Clock manufacturing yield prediction method and system based on artificial intelligence

Abstract

The invention discloses a clock manufacturing yield prediction method and system based on artificial intelligence, and the method comprises the steps: building a CMAN cross-modal attention network, extracting the local features of all modes through a self-attention mechanism, calculating the semantic association between the modes through cross-modal attention, introducing a GNN graph neural network to model the dependency between the modes, and carrying out the prediction of the clock manufacturing yield. Performing feature alignment and interaction on the multi-modal data to obtain fused feature data; a BJM Bayesian joint model is constructed to predict the fused feature data, a longitudinal sub-model is utilized to predict a time sequence track of the equipment health state, and a survival sub-model is utilized to estimate failure time distribution of key equipment and raw materials to obtain a yield demand prediction result; and generating a clock yield adjustment strategy according to the yield demand prediction result. The yield demand prediction result is more suitable for the actual production scene, and the prediction deviation is greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a method and system for predicting watch manufacturing output based on artificial intelligence. Background Technology

[0002] Traditional watchmaking production forecasting methods fail to reflect equipment status and process fluctuations during production, and the collected data generally suffers from low quality and severe noise interference. This stems not only from inconsistencies in data acquisition equipment and statistical standards but also from environmental factors and unforeseen events. Secondly, watchmaking production data often exhibits distinct time-series characteristics, including both short-term random fluctuations and long-term trends and cyclical changes. Traditional forecasting methods typically rely on linear assumptions, making it difficult to accurately capture complex nonlinear dynamic changes, resulting in significant forecast errors and failing to meet practical needs. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a method and system for predicting watch manufacturing output based on artificial intelligence.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned artificial intelligence-based watch manufacturing output prediction method: Collect production data, market data, and supply chain data during the watch manufacturing process, and preprocess the collected data to obtain multimodal data; A CMAN cross-modal attention network is established, and the self-attention mechanism of Transformer is used to extract local features of each modality. Semantic associations between modalities are calculated through cross-modal attention. A GNN graph neural network is introduced to model inter-modal dependencies. The multimodal data is then aligned and interacted to obtain fused feature data. A joint BJM Bayesian model is constructed to predict the fused feature data. The longitudinal sub-model is used to predict the time trajectory of equipment health status, and the failure time distribution of key equipment and raw materials is estimated through the survival sub-model to obtain the production demand prediction results. A watch production adjustment strategy is generated based on the production demand forecast results.

[0005] Furthermore, in the aforementioned AI-based watch manufacturing output prediction method, the collection of production data, market data, and supply chain data during the watch manufacturing process, followed by preprocessing of the collected data to obtain multimodal data, including: Collect production data, market data, and supply chain data during the watchmaking process; Numerical data is cleaned using the mean imputation method, categorical data is cleaned using the mode imputation method, and outliers are identified using box plots. Data exceeding 1.5 times the upper or lower quartile range are considered outliers and replaced with the average of adjacent data to obtain cleaned production data. The cleaning production data is standardized and converted into normally distributed data with a mean of 0 and a standard deviation of 1. The processed production data, market data, and supply chain data are linked and integrated according to time and product batch to obtain multimodal data.

[0006] Furthermore, in the aforementioned AI-based watch manufacturing output prediction method, the establishment of a CMAN cross-modal attention network involves using the Transformer's self-attention mechanism to extract local features from each modality, calculating semantic relationships between modalities through cross-modal attention, introducing a GNN graph neural network to model inter-modal dependencies, and performing feature alignment and interaction on the multimodal data to obtain fused feature data, including: The operating parameters of production equipment are divided into data segments according to the production process. The self-attention mechanism of Transformer is used to calculate the correlation between the parameters at each time point and the parameters at the same time point in the same segment, so as to obtain the production characteristic data. The weekly market demand data is divided into quarterly time windows, and the correlation between adjacent weekly data is identified through a self-attention mechanism to obtain market characteristic data; By grouping raw material procurement data by supplier, a sequence is constructed. Self-attention is used to highlight the quantity fluctuation characteristics of consecutive procurement batches from the same supplier, thus obtaining supply chain characteristic data.

[0007] Furthermore, in the aforementioned AI-based watch manufacturing output prediction method, the steps of establishing a CMAN cross-modal attention network, using the Transformer's self-attention mechanism to extract local features of each modality, calculating semantic relationships between modalities through cross-modal attention, introducing a GNN graph neural network to model inter-modal dependencies, and performing feature alignment and interaction on the multimodal data to obtain fused feature data, further include: Semantic mapping rules are constructed for production characteristic data, market characteristic data, and supply chain characteristic data respectively, generating a three-dimensional semantic space; Calculate the correlation between production data and market data, the correlation between production data and supply chain data, and the correlation between market data and supply chain data respectively.

