A method and system for intelligent management of furniture production

Abstract

This invention relates to the field of furniture production management technology, and discloses an intelligent management method and system for furniture production. The method involves real-time collection of production data such as material consumption, equipment operating parameters, and process completion time at each process node on the production line; multi-dimensional feature extraction of the data to generate a comprehensive feature set including temporal, statistical, and correlational features; division of continuous production stages according to the dynamic change patterns of the features and assignment of identifiers; reorganization of data based on the identifiers to calculate the transition probability matrix between adjacent stages; identification of potential abnormal stages and generation of marked sequences by analyzing abnormal state transition patterns in the matrix; construction of a quality prediction model combining abnormal markers and real-time data to output quality prediction scores for each process node; dynamic adjustment of the process production parameter configuration scheme based on the deviation of the score from a preset threshold; and similarity matching between the adjusted scheme and the historical optimal configuration to select a set of configurations to be verified for simulation testing.

Description

Technical Field

[0001] This invention relates to the field of furniture production management technology, specifically to an intelligent management method and system for furniture production. Background Technology

[0002] As a traditional industry, the furniture manufacturing sector has long relied on manual experience and static planning for production management. The actual production process involves multiple steps, from raw material cutting and component processing to assembly and painting. Traditional management models typically employ post-production inspections and periodic statistics, making it difficult to respond promptly to real-time situations occurring on the production line. For example, abnormal fluctuations in material consumption, subtle changes in equipment operating parameters, or sudden extensions in process completion times are often early signs of production quality problems. However, due to a lack of real-time data collection and correlation analysis capabilities across the entire process, these potential risk signals are easily overlooked, leading to delayed problem detection and increased rework costs and quality risks.

[0003] Most existing production management systems focus on order progress tracking and inventory management, lacking depth in dynamic data mining during the production process. Data collection points are scattered, information silos are common, and key parameters such as material flow, equipment status, and process sequence are not effectively integrated. Managers struggle to quickly identify core factors affecting product quality and their inherent relationships from massive amounts of data. Furthermore, the division of production stages is often based on fixed time intervals or human experience, failing to accurately reflect the inherent dynamic patterns and stage characteristics of the production process. This rigid stage division makes it difficult to accurately capture the transitions and abrupt changes between processes.

[0004] In terms of quality control, traditional methods mainly rely on finished product inspection or sampling inspection of key processes. This approach is reactive and cannot achieve process prevention. By the time quality defects are discovered during inspection, a large batch of products may have already been produced, resulting in irreparable losses. Although theoretical methods such as statistical process control have been introduced, in practical applications, due to weak data foundations and outdated analytical tools, it is often difficult to establish accurate predictive models and achieve real-time dynamic optimization of production parameters and risk warnings. Therefore, the furniture manufacturing industry urgently needs an intelligent management method that can deeply integrate real-time production data, intelligently identify production stage characteristics, predict quality trends, and support dynamic parameter optimization. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent management method and system for furniture production to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an intelligent management method for furniture production, the method comprising:

[0007] Real-time collection of production data at each process node on the furniture production line, including material consumption data, equipment operating parameters, and process completion time;

[0008] Multi-dimensional feature extraction is performed on the collected production data to generate a comprehensive feature set that includes time-series features, statistical features, and correlation features;

[0009] Based on the dynamic change patterns of each feature in the comprehensive feature set, the production process is divided into several continuous production stages, and a stage identifier is assigned to each production stage.

[0010] Production data is reorganized based on stage identifiers to form stage datasets corresponding to each production stage, and the transition probability matrix between adjacent stage datasets is calculated.

[0011] By analyzing the abnormal patterns of state transitions in the transition probability matrix, potential abnormal stages in the production process are identified, and an abnormal stage marker sequence is generated.

[0012] By combining the abnormal stage marker sequence with real-time collected production data, a production quality prediction model is constructed, and the quality prediction score of each process node is output.

[0013] Based on the degree of deviation between the quality prediction score and the preset threshold, the production parameter configuration scheme of the corresponding process node is dynamically adjusted.

[0014] The adjusted production parameter configuration scheme is matched with the historical best configuration to select the configuration set to be verified and sent to the simulation test stage;

[0015] Receive the verification results returned from the simulation test, update the production parameter configuration library, and generate the final execution command.

[0016] Preferably, the specific steps of the multi-dimensional feature extraction include:

[0017] Extract the consumption rate per unit time and the cumulative consumption from the material consumption data;

[0018] Extract peak load duration and operational stability indicators from equipment operating parameters;

[0019] Extract the standard deviation and percentage of deviation from the baseline time for each process completion time;

[0020] The three types of features are aligned according to the time window and then merged into a comprehensive feature set.

[0021] Preferably, the basis for dividing the production stages includes:

[0022] Detecting periodic fluctuation nodes in temporal features within a comprehensive feature set;

[0023] The magnitude of abrupt changes in statistical characteristics within a continuous time window exceeds a preset sensitivity threshold;

[0024] The coupling strength between associated features exhibits a step change.

[0025] Preferably, the method for calculating the transition probability matrix includes:

[0026] Frequency of occurrence of statistical stage identifiers in historical data;

[0027] Calculate the ratio of the transition frequency between adjacent stage identifiers to the total frequency;

[0028] The transition probability matrix is ​​generated after the comparison values ​​are normalized.

[0029] Preferably, the abnormal pattern recognition includes:

[0030] Detect jump paths in the transition probability matrix whose probability values ​​are below the lower tolerance limit;

[0031] A combination of stage identifiers that exhibit consecutive, unconventional transitions;

[0032] The tag combination is matched with a preset exception pattern library to output an exception stage tag sequence.

