unit: ΔE). These feature data are normalized and prepared for subsequent feature cross-validation according to certain rules (e.g., normalized to the [0,1] interval). Next, feature cross-validation is performed, setting the cross-validation factor as the product of yarn density and dyeing temperature, the product of dyeing solution pH value and dye concentration, and combining it with a weighted combination of color difference value and yarn density to generate a new generation of cross-feature sets. By calculating the correlation of each pair of features, cross-features with significant relationships are selected, ultimately forming an enhanced process feature space. The threshold for cross-validation is set to a correlation coefficient greater than 0.7; only cross-features meeting this condition are included in the process feature space.

[0081] Step S3: Based on the enhanced feature space of the processing technology, perform process fingerprinting on the textile production chain data stream to obtain a cross-chain spatiotemporal stamp fingerprint set;

[0082] In this embodiment, after establishing the enhanced feature space for the processing technology, standardized process vectors need to be extracted from each production node in the textile production chain data flow. These vectors include operation time (in seconds), tension value, dyeing temperature, etc., and a set of process quantification parameters is set, such as a standardized temperature range of 60–90℃ and a standardized tension value range of 30–45N. Each node is then encoded. Through similarity analysis of process vectors between consecutive nodes, segments with a similarity less than 0.85 are marked as potential anomalies. Combining the temporal order of the nodes, a 30-minute time window is set to screen for correctly ordered and adjacent node chains. The flow time delay between adjacent nodes is then calculated; if the delay exceeds 120 seconds, the segment is marked as a spatiotemporal anomaly. Finally, the feature values ​​of the anomaly segments are encoded, and a cross-chain spatiotemporal fingerprint set is generated using a hash method, with added time and spatial tags for subsequent traceability.

[0083] Step S4: Based on the video data of the textile production workshop, perform pixel anomaly detection on the finished textile products, and evaluate the production quality of the textile production line batches according to the anomaly detection results, thereby obtaining the production line batch data to be traced and the production line batch data to be put into the warehouse.

[0084] In this embodiment, video data from a textile production workshop captured by an industrial camera is preprocessed, including normalization, histogram equalization, and texture enhancement operations. Orientation gradient, frequency texture, and local contrast features are extracted to generate a three-channel texture feature matrix. The local anomaly factor method is used to identify significantly different regions in the image, calculating the local density distribution of each pixel to obtain a saliency score. Regions with scores higher than 0.6 are selected to generate a binary anomaly mask. Opening and closing operations and connected component analysis are performed on this image to remove noise smaller than 50 pixels, and bounding boxes are generated based on the centroid of each clustered anomaly region. The timestamps of the bounding boxes are then matched with production chain nodes to form an image anomaly label set for tracking specific process steps. For quality assessment of each batch, the anomaly pixel density per unit length of fabric in each batch is calculated based on the number, area ratio, and location distribution density of the bounding boxes of anomaly regions in the finished product image, combined with the original production chain data. Three threshold indicators are set: the proportion of anomaly regions does not exceed 5%, the anomaly concentration coefficient is less than 0.4, and the continuity of anomaly regions does not exceed three consecutive nodes. By fitting the abnormal distribution curve of batch images, the overall quality trend of the batch is evaluated, and batches that meet the above conditions are marked as "qualified"; otherwise, they enter the traceability process. Combining the image acquisition time and node ID, data and video clips of the corresponding production chain segments are extracted to form a set of batch data to be traced and to be put into storage.

[0085] Step S5: Use the cross-chain spatiotemporal stamp fingerprint set to trace the abnormal production line batch data to obtain an abnormal traceability report; evaluate the feasibility of dyeing optimization for abnormal batches based on the abnormal traceability report, and upload it to the workshop production management platform to execute the dyeing optimization task;

[0086] In this embodiment, the time period of the batch to be traced is aligned with the cross-chain spatiotemporal stamp fingerprint set, and all fingerprint nodes appearing within the corresponding time period are selected and fingerprint mapping data is established. Then, the process trajectory is reconstructed based on the hash value. If abnormally large time intervals, disordered sequences, or missing segments are found in the node sequence, they are marked as abnormal trajectory segments. Sensor and camera data for the corresponding time periods are reviewed for these abnormal segments to extract relevant operation records, equipment operating parameters, and environmental conditions. Feature analysis rules are set, such as equipment operating frequency fluctuations exceeding 0.5Hz, pH value change rate exceeding 0.3 units / minute, trajectory offset exceeding 20 pixels, and operation time difference exceeding ±5 seconds, to identify equipment oscillation-type, dyeing drift-type, and procedure skip-type anomalies, and finally, an anomaly tracing report is generated as the basis for subsequent dyeing optimization decisions. When evaluating the feasibility of dyeing optimization for abnormal batches, the feasibility of dyeing optimization is specifically assessed by comprehensively analyzing the set of abnormal cause features in the anomaly tracing report. First, based on the dyeing process characteristics and combined with the historical dyeing data of the abnormal batch, the similarity between this batch and previously successfully optimized batches is evaluated. Specifically, the process extracts dyeing conditions from historical optimized batches, including dyeing time, dye type, and liquid ratio, and calculates the characteristic distance (e.g., Euclidean or Manhattan distance) between the current batch and these historical batches. If the characteristic distance of the current batch is below a set threshold (e.g., 0.3), the batch is considered to have high dyeing optimization potential. Next, the pH fluctuation range and transition zone drift characteristics of the dyeing solution for the current batch are calculated, with a pH fluctuation range threshold set at ±0.2 units. If this value exceeds the threshold, the batch is considered to have a significant risk of dyeing optimization. Through comprehensive analysis of multiple factors such as the stability of the dyeing solution's pH value and changes in dyeing temperature, a dyeing optimization feasibility score is calculated. If the score is ≥0.6, the batch is considered to be feasible for dyeing optimization, and the dyeing optimization process is initiated; if the score is <0.6, the process enters the recovery phase. This process can accurately identify which batches need optimization during the dyeing process and formulate reasonable adjustment plans based on their characteristics to improve production efficiency and product quality.

[0087] Step S6: Transmit the batch data of the production line to be put into the warehouse and the batch data of the production line to be traced after the dyeing optimization task to the workshop warehouse management platform to execute the warehouse filing task.

[0088] In this embodiment, when production line batch data to be received is transmitted to the workshop warehouse management platform, the platform first undergoes a series of verifications to ensure the integrity and accuracy of the batch data. Basic information for each batch is extracted from the textile production chain data stream, including batch number, production date, raw material type, dyeing conditions, production process, and quality inspection report. This information is then standardized to ensure unified storage within the warehouse management platform. During data storage, each batch is assigned a unique warehouse number, and a detailed warehouse label containing batch production information, quality inspection results, and dyeing optimization status is generated. This label is automatically attached to the corresponding batch's data file. Upon receiving data on production line batches to be traced and production line batches after dyeing optimization, the warehouse management platform dynamically schedules data based on real-time inventory and warehouse space to ensure each batch is stored in the most suitable location. Through automated scheduling and optimization algorithms, a warehouse capacity threshold is set; if the storage capacity exceeds 80%, a new round of inventory optimization is automatically initiated. Finally, detailed warehousing records for each batch are uploaded to the central warehouse management system to ensure smooth data flow and accurate execution of warehousing tasks.

[0089] Optionally, step S1 specifically includes:

[0090] Step S11: Acquire sensor data and camera data from the textile production workshop;

[0091] In this embodiment, sensors for temperature, humidity, tension, pH, and conductivity, deployed at key process nodes in the textile production workshop, collect environmental and equipment status data at 1-second intervals. Simultaneously, high-definition industrial cameras installed at the fabric output sections of each production line capture color video streams with a resolution of 1920×1080 and a frame rate of 10fps. The high-definition industrial cameras installed in the workshop capture images of the production process at a rate of 25 frames per second, recording the dyeing status of the textiles, yarn movement trajectories, and fabric surface conditions. Sensor data and camera data are synchronously acquired through a data acquisition interface and transmitted to a central data storage system.

