A big data-based automobile aluminum trim part management method

Abstract

The application discloses a kind of based on big data's automobile aluminium trim management method, belong to aluminium trim management technical field, specifically include: through sensor collection raw material purity, extrusion process and environmental temperature and humidity data, transmission to cloud and according to unique product serial number are structuredly associated storage;Second, using dynamic time warping algorithm to data executes time sequence alignment, denoising and normalization processing, constructs full life cycle characteristic matrix.Subsequently, based on long short-term memory network, the characteristic matrix is iteratively trained, and a quality dynamic prediction model is established.In actual production, the model calculates the aluminum trim quality score in real time, automatically determines and marks the abnormal products and defect types below the threshold value.Finally, the application layer correlation propagation algorithm is used to analyze the model, locate the key feature dimension that causes the anomaly, and generate a closed-loop correction instruction for adjustable parameters, achieving intelligent optimization of production process.

Description

Technical Field

[0001] This invention relates to the field of aluminum trim management technology, and more specifically to a big data-based method for managing automotive aluminum trim. Background Technology

[0002] With the rapid development of the automotive industry and the increasing demands of consumers for the quality of automotive appearance and the reliability of components, automotive aluminum trim parts, due to their advantages such as lightweight, corrosion resistance, and aesthetic appeal, are widely used in automotive interior and exterior assembly. In the production management of automotive aluminum trim parts, the industry has gradually introduced sensor technology to collect data from various production stages, covering key processes such as raw material inspection, extrusion molding, and surface treatment. This involves recording and storing basic data such as raw material purity testing data, extrusion process parameters (such as extrusion temperature, extrusion speed, and holding time), and production environment temperature and humidity. Simultaneously, existing management systems often include basic quality inspection processes, determining the qualification of aluminum trim parts through finished product appearance inspection and dimensional measurement, forming a basic management process of "data collection-storage-finished product inspection," providing preliminary data and process support for the quality control of aluminum trim parts.

[0003] However, existing automotive aluminum trim management systems lack big data-driven quality correlation analysis capabilities. Quality data from each stage is stored in isolation, making it impossible to establish a quantitative correlation model between raw material purity, processing parameters, environmental temperature and humidity, and the quality of finished aluminum trim. When quality issues such as surface defects and dimensional deviations occur in aluminum trim, the root cause cannot be quickly located through data traceability. Instead, manual inspection is required, increasing manpower costs for quality control, prolonging the problem-solving cycle, reducing the timeliness of quality control, and failing to meet the stringent quality traceability standards of the automotive parts industry. Summary of the Invention

[0004] The purpose of this invention is to provide a big data-based management method for automotive aluminum trim parts, addressing the problems in the background technology:

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A big data-based management method for automotive aluminum trim components includes the following steps:

[0007] S1: Data on the purity of raw materials for aluminum decorative parts, extrusion process parameters, and ambient temperature and humidity are collected by sensors deployed at each stage of production, transmitted to the cloud server, and stored in a structured manner based on the unique product serial number;

[0008] S2: The dynamic time warping algorithm is used to perform time-series alignment on the structured associated storage data, remove noise interference and perform normalization processing to construct a full life cycle feature matrix of aluminum ornaments that includes raw material, process and environmental characteristics.

[0009] S3: Divide the full life cycle feature matrix of aluminum trim into training set and validation set, input it into long short-term memory network for iterative training, and establish a dynamic prediction model for aluminum trim quality with production process features as input and finished product quality indicators as output.

[0010] S4: Real-time collection of current production process data is input into the dynamic prediction model for aluminum trim quality, and the quality score of the current aluminum trim is calculated. If the score is lower than the standard threshold, it is automatically judged as an abnormal product and the specific defect type is marked.

[0011] S5: For aluminum trim parts that are determined to be defective, the layered class correlation propagation algorithm is used to analyze the activation state of neurons inside the model, locate the key feature dimension that causes the quality abnormality, and if the feature dimension belongs to an adjustable process parameter, a closed-loop correction instruction for that parameter is generated.

[0012] As a further aspect of the present invention: in step S1, the process of transmitting the data to the cloud server and performing structured association storage based on the unique product serial number is as follows:

[0013] The cloud server receives the encapsulated data packet uploaded by the edge gateway. The data packet contains the purity of the raw materials, extrusion process parameters, and environmental temperature and humidity values ​​that are bound to the unique product serial number.

[0014] The server parses the data packet to extract the unique product serial number as the primary key of the database, and constructs a mapping index table containing fields for raw material purity, extrusion process parameters, and environmental temperature and humidity.

[0015] The parsed data values ​​are written into the storage field of the corresponding unique product serial number in the mapping index table, forming a structured associated data entry indexed by the unique product serial number.

[0016] As a further aspect of the present invention: in step S2, the process of constructing a full life-cycle feature matrix of aluminum decorative parts that includes raw material, process, and environmental characteristics is as follows:

[0017] Read the data sequence from the structured associative storage, select the standard process timing as the benchmark, and apply the dynamic time warping algorithm to map the data sequence to a unified time axis;

[0018] Perform Kalman filtering on the data sequence on a unified time axis to remove high-frequency random noise and generate a smooth data sequence that retains trend characteristics;

[0019] The extreme values ​​are calculated by traversing the smoothed data sequence, and the range transformation formula is used to map the values ​​to a closed interval between zero and one, thus completing the data normalization process.

[0020] By splicing the normalized raw material purity, extrusion process parameters, and environmental temperature and humidity data along the time dimension, a full life cycle characteristic matrix of aluminum decorative parts is constructed.

[0021] As a further aspect of the present invention: the specific method for performing Kalman filtering on the data sequence on a unified time axis to remove high-frequency random noise and generate a smooth data sequence that retains trend characteristics is as follows:

[0022] A linear state-space model is constructed for the data sequence on a unified time axis. The system state transition matrix and observation matrix are set, and the prior state estimate and prior error covariance matrix are calculated using the posterior state estimate of the previous time step.