[0033] Preferably, the construction of the production quality prediction model includes:

[0034] Convert the abnormal phase marker sequence into a binary encoded vector;

[0035] The real-time production data is concatenated with the encoded vector to extract features.

[0036] The prediction model is trained using incremental learning and outputs a quality prediction score.

[0037] Preferably, the dynamic adjustment of production parameters includes:

[0038] When the quality prediction score is lower than the first-level threshold, the equipment operating parameter priority adjustment strategy is triggered.

[0039] When the score falls below the secondary threshold, the material ratio and equipment parameters are adjusted simultaneously.

[0040] When the score is higher than the threshold, the current configuration is retained and recorded as a candidate optimization scheme.

[0041] Preferably, the specific steps of the similarity matching include:

[0042] Calculate the Euclidean distance between the adjusted configuration and the historical best configuration;

[0043] Configurations whose distance is less than the matching radius are selected to form a set of configurations to be verified.

[0044] The selected configurations are sorted in ascending order of distance value and then sent for verification.

[0045] Preferably, the verification method for the simulation test phase includes:

[0046] Load the configuration to be verified in the digital twin environment;

[0047] Simulate the entire production cycle and collect virtual production data;

[0048] Verification results are generated by comparing the deviation between virtual data and actual historical data.

[0049] Preferably, the present invention also includes an intelligent furniture production management system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described intelligent furniture production management method.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] This invention achieves deep perception and comprehensive utilization of data throughout the entire furniture production process through real-time data acquisition and multi-dimensional feature extraction. In traditional management methods, data on materials, equipment, and working hours are often isolated. This method integrates these data into a comprehensive feature set containing rich information, providing a solid data foundation for subsequent refined management and making the production process transparent and quantifiable.

[0052] By automatically dividing production into stages and calculating transition probabilities based on dynamic data changes, this method can accurately capture the inherent rhythms and state transition characteristics of the production process. This avoids the subjectivity of manual stage division and more accurately reflects the actual operating logic of the production line. By analyzing abnormal patterns in state transitions, potential problems such as poor process connections, equipment performance degradation, or unstable material supply can be identified early, providing the possibility for proactive intervention.

[0053] The constructed quality prediction model combines anomaly markers with real-time data, realizing a shift from "post-event inspection" to "pre-event prediction." The quality prediction score output by the model enables managers to predict the probability and process of quality problems before they actually occur, thereby adjusting production parameters in a targeted manner and effectively reducing defect rates and production costs.

[0054] The mechanism of dynamically adjusting production parameters and verifying them against historical optimal configurations forms a closed-loop optimization process of "monitoring-prediction-adjustment-verification". The system can not only adjust parameters based on real-time prediction results, but also verify the effectiveness of new solutions through simulation testing, ensuring the scientific rigor and safety of the adjustments. This closed-loop optimization endows the production system with a certain degree of self-learning and self-adaptation capabilities, enabling continuous improvement and enhancing production efficiency and quality stability.

[0055] This method integrates intelligent management into all aspects of furniture production, driving the furniture manufacturing industry towards a data-driven, precision-controlled modern production model. By enhancing the precision of process control and risk warning capabilities, it helps companies optimize resource allocation, improve production efficiency, ensure product quality, and strengthen market competitiveness. Attached Figure Description

[0056] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent management method for furniture production described in this invention.

[0057] Figure 2 A flowchart of the multi-dimensional feature extraction steps;

[0058] Figure 3 A flowchart illustrating the method for calculating the transition probability matrix;

[0059] Figure 4 This is a diagram for intelligent monitoring and anomaly diagnosis analysis of production quality. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Please see Figure 1This invention provides an intelligent management method and system for furniture production. The method includes integrating dynamic data streams from various process nodes on the production line to achieve closed-loop optimization of the production process. A sensor network and data acquisition module are deployed on the furniture production line to capture production data from each process node in real time. This data includes material consumption data such as raw material usage, equipment operating parameters such as motor speed and temperature, and process completion time such as the operation time of each workstation. The collected data is transmitted to a central processing unit via an Industrial Internet of Things (IIoT) protocol for preliminary cleaning and formatting to ensure data consistency and integrity. Multi-dimensional feature extraction is performed on the preprocessed production data to generate a comprehensive feature set from temporal, statistical, and correlational perspectives. The feature extraction process employs a sliding time window mechanism, slicing and analyzing the data at fixed intervals to capture short-term fluctuations and long-term trends. The extracted features include, but are not limited to, instantaneous values ​​of material consumption rates, periodic indicators of equipment load, and the distribution characteristics of process time. These features are merged into a structured dataset for subsequent stage division.

[0062] Example 1: See Figure 2 Multi-dimensional feature extraction and production stage segmentation are achieved through the integration of data acquisition and feature analysis modules, enabling in-depth processing of production data. During multi-dimensional feature extraction, after system initialization, a real-time production data stream is loaded. This stream contains raw information such as material consumption data, equipment operating parameters, and process completion times. Material consumption data processing focuses on capturing dynamic changes. The consumption rate per unit time is calculated using a sliding window algorithm. The window size is a configurable parameter, typically adjusted based on the production rhythm. The rate value is obtained through differential operations, i.e., the difference in material usage between adjacent timestamps divided by the time interval. The difference calculation uses backward differencing to ensure real-time performance. The cumulative consumption is derived using numerical integration, with the integration interval from the production start time to the current time. The trapezoidal rule is chosen for the integration algorithm to balance accuracy and computational efficiency. The cumulative value is stored in a circular buffer for fast lookup.