[0008] Furthermore, in the aforementioned AI-based watch manufacturing output prediction method, the steps of establishing a CMAN cross-modal attention network, using the Transformer's self-attention mechanism to extract local features of each modality, calculating semantic relationships between modalities through cross-modal attention, introducing a GNN graph neural network to model inter-modal dependencies, and performing feature alignment and interaction on the multimodal data to obtain fused feature data, further include: Set the equipment group in the production data, the product category in the market data, and the raw materials in the supply chain data as core nodes, and use directed edges to represent dependencies. In the initial state, each node carries its own local features, and through the message passing mechanism of GNN, the nodes pass feature information to their neighboring nodes. After multiple rounds of message passing, the intermediate nodes are used to perform feature calibration to correct node feature deviations, resulting in fused feature data.

[0009] Furthermore, in the aforementioned AI-based watch manufacturing output prediction method, the construction of a BJM Bayesian joint model to predict the fused feature data, the use of a longitudinal sub-model to predict the time-series trajectory of equipment health status, and the estimation of the failure time distribution of key equipment and raw materials through a survival sub-model, yields the output demand prediction result, including: By analyzing the health status changes of the device over the past 12 months through a longitudinal sub-model, typical decline patterns are identified. Based on the current health level and historical patterns, the status change trajectory for the next 90 days is predicted, and a device health status calendar is generated. By using a survival sub-model to identify key influencing factors based on historical data, and by analyzing the full life cycle data of similar equipment and raw materials, failure time probability curves are plotted, and a failure risk heat map is generated.

[0010] Furthermore, in the aforementioned AI-based watch manufacturing output forecasting method, the step of generating a watch production adjustment strategy based on the output demand forecast results includes: If increased production demand is predicted and equipment is in good health with sufficient raw material supply, then production plans should be increased and equipment uptime extended.

[0011] Furthermore, in the AI-based watch manufacturing output prediction system, the watch manufacturing output prediction system includes the following modules: The multimodal data acquisition module is used to collect production data, market data, and supply chain data during the watch manufacturing process. The collected data is preprocessed to obtain multimodal data. The feature data fusion module is used to establish a CMAN cross-modal attention network. It uses the self-attention mechanism of Transformer to extract local features of each modality, calculates the semantic association between modalities through cross-modal attention, introduces GNN graph neural network to model the inter-modal dependencies, and performs feature alignment and interaction on the multimodal data to obtain fused feature data. The production demand forecasting module is used to construct a BJM Bayesian joint model to predict the fused feature data, use the longitudinal sub-model to predict the time trajectory of equipment health status, and use the survival sub-model to estimate the failure time distribution of key equipment and raw materials to obtain the production demand forecasting results. The watch production adjustment module is used to generate a watch production adjustment strategy based on the production demand forecast results.

[0012] Furthermore, in the AI-based watch manufacturing production forecasting system, the production demand forecasting module includes the following sub-modules: The analysis submodule is used to analyze the health status changes of the device over the past 12 months through a longitudinal sub-model, identify typical decline patterns, predict the status change trajectory for the next 90 days based on the current health level and historical patterns, and generate a device health status calendar. The generation submodule is used to identify key influencing factors based on historical data using the survival submodel. By analyzing the full life cycle data of similar equipment and raw materials, it draws failure time probability curves and generates a failure risk heatmap.

[0013] Furthermore, in the AI-based watch manufacturing production forecasting system, the production demand forecasting module includes the following sub-modules: The decision submodule is used to increase the production plan and extend the equipment uptime if the predicted production demand increases and the equipment is in good health and the raw material supply is sufficient.