[0023] The Kalman gain is calculated by combining the prior error covariance matrix and the observation noise covariance matrix. The Kalman gain is then used to perform a weighted correction on the residual between the current observation value and the prior state estimate in the data sequence, and the posterior state estimate at the current time is calculated.

[0024] The updated posterior error covariance matrix is ​​used for the next iteration calculation. The posterior state estimates calculated at each time step on a unified time axis are extracted to form a smooth data sequence that retains trend characteristics.

[0025] As a further aspect of the present invention: in step S3, the process of establishing a dynamic prediction model for the quality of aluminum decorative parts, with production process characteristics as input and finished product quality indicators as output, is as follows:

[0026] Based on the time series sequence, the feature matrix of the entire life cycle of aluminum ornaments is divided into non-overlapping training set matrix and validation set matrix;

[0027] Construct a long short-term memory network structure that includes an input layer, hidden layers, and a fully connected output layer, and initialize the network weight parameters and bias terms;

[0028] The training set matrix is ​​input into the Long Short-Term Memory network to perform forward propagation, and the deviation between the prediction results and the finished product quality indicators is calculated using the loss function.

[0029] The backpropagation algorithm is used to update the network weights based on the deviation gradient until the validation set error converges, outputting a dynamic prediction model for the quality of aluminum trim parts.

[0030] As a further aspect of the present invention: In step S4, the process of real-time acquisition of current production process data, inputting it into the dynamic prediction model for aluminum trim quality, calculating the current quality score of the aluminum trim, identifying abnormal products, and marking specific defect types is as follows:

[0031] The sensor values ​​in the current production process are read in real time, and the preprocessing rules in the training stage are reused to perform time alignment and normalization to generate the current feature matrix to be tested.

[0032] The current feature matrix to be tested is input into the dynamic prediction model of aluminum trim quality. The output layer of the model calculates the current quality score and defect category probability distribution of the aluminum trim.

[0033] The quality score is compared with a preset standard threshold. If the quality score is less than the preset standard threshold, an abnormal product judgment result is generated and the defect identification logic is activated.

[0034] Based on the maximum probability value index in the defect category probability distribution, the pre-stored defect type database is matched to associate the abnormal product with and label it with a specific defect type name.

[0035] As a further aspect of the present invention: in step S5, the process of generating the closed-loop correction instruction for the parameter is as follows:

[0036] The feature matrix of aluminum ornaments that are determined to be abnormal is input into the dynamic prediction model of aluminum ornament quality. The neuron nodes in the output layer that indicate the abnormality category are locked and their activation values ​​are set to the initial total correlation value.

[0037] The initial total correlation value is back-allocated according to the network weights using the layer-level correlation propagation rule. This process is then reversed step by step through the fully connected layer and hidden layer to calculate the correlation score for each feature dimension in the input layer.

[0038] Compare the relevance scores of all feature dimensions in the input layer and select the feature dimension with the highest value as the key feature dimension;

[0039] Determine whether the key feature dimension belongs to the preset list of adjustable process parameters;

[0040] If so, retrieve the preset process compensation strategy that matches the key feature dimension, calculate the physical adjustment value required for the parameter, and generate a closed-loop correction instruction containing the target device ID and the adjustment value.

[0041] As a further aspect of the present invention: the specific method for calculating the correlation score of each feature dimension in the input layer by using the layer-level correlation propagation rule to back-allocate the initial total correlation value according to the network weights, and then progressively calculating it through the fully connected layer and hidden layer, is as follows:

[0042] Calculate the weighted product of the activation value of a neuron in the fully connected layer with its connection weight, and define the proportion of this product to the sum of all weighted products in the layer as the contribution coefficient. Based on the contribution coefficient, allocate the initial total relevance value of the output layer to each neuron in the fully connected layer.

[0043] The correlation values ​​obtained from the fully connected layer neurons are used as backpropagation input. Based on the temporal unfolding structure of the long short-term memory network, the hidden layer network structure is traversed. The excitation intensity of the previous layer neurons to the next layer neurons is calculated according to the inter-layer connection weights and neuron activation states. The correlation values ​​are then passed forward layer by layer according to the excitation intensity ratio.

[0044] When the correlation value is backpropagated to the input layer, the correlation components transmitted from all nodes in the first hidden layer to the corresponding nodes in the input layer are collected, and a summation operation is performed to obtain the cumulative correlation value of that node, which is then established as the correlation score for each feature dimension of the input layer.

[0045] The beneficial effects of this invention are:

[0046] This invention achieves structured, interconnected storage of data across the entire production process through unique product serial numbers, completely resolving the pain point of isolated data storage in existing technologies. Simultaneously, the full lifecycle feature matrix constructed through time-series alignment and normalization processing using a dynamic time warping algorithm provides a high-quality data foundation for multi-dimensional data correlation analysis. The quality dynamic prediction model, trained on a long short-term memory network, can output real-time quality scores for aluminum components and accurately label defect types, enabling early warning of quality issues. Compared to traditional manual finished product inspection, this significantly reduces the outflow of defective products, improves the finished product pass rate, and eliminates the tedious process of manual piece-by-piece inspection, reducing the labor and time costs of quality control.

[0047] This invention utilizes a layered correlation propagation algorithm to accurately locate the key characteristic dimensions of abnormal products, overcoming the bottleneck of low efficiency in manual investigation and traceability in existing technologies. It can quickly pinpoint the root cause of problems in raw material purity, process parameters, or environmental temperature and humidity, shortening the quality problem resolution cycle. For the closed-loop correction instructions generated by adjustable process parameters, real-time dynamic optimization of production parameters can be achieved, preventing the recurrence of similar quality problems from the source. This significantly improves the timeliness and stability of quality control, fully meeting the stringent quality traceability standards of the automotive parts industry, and helping enterprises establish an efficient and accurate end-to-end quality control system. Attached Figure Description

[0048] The invention will now be further described with reference to the accompanying drawings.

[0049] Figure 1 This is a flowchart illustrating a big data-based management method for automotive aluminum trim parts according to the present invention. Detailed Implementation

[0050] 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.