[0063] The analysis of equipment operating parameters focuses on stability and load characteristics. Peak load duration extraction relies on a threshold detection mechanism, with the threshold dynamically set according to equipment specifications, such as through percentiles of historical data. Duration calculation involves identifying consecutive periods exceeding the threshold in the parameter sequence and calculating their length. Operational stability indicators are based on parameter variance calculation within a sliding window. Variance values ​​are updated online to avoid full recalculation. The indicators also incorporate the ratio of moving averages as an auxiliary feature, calculated using an exponentially weighted average to emphasize recent data. Process completion time handling involves quantifying distribution characteristics. Standard deviation calculation uses an unbiased estimation method, with the window size aligned with the production batch. The percentage deviation from the baseline duration is obtained by comparing actual values ​​with preset standard values, and the percentage value undergoes a logarithmic transformation to mitigate the impact of skewed distributions.

[0064] The processing after feature extraction begins with time window alignment. The alignment mechanism employs a global timestamp synchronization system, which uses the production line's master clock signal as a reference and coordinates the timing of all data acquisition devices via a network time protocol, ensuring the timestamp accuracy of each data point is at the millisecond level. The default time window size is set to a 5-minute interval, but the system monitors the production line's throughput metrics in real time, such as the number of processes completed per unit time. When the throughput exceeds a preset threshold, the window is automatically reduced to 1 minute to accommodate high-frequency data streams. During window alignment, the system divides the continuous time axis into equal-length intervals, each corresponding to a window. Data points are assigned to the corresponding windows based on their timestamps. For missing data points within a window, a linear interpolation algorithm is used to fill them. The interpolation calculation is based on the numerical change trend of adjacent time points. For example, if a device parameter value is missing in the middle of a window, the arithmetic mean of the previous and next valid values ​​is taken as the estimate, thus ensuring a uniform distribution of data points.

[0065] The feature merging stage combines the preprocessed feature vectors into a comprehensive feature set. Material consumption features include consumption rate per unit time and cumulative consumption; equipment parameter features include peak load duration and operational stability indicators; and process time features include standard deviation and percentage deviation from the baseline duration. Each feature dimension is independently normalized before concatenation using a min-max scaling technique. The system extracts the minimum and maximum values ​​of each feature from historical data and maps the current feature value to the 0-1 range. For example, for the percentage deviation feature of process completion time, the historical minimum might be -10% (indicating early completion), and the historical maximum might be +20% (indicating delay). The current value of -5% will be scaled to 0.25. The normalized feature vectors are concatenated in a fixed order: material features first, then equipment features, and finally process features. The dimension of the concatenated comprehensive feature vector is the sum of all feature dimensions, and vector elements are stored in floating-point format to maintain precision.

[0066] The comprehensive feature set is persistently stored in matrix form. Rows in the matrix correspond to the center time point of each time window, and columns correspond to each normalized feature dimension. The matrix data structure uses a sparse storage format, recording only the positions and values ​​of non-zero elements, and a multi-dimensional index is built to support fast range queries. Matrix updates are incremental. When a new batch of data arrives, the system verifies the continuity of timestamps, adding new rows or modifying changed values ​​in existing rows, while simultaneously updating the index structure of the sparse matrix, for example, using an incremental B-tree to maintain the timestamp index. The entire process runs automatically without manual intervention, ensuring real-time data availability.

[0067] In the production phase segmentation stage, the system analyzes the temporal features of the comprehensive feature set to identify periodic fluctuation nodes. Periodic detection employs a Fourier transform algorithm. Before the transform, the sequence undergoes denoising preprocessing, such as using wavelet transform to filter out high-frequency noise. The dominant frequency component is identified through power spectrum analysis. Fluctuation nodes are defined as moments when the spectral peak exceeds the background noise threshold, which is dynamically calibrated based on the signal-to-noise ratio. Monitoring the abrupt change amplitude of statistical features is based on comparisons within continuous time windows. The window size is consistent with that in the feature extraction stage. The abrupt change amplitude is calculated using the relative rate of change, which employs a percentage difference method. A flag is triggered when the difference exceeds a preset sensitivity threshold, initially set at 10%, but can be adaptively optimized through a machine learning model, for example, by adjusting based on historical abnormal frequencies. The coupling strength between associated features is evaluated using mutual information calculation. The mutual information value is derived through probability distribution estimation, which uses the kernel density method to handle continuous variables. Step change detection is based on gradient analysis, with gradient values ​​calculated using the Sobel operator. A boundary point is identified when the absolute value of the gradient exceeds twice the standard deviation of the historical mean.

[0068] The stage partitioning algorithm employs a hierarchical clustering method. The clustering input consists of a set of fluctuating nodes, abrupt change points, and step points. The point sets are initially grouped using Euclidean distance, and cosine similarity is chosen as the intra-group similarity metric to emphasize directional consistency. The clustering process is iterative, merging the closest groups each time until the inter-group distance exceeds a merging threshold, which is set based on the complexity of the production process. A greedy strategy is used to optimize stage boundaries, prioritizing the merging of adjacent segments with smaller feature variances. Variance is calculated based on the aggregated value of all features within a segment. Each production stage is assigned a unique stage identifier, encoded as a fixed-length string, e.g., "S001" represents the first stage. The encoding rule includes the stage number and type code. Identifiers are bound to feature statistics and stored in a metadata database, which uses a relational model to support complex queries. The stage partitioning results are verified in real-time by cross-referencing production log data to ensure accuracy and consistency. The entire implementation emphasizes the robustness of the algorithm, for example, by introducing fault-tolerance mechanisms to handle sensor data anomalies and prevent the propagation of partitioning errors.