[0014] Its beneficial effects are as follows: 1. It fully considers uncertainties such as changes in equipment health status and the risk of critical resource failure, making the output demand forecast more closely aligned with actual production scenarios, significantly reducing forecast bias and providing a reliable basis for subsequent production decisions. 2. It helps companies prepare for equipment replacement and raw material reserves in advance, avoiding production delays due to resource shortages. This forward-looking management approach significantly improves the smoothness of the production process, reduces production management costs, and increases overall production efficiency. 3. It can prevent the accumulation or shortage of raw materials. When raw materials are expected to expire early or the supply is unstable, the strategy of promptly changing suppliers or adjusting storage methods enhances the supply chain's ability to cope with emergencies and ensures a stable supply of raw materials needed for production. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the watch manufacturing output prediction method based on artificial intelligence in this invention. Figure 2 This is a schematic diagram of the second embodiment of the watch manufacturing output prediction method based on artificial intelligence in this invention. Figure 3 This is a schematic diagram of the first embodiment of the watch manufacturing output prediction system based on artificial intelligence in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, an artificial intelligence-based method for predicting watch manufacturing output includes the following steps: Step 101: Collect production data, market data, and supply chain data during the watch manufacturing process, and preprocess the collected data to obtain multimodal data; Specifically, in this embodiment, production data, market data, and supply chain data are collected during the watch manufacturing process; Numerical data is cleaned using the mean imputation method, categorical data is cleaned using the mode imputation method, and outliers are identified using box plots. Data exceeding 1.5 times the upper or lower quartile range are considered outliers and replaced with the average of adjacent data to obtain cleaned production data. The cleaning production data is standardized and converted into normally distributed data with a mean of 0 and a standard deviation of 1. The processed production data, market data, and supply chain data are linked and integrated according to time and product batch to obtain multimodal data.

[0020] Specifically, (a) Data collection; Production data: Production equipment operating parameters: These are collected by various sensors installed on the production equipment, such as temperature sensors to collect the equipment's operating temperature, vibration sensors to collect the equipment's vibration frequency, and current sensors to collect the equipment's operating current. The sampling frequency is set to once per minute.

[0021] Production progress: Obtained from the production management system, including the completed quantity and production hours of each production stage, and updated in real time.

[0022] Product quality inspection data: obtained by quality inspectors using professional testing equipment, such as data on watch timekeeping accuracy and appearance defects. A comprehensive inspection is conducted and recorded for each batch of products produced.

[0023] Market data: Market demand: Relevant reports were purchased from market research institutions (Nielsen, Euromonitor International, etc.), and sales data from e-commerce platforms (JD.com, Taobao, etc.) were used for auxiliary statistics, collected weekly.

[0024] Consumer preferences: By using social media analytics tools (Weibo Index, WeChat Index, etc.) to explore consumer discussions and evaluations of watch styles, functions, prices, etc., and combining them with questionnaire survey data, the data is summarized monthly.

[0025] Competitor product information: Market research personnel are assigned to regularly collect information on competitors' products, including product styles, prices, sales volume, etc., and this information is updated quarterly.

[0026] Supply chain data: Raw material procurement quantity: Extracted from the enterprise's procurement management system, recording the quantity and time of each raw material procurement, and updated in real time.

[0027] Supplier information, including supplier name, contact information, production capacity, delivery cycle, product quality, etc., is stored in the supplier management database and is reviewed and updated every six months.

[0028] Logistics and transportation status: This information is obtained through a logistics tracking system, including the transportation routes, transportation times, and on-the-go status of raw materials, and is monitored in real time.

[0029] (ii) Data preprocessing; Data cleaning: Missing value imputation methods are used. For numerical data (equipment operating parameters), the mean imputation method is used; for categorical data (supplier names), the mode imputation method is used. Outliers are identified using box plots, and data exceeding 1.5 times the upper or lower quartile range are considered outliers and replaced with the average of adjacent data.

[0030] Data conversion: Numerical data: Standardize the data to convert it into normally distributed data with a mean of 0 and a standard deviation of 1. Text-based data, such as consumer reviews and supplier introductions, is segmented using Jieba word segmentation, and then converted into word vectors using the Word2Vec model.

[0031] Image-based data, such as product appearance images, can be processed using convolutional neural networks (CNNs) to extract features and convert the images into fixed-dimensional feature vectors.

[0032] Multimodal data integration: Processed production data, market data, and supply chain data are linked and integrated according to time dimension or product batch to form multimodal data.

[0033] Step 102: Establish a CMAN cross-modal attention network, use the self-attention mechanism of Transformer to extract local features of each modality, calculate the semantic association between modalities through cross-modal attention, introduce GNN graph neural network to model the inter-modal dependencies, and perform feature alignment and interaction on multimodal data to obtain fused feature data; Specifically, in this embodiment, the operating parameters of the production equipment are divided into data segments according to the production process. Through the self-attention mechanism of Transformer, the correlation between the parameters at each time point and the parameters at the same time point in the same segment is calculated to obtain production feature data. The weekly market demand data is divided into quarterly time windows, and the correlation between adjacent weekly data is identified through a self-attention mechanism to obtain market characteristic data; By grouping raw material procurement data by supplier, a sequence is constructed. Self-attention is used to highlight the quantity fluctuation characteristics of consecutive procurement batches from the same supplier, thus obtaining supply chain characteristic data.