[0051] Please see Figure 1 As shown, this invention is a big data-based management method for automotive aluminum trim parts, comprising the following steps:

[0052] S1: Data on the purity of raw materials for aluminum decorative parts, extrusion process parameters, and ambient temperature and humidity are collected by sensors deployed at each stage of production, transmitted to the cloud server, and stored in a structured manner based on the unique product serial number;

[0053] S2: The dynamic time warping algorithm is used to perform time-series alignment on the structured associated storage data, remove noise interference and perform normalization processing to construct a full life cycle feature matrix of aluminum ornaments that includes raw material, process and environmental characteristics.

[0054] S3: Divide the full life cycle feature matrix of aluminum trim into training set and validation set, input it into long short-term memory network for iterative training, and establish a dynamic prediction model for aluminum trim quality with production process features as input and finished product quality indicators as output.

[0055] S4: Real-time collection of current production process data is input into the dynamic prediction model for aluminum trim quality, and the quality score of the current aluminum trim is calculated. If the score is lower than the standard threshold, it is automatically judged as an abnormal product and the specific defect type is marked.

[0056] S5: For aluminum trim parts that are determined to be defective, the layered class correlation propagation algorithm is used to analyze the activation state of neurons inside the model, locate the key feature dimension that causes the quality abnormality, and if the feature dimension belongs to an adjustable process parameter, a closed-loop correction instruction for that parameter is generated.

[0057] In one embodiment of the present invention, the process of collecting data on the purity of aluminum decorative parts raw materials, extrusion process parameters, and ambient temperature and humidity through sensors deployed at various stages of production, transmitting the data to a cloud server, and storing it in a structured association based on a unique product serial number is as follows:

[0058] The cloud server continuously listens for upload requests from the edge gateway at the production site via a high-bandwidth network interface. The edge gateway acts as a central node for data aggregation at the production site. It first connects to various independent acquisition probes via wired or wireless means, converting the acquired analog signals into computer-readable digital signals. The edge gateway then packages and processes the multi-dimensional data collected within the same time window. This data specifically includes the percentage of aluminum in the raw materials for aluminum trim and the values ​​of trace elements such as iron and silicon, as the chemical composition of the raw materials directly determines the mechanical strength and corrosion resistance of the final product. It also includes the real-time pusher speed of the extruder during operation and the real-time temperature at the die exit. These process parameters determine the physical state of the aluminum trim during the forming process. Additionally, the ambient temperature and relative humidity within the production workshop are recorded simultaneously. Environmental factors affect the natural cooling rate of the aluminum material through heat exchange, thus altering its crystal structure. The edge gateway forcibly binds all the data from these different sources to the unique identification code of the currently processed aluminum trim, encapsulating them in a unified data transmission unit. This encapsulation mechanism utilizes the principle of data packet integrity, ensuring strict time synchronization and logical correspondence of all parameters during transmission, preventing data misalignment due to network latency. Once the cloud server receives this complete data packet, it has completed the acquisition of the original data.

[0059] After receiving the encapsulated data packet, the cloud server initiates a memory parsing program. It reads the binary data stream bit by bit using a preset decoding protocol. This process aims to restore the original information encoded by the edge gateway. The server first retrieves the unique identification code of the aluminum trim piece from the header or a specific identifier bit of the data stream. Because this code is globally unique and accompanies the entire product lifecycle, the server establishes it as the core benchmark for data indexing. Next, based on the number and attributes of the data types contained in the data packet, the server dynamically plans a logical mapping table in the database storage space. This mapping table is similar to a multi-dimensional information storage grid. The server assigns a dedicated logical column to each parsed data category; for example, a column specifically for storing... The system includes columns for raw material chemical composition values, columns for recording extrusion speed and temperature, and columns for recording ambient temperature and humidity. This structure is based on relational data management principles. Predefined data types and storage lengths ensure the standardization of written data. The server establishes a row-column interleaved index structure, creating a one-to-one logical correspondence between abstract unique identifiers and specific physical parameter categories. This ensures that every piece of data to be written can find its accurate location, laying the foundation for rapid retrieval and correlation analysis of large-scale data. This primary key-based indexing method greatly improves data query efficiency and avoids data storage chaos and redundancy.

[0060] After the index structure was built, the cloud server began executing the data entity writing operation. The server extracted the specific values ​​parsed from the data packets one by one; for example, it precisely matched the parsed raw material aluminum content value to the raw material purity storage location in the mapping table; it filled the extruder pusher speed value and die temperature value into the corresponding process parameter storage location; at the same time, it wrote the workshop environment temperature and humidity values ​​into the environmental parameter storage location. This process was strictly carried out according to the previously established logical pointing relationship. The server performed data type verification when writing each piece of data to ensure that the written value met the predefined precision requirements. When all the categorized data accurately fell into their corresponding storage locations, these originally discrete... Data from different sources are combined in the mapping table to form a complete data record. This record is headed by the unique identification code of the aluminum trim piece, connecting all the status information of the product at the moment of production, forming a tightly structured data entry. The principle of this structured storage is to use the rows and columns of a two-dimensional table to solidify the logical relationship between data, so that any product code can be directly indexed to its corresponding set of production and environmental parameters, realizing the transformation of data from fragmented to ordered. The stored data entries not only contain the numerical values ​​themselves, but also implicitly contain the spatiotemporal correlation attributes between these values, thus building a high-quality data foundation to support quality traceability and model training, and completing the transformation from raw signals to structured assets.