[0069] Example 2: See Figure 3The transition probability matrix calculation and anomaly pattern recognition construct a state transition model based on historical production stage sequence data, and identify potential production problems by analyzing anomalous behaviors in transition patterns. The calculation of the transition probability matrix begins with the preprocessing of historical stage identifier sequences. Sequence data is extracted from the production database, with each sequence representing a complete batch of stage transition records, sorted by timestamps to ensure time-series accuracy. The system traverses these sequences to count the frequency of each stage identifier. Frequency calculation uses a counting algorithm, storing the mapping between identifiers and occurrence counts through a hash table, supporting fast querying and updating. The continuity of sequences is considered during traversal to avoid statistical bias caused by missing data. The transition frequency between adjacent stage identifiers is obtained through sliding window analysis, with a window size of 2 to capture direct transition relationships. The transition frequency record is the number of jumps from stage A to stage B, stored in a two-dimensional frequency matrix. The matrix dimension equals the total number of stages, and the elements are initialized to zero and incremented as data is processed.

[0070] The calculation of the transition probability matrix begins with the preprocessing of historical production stage sequences. The system loads a time-sorted sequence of stage identifiers from the database. Each sequence represents a complete production batch's stage transition record. The sequence data is cleaned to remove invalid or duplicate entries, ensuring the continuity and consistency of the sequence. When calculating the transition frequency, the system traverses all sequences using a sliding window size of 2, sequentially recording the occurrence count of each stage identifier pair. For example, a transition event where stage A immediately follows stage B. Frequency counting is implemented using a hash table data structure, where the key is the stage pair and the value is the frequency value, supporting fast insertion and querying. The calculation of the total transition frequency is performed for each starting stage, accumulating the sum of the frequencies of all transition events originating from that stage. For example, the sum of the frequencies of all transitions originating from stage A. The calculation process is parallelized, dividing the sequence into multiple subsets, calculating the frequencies separately, and then merging the results to improve efficiency with large datasets.

[0071] To prevent the zero-frequency problem, Laplace smoothing is introduced. The smoothing parameter is set to a small constant, typically 1, but can be dynamically adjusted based on data sparsity. For example, if certain transition events never occur in historical data, the system automatically increases or decreases the smoothing value based on the overall data density. The ratio calculation adds the smoothing constant to each transition frequency and divides it by the total transition frequency plus the smoothing constant multiplied by the total number of stages to obtain the original transition probability. This ensures that even unobserved transitions have a non-zero probability, avoiding model overfitting. Since the original probability values ​​may not be within the standard range, normalization is required. Row normalization is chosen. For each probability row corresponding to the initial stage, the sum of all probability values ​​in that row is calculated, and each probability value is divided by this sum, ensuring that the sum of probabilities in each row is 1. Row summation uses an accumulation method, processing the probability matrix row by row. The sum of elements in each row is used as the divisor. The normalization operation is completed by iterating through each element. Parallel processing utilizes multi-threading to process multiple rows simultaneously, significantly improving computation speed. The generated transition probability matrix is ​​stored in a sparse matrix format because stage transitions in actual production often exhibit locality, with most probability values ​​being zero or close to zero. Sparse storage only records the positions and values ​​of non-zero elements, saving memory space. Matrix elements represent the likelihood of transitioning from stage i to stage j. The matrix dimension equals the total number of stages, and the elements are arranged in a square matrix form for easy subsequent matrix operations.

[0072] The matrix periodic update mechanism is based on new production batch data, with a configurable update cycle. The default setting is to trigger an update every 100 production batches completed or every 24 hours to balance real-time performance and computational overhead. The update process is incremental; the system only processes newly added sequence data, recalculates the transformation frequency of affected stages, and partially updates the corresponding rows and columns in the matrix, avoiding full matrix recalculation. During incremental updates, the system maintains a change log, recording transformation events for new data, merging them into a historical frequency table, and then recalculating the probability values. This partial update method reduces computational resource consumption. The updated matrix version control mechanism ensures data consistency; old versions are backed up for rollback, and new versions take effect after verification. The entire process runs automatically without manual intervention.

[0073] The anomaly pattern recognition module monitors low-probability transition events in the transition probability matrix. The tolerance lower limit is dynamically set based on historical probability distributions, using percentiles (e.g., 5th percentile) of probability values ​​as thresholds. Quantile estimation employs a sorting-based algorithm, and approximation methods such as T-Digest are used to balance accuracy and performance when handling large-scale data. Transition paths are defined as consecutive stage transition sequences with configurable path lengths, defaulting to 2 to 5 stages. The system scans elements in the matrix with probability values ​​below the tolerance lower limit, marking the corresponding transition pairs. The scanning process uses a breadth-first search algorithm, exploring possible paths from each stage. Detection of consecutive non-routine transitions is based on a time window mechanism, with the window size related to production rhythm, for example, set to 10 consecutive transition events. The system counts the density of low-probability transitions within the window; when the density exceeds a preset threshold (e.g., 30%), it is marked as an anomalous combination. This threshold is calibrated through historical anomaly data analysis.