[0034] Semantic mapping rules are constructed for production characteristic data, market characteristic data, and supply chain characteristic data respectively, generating a three-dimensional semantic space; Calculate the correlation between production data and market data, the correlation between production data and supply chain data, and the correlation between market data and supply chain data respectively.

[0035] Set the equipment group in the production data, the product category in the market data, and the raw materials in the supply chain data as core nodes, and use directed edges to represent the dependencies. In the initial state, each node carries its own local features, and through the message passing mechanism of GNN, the nodes pass feature information to their neighboring nodes. After multiple rounds of message passing, the feature calibration is performed through intermediate nodes to correct the node feature deviations, resulting in fused feature data.

[0036] Specifically, (a) Extraction of local features for each modality; Local feature extraction from production data: For the operating parameters of the production equipment (time-series data such as temperature and vibration frequency), the data is first divided into segments according to the production process (30 minutes of data from the movement assembly process is one segment).

[0037] The Transformer's self-attention mechanism calculates the correlation between parameters at each time point and parameters at other time points within the same segment. For example, when the equipment's vibration frequency suddenly increases, the system automatically strengthens the correlation weight between that time point and temperature parameters within the preceding and following 5 minutes, thereby capturing local features of abnormal equipment operation.

[0038] For production progress data (the quantity completed in each stage), the sequence is divided by product batch, and the impact of the progress data of key processes (precision gear machining) on ​​the overall production rhythm is highlighted through self-attention.

[0039] Local feature extraction from market data: For weekly market demand data, time windows are divided by quarter, and a self-attention mechanism is used to identify the correlation between adjacent weekly data. For example, demand data before and after holidays are given higher correlation weights.

[0040] For consumer preference text (social media reviews), we first break it down into thematic segments such as style, function, and price. Then, we use self-attention to calculate the correlation between words within the same theme (the correlation between ultra-thin dial and lightweight design) and extract core preference features.

[0041] Local feature extraction from supply chain data: For raw material procurement data, a sequence is constructed by grouping suppliers, and the quantity fluctuation characteristics of consecutive procurement batches from the same supplier are highlighted through self-attention.

[0042] For logistics and transportation data, segments are divided based on transportation routes, and the local correlation between transportation time and factors such as weather and road conditions is calculated (the correlation weight between transportation time and road congestion during heavy rain).

[0043] (ii) Cross-modal attention computation; Semantic space mapping: Semantic mapping rules are constructed for production data, market data, and supply chain data respectively: production data takes "equipment status-production efficiency" as the core dimension, market data takes "demand intensity-consumer preference" as the core dimension, and supply chain data takes "supply stability-cost fluctuation" as the core dimension, and the three types of data are uniformly mapped to the three-dimensional semantic space of "production-market-supply".

[0044] Intermodal correlation weight calculation: Calculate the correlation between production data and market data: For example, when the market data shows a "surge in demand for mechanical watches", the system will automatically increase the correlation weight of the "operating parameters of movement processing equipment" in the production data.

[0045] Calculate the correlation between production data and supply chain data: If the supply chain data shows "sapphire glass raw material shortage", then increase the correlation weight of "downtime of watch case polishing equipment" in the production data.

[0046] Calculate the correlation between market data and supply chain data: When market data captures the trend of "retro styles in fashion", enhance the correlation weight of "brass raw material procurement volume" in supply chain data.

[0047] Dynamic weight adjustment: The correlation weight is updated weekly based on the actual production-sales matching degree: if the matching error between "smart meter demand forecast" and "chip purchase volume" exceeds 5% in a certain week, the correlation calculation rules between the two will be automatically corrected.

[0048] (III) GNN Graph Neural Network Modeling; Graph structure construction; Node settings: Set the equipment groups (movement production line, watch case production line) in the production data, the product categories (mechanical watches, electronic watches) in the market data, and the core raw materials (sapphire glass, precision gears) in the supply chain data as core nodes.

[0049] Edge definition: Dependencies are represented by directed edges. For example, the mechanical watch category node points to the movement production line node (indicating that market demand drives production), and the precision gear node points to the movement production line node (indicating that raw material supply supports production).