[0061] In one embodiment of the present invention, the process of constructing a full life-cycle feature matrix of aluminum decorative parts, including raw material, process, and environmental characteristics, by using a dynamic time warping algorithm to perform time-series alignment on the structured associated storage data, remove noise interference, and perform normalization processing, is as follows:

[0062] The server first retrieves all the original data sequences of the aluminum trim pieces to be processed from the database in batches based on the index. These data sequences often exhibit non-linear characteristics of varying lengths over time. This is because the operating speed of the equipment during actual extrusion production is slightly affected by voltage fluctuations or hydraulic instability, causing the duration of the same process to change. To eliminate this misalignment on the time axis, the server selects a golden batch production curve certified by experts as the standard process sequence. The core of applying the dynamic time warping algorithm lies in finding an optimal warping path. This algorithm constructs a cumulative distance matrix to measure the relationship between the data sequence to be processed and the standard process sequence. The algorithm calculates the similarity between the two sequences by calculating the Euclidean distance between each corresponding point in the two sequences and finding the path with the minimum cumulative distance. The algorithm can intelligently identify the characteristic peaks and valleys in the data waveform. For time-stretched segments caused by slow production speed, the algorithm will perform compression processing. For time-shortened segments caused by fast production speed, the algorithm will perform interpolation stretching processing. This non-linear mapping process can ensure that every process action in the data to be processed is strictly aligned to the scale of the standard process sequence on the time axis. This gives the mathematical basis for horizontal comparison of product data from different batches under the same time reference. The algorithm also completes the geometric correction of the data in the time dimension.

[0063] Kalman filtering is applied to the data sequence on a unified time axis to remove high-frequency random noise and generate a smooth data sequence that retains trend characteristics. The specific method is as follows:

[0064] A linear state-space model is constructed for the aluminum trim production data sequence that has been unified to the same time axis. The core of this model lies in defining two key matrices: the state transition matrix and the observation matrix. The state transition matrix describes the inherent law of the evolution of aluminum trim extrusion process parameters over time under ideal physical conditions; for example, the extrusion cylinder temperature has thermal inertia and will not undergo abrupt changes. The observation matrix describes the linear mapping relationship between sensor readings and the actual physical state. Using the posterior state estimate calculated at the previous time step, which represents the judgment closest to the true value after combining all information from the previous time step, this value is input into the state evolution law for deduction, thereby calculating the prior state estimate at the current time step. This prior value is essentially a theoretical prediction of the current state based on physical laws and does not yet include the actual sensor reading information at the current time step. At the same time, based on the error propagation law, combined with the posterior error covariance matrix and process noise covariance matrix of the previous time step, the prior error covariance matrix at the current time step is calculated. This matrix quantifies the degree of uncertainty brought about by relying solely on theoretical predictions and provides a mathematical basis for the next step of weighted correction based on actual observation values. The entire process simulates the natural evolution of the physical process.

[0065] After obtaining the theoretical prediction results, the crucial correction stage begins. First, the Kalman gain, a core weighting coefficient, needs to be calculated. This coefficient's calculation relies on the ratio of the prior error covariance matrix to the observation noise covariance matrix. Its physical meaning lies in assessing the relative magnitude between the uncertainty of the theoretical prediction and the sensor measurement noise. If the sensor noise is high, the gain value will decrease, thus favoring the theoretical prediction; conversely, the gain value will increase to utilize high-precision measurements. Next, the difference between the actual observation value and the prior state estimate in the data sequence is calculated. This difference is called the residual, representing the degree of deviation between the actual detection data and the theoretical prediction data. The calculated Kalman gain is used to weight this residual. The weighted residual is then superimposed on the prior state estimate, thus obtaining the posterior state estimate for the current moment. This process is essentially a data fusion for optimal estimation. Through mathematical methods, a balance point with minimal error is found between noisy observations and predictions based on physical laws. Errors caused by model bias or instantaneous sensor interference are corrected, ensuring that the final estimated value conforms to both the physical evolution trend and closely follows the actual measurement results.

[0066] After completing the state correction at the current moment, the statistical parameters inside the algorithm need to be updated to prepare for the iterative calculation at the next moment. The prior error covariance matrix is ​​corrected and transformed using Kalman gain and identity matrix. The posterior error covariance matrix at the current moment is calculated. The principle of this step is that by introducing the current observation information, the uncertainty of the state estimation is mathematically reduced. The updated covariance matrix will serve as the basic input parameter for the prediction step at the next moment, thus maintaining the recursive operation mechanism of the algorithm. As the time axis progresses, the algorithm will repeat the above prediction and correction calculation process for each discrete time point. The server extracts the posterior state estimate calculated at each time point in sequence. These estimates have been freed from random high-frequency noise interference. They are rearranged and combined according to the original time order to form a complete smooth data sequence that retains the characteristics of the process change trend. This sequence retains the real process fluctuations such as extrusion speed changes or temperature rises and falls, while filtering out glitch signals caused by circuit interference or vibration. It provides high signal-to-noise ratio basic data support for the construction of the feature matrix.

[0067] Although the smoothed data sequence after filtering has a clear waveform, its numerical magnitudes vary greatly. For example, the extrusion temperature may be as high as several hundred degrees, while the trace element content in the raw material is only a few tenths of a degree. This disparity in magnitude can cause the neural network to fail to converge during training or become overly reliant on large numerical features. Therefore, the server needs to traverse each smoothed data sequence to perform a normalization operation. During the traversal, the algorithm compares and records the maximum and minimum values ​​of the sequence in real time throughout its entire lifecycle to determine the range of values ​​for this parameter. Then, the range transformation principle is used to recalculate each sampling point in the sequence. The specific calculation logic is to recalculate the current sampling point... The relative offset is obtained by subtracting the minimum value of the sequence from the sample value; then, this relative offset is divided by the difference between the maximum and minimum values. Through this mathematical transformation, the data originally distributed at different orders of magnitude and different physical units are forcibly mapped to a closed interval between zero and one. In this new numerical space, whether it is a temperature of several hundred degrees or a pressure of several megapascals, they all become equally weighted. This processing method eliminates the interference of physical dimensions on the model weights, making all features equal at the mathematical level, thereby ensuring that the neural network can fairly extract information from each feature dimension and completing the standardization preparation before the data is input into the model.