[0074] After the marker combination is generated, it is matched against a preset anomaly pattern library. The anomaly pattern library stores the transition sequences corresponding to common failure scenarios. The pattern library is built based on domain expert knowledge and historical failure records. Each pattern is encoded as a string sequence, for example, "S001->S003->S005" represents a specific equipment failure path. The matching algorithm uses edit distance calculation, allowing insertion, deletion, or replacement operations. The fault tolerance threshold is set to 1 or 2 to handle minor variations. The matching degree is quantified by similarity score; patterns with scores higher than the threshold (e.g., 0.8) are considered successful matches. The matching results are used to generate anomaly stage marker sequences. The sequences are binary vectors with a length equal to the number of production stages. Vector elements are selected by querying the stage settings involved in the matching pattern; "1" indicates anomaly, and "0" indicates normal. The sequences are stored in a time-series database in association with the production timeline, supporting real-time querying and backtracking analysis. The entire implementation process emphasizes computational efficiency and robustness, for example, by using caching mechanisms to accelerate frequent queries and anomaly handling logic to deal with data noise, ensuring the stability of the system under high-load production environments.

[0075] Example 3: Model construction begins with the processing of the anomaly stage marker sequence. The sequence data is received from the anomaly detection module in binary marker list format. Each marker corresponds to a production stage, with a marker value of "1" indicating an anomaly and "0" indicating normal operation. Sequence conversion uses a direct mapping method to generate binary encoded vectors. The vector dimension equals the total number of production stages; for example, if the production line has 50 stages, the vector length is 50. Each element stores the marker value for the corresponding stage. The mapping process is implemented using a lookup table. The marker sequence is arranged chronologically, and the vector is initialized to all zeros. The corresponding position is set to 1 based on the anomaly stage identifier. The correspondence between the identifier and the vector index is based on the stage number; for example, stage S001 corresponds to index 0, S002 corresponds to index 1, and so on. After vector generation, normalization is performed using L2 normalization to scale the vector magnitude to 1 to eliminate the impact of dimensional differences on the model. Normalization is calculated by dividing by the vector's Euclidean norm, which is calculated using the sum of squares and square root algorithm. The steps for norm calculation are as follows: Square each element in the vector to obtain its square value; sum these square values ​​one by one to obtain their total; then take the square root of this sum to obtain the Euclidean norm value of the vector. After the norm calculation is completed, normalization is performed by dividing each element in the original vector by the calculated norm value, ensuring that the overall magnitude of the processed vector is uniformly 1, thus guaranteeing the consistency and comparability of the feature data in subsequent production quality prediction models.

[0076] The feature concatenation operation between real-time production data and encoded vectors is performed in the data fusion module. Real-time production data includes features such as material consumption, equipment parameters, and process times. These features are preprocessed and converted into numerical vectors. The vector dimension depends on the number of features; for example, if 20 features are extracted, the vector length is 20. Before concatenation, the real-time production data vectors are standardized by scaling each feature value using the following formula:

[0077]

[0078] in: It is the k-th eigenvalue after standardization. It is the original value of the k-th feature in the real-time production data. It is the mean of this feature in historical data. This is the standard deviation of this feature in historical data. Mean Calculated using a sliding window, with the window size aligned to the production batch, the standard deviation... The derivation uses an unbiased estimation method, and standardization ensures that the eigenvalues ​​conform to zero mean and unit variance. The encoded vector is horizontally concatenated with the standardized real-time production data vector to generate an extended feature vector. The concatenation operation is implemented through array joins; for example, if the real-time vector length is 20 and the encoded vector length is 50, then the extended vector length is 70. The extended vector is stored in a dense array format, supporting efficient matrix operations.

[0079] The incremental learning method for training the prediction model begins with the preprocessing of the data stream. Extended feature vectors and anomaly stage marker sequences collected in real-time from the production line are sorted by timestamp and then enter the training buffer. The buffer uses a first-in, first-out queue structure, with a capacity to hold multiple batches of data; the default size is 1000 sample points. Each sample point contains a 70-dimensional extended feature vector and a corresponding actual quality label. The quality label originates from manual inspection records in the offline quality inspection process or the output of automated inspection equipment. The label value is mapped to a continuous interval of 0 to 1 using a min-max normalization method, representing the quality pass rate of that production unit.

[0080] The model architecture employs a multilayer perceptron network with two hidden layers. The number of neurons in the input layer strictly corresponds to the dimension of the extended feature vector. The number of nodes in the hidden layer is determined through a grid search on the validation set, typically ranging from 50 to 200 nodes. Each layer uses the ReLU activation function to introduce nonlinear transformation capability. During parameter initialization, the Xavier uniform distribution method is used, automatically calculating the initialization range of the weight matrix based on the input and output dimensions. The bias term is initialized to zero, a strategy beneficial for maintaining the stability of the signal variance during forward propagation. The output layer is designed as a single-neuron structure, and the linear output is ultimately compressed to the 0-1 interval using the sigmoid function, directly corresponding to the probabilistic interpretation of the quality prediction score.