[0092] Step S12: Perform data preprocessing and standardization on the sensor data and camera data of the textile production workshop respectively to obtain standardized sensor data and standardized camera data;

[0093] In this embodiment, the acquired sensor data undergoes denoising processing by removing high-frequency noise through a filter (e.g., a low-pass filter with a cutoff frequency of 0.1Hz). Subsequently, different types of data, such as temperature and humidity, are standardized to a range of [0,1]. For example, temperature data ranging from 20℃ to 90℃ is standardized to the [0,1] interval using a linear mapping. When preprocessing the camera data, image denoising and contrast enhancement are performed first, improving image clarity by setting contrast enhancement parameters (e.g., an enhancement factor of 1.5). Then, image normalization is used to adjust the pixel values ​​of each image to between 0 and 255, ensuring the uniformity and comparability of images in subsequent analysis. The preprocessed sensor data and camera data are stored as standardized sensor datasets and standardized camera datasets, respectively, for subsequent analysis.

[0094] Step S13: Perform dynamic time warping based on standardized sensor data and standardized camera data to obtain a time-aligned dataset;

[0095] In this embodiment, dynamic time warping (DTW) is performed on standardized sensor data and standardized camera data. First, the timestamp information in each dataset is identified to ensure consistency between the sensor and camera data in the time dimension. The sensor data is acquired once per second, while the camera data is acquired at a frequency of 25 frames per second. Therefore, a dynamic time warping algorithm is used to align the two data sources. During this process, sensor data and image frames are matched based on their timestamps, aligning misaligned data to the same time step, ensuring time synchronization between the two data sources. Specifically, the camera data is interpolated according to the 25 frames per second acquisition rate and mapped to the time step of the sensor data. The resulting time-aligned dataset ensures that each moment has corresponding data from both the sensor and the camera.

[0096] Step S14: Extract the joint context features of the temporal alignment dataset, construct a multimodal fusion feature representation, and thus obtain windowed fusion feature data;

[0097] In this embodiment, after the time-series aligned dataset is generated, joint contextual features are extracted from the data. First, key process features are extracted from the sensor data, such as temperature fluctuations of the dyeing solution, pH change rate, and yarn movement state. For camera data, image recognition technology is used to extract information such as the dyeing state of the fabric surface, defect detection, and movement trajectory. Then, a multimodal fusion feature representation is constructed by fusing the features of the sensor data and camera data. Specifically, the time-series features and image features are windowed, and the data within each time window is merged and weighted averaged to obtain the fusion features for each window. The window size is set to 30 seconds and adjusted according to the feature weights of the data sources. For example, the weight for sensor data is set to 0.6, and the weight for camera data is set to 0.4. Finally, these fusion features form a comprehensive windowed fusion feature dataset, providing a comprehensive, cross-modal characterization of the textile production process.

[0098] Step S15: Logically reconstruct the windowed feature data, connect the data links of production flow relationship, and thus obtain the textile production chain data flow.

[0099] In this embodiment, after the windowed fusion feature dataset is constructed, it is logically reconstructed. Specifically, firstly, based on the feature information of each data window, the production flow relationship of each production node is identified. For example, processes such as dyeing, washing, and drying are identified, and the temporal relationship of these processes in the production process is determined. Next, based on these node features, a complete production data link is constructed. This link includes all production stages from raw material input to finished product output and their corresponding data flows. According to the temporal information in sensor data and camera data, the data links are connected in sequence according to the production process. For example, the sensor data and image data corresponding to the dyeing process are connected together to form a complete dyeing process data link. Finally, the logically reconstructed data link becomes the textile production chain data flow, which is used for subsequent production monitoring, quality inspection, and optimization tasks.

[0100] Optionally, step S14 specifically includes:

[0101] Step S141: Set the sliding time window length to [30s, 120s], and synchronously extract the sensor data and camera data frame sequences of the same time period in the time-aligned dataset within each sliding time window to obtain the time window structure set;

[0102] In this embodiment, the length of the sliding time window is set to [30 seconds, 120 seconds], and an appropriate time window is selected according to the actual situation. For rapidly changing parameters (such as temperature fluctuations), a 30-second window can be selected. For slower-changing processes (such as dyeing), a 2-minute or 5-minute window can be selected. Because different production processes have different cycles and rhythms, and the frequency of process flow changes is relatively slow, the sliding time window length is generally set to be relatively long (e.g., 60 seconds). Within each time window, sensor data and camera data frame sequences are synchronously extracted from the time-aligned dataset to form a time window structure set. For example, assuming the sampling frequency of camera data is 1 frame per second and the sampling frequency of sensor data is 10 data points per second, then for a time window of 30 seconds, 30 frames of camera data and 300 sensor data points will be collected within this window. If the time window is 120 seconds, the amount of data collected will increase accordingly.

[0103] Step S142: Perform trend gradient encoding on the sensor data in the time window structure set, and extract principal component features from the camera data frame sequence to obtain the preliminary feature vectors in the sensor mode and the preliminary feature vectors in the camera mode, respectively.

[0104] In this embodiment, for sensor data with a time window structure, trend gradient encoding is first performed. This step aims to capture the trend characteristics of sensor data over time, such as the rate of change of parameters like temperature, humidity, and pressure. Trend gradient encoding calculates the magnitude of change between consecutive time points to obtain gradient values ​​reflecting time changes. Assuming the sensor data is temperature values, trend gradient encoding yields the rate of temperature change at each time point. For the sequence of camera data frames, principal component feature extraction is performed. This process uses PCA (Principal Component Analysis) to extract the main features of the data, such as texture, color, and contrast, compressing the data dimensionality and retaining the most representative features. Taking a photographed textile surface as an example, PCA can extract the most significant texture patterns or color changes in the image, providing a clear feature representation for subsequent analysis. Finally, preliminary feature vectors are obtained for the sensing modality and the camera modality, respectively.

[0105] Step S143: Align the preliminary feature vectors in the sensing mode and the preliminary feature vectors in the camera mode between modes, perform intermodal spatiotemporal correlation modeling, and enhance the signal fusion strength to obtain spatiotemporal aligned embedding features;

[0106] In this embodiment, the preliminary feature vectors in the sensing modality and the preliminary feature vectors in the camera modality are aligned intermodally. This means that the features of the two data modalities need to be paired to ensure their temporal consistency. After alignment, intermodal spatiotemporal correlation modeling is performed, that is, by establishing a model of the spatial and temporal relationship between the data, the features of the sensing data and the camera data are effectively fused. For example, suppose the sensing data records temperature changes, while the camera data reflects the texture changes on the surface of textiles. When performing spatiotemporal correlation modeling, the changes in surface texture can be predicted based on temperature changes, thereby improving the strength of signal fusion. Finally, the obtained spatiotemporally aligned embedded features will contain comprehensive information from the sensing modality and the camera modality, effectively capturing the relationship between temperature changes and surface texture changes during the production process.

[0107] Step S144: Perform upstream and downstream semantic path graph construction and semantic smoothing on the spatiotemporally aligned embedding features of the continuous sliding time window to obtain a semantic fusion vector group;

[0108] In this embodiment, the spatiotemporal alignment embedding features within a continuous sliding time window are first used as the basis for constructing the upstream and downstream semantic path graph. Each spatiotemporal embedding feature contains the production stage status, operational features, and process association information within the current time window, expressed in the form of nodes, and the upstream and downstream path relationships of the graph structure are established according to the task flow sequence in the textile production chain. For example, if three process segments, "fabric pretreatment," "dye injection," and "drying treatment," are identified within a time window, then in the constructed semantic path graph, these three nodes are connected into directed edges according to the physical process execution order. The edge weights are calculated by weighting the similarity of features of adjacent nodes, timestamp intervals, and equipment synchronization; higher weights indicate closer interdependence in actual production. The semantic smoothing process corrects locations with process jumps or path breaks through moving averages, abnormal connection removal, and replacement of low-confidence nodes. For example, if "fabric pretreatment" is detected and the process jumps directly to "inspection and packaging," the jump path is marked as abnormal, and a "dyeing process" node is inserted based on a node with higher semantic confidence in the adjacent window to repair the link logic. After processing, the output semantic fusion vector group retains the contextual information of the window embedding features and integrates the structural semantics of upstream and downstream process links, which can be used for subsequent process alignment and quality assessment modeling.