[0068] Finally, the server performs a matrix construction operation for multi-dimensional features to adapt to the input format of the deep learning model. This process involves physically concatenating all normalized independent data streams along the time dimension. The server creates a two-dimensional matrix structure with time steps as rows and feature dimensions as columns. For extrusion process parameters and environmental temperature and humidity data that change continuously over time, the server directly fills the corresponding normalized values ​​into the columns of the matrix corresponding to the time step. For static data such as raw material purity that remains constant throughout the production process, the server uses a numerical broadcasting strategy to copy and fill the corresponding columns of each time step, ensuring that the data is displayed correctly in the matrix. Unfolding into a constant value sequence in the time dimension; through this splicing method, the originally scattered aluminum content data, temperature data, pressure data, and environmental humidity data are integrated into a unified mathematical tensor; each row in this aluminum trim full life cycle feature matrix represents a complete state snapshot at a specific millisecond moment in the production process; and each column represents the entire process of the evolution of a specific monitoring indicator over time; this high-dimensional, high-density information organization form completely preserves the spatiotemporal coupling relationship between features; and provides a standardized data carrier for Long Short-Term Memory Networks to capture the dynamic temporal dependencies in the production process.

[0069] In one embodiment of the present invention, step S3, which involves dividing the full lifecycle feature matrix of aluminum trim into a training set and a validation set, inputting it into a long short-term memory network for iterative training, and establishing a dynamic prediction model for the quality of aluminum trim with production process features as input and finished product quality indicators as output, is as follows:

[0070] The server first retrieves all the original data sequences of the aluminum trim pieces to be processed from the database in batches based on the index. To ensure that the model can learn the dynamic laws of the production process over time, the server strictly follows the physical chronological order of the time series to arrange the data. All samples in the full life cycle feature matrix of the aluminum trim pieces are sorted unidirectionally according to the production date. The server sets a specific time split point or quantity ratio split point and cuts the sorted data sequence into two non-overlapping parts. The historical data in the first half of the time axis is divided into the training set matrix, mainly used for learning and fitting the model's internal parameters. The data in the second half of the time axis is divided into the validation set matrix, specifically used to evaluate the model's generalization ability. The principle behind this time-series-based segmentation strategy is that there are long-term trends such as equipment wear and sensor drift in the industrial production environment. If the data is randomly shuffled, it will destroy this time dependence, causing the model to fail when facing new data in the future. Through this physical isolation, it is ensured that the data distribution of the validation set is completely unknown to the model, thus enabling a realistic simulation of the model's prediction performance when facing new production batches after going live.

[0071] After preparing the data, the server begins building the basic architecture of the deep learning network. A long short-term memory (LSTM) network topology is constructed, comprising an input layer, hidden layers, and a fully connected output layer. The number of nodes in the input layer is set to be exactly the same as the number of columns in the feature matrix, ensuring that each normalized raw material purity, extrusion process parameter, and environmental temperature and humidity feature can find a corresponding input channel. The hidden layer consists of multiple stacked LSM units, which integrate complex gating mechanisms such as input gates, forget gates, and output gates. The principle is that it can autonomously decide which historical information needs to be retained for a long time and which short-term fluctuations need to be forgotten during the time-series transmission process, thereby capturing the lag effect of production parameters on quality. The fully connected output layer is responsible for mapping the high-dimensional abstract features extracted by the hidden layer back to specific quality indicator dimensions. After the network structure is established, the server performs parameter initialization, using a random number generation algorithm to assign extremely small random initial values ​​to thousands of weight parameters and bias terms in the network. The physical significance of this step is to break the symmetry of neurons, preventing all nodes from learning the exact same feature patterns in the early stages of training, thus providing mathematical possibilities for the network to capture complex nonlinear production laws.

[0072] After network initialization, the server initiates the forward propagation computation process; the training set matrix is ​​input into the input layer of the Long Short-Term Memory network in batches according to the time step; the data flows forward along the time axis inside the network; the input data at each time step is fused with the cell state passed down from the previous time step; the gating unit dynamically adjusts the information throughput according to the fluctuation of the current process parameters; after multiple nonlinear transformations, the data finally reaches the fully connected output layer; the output layer outputs the predicted quality value corresponding to each sample; the server then retrieves the actual finished product quality index recorded in actual production for that sample; the predicted value is compared with the actual value using loss functions such as mean square error or cross-entropy; the geometric distance or probability deviation between the two is calculated; this deviation value quantifies the accuracy of the current model's judgment on the quality of aluminum trim parts; the underlying principle is to transform the abstract quality prediction problem into a specific mathematical optimization problem; the magnitude of the quantified error is used to measure the quality of the current network parameter configuration; providing a clear numerical guide for the next step of parameter correction.

[0073] Based on the calculated deviation value, the server activates the backpropagation algorithm to optimize the model; using the chain rule in calculus, it derives the model layer by layer from the output layer to the input layer; it calculates the gradient value of the loss function with respect to each weight parameter and bias term; this gradient value mathematically represents the direction of the fastest error descent; the server makes minor adjustments to the weights in the network based on this gradient direction and the preset learning rate; this allows the error generated by the network in the next prediction to be slightly reduced; this process is called a training iteration; the server continuously repeats the cycle of forward calculation and backward adjustment; after each iteration, the server inputs the validation set matrix into the current model for testing; it observes the error change curve of the model on unseen validation data; as training progresses, the validation error will first decrease and then tend to stabilize; when the server detects that the validation set error no longer decreases significantly with the number of iterations or even begins to increase, it determines that the model has reached convergence according to the early stopping mechanism; at this time, it stops training and freezes all weight parameters; and outputs the finalized dynamic prediction model for aluminum trim quality; this mechanism ensures that the model not only remembers historical data, but also has the ability to reason about future production quality.

[0074] In one embodiment of the present invention, step S4 involves real-time acquisition of current production process data, inputting it into the dynamic prediction model for aluminum trim quality, calculating the current quality score of the aluminum trim, and automatically identifying it as an abnormal product and marking the specific defect type if the score is lower than the standard threshold.