[0081] The training process employs an online learning mechanism based on mini-batch stochastic gradient descent. The optimizer uses an adaptive moment estimation algorithm, which dynamically maintains the estimates of the first and second moments for each parameter and calculates the individualized learning rate accordingly. The loss function is defined as mean squared error, calculating the mean squared difference between the model output and the true quality label. During training, every 100 new samples constitute a batch, and forward and backward propagation calculations are performed immediately. Parameter updates follow the gradient descent principle, but are based solely on the gradient direction of the current batch of data. After the update, the batch of data is discarded immediately to save storage space, and new batch data overwrites the old data in the buffer. Model performance evaluation is synchronized with the production rhythm. After every three training batches, a separate reserved dataset is used for validation. The reserved dataset is randomly sampled from historical data and is not used for training. Evaluation metrics include accuracy and recall calculations, but the production line is not interrupted. When model performance degradation exceeds a preset tolerance, the system triggers a model rollback mechanism, automatically loading previously stored stable parameter versions and recording the current abnormal state for technical personnel analysis. The entire training system is designed to be fault-tolerant, which keeps the model parameters frozen when the data flow is temporarily interrupted, and continues to update when new data arrives, ensuring continuous and stable operation in the production environment.

[0082] The quality prediction score is calculated through the model's forward propagation, expanding the feature vector input network. This is then processed by the hidden layer activation function, with ReLU chosen to introduce non-linearity. The output layer applies the sigmoid function to map the values ​​between 0 and 1, serving as the score. Score calculation is performed in real-time, synchronized with the production data stream. A score is generated immediately upon completion of each process node, and the results are cached in an in-memory database for use by the dynamic adjustment module. Model maintenance mechanisms include periodic pruning and bloat control to prevent overfitting. Pruning is achieved by removing unimportant connection weights, with importance evaluated based on gradient magnitude. The entire implementation emphasizes low latency and high throughput, for example, by accelerating feature concatenation through parallel computing and supporting large-scale data processing with distributed storage, ensuring the model's usability in continuous production environments.

[0083] See Figure 4Subfigure (a) illustrates the performance of the production quality prediction system in a real production environment. The figure contains three main data curves: the actual quality score curve (solid blue line) shows the real quality assessment results obtained based on offline quality inspection; the predicted quality score curve (dashed red line) shows the quality prediction values ​​calculated by the model based on anomaly stage markers and real-time production data; and the orange filled area reflects the number of anomaly stages detected in each production sample. A good consistency trend can be observed between the predicted curve and the actual quality curve, indicating that the model can effectively capture the quality change patterns in the production process. When the number of anomaly stages increases (the filled area rises), the corresponding quality score shows a downward trend, which verifies the importance of the anomaly stage marker sequence for quality prediction.

[0084] Subplot (b) uses a heatmap to illustrate the distribution of abnormal stages in the production sample, while overlaying the changing trends of key production characteristics. Darker colors in the heatmap indicate a higher probability that a particular production stage is marked as abnormal, clearly showing the spatial distribution pattern of abnormalities in the production process. The overlaid lines represent the normalized values ​​of different production characteristics, and the correlation between these characteristics and abnormal patterns can be analyzed through the correspondence between the lines and the heatmap regions. This visualization method helps to understand the potential relationship between changes in production characteristics and the occurrence of abnormalities, providing data support for optimizing production parameters.

[0085] Example 4: Dynamic Adjustment of Production Parameters. Production line configuration is dynamically optimized based on real-time comparison of quality prediction scores and preset thresholds. A hierarchical response mechanism ensures the accuracy and timeliness of adjustments. The specific implementation process uses a furniture production line as an example. This production line includes four main process nodes: cutting, sanding, assembly, and painting. Each node is equipped with sensors to monitor production data. The quality prediction score is received from the prediction model module, with the score range normalized to between 0 and 1, representing a probability estimate of the process completion quality. The preset thresholds are divided into two levels: a first-level threshold set at 0.7 and a second-level threshold set at 0.5. These thresholds are calibrated through historical quality data analysis, for example, referencing the 30th and 50th percentiles of the pass rate data distribution over the past year as benchmarks. The thresholds are stored in the configuration database and support dynamic updates.

[0086] When the system detects that the quality prediction score of a certain process node is lower than the first-level threshold, it triggers a priority adjustment strategy for equipment operating parameters. The strategy focuses on modifying the operating parameters of the equipment associated with that process. Taking the cutting process as an example, assuming the real-time score drops to 0.65, the system queries the equipment configuration library for that process to obtain the current parameters, such as the saw blade speed set to 3000 RPM and the feed rate set to 5 m / min. The system then calculates adjustment suggestions through the rule engine, which has a built-in expert knowledge base containing common fault modes and parameter mappings. For example, the score reduction may be due to saw blade wear, so it is recommended to increase the speed to compensate for efficiency loss. The adjustment range is derived based on the PID control principle, with the goal of raising the score above the threshold, but limiting the single adjustment range to no more than 10% to avoid drastic fluctuations. In specific implementation, the system adjusts the saw blade speed to 3300 RPM, while keeping the feed rate unchanged. The adjustment command is sent to the PLC controller for execution through the industrial network. After execution, the system monitors subsequent score changes, and if the score does not improve, it iterates and adjusts.