[0109] Step S145: Reorganize the semantic fusion vector group into time series data according to the time window order of the time window structure set, and perform structured encapsulation to obtain windowed fusion feature data.

[0110] In this embodiment, the semantic fusion vector group is reorganized into time series data according to the time window order of the time window structure set. This time series data will contain feature information of each time window in the production process, ensuring the temporality and continuity of the data. After obtaining the time series data, it will be structured and encapsulated, that is, the data will be formatted to meet the requirements of subsequent analysis and application. For example, if the data in each time window contains temperature, humidity, camera data, and texture information, then through structured encapsulation, the final data will be transformed into a format with clear fields and structure, facilitating subsequent processing and storage.

[0111] Optionally, step S15 specifically includes:

[0112] Step S151: Perform semantic matching between the semantic features in the windowed fusion feature data and the preset textile production process framework, and construct a process node transmission relationship diagram based on the semantic matching results.

[0113] In this embodiment, the generated windowed fusion feature data is further semantically matched with the preset process flow framework in the textile production management system. The preset process flow exists in the form of a "process node template library," where each node defines a standardized name (such as "fabric pretreatment," "dyeing process," "drying treatment," "setting and ironing," etc.), typical equipment type, operation time window range, and feature description. The semantic matching process uses each vector in the windowed fusion feature data as a matching item and compares it with each node in the template library using cosine similarity calculation. A semantic similarity threshold of 0.75 is set during matching. If the semantic features in the windowed fusion feature data have a similarity to the semantic description of the template node exceeding this threshold, it is considered that the corresponding process step can be correctly matched. When multiple candidate nodes meet the matching conditions, the node with the closest timestamp is selected for binding first. After matching, the semantic path structure is reconstructed according to the time sequence to generate a process node transmission relationship graph, where nodes in the graph represent the actually identified process links, and edges represent the data flow relationship and equipment interaction channels between links. For example, the weight of the edge between the "dyeing process" node and the "drying process" node will reflect whether these two steps are completed by the same equipment in actual production, whether there is a direct data transmission path, etc., which will facilitate subsequent process consistency verification and abnormal link location.

[0114] Step S152: Perform topological sorting on the process node transmission relationship graph, and perform data flow dependency modeling on the sorted node transmission relationship graph to obtain the data flow dependency chain;

[0115] In this embodiment, a topological sort is performed on the process node transmission relationship graph. The topological sort is based on the execution order of each process node. For example, if one process node represents "raw material preparation" and another represents "dyeing process," the topological sort ensures that "raw material preparation" precedes "dyeing process." During the sorting process, nodes with an in-degree of 0 (i.e., nodes without a preceding process node) are first identified. These nodes are then processed sequentially in chronological order, gradually reducing the in-degree of subsequent nodes until all nodes have been processed. After sorting, data flow dependency modeling is performed. This involves analyzing the input and output data of each process node and establishing the data transmission relationships between each node. Based on the dependencies between nodes, dependency chains are established to ensure that the data source for each node can only be passed to the next node after the previous node has completed its processing.

[0116] Step S153: Construct a cross-node data flow connector based on the data flow dependency chain;

[0117] In this embodiment, cross-node data flow connectors are constructed based on the data flow dependency chain. Each cross-node data flow connector is used to connect the data flow channels between two or more process nodes. For example, a connector connecting the "dyeing process" node and the "quality inspection" node will allow the data flow after dyeing to be transmitted to the inspection system. During implementation, bandwidth and latency requirements are specified for each connector; for example, the data transmission rate of each connector is 10MB / s, and the latency is no more than 5ms, to simulate the communication conditions between different nodes in the production process. The cross-node data flow connectors are configured according to the actual needs in the process flow to ensure efficient data flow transmission.

[0118] Step S154: Use cross-node data flow connectors to connect the data flows of each production node in the process node transmission relationship diagram, and perform spatiotemporal consistency verification to obtain the textile production chain data flow.

[0119] In this embodiment, cross-node data flow connectors are used to chain the data flows of various process nodes. The data flow connectors transmit data between process nodes and ensure that data is transmitted correctly within a specified time. For example, between the "raw material preparation" node and the "dyeing process" node, the data flow connector will transmit raw material information and production plan data to ensure that the subsequent dyeing process receives the correct input data. For each connector, a spatiotemporal consistency check is set to ensure that data transmission meets time and space requirements. A time difference threshold of ±30 seconds is set to ensure that the timestamp of data transmission is within the correct time window. If data fails to be transmitted within the specified time, the connector will be marked as abnormal. In this way, the data flow of all process nodes is ensured to follow the production sequence, thereby ultimately constructing a complete textile production chain data flow.

[0120] Optionally, the analysis of yarn dyeing characteristic patterns in step S2 specifically includes:

[0121] Extract yarn characteristic data and pH value data of the dyeing process from the data stream of the textile production chain;

[0122] In this embodiment, yarn characteristic data and pH data from the dyeing process are extracted from the textile production chain data stream. Yarn characteristic data includes the physical properties of the yarn, such as yarn density, yarn strength, and twist. The pH data from the dyeing process is derived from real-time pH measurements recorded in the dyeing tank. The textile production chain data stream acquires this information via sensors and records it in a time-series format. The data acquisition frequency is set to once per second to ensure high-precision real-time monitoring. During this process, the yarn characteristic data and pH data are synchronized to ensure that the data from both are consistent at the same point in time.

[0123] Obtain the textile yarn property library, and extract the corresponding yarn properties from the textile yarn property library based on the yarn characteristic data to obtain yarn property data;

[0124] In this embodiment, a textile yarn property database is obtained through expert experience or by consulting relevant experimental data. This database contains property information of different types of yarn, such as yarn density, strength, fiber length, and surface roughness. The data in the database is classified according to different yarn types, and each type of yarn has unique production processes and physical properties. Based on the yarn characteristic data extracted from the production chain data stream, the matching yarn properties are extracted by comparing them with the data in the yarn property database. For example, if the yarn density in the production data is 0.8 g / cm³... 3 It will then search the yarn property library for yarns with a density of 0.8 g / cm³. 3 Relevant characteristic data, such as yarn tensile strength and elongation, are collected to construct a yarn characteristic dataset. Data matching in this process employs a combination of exact matching and fuzzy matching, with a fuzzy matching tolerance set at ±5%.

[0125] Spearman rank correlation analysis was performed on yarn characteristic data and pH data during the dyeing process, and feature pairs with Spearman correlation coefficients less than 0.7 were removed to obtain yarn-pH highly correlated feature pairs.

[0126] In this embodiment, Spearman rank correlation analysis is performed on the extracted yarn characteristic data and the pH data during the dyeing process. Spearman rank correlation is used to analyze monotonic relationships between variables, and is particularly suitable for nonlinear relationships. Specifically, the correlation between yarn density, yarn strength, and pH during the dyeing process is selected for evaluation. For example, the Spearman correlation coefficient between yarn density and dyeing pH is selected; if the value is greater than 0.7, a strong correlation is considered to exist; if the correlation coefficient is less than 0.7, the feature pair is removed from subsequent analysis. After data removal, feature pairs with high correlation are ultimately retained to ensure the accuracy and validity of the analysis results.

[0127] Based on the high correlation between yarn and pH value, variable relationship pattern recognition is performed. If the yarn density in the feature pair is strongly positively correlated with the dyeing pH value, then a high-density yarn dyeing feature pair is obtained; if the yarn strength in the feature pair is negatively correlated with the dyeing pH value, then a high-strength yarn dyeing feature pair is obtained.

[0128] In this embodiment, based on the results of Spearman rank correlation analysis, variable relationship pattern recognition is performed on highly correlated yarn characteristics and pH value feature pairs. If a strong positive correlation (e.g., correlation coefficient above 0.85) is observed between yarn density and dyeing pH value in a feature pair, it is identified as a "high-density yarn dyeing feature pair"; if a negative correlation (e.g., correlation coefficient -0.72) is observed between yarn strength and dyeing pH value, it is identified as a "high-strength yarn dyeing feature pair". During implementation, for different types of feature pairs, each pair of data is further verified using a pattern recognition algorithm to ensure the scientific validity of the variable relationship pattern. For example, high-density yarn may require a lower pH value during the dyeing process, while high-strength yarn may exhibit a negative response to pH changes.