[0050] Node feature interaction: In the initial state, each node carries its own local characteristics (the nodes on the movement production line carry the characteristics of equipment operation stability).

[0051] Through the message passing mechanism of GNN, nodes pass feature information to neighboring nodes: for example, the sapphire glass node passes the inventory balance feature to the watch case production line node, and at the same time receives the production consumption feature fed back by the node.

[0052] Feature alignment and fusion: After 3-5 rounds of message transmission, the system automatically corrects the node feature deviation: if the deviation between the "electronic watch demand" node feature and the "chip supply" node feature exceeds 10%, feature calibration is performed through the intermediate node (smart watch production line).

[0053] The final output is a fusion feature package that includes comprehensive feature data encompassing "equipment operating status, market demand trends, and raw material supply capacity," such as "mechanical watch demand growth of 30% + movement equipment stability of 92% + sufficient gear raw material inventory."

[0054] Step 103: Construct a BJM Bayesian joint model to predict the fused feature data, use the longitudinal sub-model to predict the time trajectory of equipment health status, and use the survival sub-model to estimate the failure time distribution of key equipment and raw materials to obtain the production demand prediction results. Specifically, in this embodiment, the longitudinal sub-model analyzes the health status change patterns of the device over the past 12 months, identifies typical decline patterns, predicts the status change trajectory for the next 90 days based on the current health level and historical patterns, and generates a device health status calendar. By using a survival sub-model to identify key influencing factors based on historical data, and by analyzing the full life cycle data of similar equipment and raw materials, failure time probability curves are plotted, and a failure risk heat map is generated.

[0055] Specifically, (a) Overall Model Architecture; The BJM Bayesian joint model employs a "dual-track parallel + dynamic fusion" architecture. The longitudinal sub-model and the survival sub-model operate independently and exchange key parameters in real time. Finally, the outputs of the two are integrated through a Bayesian inference mechanism. The core of the model lies in incorporating the dynamic changes in equipment health status and the failure risk of key resources into a unified prediction framework, solving the problem that a single model cannot simultaneously take into account both time-series trends and sudden risks.

[0056] (II) Vertical Sub-model: Prediction of Equipment Health Status Time-Series Trajectory; Data input: The system receives integrated characteristics of production equipment output from the CMAN network, including historical operating parameters of the equipment (temperature fluctuations and vibration peaks over the past 3 months), maintenance records (number of repairs and types of parts replaced), and production load data (average daily operating time and number of parts processed).

[0057] Supplement basic equipment information, such as equipment model, service life, manufacturer-recommended maintenance cycle, and other static parameters.

[0058] Trajectory prediction process: Status classification: The health status of the equipment is divided into four levels: "excellent, good, medium and poor". Each level corresponds to a clear judgment standard ("excellent" means that the fluctuation of the equipment operating parameters is within ±5% and there is no fault record).

[0059] Trend learning: By analyzing the changes in the health status of equipment over the past 12 months, typical degradation patterns can be identified. For example, a certain model of movement processing equipment typically exhibits a progressive degradation trajectory in its third year of use: "slowly increasing vibration frequency → increased temperature fluctuations → decreased precision".

[0060] Dynamic prediction: Based on the current health level and historical patterns, predict the trajectory of the equipment's condition changes over the next 90 days. If the equipment is currently at the "Good" level and its vibration frequency has increased by 10% in the past month, it is predicted that it may drop to the "Medium" level in 30 days, and there is a 50% probability that it will enter the "Poor" level after 60 days.

[0061] Anomaly Correction: When the equipment experiences a sudden anomaly (single over-temperature alarm), the sub-model will adjust the trajectory slope in real time to avoid prediction deviations caused by the inertia of historical data.

[0062] Output format: Generate a "Device Health Status Calendar", mark the predicted level for the next 3 months by day, and highlight key nodes (the date when it is expected to first enter the "poor" level).

[0063] The accompanying status impact description states, for example, "When the movement polishing machine enters the 'medium' level, the daily output of qualified parts is expected to decrease by 8%."

[0064] (III) Survival Sub-model: Estimation of the failure time distribution of critical resources; Data input: For critical equipment: Receive equipment health status trajectory (from the longitudinal sub-model), historical fault records (downtime and cause of fault in the past 5 years), and environmental parameters (workshop humidity and dust concentration).

[0065] For raw materials: receive supply chain integration characteristics, including raw material storage conditions (temperature stability of constant temperature warehouses), shelf life data (usage time of lubricating oil after opening), and historical loss records (breakage rate of a certain batch of springs).