[0075] The server opens a real-time listening port to read the latest data streams uploaded by various sensors on the production site at an extremely high frequency. These data streams cover key physical quantities such as real-time temperature, pressure, and speed of the aluminum trim pieces currently being extruded. To ensure that the data input to the model maintains strict consistency in mathematical distribution with the data encountered during model training, the server must call the training phase preprocessing rule configuration file stored in the storage module. This file records the mean and variance of the training set data and the baseline curve of the standard process timing. The server executes the exact same preprocessing logic based on these parameters. First, a dynamic time warping algorithm is applied to map the currently acquired nonlinear time series onto the standard baseline curve. The principle of this step is to eliminate the time axis misalignment caused by production cycle fluctuations and ensure that each sampling point is aligned to the correct process stage. Next, the extreme value parameters in the configuration file are read to perform normalization calculations, compressing all physical quantities into the numerical range between zero and one. This reuse of rules is to ensure the isomorphism of the feature space and prevent model weight failure due to data dimension or distribution offset. Finally, the processed multidimensional data is concatenated in the time dimension to generate a current test feature matrix that meets the requirements of the model input tensor.

[0076] The server loads the constructed feature matrix to be tested into memory and pushes it into the input layer of the aluminum trim quality dynamic prediction model, which has been trained and has its parameters frozen. The data propagates forward through the hierarchical structure of the Long Short-Term Memory (LSTM) network. Each neuron in the layer performs weighted summation and nonlinear activation on the input features based on the pre-learned weight matrix. The LSM unit uses its internal gating mechanism to dynamically analyze the temporal dependencies in the production process and capture the cumulative impact of process parameter fluctuations on the final quality. When the data flows to the output layer of the model, the network structure is designed as a dual-path parallel output mode. One path of neurons performs a regression task, calculating a continuous scalar value through a linear activation function. This value represents the overall quality score of the aluminum trim. The other path of neurons performs a classification task, calculating a vector through a probabilistic activation function. Each dimension of the vector corresponds to a predefined defect type, and the value of each dimension represents the confidence probability that the sample belongs to the corresponding defect category. Thus, in one forward inference process, both quality scoring and probability prediction of potential defects are completed simultaneously.

[0077] The server extracts the scalar value of the quality score output by the model and reads the pre-set standard threshold from the configuration database. This threshold is a critical line set by the quality management department based on historical yield rates and customer acceptance standards. The server executes numerical comparison logic to compare the real-time calculated quality score with the standard threshold. The principle is to use quantitative indicators to define the boundary between qualified and unqualified products. If the current quality score is significantly lower than the preset standard threshold, it indicates that the production process parameters of the aluminum trim have deviated from the ideal state according to the model's comprehensive judgment, and there is a high possibility of quality problems. At this time, the logic judgment module will immediately generate an abnormal product judgment result, mark the digital twin file of the aluminum trim as abnormal, and trigger an alarm signal to notify the on-site control unit. At the same time, the defect identification logic module in standby state is activated. This step realizes the logical jump from macro quality assessment to micro defect analysis, ensuring that only products with real problems will enter the complex attribution analysis process, thereby saving computing resources and improving response speed.

[0078] After confirming a product as defective, the defect identification logic module immediately analyzes the defect category probability distribution vector output by the model. This vector contains probability values ​​for various defects such as cracks, bubbles, and dimensional deviations. The server traverses the entire vector to find the element with the largest value. The mathematical principle behind this operation is maximum likelihood estimation, which assumes that the category with the highest probability value is most likely to reflect the current true situation. The server obtains the index position of the maximum probability value; for example, if the maximum value appears in the third position of the vector, then the index value 3 is extracted. The server uses this index value as a query keyword to search in a pre-stored defect type database. The database stores the mapping relationship between index values ​​and specific text descriptions. Through matching, the server finds the specific defect name corresponding to the index value, such as surface scratches or internal pores, and associates this specific text label with the electronic record of the current defective product. This process transforms abstract mathematical probabilities into technical terms that workshop operators can directly understand, providing clear semantic guidance for manual review and process adjustments.

[0079] In one embodiment of the present invention, in step S5, the process of using a layered class correlation propagation algorithm to analyze the activation state of neurons within the model for aluminum trim parts determined to be defective, locating the key feature dimension causing the quality defect, and generating a closed-loop correction instruction for that parameter if the feature dimension is an adjustable process parameter, is as follows:

[0080] The server first reloads the feature matrix of the aluminum trim pieces previously identified as having quality defects into the memory buffer. This matrix records all time-series data and environmental parameters of the product throughout the entire production process. The server then inputs this matrix back into the already trained dynamic prediction model for aluminum trim piece quality; however, the purpose this time is not to make new predictions but to perform reverse attribution analysis. The program logic directly locks the fully connected output layer in the model architecture, finds the neuron node representing the specific anomaly category; for example, if the model determines it to be a surface crack, it locks the node corresponding to the crack, reads the activation value generated by that node during the inference process, and directly sets this value as the initial total correlation value. The theoretical basis for this operation lies in the correlation conservation law in interpretability algorithms, which states that the final judgment made by the model's output layer is the sum of the contributions of all features in the input layer. Therefore, the output value is used as the total amount of evidence to be assigned, establishing the mathematical starting point for inferring the cause from the result, and laying the foundation for accurately locating the root cause of the quality problem.

[0081] The initial total correlation value is back-allocated according to the network weights using the layer-level correlation propagation rule. This process is repeated step-by-step through fully connected layers and hidden layers to calculate the correlation score for each feature dimension in the input layer. The specific method is as follows:

[0082] The server first initiates a correlation allocation calculation program for the fully connected layer. This program reads the activation values ​​generated by each neuron in the fully connected layer during forward propagation, as well as the connection weight values ​​between that neuron and the output layer nodes. It then calculates the weighted product of these two values ​​using multiplication. This product physically represents the direct contribution of that neuron to the final output. To determine the relative position of each neuron in the overall contribution, the server sums the weighted products of all neurons in the layer to obtain a total value. Next, it uses division to calculate the percentage of each individual weighted product relative to this total value. This calculated percentage is defined as the contribution coefficient. The principle behind this is that normalization converts absolute values ​​into relative weights, ensuring that the sum of all coefficients is strictly equal to 1. Based on the calculated contribution coefficients, the server divides the initially set total correlation value of the output layer into several parts and assigns them proportionally to each neuron in the fully connected layer. This process follows the law of correlation conservation, ensuring that the total amount of evidence transmitted from the output layer back to the fully connected layer remains numerically unchanged. This completes the first attribution from macroscopic judgment results to higher-level abstract features.