[0087] If the score further drops below the secondary threshold, for example, if the assembly process score drops to 0.45, the system initiates a strategy to simultaneously adjust the material ratio and equipment parameters. The material ratio adjustment involves recalculating the raw material usage. Taking the adhesive and wood used in the assembly process as an example, the current ratio is 5% adhesive. The system calculates the new ratio through an optimization algorithm based on a linear programming model. The objective function is to minimize cost while satisfying quality constraints. The constraints are derived from historical experimental data. For example, if the adhesive usage is less than 3%, it may lead to insufficient bonding strength. After adjustment, the ratio is set to 6% adhesive. At the same time, equipment parameters such as the air compressor pressure are increased from 0.5MPa to 0.55MPa. The adjustment decision is achieved through multi-objective optimization, balancing quality, cost, and efficiency. The system retrieves material inventory data in real time to ensure the feasibility of the new ratio. The adjustment command is simultaneously issued to the material distribution system and equipment controller to achieve collaborative operation. When the score exceeds a threshold, for example, if the spraying process score consistently stays above 0.8, the system retains the current configuration and records it as a candidate optimization solution. These candidate solutions are stored in a dedicated database, along with contextual information such as timestamps, production batch numbers, and environmental conditions. The records are formatted using JSON for easy querying, including parameters such as the nozzle diameter and paint flow rate of the current spraying machine. These solutions are marked as "high-scoring configurations" for subsequent analysis or wider adoption. The system periodically evaluates the effectiveness of the candidate solutions, confirming their stability by comparing the scoring trends of multiple batches. See Table 1 for the correspondence between the quality prediction score threshold and the adjustment strategy.

[0088] Table 1: Correspondence between Quality Prediction Scoring Thresholds and Adjustment Strategies

[0089] Rating range Threshold level Adjust strategy Adjusting the focus Adjusting the action [0.7,1.0] Above the threshold Preserve configuration No adjustment Record the current parameters as candidate solutions [0.5,0.7] Level 1 threshold Equipment parameters should be adjusted first. Modify equipment operating parameters Cutting process: Increase saw blade speed by 10%. [0.0,0.5] Secondary threshold Synchronous adjustment of materials and equipment Optimize material ratios and equipment parameters Assembly process: Adhesive content increased to 6%, press pressure increased by 10%.

[0090] The execution of the adjustment strategy relies on real-time data streams. The system integrates a monitoring module to track production data after adjustments. For example, after adjustments in the cutting process, the vibration frequency of the saw blade and the product dimensional accuracy are collected. This data is fed back to the scoring model to verify the adjustment effect. If the score does not recover within three production cycles after the adjustment, the system upgrades the adjustment strategy, such as introducing backup equipment or notifying manual intervention. All adjustment operations are recorded in the audit log, which includes parameters before and after adjustment, timestamps, and operator IDs, supporting backtracking analysis. During implementation, the system emphasizes response speed and security. For example, simulation verification is performed before issuing adjustment instructions to avoid equipment failure due to conflicting parameters. At the same time, high availability is ensured through redundancy design, such as automatic switchover to the backup system in case of main controller failure. The entire adjustment cycle achieves closed-loop control, continuously optimizing the production process.

[0091] Example 5: Similarity Matching and Simulation Testing. The optimization candidates are screened by calculating the geometric distance between configuration schemes, and their effectiveness is verified in a virtual environment. A furniture production line is used as an example. This production line includes four core processes: cutting, sanding, assembly, and painting. Each process is associated with a series of adjustable parameters, such as the saw blade speed in the cutting process, the sanding belt speed in the sanding process, the adhesive flow rate in the assembly process, and the paint pressure in the painting process. Parameter values ​​are monitored in real time by sensors and stored in the configuration database. The similarity matching process begins with the generation of adjusted production parameter configuration schemes. These schemes come from the output of the dynamic adjustment module. For example, when the quality prediction score triggers an adjustment, the system may generate multiple parameter combinations, such as increasing the cutting speed from 3000 RPM to 3200 RPM, or adjusting the adhesive ratio from 5% to 5.5%. Each configuration is represented as a numerical vector, with vector elements corresponding to the normalized values ​​of each parameter. The normalization process is based on the historical range of the parameters, mapping the values ​​to the [0,1] interval to eliminate the influence of dimensions.

[0092] When calculating the Euclidean distance between the adjusted configuration and the historical best configuration, the system retrieves the historical best configuration from the configuration library. The historical best configuration is selected from the parameter set with the highest quality score and good stability in past production batches, such as configuration number C-2023-05, which has a cutting speed of 3100 RPM, a grinding speed of 25 m / s, an adhesive ratio of 5.2%, and a coating pressure of 0.6 MPa. The retrieval process is based on tag filtering, such as matching by product type and equipment model, to ensure comparability. The Euclidean distance is calculated through vector difference operations. Each configuration vector contains the concatenation of all process parameters. For example, if the production line has 10 parameters, the vector length is 10. The distance formula is the square root of the sum of the squares of the differences between the parameters. Floating-point operations are used during calculation to ensure accuracy. The system processes multiple configuration pairs in parallel to improve efficiency. The matching radius is set as an adjustable parameter, with a default value of 0.1 determined based on the historical distance distribution. It is calibrated by analyzing the distance clustering of successful and failed configurations, and the radius value is reviewed periodically to adapt to changes in the production environment. Configurations with a Euclidean distance less than the matching radius are selected to form a set of configurations to be verified. The selection algorithm uses a range query, traversing all adjusted configurations and their distance values ​​from the historical best configuration. Results meeting the criteria are stored in a temporary set, the size of which is configurable to avoid resource strain from too many candidates. The selected configurations are sorted in ascending order of distance value using either quicksort or heapsort to ensure performance under large datasets. After sorting, the configuration with the smallest distance is processed first, as it is more similar to the historical best and theoretically more likely to succeed. The sorted results are encapsulated into a data structure containing the configuration ID, distance value, and parameter list, and sent to the simulation test phase via a message queue. Format validation is performed before sending to prevent data corruption.