[0066] Failure distribution estimation process: Failure event definition: Equipment failure is defined as requiring downtime for maintenance for more than 4 hours, and raw material failure is defined as "not meeting the minimum quality standard (insufficient gear hardness)".

[0067] Weighting of influencing factors: Key influencing factors are determined based on historical data. For example, for battery raw materials, the weight of storage temperature (60%) is higher than that of transportation time (30%) and batch variation (10%).

[0068] Distribution curve construction: By analyzing the full life cycle data of similar equipment / raw materials, failure time probability curves are plotted. For example, after 18 months of operation, the failure probability of a certain model of watch case die-casting machine suddenly increased from 5% to 30%, showing a "step-like increase" characteristic.

[0069] Dynamic updates: Curve parameters are adjusted every two weeks based on the latest failure records. If scratches appear in a batch of sapphire glass during the 6th month of storage, the peak failure probability is advanced by one month.

[0070] Output format: Generate a "failure risk heat map" to show the failure probability range of each key piece of equipment and raw material in the next 6 months on a weekly basis (high risk indicates a probability >50%, low risk indicates a probability <10%).

[0071] The text is labeled with a "vulnerability window," such as "Precision bearings have a 20% increased probability of failure during the high-temperature period of summer (June-August)."

[0072] (iv) Bayesian Joint Inference: Fusion of Output and Demand Forecasts; Parameter interaction mechanism: The longitudinal sub-model passes the "expected date when the equipment condition deteriorates to the 'poor' level" to the survival sub-model to correct the equipment's failure time distribution (entering the "high-risk" range earlier).

[0073] The survival sub-model feeds back the "time period when the probability of raw material failure is greater than 30%" to the longitudinal sub-model, adjusting the equipment load forecast for the corresponding production link (reducing the planned output during that time period).

[0074] Multi-scenario prediction: Baseline scenario: Under ideal conditions where equipment is fault-free and raw materials are sufficient, predict the normal production range based on the matching pattern of historical output and market demand.

[0075] Risk scenario: Simulate the output change when a single risk occurs, such as "If the movement processing equipment fails on the 45th day, the monthly output is expected to decrease by 12%".

[0076] Complex scenarios: Assess the combined impact of multiple risks, such as "simultaneous shortage of gear raw materials and failure of watch case polishing machine, which may lead to a 25%-30% decrease in production."

[0077] Prediction results output: Generate a "production demand probability distribution table" that lists the predicted production range and corresponding probability for the next 3 months by week (production in week 5 is 8000-8500 units, with a probability of 60%).

[0078] Key influencing factors should be highlighted, such as "The production fluctuation in week 7 was mainly affected by the risk of battery raw material failure".

[0079] Step 104: Generate a watch production adjustment strategy based on the production demand forecast results.

[0080] Specifically, in this embodiment, if the predicted increase in production demand and the equipment is in good health and the raw material supply is sufficient, the production plan will be increased and the equipment operating time will be extended.

[0081] Specifically, (a) Factors to consider; Based on production demand forecasts, combined with equipment health status forecasts, failure time distribution of key equipment and raw materials, and factors such as market demand changes and supply chain stability, production adjustment strategies are formulated.

[0082] (II) Specific Strategies; If increased production demand is predicted, and equipment is in good health with sufficient raw material supply, then production plans should be increased, equipment operating time should be appropriately extended, and the amount of raw materials procured should be increased to ensure production continuity.

[0083] If production demand is predicted to decrease, or equipment is in poor health or raw materials may be in short supply, production plans should be reduced, equipment maintenance schedules should be arranged reasonably, and raw material procurement plans should be adjusted to avoid inventory backlog.

[0084] When critical equipment is expected to fail in the near future, schedule equipment repair or replacement in advance and adjust production plans to ensure that production is not significantly affected during equipment repair.

[0085] If the raw materials are expected to expire early and market demand is relatively stable, the raw material supplier should be changed in a timely manner or the raw material storage method should be adjusted to ensure the normal supply of raw materials and maintain normal production.

[0086] Its beneficial effects are as follows: 1. It fully considers uncertainties such as changes in equipment health status and the risk of critical resource failure, making the output demand forecast more closely aligned with actual production scenarios, significantly reducing forecast bias and providing a reliable basis for subsequent production decisions. 2. It helps companies prepare for equipment replacement and raw material reserves in advance, avoiding production delays due to resource shortages. This forward-looking management approach significantly improves the smoothness of the production process, reduces production management costs, and increases overall production efficiency. 3. It can prevent the accumulation or shortage of raw materials. When raw materials are expected to expire early or the supply is unstable, the strategy of promptly changing suppliers or adjusting storage methods enhances the supply chain's ability to cope with emergencies and ensures a stable supply of raw materials needed for production.