[0129] The high-density yarn dyeing feature pairs and the high-intensity yarn dyeing feature pairs are merged to obtain the yarn dyeing feature pattern.

[0130] In this embodiment, data from the identified "high-density yarn dyeing feature pairs" and "high-strength yarn dyeing feature pairs" are merged. The merging process involves combining related data from both pairs to construct a comprehensive yarn dyeing feature pattern. This pattern reflects the reaction characteristics of different types of yarn during the dyeing process, facilitating data analysis and optimization in production. For example, the dataset for the high-density yarn dyeing feature pairs includes data on the relationship between yarn density and dyeing pH value, while the dataset for the high-strength yarn dyeing feature pairs includes data on the negative correlation between yarn strength and pH value. During implementation, the specific data merging operation is processed according to the feature type to ensure accurate data alignment, ultimately generating a yarn dyeing feature pattern containing both features. This pattern can be applied to predict and optimize yarn dyeing behavior during production.

[0131] Optionally, step S3 specifically includes:

[0132] Step S31: Enhance the feature space according to the processing technology, set the set of process feature quantification parameters, and encode the process feature vectors of each production node in the textile production chain data flow to generate a node feature encoding set.

[0133] In this embodiment, within the existing enhanced feature space of the processing technology, a set of parameters for quantization encoding is determined, including five key process parameters: dye liquor temperature (°C), dyeing duration (min), weaving tension (N), drying temperature (°C), and finishing pressure (kPa). Encoding ranges are set for each parameter; for example, the dye liquor temperature is between 60°C and 130°C, with each 10°C increment; the tension range is set to 5–25N, with each 5N increment. Then, for each production node in the textile production chain data stream, its original process parameters are read, and their values ​​are converted into corresponding integer labels according to the above encoding ranges. Taking the "dyeing node" as an example, if its dye liquor temperature is recorded as 110°C, dyeing time as 35 min, and tension as 15N, it is mapped to "5", "3", and "3" respectively. All converted labels are merged into a node feature vector [5,3,3,...]. This process traverses all nodes, ultimately generating a node feature encoding set containing the corresponding process feature vector for each node. The structure is stored as a JSON array, with each item containing the node ID and its corresponding feature vector.

[0134] Step S32: Perform node comparison learning based on the node feature encoding set, evaluate the similarity of process feature transfer between adjacent nodes, and if the similarity is less than 0.85, mark it as a potential process mutation point, thereby generating a set of potential mutation node pairs;

[0135] In this embodiment, after obtaining the node feature encoding set, the feature vectors of each pair of adjacent nodes in the production chain are read sequentially. For example, for node A (ID = ND101) and node B (ID = ND102), their feature vectors are [5,3,3,4,2] and [5,3,2,4,2], respectively. Using the Euclidean distance of the vectors as the difference metric, the similarity score S = 1 / (1 + √∑(xi - yi)) is calculated. 2 If S is lower than the set threshold of 0.85, the node pair (ND101-ND102) is recorded as a potential process variation pair. This operation performs batch processing on all sequential nodes in the entire production chain, and finally filters out a set of potential variation node pairs that meet the conditions. Each record saves two node IDs, corresponding feature vectors and similarity values, and is stored in a CSV file for subsequent processing.

[0136] Step S33: Perform temporal consistency analysis on the potential variant node pair set. Within a preset time window of 30 minutes, select node pairs that conform to the transmission order of the textile production chain data flow and generate a preliminary spatiotemporally consistent node chain.

[0137] In this embodiment, time sequence and data flow logic verification is performed on the obtained set of potential mutated node pairs. First, the timestamp field of each node pair in the production data record table is read. For example, node ND101's completion time is 2024-11-12 10:30:00, and node ND102's start time is 2024-11-12 10:42:00. A time window of 30 minutes is set. If the time difference between two nodes is between [0, 1800] seconds, and the node numbers have a clear upstream / downstream relationship in the production chain (e.g., from the same production line number Line_05), then the flow order of the node pair is considered to meet the requirements. These node pairs that meet the conditions are written into a new data table as a preliminary spatiotemporally consistent node chain. Each record includes the starting node ID, the target node ID, the time difference, and the production line number.

[0138] Step S34: Based on the preliminary spatiotemporally consistent node chain, extract the position information and timestamp of each pair of nodes in the spatial layout, and calculate the node transfer delay. If the node transfer delay exceeds 120s, mark the node pair as having a transfer anomaly, thereby generating cross-chain spatiotemporal feature pairs.

[0139] In this embodiment, the spatial location information of each node is extracted from the initially selected spatiotemporally consistent node chain, namely the device number (e.g., DYE_M01, WOVEN_T03) and the workshop number (e.g., Zone_A, Zone_B). Then, the time difference between each node pair is calculated in seconds to obtain the flow delay value. For example, the delay between ND101 and ND102 is 720 seconds. A threshold of 120 seconds is set, and any node pair with a flow delay exceeding this threshold is marked as "flow abnormal". The specific marking information includes four fields: "Node Start ID", "Node End ID", "Delay Time (seconds)", and "Abnormal Type = Flow Delay". Finally, a cross-chain spatiotemporal feature pair set is generated and stored in the database table Anomaly_Feature_Pairs for subsequent use.

[0140] Step S35: Perform feature hash encoding on the cross-chain spatiotemporal feature pairs, integrate all fingerprint identifiers, and construct a cross-chain spatiotemporal stamp fingerprint set.

[0141] In this embodiment, for each feature pair in the generated cross-chain spatiotemporal feature pair set, its key fields are combined and encoded. These combined fields include the node start-point ID (e.g., ND101), end-point ID (ND102), flow delay time (seconds), and anomaly type code (e.g., 01 represents "delay"). These fields are concatenated into a raw string, such as "ND101_ND102_720_01," and then converted into a fixed-length 128-bit hash fingerprint using an MD5 hashing tool, such as "3e4a...b8c7." This operation is performed sequentially on all anomalous node pairs, and the generated hash fingerprints are uniformly stored in a cross-chain spatiotemporal stamp fingerprint set. The set is organized in key-value pair format, where the key is the fingerprint code and the value is its corresponding original record data. Finally, it is encapsulated into a structured JSON document for subsequent querying and comparison.

[0142] Optionally, the anomaly detection of finished textile pixels in step S4 specifically includes:

[0143] Textile production workshop camera data that has been standardized and dynamically time-normalized is subjected to texture feature enhancement processing to extract texture direction, frequency and contrast features, resulting in an enhanced texture image feature matrix.

[0144] In this embodiment, based on the standardized and dynamically time-warped video data from a textile production workshop, an image sequence of 1 frame per second is selected as the analysis sample. The gray-level co-occurrence matrix method is used to calculate the texture direction features of each frame in four directions: 0°, 45°, 90°, and 135°. The frequency features of the main texture region in the image are extracted, with the frequency distribution calculated at a granularity of 5 pixels. Simultaneously, the contrast value of each window region in the image is obtained using a local contrast measurement method, with the window size set to 8×8 pixels. Finally, an enhanced texture image feature matrix is ​​constructed by fusing texture direction, frequency, and contrast indicators.

[0145] A pixel-level image anomaly detection model is constructed based on the enhanced texture image feature matrix. The model detection threshold is set to a significance score greater than 0.6 to identify abnormal pixel regions and output a preliminary abnormal overlay image.