[0093] The simulation testing phase loads the configuration to be verified in a digital twin environment. This digital twin environment is a virtual copy of the production line, built on a physics engine such as Unity or a dedicated industrial simulation platform. The environment model is generated through 3D scanning and training with historical data, accurately replicating the dynamic characteristics of actual equipment, such as the inertial effect of a cutting saw or the fluid behavior of a spraying machine. During configuration loading, the system parses the received parameter vectors and maps them to virtual equipment instances, such as setting the rotational speed parameters of a virtual saw blade. The loading process includes consistency checks to ensure parameters are within reasonable ranges and to avoid simulation crashes. The simulation runs a complete production cycle with a duration consistent with actual production; for example, a batch takes approximately 2 hours. The simulation time is shortened using time acceleration technology while maintaining the realism of the physical process. During operation, virtual sensors collect data such as virtual material consumption, equipment status changes, and process completion times. Data is recorded at a high frequency, supporting fine-grained analysis.

[0094] After virtual production data is generated, the system compares its deviation from actual historical data. Actual historical data is extracted from the production database, selecting real batch records under conditions similar to the simulation configuration, such as the same product and the same environmental temperature and humidity. Deviation is calculated using multiple indicators, such as mean absolute error (MAE) or relative deviation. MAE is obtained by comparing the average of the absolute differences between virtual and actual data point-by-point. Relative deviation is calculated as the percentage of the difference relative to the actual value. These indicators are combined to generate a deviation score, which is normalized to a range of 0-1; a lower score indicates better consistency. Verification results are based on the deviation score. For example, a threshold of 0.05 is set; a score below the threshold is marked as successful, otherwise as a failure. The results are accompanied by a detailed analysis report, such as parameter sensitivity analysis, indicating which parameters have the greatest impact.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent management of furniture production, characterized in that, The application relates to a production parameter dynamic adjustment method for a furniture production line. Real-time production data of each process node on the furniture production line is collected, and the production data comprises material consumption data, equipment operation parameters and process completion time; Multi-dimensional feature extraction is performed on the collected production data to generate a comprehensive feature set containing time sequence features, statistical features and correlation features; According to the dynamic change law of each feature in the comprehensive feature set, the production process is divided into a plurality of continuous production stages, and a stage identifier is assigned to each production stage; Based on the stage identifier, the production data is reorganized to form a stage data set corresponding to each production stage, and a transition probability matrix between adjacent stage data sets is calculated; By analyzing the abnormal mode of state jump in the transition probability matrix, potential abnormal stages in the production process are identified, and an abnormal stage marker sequence is generated; A production quality prediction model is constructed by combining the abnormal stage marker sequence and the real-time collected production data, and quality prediction scores of each process node are outputted; According to the deviation degree of the quality prediction score from a preset threshold value, the production parameter configuration scheme of the corresponding process node is dynamically adjusted; The adjusted production parameter configuration scheme is matched with a historical optimal configuration in terms of similarity, a to-be-verified configuration set is screened out and sent to a simulation test link; Verification results returned by the simulation test link are received, a production parameter configuration library is updated, and a final execution instruction is generated; The abnormal mode identification comprises: detecting a jump path with a probability value lower than a tolerance lower limit in the transition probability matrix; labeling a stage identifier combination with continuously appearing abnormal jumps; matching the labeled combination with a preset abnormal mode library to output an abnormal stage marker sequence; The construction of the production quality prediction model comprises: converting the abnormal stage marker sequence into a binary coding vector; performing feature splicing on the real-time production data and the coding vector; training the prediction model in an incremental learning mode and outputting a quality prediction score.

2. The intelligent management method for furniture production according to claim 1, characterized in that, The specific steps of the multi-dimensional feature extraction comprise: extracting unit time consumption rate and cumulative consumption amount from the material consumption data; extracting peak load duration and operation stability indicators from the equipment operation parameters; extracting standard deviation and percentage of deviation from the reference time length from the process completion time; aligning the above three types of features according to a time window and combining them into a comprehensive feature set.

3. The intelligent management method for furniture production according to claim 1, characterized in that, The basis for dividing the production stages comprises: detecting periodic fluctuation nodes of the time sequence features in the comprehensive feature set; statistical features with a mutation amplitude exceeding a preset sensitivity threshold in a continuous time window; a step change in coupling strength between correlation features.

4. The intelligent management method for furniture production according to claim 3, characterized in that, The calculation method of the transition probability matrix comprises: counting the occurrence frequency of the stage identifier in historical data; calculating the ratio of the conversion frequency to the total frequency between adjacent stage identifiers; generating the transition probability matrix after normalizing the ratio.

5. The intelligent management method for furniture production according to claim 1, characterized in that, The production parameter dynamic adjustment comprises: when the quality prediction score is lower than a first threshold value, triggering a device operation parameter priority adjustment strategy; when the score is lower than a second threshold value, synchronously adjusting material proportioning and equipment parameters; when the score is higher than the threshold value, retaining the current configuration and recording it as a candidate optimization scheme.

6. The intelligent management method for furniture production according to claim 5, characterized in that, The specific steps of the similarity matching comprise: calculating the Euclidean distance between the adjusted configuration and the historical optimal configuration; Screening the configuration set with distance less than the matching radius to form a configuration set to be verified; Sending the screened configuration in ascending order of distance value for verification.

7. The intelligent management method for furniture production according to claim 1, characterized in that, The verification method of the simulation test link comprises: Loading the configuration to be verified in the digital twin environment; Simulating a complete production cycle and collecting virtual production data; Generating a verification result by comparing the deviation of virtual data from actual historical data.

8. An intelligent management system for furniture production, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, implements the steps of the intelligent management method of furniture production according to any one of claims 1 to 7.

Images