[0083] After each neuron in the fully connected layer obtains its own correlation value, the server uses this as a new starting point to initiate backpropagation computation on the Long Short-Term Memory (LSTM) network structure. Because the LSM network exhibits cyclic dependencies in the time dimension when processing sequential data, the server first logically unfolds the folded recurrent neural network into a chain structure arranged by time steps. The algorithm traverses the entire hidden layer network structure backward from the last time step. At each time point, the server reads the current neuron state and the hidden state passed from the previous time step. Combining this with the fixed connection weight matrix between layers, the server calculates the activation intensity of neurons in the previous layer on neurons in the next layer. This activation intensity quantifies the smoothness of information flow from historical moments to the current moment. If the connection weight is large and the activation value in the previous time step is high, it indicates that the path is the main channel for information transmission. Based on the relative proportions of these activation intensities, the server distributes the correlation value held at the current moment to neurons in the previous time step or the next layer. This layer-by-layer forward propagation process simulates the reverse flow of signals in the neural network. Through continuous weighted allocation, the correlation value is pushed back to the beginning of the network step by step along the time path that has the greatest impact on the result.

[0084] When the correlation values ​​finally reach the boundary between the input layer and the first hidden layer after a long backpropagation path, the server begins the final aggregation and summation operation. Each node in the input layer corresponds to a specific data point in the feature matrix, such as the extrusion temperature value at a specific moment. Since the first hidden nodes of the Long Short-Term Memory network usually have fully connected or densely connected relationships with the input layer, this means that an input data point may send signals to multiple hidden layer nodes simultaneously. Conversely, during backpropagation, this input node will also receive the correlation components returned from multiple hidden layer nodes. The server traverses all nodes in the first hidden layer, filters out all connection paths pointing to the current input node, and sums the correlation values ​​returned along these paths. The mathematical principle of this summation process lies in the superposition principle, which assembles the evidence fragments returned from different paths into a complete chain of evidence. The calculated sum is established as the cumulative correlation value of that node. This value directly reflects the contribution of that specific physical feature to the final quality anomaly determination. The server repeats this operation for each feature dimension of the input layer, ultimately generating a correlation heatmap that is completely consistent with the dimensions of the original input matrix.

[0085] After obtaining the relevance scores of all dimensions in the input layer, the server performs a full-domain scan and filtering logic. This step aims to extract the core causes of quality problems from massive amounts of production data. The algorithm compares the numerical values ​​of the relevance scores of all feature dimensions in the input layer. The larger the relevance score, the stronger the positive correlation between the feature and the abnormal result, which is the primary factor leading to the model's judgment of a defective product. The server uses bubble sort or quick sort algorithm to find the feature dimension with the largest value and, based on the index mapping table constructed during feature matrix construction, restores this abstract matrix coordinate to a process name with actual physical meaning. For example, the maximum score corresponds to the temperature data of the third temperature zone of the extruder at the 200th second of the production process. This method of finding physical extreme values ​​through mathematical extreme values ​​can quickly and accurately identify the culprit from hundreds or thousands of complex variables, realizing the logical transformation from vague quality alarms to precise parameter positioning.

[0086] After identifying the key feature dimensions, the server immediately initiates a parameter attribute verification process. This step ensures that the subsequently generated control commands are physically executable. The server reads a pre-configured list of adjustable process parameters, which defines in detail which parameters on the production line can be intervened in real time by the equipment. For example, extrusion speed and heating power are adjustable parameters, while ambient humidity and raw material impurity content are unadjustable objective conditions. The server performs a Boolean matching operation between the selected key feature dimensions and this list to determine whether they fall within the controllable range. The principle behind this judgment logic is to prevent the control loop from entering a dead end. If the cause of the anomaly is found to be an uncontrollable environmental factor, the server will not generate invalid adjustment commands to avoid causing equipment logic confusion. Only when it is confirmed that the parameter can indeed be physically changed by the actuator will the process proceed, ensuring the logical rigor and engineering feasibility of the closed-loop control.

[0087] Once the key feature dimension is confirmed to be an adjustable process parameter, the server immediately invokes the process compensation strategy module stored in the knowledge base. This module presets corresponding adjustment algorithms according to different parameter types; for example, it sets a strategy to gradually reduce heating power for cases of excessively high temperatures. Based on the deviation between the current actual parameter value and the ideal standard value, the server calculates the physical adjustment value required to eliminate the deviation, combining PID control algorithms or fuzzy control logic; for example, it calculates that the heater power needs to be reduced by 5%. Then, the server constructs a control instruction containing a unique identification code for the target device and a specific adjustment value, according to the format specifications of the industrial communication protocol. This instruction clearly specifies which PLC controller will perform the operation and the specific magnitude of the operation, forming a complete closed-loop correction instruction package. After the instruction is sent to the underlying actuator, it will directly drive the physical equipment to move, thereby bringing the production process back to the normal process window.