[0146] In this embodiment, after obtaining the enhanced texture image feature matrix, a pixel-level anomaly detection model for production images is constructed based on it. This model employs a feature-driven, pixel-by-pixel classification approach. First, the image is divided into 16×16 pixel grid cells. The directional, frequency, and contrast features extracted from each grid cell are normalized using standard deviation and then concatenated into a 96-dimensional feature vector. Subsequently, normal and abnormal regions from 1000 manually selected production sample images are used as training data. A pixel-by-pixel sliding window scanning method is used to statistically analyze the feature distribution range of each category, and a local feature density table is constructed during the training phase as a reference for pixel scoring. The model uses a discrimination mechanism based on the degree of density deviation to score each pixel, determining whether its features deviate from the normal distribution range of its category. The scoring method is based on the standardized deviation of Euclidean distance and feature center, with a score range of [0,1]. During the inference phase, for any input frame, the model compares the enhanced texture features within each pixel sliding window with the constructed feature density table and outputs the corresponding anomaly saliency score map. The aforementioned feature matrix is ​​input into an image anomaly detection model, which is implemented based on pixel block saliency calculation. This model generates a saliency score map for each region in the image, with a scoring threshold set to 0.6. This means that a pixel region is considered potentially anomaly only if its saliency score is higher than this value. The saliency score calculation is based on the statistical bias of the texture distribution, ultimately outputting a preliminary anomaly mask map for each frame. This mask map labels suspected anomaly regions in binary form.

[0147] Morphological structure analysis and pixel clustering are performed on the preliminary anomaly mask image to remove pseudo-anomaly noise areas with an area of ​​less than 50 pixels, and a set of anomaly region bounding boxes is generated based on the pixel cluster center as the annotation basis.

[0148] In this embodiment, the preliminary anomaly mask image is fed into the morphological processing module, where dilation-erosion operations are performed sequentially to remove isolated pixels at the edges. The kernel size is set to 3×3, and a pixel-distance-based clustering method is used to perform cluster analysis on the mask region. The cluster centers are merged based on a Euclidean distance of less than 15 pixels, and mask connected regions with an area of ​​less than 50 pixels are considered pseudo-anomalies and are discarded. Each final clustered region generates a bounding box, which is output using the minimum bounding rectangle method, forming a set of anomaly region bounding boxes.

[0149] The set of boundary boxes of abnormal regions is mapped back to the process nodes in the corresponding textile production chain data stream, and the timestamps of the video data from the textile production workshop are combined to generate a set of finished product pixel anomaly detection labels.

[0150] In this embodiment, the position index of each bounding box in the abnormal region bounding box set is synchronized with the timestamp of the camera data, and the corresponding process node code in the textile production chain data stream is retrieved for that time point. By matching the bounding box time with the production node time interval, a set of finished product pixel anomaly detection tags is generated. Each tag contains the node number, the boundary coordinates of the abnormal region, and the time point of its occurrence, which are used for subsequent abnormal node location and source tracing.

[0151] Most importantly, the initial abnormal flooding image is output as follows:

[0152] Local anomaly factor scores are evaluated for the local neighborhood of each pixel in the feature matrix of the enhanced texture image. The neighborhood size is set to k=20, and the corresponding local anomaly factor score distribution map is obtained.

[0153] In this embodiment, after extracting the enhanced texture image feature matrix, a local feature neighborhood is constructed for each pixel in the matrix, centered on that pixel and containing 20 neighboring pixels within a radius. The neighborhood size is fixed at k=20. Within this local neighborhood, the local outlier degree of each pixel is calculated based on indicators such as texture direction consistency, contrast difference, and frequency variation amplitude between each pixel and its neighboring pixels, and a local anomaly factor score is generated accordingly. This score is calculated by statistically analyzing the average relative distance between the pixel and other pixels in the neighborhood in the multidimensional texture feature space, while also assessing whether the pixel is a structural anomaly point within the neighborhood. This processing flow is executed once for each pixel in the entire image, thereby generating a local anomaly factor score distribution map for the entire image, reflecting the locations where there are significant structural abrupt changes in texture features.

[0154] The local anomaly factor scores in the local anomaly factor score distribution map are standardized to the [0,1] interval and used as pixel-level significance scores.

[0155] In this embodiment, all scores in the generated local anomaly factor score distribution map are standardized to the [0,1] interval using min-max normalization to maintain score consistency and comparability across different images. Specifically, the maximum and minimum local anomaly factor scores appearing in the entire distribution map are first counted and denoted as LOF_max and LOF_min. Then, the original local anomaly factor score of each pixel is standardized using the formula (score-LOF_min) / (LOF_max-LOF_min). The score of each pixel after the transformation is its significance score, which numerically reflects the degree of feature difference of the pixel in its local region. The closer the score is to 1, the more likely the pixel is to be part of an anomaly region; the closer it is to 0, the more likely it belongs to a normal region with consistent texture structure.

[0156] Set the saliency score threshold to 0.6, filter all pixels with a saliency score greater than 0.6, and generate a preliminary binary anomaly mask map, where 1 represents a potential anomaly region and 0 represents a normal background region.

[0157] In this embodiment, after obtaining the saliency score map, it is further binarized to generate a preliminary anomaly mask map. During the processing, a saliency score threshold of 0.6 is set. Each pixel in the score map is traversed; if its saliency score is higher than 0.6, the pixel is marked as a potential anomaly pixel with a value of 1; otherwise, it is marked as a background pixel with a value of 0. In the final generated binary image, regions with a pixel value of 1 constitute potential anomaly region candidates, reflecting a significant difference in texture features between the region and its surrounding environment, while regions with a pixel value of 0 are considered background. The generated anomaly mask map will serve as the input basis for subsequent anomaly region morphology analysis and process localization.

[0158] Optionally, the evaluation of batch production quality of textile production line in step S4 specifically includes:

[0159] The finished product pixel anomaly detection tag set is timestamped and batch-aligned according to the node time window in the textile production chain data flow to generate a batch-level anomaly tag matrix.

[0160] In this embodiment, the previously constructed finished product pixel anomaly detection tag set is batch-merged according to the time windows of each production node in the textile production chain data flow. An aggregation cycle of 30 minutes is used, and anomaly tags within the same time period are categorized based on their timestamps to complete aggregation along the timeline. Subsequently, batch alignment is performed on the aggregated tags at the process node level. This involves integrating the tags based on the production line number and batch serial number of the corresponding node in the textile production chain, generating a two-dimensional batch-level anomaly tag matrix. The rows of the matrix represent batch numbers, and the columns represent anomaly label types.

[0161] Abnormal pixel features are extracted from the batch-level abnormal label matrix to obtain abnormal pixel density, abnormal distribution dispersion and abnormal region coverage.

[0162] In this embodiment, the constructed batch-level anomaly label matrix will serve as the data foundation for further extraction of batch quality-related image features. To this end, it is necessary to separately calculate the proportion of abnormal pixels in each batch of images, defining this proportion as "abnormal pixel density," and setting the calculation method for this indicator as the ratio of the number of abnormal pixels to the total number of pixels. Simultaneously, the spatial discreteness of the centroid distribution of each abnormal region is measured, using the discreteness index to reflect the spatial diffusion degree of each abnormal block in the finished image. Furthermore, the ratio of the sum of the areas of all abnormal regions to the image area is used as the abnormal region coverage rate, quantifying the spatial feature distribution of anomalies in each batch of images from three perspectives.

[0163] A batch quality quantitative evaluation system is constructed based on abnormal pixel density, abnormal distribution dispersion, and abnormal region coverage.

[0164] In this embodiment, after feature extraction is completed, the three types of abnormal image features mentioned above are used as input to establish a batch quality quantification evaluation system. During the evaluation process, weight parameters are set for each feature. For example, the weights of abnormal pixel density, abnormal distribution dispersion, and abnormal region coverage are set to 0.4, 0.3, and 0.3, respectively, and a weighted scoring function is constructed based on a normalized weighting mechanism.

[0165] The batch quality quantitative evaluation system is used to score the batch quality of each batch in the batch-level anomaly label matrix. The scoring range is set to [0,1]. Batches with a quality score lower than 0.6 are marked as batches to be traced. A preliminary quality score result set is generated.

[0166] In this embodiment, a batch quality quantification evaluation system is used to score the quality of each batch of images. The score value is standardized to the range [0,1], and a scoring threshold of 0.6 is set. If the score value of a batch is lower than this threshold, the batch is automatically classified into the "batch to be traced" label, and a preliminary quality score result set is compiled and output.