[0087] Please see Figure 2In the AI-based method for predicting watch manufacturing output, a CMAN cross-modal attention network is established. The Transformer's self-attention mechanism is used to extract local features from each modality. Semantic relationships between modalities are calculated through cross-modal attention. A GNN graph neural network is introduced to model inter-modal dependencies. Multimodal data is then aligned and interacted to obtain fused feature data. The steps include: Step 201: Set the equipment group in the production data, the product category in the market data, and the raw materials in the supply chain data as core nodes, and use directed edges to represent the dependencies. Step 202: In the initial state, each node carries its own local features, and through the message passing mechanism of GNN, the nodes pass feature information to their neighboring nodes. Step 203: After multiple rounds of message passing, feature calibration is performed through intermediate nodes to correct node feature deviations and obtain fused feature data.

[0088] The above describes embodiments of the watch manufacturing output prediction method based on artificial intelligence of the present invention. Please refer to [link / reference]. Figure 3 In the AI-based watch manufacturing output prediction system, the system includes the following modules: The multimodal data acquisition module is used to collect production data, market data, and supply chain data during the watch manufacturing process. The collected data is preprocessed to obtain multimodal data. The feature data fusion module is used to establish a CMAN cross-modal attention network. It uses the self-attention mechanism of Transformer to extract local features of each modality, calculates the semantic association between modalities through cross-modal attention, introduces GNN graph neural network to model the inter-modal dependencies, and performs feature alignment and interaction on multimodal data to obtain fused feature data. The production demand forecasting module is used to construct a BJM Bayesian joint model to predict fused feature data, use the longitudinal sub-model to predict the time trajectory of equipment health status, and use the survival sub-model to estimate the failure time distribution of key equipment and raw materials to obtain the production demand forecasting results. The watch production adjustment module is used to generate watch production adjustment strategies based on production demand forecasts.

[0089] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting watch manufacturing output based on artificial intelligence, characterized in that, The method for predicting watch manufacturing output includes the following steps: Collect production data, market data, and supply chain data during the watch manufacturing process, and preprocess the collected data to obtain multimodal data; A CMAN cross-modal attention network is established, and the self-attention mechanism of Transformer is used to extract local features of each modality. Semantic associations between modalities are calculated through cross-modal attention. A GNN graph neural network is introduced to model inter-modal dependencies. The multimodal data is then aligned and interacted to obtain fused feature data. A joint BJM Bayesian model is constructed to predict the fused feature data. The longitudinal sub-model is used to predict the time trajectory of equipment health status, and the failure time distribution of key equipment and raw materials is estimated through the survival sub-model to obtain the production demand prediction results. A watch production adjustment strategy is generated based on the production demand forecast results.

2. The watch manufacturing output prediction method based on artificial intelligence as described in claim 1, characterized in that, The process of collecting production data, market data, and supply chain data during the watch manufacturing process involves preprocessing the collected data to obtain multimodal data, including: Collect production data, market data, and supply chain data during the watchmaking process; Numerical data is cleaned using the mean imputation method, categorical data is cleaned using the mode imputation method, and outliers are identified using box plots. Data exceeding 1.5 times the upper or lower quartile range are considered outliers and replaced with the average of adjacent data to obtain cleaned production data. The cleaning production data is standardized and converted into normally distributed data with a mean of 0 and a standard deviation of 1. The processed production data, market data, and supply chain data are linked and integrated according to time and product batch to obtain multimodal data.

3. The watch manufacturing output prediction method based on artificial intelligence as described in claim 1, characterized in that, The CMAN cross-modal attention network is established by using the Transformer's self-attention mechanism to extract local features of each modality, calculating semantic relationships between modalities through cross-modal attention, introducing a GNN graph neural network to model inter-modal dependencies, and performing feature alignment and interaction on the multimodal data to obtain fused feature data, including: The operating parameters of production equipment are divided into data segments according to the production process. The self-attention mechanism of Transformer is used to calculate the correlation between the parameters at each time point and the parameters at the same time point in the same segment, so as to obtain the production characteristic data. The weekly market demand data is divided into quarterly time windows, and the correlation between adjacent weekly data is identified through a self-attention mechanism to obtain market characteristic data; By grouping raw material procurement data by supplier, a sequence is constructed. Self-attention is used to highlight the quantity fluctuation characteristics of consecutive procurement batches from the same supplier, thus obtaining supply chain characteristic data.