[0088] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A big data-based management method for automotive aluminum trim parts, characterized in that, Includes the following steps: S1: Data on the purity of raw materials for aluminum decorative parts, extrusion process parameters, and ambient temperature and humidity are collected by sensors deployed at each stage of production, transmitted to the cloud server, and stored in a structured manner based on the unique product serial number; S2: The dynamic time warping algorithm is used to perform time-series alignment on the structured associated storage data, remove noise interference and perform normalization processing to construct a full life cycle feature matrix of aluminum ornaments that includes raw material, process and environmental characteristics. S3: Divide the full life cycle feature matrix of aluminum trim into training set and validation set, input it into long short-term memory network for iterative training, and establish a dynamic prediction model for aluminum trim quality with production process features as input and finished product quality indicators as output. S4: Real-time collection of current production process data is input into the dynamic prediction model for aluminum trim quality, and the quality score of the current aluminum trim is calculated. If the score is lower than the standard threshold, it is automatically judged as an abnormal product and the specific defect type is marked. S5: For aluminum trim pieces identified as defective, a layered class correlation propagation algorithm is used to analyze the activation state of neurons within the model, locate the key feature dimension causing the quality anomaly, and if the feature dimension belongs to an adjustable process parameter, a closed-loop correction instruction for that parameter is generated; the specific process is as follows: The feature matrix of aluminum ornaments that are determined to be abnormal is input into the dynamic prediction model of aluminum ornament quality. The neuron nodes in the output layer that indicate the abnormality category are locked and their activation values ​​are set to the initial total correlation value. The initial total correlation value is back-allocated according to the network weights using the layer-level correlation propagation rule. This process is then reversed step by step through the fully connected layer and hidden layer to calculate the correlation score for each feature dimension in the input layer. Compare the relevance scores of all feature dimensions in the input layer and select the feature dimension with the highest value as the key feature dimension; Determine whether the key feature dimension belongs to the preset list of adjustable process parameters; If so, retrieve the preset process compensation strategy that matches the key feature dimension, calculate the physical adjustment value required for the parameter, and generate a closed-loop correction instruction containing the target device ID and the adjustment value. The specific method for calculating the relevance score of each feature dimension in the input layer is as follows: Calculate the weighted product of the activation value of a neuron in the fully connected layer with its connection weight, and define the proportion of this product to the sum of all weighted products in the layer as the contribution coefficient. Based on the contribution coefficient, allocate the initial total relevance value of the output layer to each neuron in the fully connected layer. The correlation values ​​obtained from the fully connected layer neurons are used as backpropagation input. Based on the temporal unfolding structure of the long short-term memory network, the hidden layer network structure is traversed. The excitation intensity of the previous layer neurons to the next layer neurons is calculated according to the inter-layer connection weights and neuron activation states. The correlation values ​​are then passed forward layer by layer according to the excitation intensity ratio. When the correlation value is backpropagated to the input layer, the correlation components transmitted from all nodes in the first hidden layer to the corresponding nodes in the input layer are collected, and a summation operation is performed to obtain the cumulative correlation value of that node, which is then established as the correlation score for each feature dimension of the input layer.

2. The method for managing automotive aluminum trim parts based on big data according to claim 1, characterized in that, In step S1, the process of transmitting the data to the cloud server and storing it in a structured association based on the unique product serial number is as follows: The cloud server receives the encapsulated data packet uploaded by the edge gateway. The data packet contains the purity of the raw materials, extrusion process parameters, and environmental temperature and humidity values ​​that are bound to the unique product serial number. The server parses the data packet to extract the unique product serial number as the primary key of the database, and constructs a mapping index table containing fields for raw material purity, extrusion process parameters, and environmental temperature and humidity. The parsed data values ​​are written into the storage field of the corresponding unique product serial number in the mapping index table, forming a structured associated data entry indexed by the unique product serial number.

3. The method for managing automotive aluminum trim parts based on big data according to claim 1, characterized in that, In step S2, the process of constructing the full life-cycle feature matrix of aluminum decorative parts, which includes raw material, process, and environmental characteristics, is as follows: Read the data sequence from the structured associative storage, select the standard process timing as the benchmark, and apply the dynamic time warping algorithm to map the data sequence to a unified time axis; Perform Kalman filtering on the data sequence on a unified time axis to remove high-frequency random noise and generate a smooth data sequence that retains trend characteristics; The extreme values ​​are calculated by traversing the smoothed data sequence, and the range transformation formula is used to map the values ​​to a closed interval between zero and one, thus completing the data normalization process. By splicing the normalized raw material purity, extrusion process parameters, and environmental temperature and humidity data along the time dimension, a full life cycle characteristic matrix of aluminum decorative parts is constructed.

4. The method for managing automotive aluminum trim parts based on big data according to claim 3, characterized in that, The specific method for performing Kalman filtering on the data sequence on a unified time axis to remove high-frequency random noise and generate a smooth data sequence that retains trend characteristics is as follows: A linear state-space model is constructed for the data sequence on a unified time axis. The system state transition matrix and observation matrix are set, and the prior state estimate and prior error covariance matrix are calculated using the posterior state estimate of the previous time step. The Kalman gain is calculated by combining the prior error covariance matrix and the observation noise covariance matrix. The Kalman gain is then used to perform a weighted correction on the residual between the current observation value and the prior state estimate in the data sequence, and the posterior state estimate at the current time is calculated. The updated posterior error covariance matrix is ​​used for the next iteration calculation. The posterior state estimates calculated at each time step on a unified time axis are extracted to form a smooth data sequence that retains trend characteristics.

5. The method for managing automotive aluminum trim parts based on big data according to claim 1, characterized in that, In step S3, the process of establishing a dynamic prediction model for the quality of aluminum decorative parts, with production process characteristics as input and finished product quality indicators as output, is as follows: Based on the time series sequence, the feature matrix of the entire life cycle of aluminum ornaments is divided into non-overlapping training set matrix and validation set matrix; Construct a long short-term memory network structure that includes an input layer, hidden layers, and a fully connected output layer, and initialize the network weight parameters and bias terms; The training set matrix is ​​input into the Long Short-Term Memory network to perform forward propagation, and the deviation between the prediction results and the finished product quality indicators is calculated using the loss function. The backpropagation algorithm is used to update the network weights based on the deviation gradient until the validation set error converges, outputting a dynamic prediction model for the quality of aluminum trim parts.

6. The method for managing automotive aluminum trim parts based on big data according to claim 1, characterized in that, In step S4, the process of real-time acquisition of current production process data, inputting it into the dynamic prediction model for aluminum trim quality, calculating the current quality score of the aluminum trim, identifying abnormal products, and marking specific defect types is as follows: The sensor values ​​in the current production process are read in real time, and the preprocessing rules in the training stage are reused to perform time alignment and normalization to generate the current feature matrix to be tested. The current feature matrix to be tested is input into the dynamic prediction model of aluminum trim quality. The output layer of the model calculates the current quality score and defect category probability distribution of the aluminum trim. The quality score is compared with a preset standard threshold. If the quality score is less than the preset standard threshold, an abnormal product judgment result is generated and the defect identification logic is activated. Based on the maximum probability value index in the defect category probability distribution, the pre-stored defect type database is matched to associate the abnormal product with and label it with a specific defect type name.

Images