[0167] The preliminary quality score result set is subjected to abnormal fluctuation identification. If the score fluctuation of three consecutive batches exceeds 0.15, it is marked as a batch group with trend quality deviation. The process node identifiers of the textile production chain data flow corresponding to the abnormal pixel area are integrated to output the final production line batch quality assessment report.

[0168] In this embodiment, after the quality scoring is completed, abnormal fluctuation trends are identified in the preliminary result set. In this step, for every three consecutive batches of scores in time, the score difference between the preceding and following batches is calculated. If the score fluctuation exceeds 0.15, it is marked as a trend quality shift, and the corresponding three consecutive batches are grouped into a "trend shift batch group". Subsequently, the production chain node identifiers corresponding to the abnormal pixel areas covered by this batch group are integrated into the trend shift information and output together as the final production line batch quality assessment report.

[0169] Batches marked as pending traceability in the production line batch quality assessment report are classified as pending traceability production line batch data; batches with a quality score higher than 0.6 and not marked as trending quality deviation batches are classified as pending warehousing production line batch data, and the tag status is synchronized to the production chain data stream.

[0170] In this embodiment, batch-level classification is performed based on the scores and status labels of each batch in the report. Batches with scores below 0.6 and marked as "pending traceability" are uniformly classified as "production line batch data pending traceability"; while batches with scores above 0.6 and not marked as having a trend of quality deviation are classified as "production line batch data pending warehousing". Simultaneously, the label status information for these two types of batches is updated in the production chain data stream to ensure that subsequent traceability and warehousing processes can directly call the evaluation labels for judgment and processing.

[0171] Optionally, step S5 specifically includes:

[0172] Step S51: Perform fingerprint matching between the batch data of the production line to be traced and the cross-chain spatiotemporal stamp fingerprint set. Compare the time period of the textile production chain data chain corresponding to the batch data of the production line to be traced with the spatiotemporal stamp distribution range in the fingerprint set, extract the batch fingerprint subset, and generate batch-level fingerprint mapping data.

[0173] In this embodiment, the recorded production line batch data to be traced is classified by batch number. The start and end timestamps of each batch are extracted and matched against all fingerprint items in the cross-chain spatiotemporal stamp fingerprint set. Each fingerprint item contains a node number, timestamp, spatial identifier, and hash code. The time matching threshold is set to ±60 seconds, and the spatial location error tolerance is 5 meters. During the matching process, by comparing the overlap between the node event occurrence time of the batch data to be traced and the timestamp field in the fingerprint set, spatiotemporally matching fingerprint items are selected to form a subset of the batch's fingerprints. For each matching item, the associated node number, matching error range, and hash value are recorded, ultimately generating batch-level fingerprint mapping data for subsequent process trajectory reconstruction.

[0174] Step S52: Based on the node feature hash values ​​in the batch-level fingerprint mapping data, reconstruct the process flow trajectory of the corresponding textile production chain data flow to obtain the reconstructed trajectory node sequence.

[0175] In this embodiment, all node feature hash values ​​extracted from the batch-level fingerprint mapping data are arranged in chronological order, duplicate or discontinuous hashes are removed, and the hashes are mapped back to the original textile production chain structure according to the binding rules between the hashes and node process numbers. A sequential lookup table method is used to locate the production node represented by the hash value, such as sizing, spinning, desizing, dyeing, and finishing. These nodes are then constructed into a process trajectory sequence in ascending order of their timestamps, and the start time, duration, process type code, and spatial coordinates of each node are marked. After construction, the reconstructed trajectory node sequence is output for subsequent anomaly localization operations.

[0176] Step S53: Perform anomaly interpolation fitting analysis on the reconstructed trajectory node sequence to identify missing, skipped, or time-reversed node segments in the trajectory and generate anomaly trajectory segment marker set;

[0177] In this embodiment, for the reconstructed trajectory node sequence, the time interval between every two adjacent nodes and the continuity of the process flow are examined. If the time interval between a node and the previous node exceeds the normal process standard interval (set to 300 seconds), or the node number does not conform to the preset continuous process coding rules (e.g., the desizing process should precede the dyeing process), it is marked as a "jump segment" or "process breakpoint". If the time sequence shows that the time of the later node is less than the time of the earlier node, it is marked as a "time sequence reversal segment". All marked items are summarized to form an abnormal trajectory segment mark set, and each abnormal segment is accompanied by fields such as the preceding and following node numbers, time difference, abnormality type, and duration, as the basis for abnormal location.

[0178] Step S54: Perform cross-source data tracing on the abnormal trajectory segment marker set, and use the textile production chain data chain to retrieve the environmental status, operation records and equipment operating parameters of the corresponding node segment, and extract the abnormal cause feature set;

[0179] In this embodiment, for each node segment involved in the abnormal trajectory segment marker set, the corresponding time period's process logs, environmental status records, and equipment operating parameters are retrieved one by one from the textile production chain data chain. Retrieval items include, but are not limited to, process tank temperature, pH value, workshop humidity, motor load, equipment start / stop status, and operator employee number. A time window is set to ±120 seconds before and after the abnormal segment, and corresponding logs and data streams are matched according to the node number. After all data is converted to a unified format, key abnormal feature fields are extracted, such as tank temperature jumps greater than 5°C, dye liquor pH fluctuations exceeding 0.8, and equipment speed drops of more than 20%. These fields are ultimately summarized into an abnormal cause feature set for reference in the pattern recognition stage.

[0180] Step S55: Perform pattern recognition based on the set of abnormal cause features to identify the abnormal patterns of equipment oscillation deviation and dyeing transition zone drift, and obtain an abnormal source tracing report;

[0181] In this embodiment, based on the obtained set of abnormal cause features, a comparison is performed according to the defined abnormal feature pattern template. If the equipment speed fluctuations exceed 20% three times consecutively, the equipment load alternates periodically between high and low, or frequent starts and stops occur, it is classified as an "equipment oscillation-type deviation abnormal pattern." If the pH value in the dyeing section rises slowly and continuously at multiple nodes and matches the yarn density characteristics, it can be classified as a "dyeing transition zone drift abnormal pattern." A unique identifier is generated for each type of abnormality using a pattern matching table, and an abnormality tracing report is compiled. The report includes information such as the abnormal segment number, abnormal pattern type, corresponding node and time, and affected process segment.

[0182] Step S56: Combine the anomaly traceability report with the yarn dyeing characteristic pattern to establish a batch-level dyeing process reversibility assessment model to evaluate the feasibility of dyeing optimization. If the dyeing optimization feasibility is higher than or equal to 0.6, the anomaly traceability report of the corresponding batch is uploaded to the workshop production management platform, and the corresponding batch status is marked as pending dyeing optimization. The workshop production management platform then schedules the dyeing optimization process. If the dyeing optimization feasibility is lower than 0.6, the anomaly traceability report of the corresponding batch is uploaded to the workshop production management platform, and the corresponding batch status is marked as pending recycling. The workshop production management platform then schedules the batch recycling process.

[0183] In this embodiment, after generating an anomaly tracing report, the yarn dyeing characteristic mode is invoked, and a reversibility assessment is performed based on the anomaly type and impact characteristics recorded in the report. The assessment process parameters include process adjustment tolerance (pH adjustment within 0.15 is controllable), dye liquor circulation time, and yarn dye absorption delay threshold. A score is assigned based on whether the process conditions allow for adjustment. If the reversibility score is higher than or equal to 0.6, the anomaly tracing report is uploaded through the workshop MES platform interface, and the batch is marked as "to be dyed optimized," and pushed to the scheduling process. Conversely, if the score is lower than 0.6, it is marked as "to be recycled," and the backend executes the batch recycling process. Simultaneously, the status is updated and synchronized to the production chain data stream to maintain production line consistency.

[0184] Of particular importance is that the identification of the oscillation-type deviation anomaly pattern and the staining transition zone drift anomaly pattern in step S55 are specifically as follows:

[0185] Frequency domain analysis is performed on the equipment operating parameters in the abnormal cause feature set to extract the main peak of the spectrum and set the frequency fluctuation threshold to ±0.5Hz. If the main peak of the frequency is detected to fluctuate more than twice within a unit time, it is determined to be an abnormal mode of equipment oscillation deviation.