4. The watch manufacturing output prediction method based on artificial intelligence as described in claim 3, characterized in that, The process of establishing a CMAN cross-modal attention network, using the Transformer's self-attention mechanism to extract local features of each modality, calculating semantic relationships between modalities through cross-modal attention, introducing a GNN graph neural network to model inter-modal dependencies, and performing feature alignment and interaction on the multimodal data to obtain fused feature data, further includes: Semantic mapping rules are constructed for production characteristic data, market characteristic data, and supply chain characteristic data respectively, generating a three-dimensional semantic space; Calculate the correlation between production data and market data, the correlation between production data and supply chain data, and the correlation between market data and supply chain data respectively.

5. The watch manufacturing output prediction method based on artificial intelligence as described in claim 3, characterized in that, The process of establishing a CMAN cross-modal attention network, using the Transformer's self-attention mechanism to extract local features of each modality, calculating semantic relationships between modalities through cross-modal attention, introducing a GNN graph neural network to model inter-modal dependencies, and performing feature alignment and interaction on the multimodal data to obtain fused feature data, further includes: Set the equipment group in the production data, the product category in the market data, and the raw materials in the supply chain data as core nodes, and use directed edges to represent dependencies. In the initial state, each node carries its own local features, and through the message passing mechanism of GNN, the nodes pass feature information to their neighboring nodes. After multiple rounds of message passing, the intermediate nodes are used to perform feature calibration to correct node feature deviations, resulting in fused feature data.

6. The watch manufacturing output prediction method based on artificial intelligence as described in claim 1, characterized in that, The constructed BJM Bayesian joint model is used to predict the fused feature data. The longitudinal sub-model predicts the time-series trajectory of equipment health status, and the survival sub-model estimates the failure time distribution of key equipment and raw materials to obtain production demand prediction results, including: By analyzing the health status changes of the device over the past 12 months through a longitudinal sub-model, typical decline patterns are identified. Based on the current health level and historical patterns, the status change trajectory for the next 90 days is predicted, and a device health status calendar is generated. By using a survival sub-model to identify key influencing factors based on historical data, and by analyzing the full life cycle data of similar equipment and raw materials, failure time probability curves are plotted, and a failure risk heat map is generated.

7. The watch manufacturing output prediction method based on artificial intelligence as described in claim 1, characterized in that, The process of generating a watch production adjustment strategy based on the production demand forecast includes: If increased production demand is predicted and equipment is in good health with sufficient raw material supply, then production plans should be increased and equipment uptime extended.

8. A watch manufacturing output prediction system based on artificial intelligence, characterized in that, The watch manufacturing output prediction system includes the following modules: The multimodal data acquisition module is used to collect production data, market data, and supply chain data during the watch manufacturing process. The collected data is preprocessed to obtain multimodal data. The feature data fusion module is used to establish a CMAN cross-modal attention network. It uses the self-attention mechanism of Transformer to extract local features of each modality, calculates the semantic association between modalities through cross-modal attention, introduces GNN graph neural network to model the inter-modal dependencies, and performs feature alignment and interaction on the multimodal data to obtain fused feature data. The production demand forecasting module is used to construct a BJM Bayesian joint model to predict the fused feature data, use the longitudinal sub-model to predict the time trajectory of equipment health status, and use the survival sub-model to estimate the failure time distribution of key equipment and raw materials to obtain the production demand forecasting results. The watch production adjustment module is used to generate a watch production adjustment strategy based on the production demand forecast results.

9. The watch manufacturing output prediction system based on artificial intelligence as described in claim 8, characterized in that, The production demand forecasting module includes the following sub-modules: The analysis submodule is used to analyze the health status changes of the device over the past 12 months through a longitudinal sub-model, identify typical decline patterns, predict the status change trajectory for the next 90 days based on the current health level and historical patterns, and generate a device health status calendar. The generation submodule is used to identify key influencing factors based on historical data using the survival submodel. By analyzing the full life cycle data of similar equipment and raw materials, it draws failure time probability curves and generates a failure risk heatmap.

10. The watch manufacturing output prediction system based on artificial intelligence as described in claim 8, characterized in that, The production demand forecasting module includes the following sub-modules: The decision submodule is used to increase the production plan and extend the equipment uptime if the predicted production demand increases and the equipment is in good health and the raw material supply is sufficient.

Images