[0186] In this embodiment, when performing frequency domain analysis on the equipment operating parameters in the abnormal cause feature set, continuous time-series data including motor current, motor speed, and mechanical vibration sensor signals are selected. The sampling frequency is set to 100Hz, and the sampling duration is 120 seconds. After time-window weighting and signal denoising, the time-series data is converted to the frequency domain, and the spectral curve is extracted and the main peak frequency is detected. The frequency fluctuation threshold is set to ±0.5Hz, and a 1-minute unit time window is divided. If the main peak frequency is detected to exceed the threshold range twice or more within a certain unit time, combined with the number and amplitude of peaks in the sensor vibration waveform, it is determined that the equipment has unstable operating behavior. Such fluctuations are marked as "equipment oscillation-type deviation anomaly mode", and the specific time, equipment number, main peak frequency value, and number of fluctuations are recorded as the core parameters for tracing the anomaly mode.

[0187] A joint analysis of the pH value of the dyeing solution and the movement trajectory of the fabric surface in the abnormal cause feature set is performed. If the pH value of the dyeing solution changes by more than 0.3 units / minute and the deviation of the fabric surface movement trajectory exceeds 20 pixels within 3 consecutive nodes in the node segment, it is determined to be an abnormal mode of dyeing transition zone drift.

[0188] In this embodiment, for the joint analysis and processing of the dyeing solution pH value and the fabric movement trajectory, the real-time pH monitoring value of the dyeing bath solution and the coordinates of the fabric movement trajectory acquired by the camera within a continuous time period are first selected from the abnormal cause feature set. The pH data sampling interval is set to 10 seconds, and the fabric movement trajectory is calculated for inter-frame displacement using the optical flow method, with a displacement pixel quantization accuracy of 5 pixels. According to the trajectory node segment division, every three consecutive nodes form an analysis window. Within this window, the pH change rate is calculated. If it exceeds 0.3 units / minute, and the maximum amplitude of the fabric displacement exceeds 20 pixels, it is marked as "dyeing transition zone drift anomaly mode". At the same time, the start and end time, trajectory offset direction, maximum pH change amplitude, and dyeing node number where the anomaly occurs are recorded within this window and included in the anomaly tracing mode identifier.

[0189] The abnormal patterns of equipment oscillation deviation and abnormal patterns of dyeing transition zone drift are classified and labeled to form an abnormality tracing report.

[0190] In this embodiment, the "equipment oscillation-type deviation anomaly mode" and the "dyeing transition zone drift anomaly mode" identified in the aforementioned analysis are each assigned a unique anomaly mode code, and are archived and organized according to the time, location, and corresponding production line batch number of the anomaly. Each anomaly mode data item includes the anomaly type, trigger parameters (such as frequency peak, trajectory offset, pH change rate), corresponding node number, anomaly duration, and number of fluctuations. When generating an anomaly tracing report, the system sorts the above classification information in chronological order to form a structured report document. This report file supports output in both JSON and CSV formats and is simultaneously uploaded to the production data analysis platform to provide data support for subsequent reversibility analysis and optimization of process scheduling.

[0191] Optionally, this specification also provides a data fusion system for textile production traceability, used to execute the data fusion method for textile production traceability as described above. The data fusion system for textile production traceability includes:

[0192] The data fusion module is used to acquire sensor data and camera data from the textile production workshop, and to perform data fusion of the textile production chain data stream to obtain the textile production chain data stream.

[0193] The yarn feature pattern recognition module is used to analyze yarn dyeing feature patterns based on the data flow of the textile production chain, and to perform feature cross calculations based on the yarn dyeing feature patterns to generate a processing technology enhancement feature space.

[0194] The fingerprinting module is used to perform process-based fingerprinting of textile production chain data streams based on the enhanced feature space of the processing technology, resulting in a cross-chain spatiotemporal stamp fingerprint set.

[0195] The anomaly detection module is used to detect anomalies in the pixels of finished textile products based on camera data from the textile production workshop, and to evaluate the production quality of textile production line batches based on the anomaly detection results, thereby obtaining batch data of production lines to be traced and batch data of production lines to be put into storage.

[0196] The abnormal production traceability module is used to trace abnormal production data of production line batches by using cross-chain spatiotemporal stamp fingerprint set to obtain an abnormal traceability report; evaluate the feasibility of dyeing optimization for abnormal batches based on the abnormal traceability report, and upload it to the workshop production management platform to execute the dyeing optimization task;

[0197] The warehouse transmission module is used to transmit batch data of production lines to be put into storage and batch data of production lines to be traced after dyeing optimization to the workshop warehouse management platform to perform warehouse filing tasks.

[0198] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0199] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A data fusion method for textile production traceability, characterized by, The method comprises the following steps: Step S1: acquiring textile production workshop sensor data and textile production workshop camera data, and performing textile production chain data stream data fusion to obtain textile production chain data stream; Step S2: analyzing yarn dyeing feature mode according to the textile production chain data stream, and generating processing technology strengthening feature space according to the yarn dyeing feature mode; Step S3: performing textile production chain data stream process fingerprinting on the textile production chain data stream based on the processing technology strengthening feature space to obtain a cross-chain space-time stamp fingerprint set; Step S4: performing textile finished product pixel point anomaly detection based on the textile production workshop camera data, and evaluating textile production line batch production quality according to the anomaly detection result, thereby obtaining to-be-traced production line batch data and to-be-warehoused production line batch data; Step S5: performing production line abnormal production tracing on the to-be-traced production line batch data using the cross-chain space-time stamp fingerprint set to obtain an abnormal tracing report; evaluating abnormal batch dyeing optimization feasibility according to the abnormal tracing report, and uploading to a workshop production management platform to execute a dyeing optimization task; Step S6: transmitting the to-be-warehoused production line batch data and the to-be-traced production line batch data after the dyeing optimization task to a workshop warehouse management platform to execute a warehouse recordation task.

2. The data fusion method for textile production traceability according to claim 1, characterized in that, Step S1 is specifically: Step S11: acquiring textile production workshop sensor data and textile production workshop camera data; Step S12: performing data preprocessing and standardization on the textile production workshop sensor data and the textile production workshop camera data respectively to obtain standardized sensor data and standardized camera data; Step S13: performing dynamic time warping based on the standardized sensor data and the standardized camera data to obtain a time series alignment data set; Step S14: extracting joint context features of the time series alignment data set, constructing a multi-modal fusion feature representation, and thereby obtaining windowed fusion feature data; Step S15: logically reconstructing the windowed feature data, concatenating the data link of the production flow relationship, and thereby obtaining the textile production chain data stream.

3. The data fusion method for textile production traceability according to claim 2, characterized in that, Step S14 is specifically: Step S141: setting a sliding time window length of [30s, 120s], synchronously intercepting sensor data and camera data frame sequences in the same time period in each sliding time window in the time series alignment data set to obtain a time window structure set; Step S142: performing trend gradient coding on the sensor data in the time window structure set, and simultaneously performing principal component feature extraction on the camera data frame sequence to respectively obtain a preliminary feature vector in the sensor modal and a preliminary feature vector in the camera modal; Step S143: aligning the preliminary feature vector in the sensor modal and the preliminary feature vector in the camera modal, performing inter-modal space-time correlation modeling, and strengthening signal fusion strength, thereby obtaining space-time alignment embedded features; Step S144: performing upstream and downstream semantic path graph construction and semantic smoothing on the space-time alignment embedded features of the continuous sliding time window to obtain a semantic fusion vector group; Step S145: reorganize the semantic fusion vector group in the time window order of the time window structure set into time series data, and perform structured packaging to obtain windowed fusion feature data.

4. The data fusion method for textile production traceability of claim 2, wherein, Step S15 is specifically: Step S151: perform semantic matching of the semantic features in the windowed fusion feature data with a preset textile production process flow framework, and construct a process node transmission relationship graph based on the semantic matching result; Step S152: topologically sort the process node transmission relationship graph, and model the data flow dependency of the sorted node transmission relationship graph to obtain a data flow dependency chain; Step S153: construct a cross-node data flow connector according to the data flow dependency chain; Step S154: use the cross-node data flow connector to connect the data flow of each production node in the process node transmission relationship graph, and perform space-time consistency verification to obtain a textile production chain data flow.

5. The data fusion method for textile production traceability of claim 1, wherein, The analysis of the yarn dyeing feature mode in step S2 is specifically: According to the textile production chain data flow, extract yarn feature data and dyeing process pH value data; Obtain a textile yarn characteristic library, and extract corresponding yarn characteristics from the textile yarn characteristic library according to the yarn feature data to obtain yarn characteristic data; Perform Spearman rank correlation analysis on the yarn characteristic data and the dyeing process pH value data, and eliminate feature pairs with a Spearman correlation coefficient less than 0.7 to obtain yarn-pH value high correlation feature pairs; According to the yarn-pH value high correlation feature pairs, perform variable relationship mode recognition, if the yarn density in the feature pair is strongly positively correlated with the dyeing pH value, a high-density yarn dyeing feature pair is obtained; If the yarn strength in the feature pair is negatively correlated with the dyeing pH value, a high-strength yarn dyeing feature pair is obtained; Merge the high-density yarn dyeing feature pair and the high-strength yarn dyeing feature pair to obtain a yarn dyeing feature mode.

6. The data fusion method for textile production traceability of claim 5, wherein, Step S3 is specifically: Step S31: According to the processing technology enhancement feature space, set the process feature quantization parameter set, and encode the process feature vectors of each production node in the textile production chain data flow to generate a node feature encoding set; Step S32: Perform node comparison learning based on the node feature encoding set to evaluate the similarity of process feature transfer between adjacent nodes. If the similarity is less than 0.85, mark it as a potential process variation point to generate a set of potential variation node pairs; Step S33: Perform time sequence consistency analysis on the set of potential variation node pairs. Within a preset time window of 30 minutes, filter node pairs that meet the transmission order of the textile production chain data flow to generate a preliminary space-time consistent node chain; Step S34: According to the preliminary space-time consistent node chain, extract the position information and time stamp of each node pair in the spatial layout, and calculate the node flow transfer delay. If the node flow transfer delay exceeds 120s, perform flow transfer anomaly labeling on the node pair to generate a cross-chain space-time feature pair; Step S35: Feature hash encoding is performed on the cross-chain space-time feature pair, all fingerprint identifiers are integrated, and a cross-chain space-time stamp fingerprint set is constructed.

7. The data fusion method for textile production traceability of claim 1, wherein, The textile finished product pixel point anomaly detection in step S4 is specifically: The textile production workshop camera data which has been standardized and dynamic time warping is executed texture feature enhancement processing, the texture direction, frequency and contrast features are extracted, and an enhanced texture image feature matrix is obtained; According to the enhanced texture image feature matrix, a pixel-level image anomaly detection model is constructed, the model detection threshold is set to be greater than 0.6, and the abnormal pixel point area is identified, so as to output a preliminary abnormal mask image; The preliminary abnormal mask image is executed morphological structure analysis and pixel clustering processing, the pseudo abnormal noise area with an area less than 50 pixels is removed, and the abnormal area boundary box set is generated based on the pixel clustering center as the labeling basis; The abnormal area boundary box set is mapped back to the process node in the textile production chain data stream, and the finished product pixel anomaly detection label set is generated combined with the time stamp of the textile production workshop camera data.

8. The data fusion method for textile production traceability of claim 1, wherein, The evaluation of the textile production line batch quality in step S4 is specifically: The finished product pixel anomaly detection label set is timestamp aggregated, and the batch is aligned according to the node time window in the textile production chain data stream, and a batch-level abnormal label matrix is generated; Abnormal pixel feature extraction is performed on the batch-level abnormal label matrix to obtain abnormal pixel density, abnormal distribution dispersion and abnormal area coverage rate; Based on the abnormal pixel density, the abnormal distribution dispersion and the abnormal area coverage rate, a batch quality quantitative evaluation system is constructed; The batch quality quantitative evaluation system is used to score the quality of each batch in the batch-level abnormal label matrix, and the score range is set to [0, 1], wherein the batch with a quality score less than 0.6 is marked as a batch to be traced, and a preliminary quality score result set is generated; Abnormal fluctuation identification is performed on the preliminary quality score result set, if the score fluctuation amplitude of three consecutive batches exceeds 0.15, it is marked as a trend quality deviation batch group, and the abnormal pixel area corresponding to the process node identification of the textile production chain data stream is fused, and a final production line batch quality evaluation report is output; The batches marked as batches to be traced in the production line batch quality evaluation report are classified as batches to be traced in the production line batch data; the batches with a quality score higher than 0.6 and not marked as a trend quality deviation batch group are classified as batches to be in the warehouse in the production line batch data, and the label state is synchronized to the production chain data stream.

9. The data fusion method for textile production traceability of claim 1, wherein, Step S5 is specifically: Step S51: Perform fingerprint matching on the batches to be traced in the production line batch data and the cross-chain space-time stamp fingerprint set, compare the time period of the textile production chain data chain corresponding to the batches to be traced in the production line batch data with the space-time stamp distribution interval in the fingerprint set, extract a batch fingerprint subset, and generate batch-level fingerprint mapping data; Step S52: Based on the node feature hash value in the batch-level fingerprint mapping data, the process flow track of the corresponding textile production chain data stream is reconstructed, and a reconstructed track node sequence is obtained; Step S53: Abnormal interpolation fitting analysis is performed on the reconstructed track node sequence, the node segments with missing, jumping or time reversal in the track are identified, and an abnormal track segment label set is generated; Step S54: Cross-source data tracing is performed on the abnormal trajectory segment label set, the environment state, operation record and equipment operation parameter of the corresponding node segment are searched by using the textile production chain data chain, and an abnormal reason feature set is extracted; Step S55: Mode recognition is performed based on the abnormal reason feature set, a device shock type deviation abnormal mode and a dyeing transition zone drift abnormal mode are recognized, and an abnormal source report is obtained; Step S56: A batch-level dyeing process reversibility evaluation model is established in combination with the abnormal source report and the yarn dyeing feature mode, so as to evaluate the dyeing optimization feasibility, if the dyeing optimization feasibility is higher than or equal to 0.6, the abnormal source report of the corresponding batch is uploaded to the workshop production management platform, and the state of the corresponding batch is marked as to be dyed and optimized, and the dyeing optimization process is dispatched by the workshop production management platform; If the dyeing optimization feasibility is lower than 0.6, the abnormal source report of the corresponding batch is uploaded to the workshop production management platform, and the state of the corresponding batch is marked as to be recycled, and the batch recycling process is dispatched by the workshop production management platform.

10. A data fusion system for textile production traceability, characterized in that, The textile production oriented data fusion system for performing the textile production oriented data fusion method of claim 1 comprises: a data fusion module for acquiring textile production workshop sensing data and textile production workshop camera data, and performing textile production chain data stream data fusion to obtain textile production chain data stream; a yarn feature mode recognition module for analyzing yarn dyeing feature mode according to the textile production chain data stream, and performing feature cross calculation according to the yarn dyeing feature mode to generate processing process enhanced feature space; a fingerprinting module for performing textile production chain data stream process fingerprinting on the textile production chain data stream based on the processing process enhanced feature space to obtain cross-chain space-time stamp fingerprint set; an abnormality detection module for performing textile product pixel point abnormality detection based on the textile production workshop camera data, and evaluating textile production line batch production quality according to the abnormality detection result, so as to obtain to-be-traced production line batch data and to-be-warehouse production line batch data; an abnormal production source tracing module for performing production line abnormal production source tracing on the to-be-traced production line batch data by using the cross-chain space-time stamp fingerprint set to obtain an abnormal source report; evaluating abnormal batch dyeing optimization feasibility according to the abnormal source report, and uploading to the workshop production management platform to perform dyeing optimization task; a warehouse transmission module for transmitting the to-be-warehouse production line batch data and the to-be-traced production line batch data after the dyeing optimization task to the workshop warehouse management platform to perform warehouse record task.

Images