Injection mold pressure monitoring process and system

Abstract

This invention discloses a method and system for monitoring and processing injection mold pressure, comprising: firstly, acquiring a pressure time-series sample set, wherein each sample contains a mold pressure time-series signal and corresponding sample pressure state diagnostic data; training a pressure state analysis model based on the sample set, the model including a pre-trained signal representation extraction network, a working condition decoding network, and an untrained diagnostic feature generation network, wherein during training, pressure distribution features are extracted, converted into diagnostic semantic features, decoded to generate inferred diagnostic data, and the parameters of the diagnostic feature generation network are optimized based on the error between the inference and the sample data; acquiring the current mold pressure time-series signal in real time, inputting it into the trained model, and outputting the current pressure state diagnostic data. This invention improves the accuracy and real-time performance of pressure state diagnosis by accurately extracting the local and coupling features of the pressure signal and dynamically optimizing the parameters of the diagnostic feature generation network, effectively supporting injection molding process optimization and mold maintenance early warning.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a method and system for monitoring and processing pressure in injection molds. Background Technology

[0002] The pressure state of injection molds is a key factor affecting the quality of injection molded products, and its monitoring and diagnosis are of great significance for ensuring production stability and reducing defect rates. Current technologies for injection mold pressure monitoring largely rely on manual experience or simple threshold judgments, which have the following shortcomings: First, pressure timing signals contain dynamic characteristics across multiple stages, such as filling, holding pressure, and cooling. Traditional methods struggle to accurately extract the local distribution characteristics and inter-stage coupling relationships of the signals, resulting in incomplete feature representation. Second, there is insufficient correlation analysis of multi-dimensional pressure condition indicators, leading to low matching between diagnostic semantic features and actual pressure states. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for monitoring and processing pressure in injection molds.

[0004] In a first aspect, embodiments of the present invention provide a method for monitoring and processing pressure in injection molds, comprising:

[0005] Obtain a pressure time series sample set; wherein each pressure time series sample includes a corresponding mold pressure time series signal, and sample pressure state diagnostic data containing multi-dimensional sample pressure condition indicators of the mold pressure time series signal;

[0006] Based on the aforementioned pressure time-series sample set, multiple batches of parameter tuning are performed on the untrained pressure state analysis model; wherein, the pressure state analysis model includes: a pre-trained signal representation extraction network and a working condition decoding network, as well as an untrained diagnostic feature generation network; each training cycle includes:

[0007] The signal characterization extraction network is used to extract the pressure distribution characteristics represented by the mold pressure time-series signal;

[0008] The obtained pressure distribution features are converted into diagnostic semantic features using the diagnostic feature generation network; wherein, the diagnostic semantic features are used to characterize the multidimensional inferred pressure condition index of the mold pressure time series signal.

[0009] Using the aforementioned working condition decoding network, the diagnostic semantic features are decoded to generate inferred stress state diagnostic data;

[0010] Based on the error between the obtained inferred pressure state diagnostic data and the corresponding sample pressure state diagnostic data, the network parameters of the diagnostic feature generation network are optimized.

[0011] The system acquires the current mold pressure timing signal during the injection process and outputs the corresponding current pressure state diagnostic data based on the trained pressure state analysis model.

[0012] In one possible implementation, acquiring the pressure time series sample set includes:

[0013] Collect pressure monitoring datasets for multiple injection molds corresponding to the injection process, and collect the process parameter records corresponding to each pressure monitoring dataset;

[0014] Analyze the signal energy changes or frequency component changes in the pressure monitoring dataset to determine multiple signal mode transition points;

[0015] Based on the multiple signal mode transition points, the segmentation points are determined, and the pressure monitoring dataset is divided into multiple pressure signal segments with inherent consistency; wherein, the inherent consistency indicates that the signal modes within the same pressure signal segment remain consistent.

[0016] Multiple pressure signal segments corresponding to the corresponding pressure monitoring dataset are obtained, and each obtained pressure signal segment is used as the corresponding mold pressure time sequence signal.

[0017] Using a pre-trained pressure feature extraction model, pressure condition indicators are identified for each mold pressure time series signal, resulting in multi-dimensional sample pressure condition indicators of the corresponding mold pressure time series signal.

[0018] Using the pre-trained diagnostic generation model, based on each obtained process parameter record and each sample pressure condition index, sample pressure state diagnostic data corresponding to each mold pressure time sequence signal is generated.

[0019] Based on the pressure timing signal of each mold and its corresponding sample pressure state diagnostic data, a pressure timing sample set is obtained.

[0020] In one possible implementation, prior to performing multiple batch parameter tuning on the untrained pressure state analysis model based on the pressure time-series sample set, the method further includes:

[0021] Based on the model architecture configuration scheme of the pressure state analysis model, an untrained signal representation extraction network and a diagnostic feature generation network are constructed, as well as a pre-trained working condition decoding network is obtained.

[0022] In each training cycle, local signal suppression is performed on the pressure monitoring time series data to obtain locally distorted data of locally shielded random signal segments;

[0023] Feature extraction is performed on the locally distorted data to obtain signal characterization results;

[0024] Based on the signal characterization results, the locally masked random signal segments in the locally distorted data are inferred to obtain the inferred signal segments;

[0025] Based on the error between the obtained inferred signal segment and the random signal segment, the network parameters of the signal representation extraction network are optimized, and the pre-trained signal representation extraction network is output.

[0026] Based on the pre-trained signal representation extraction network and working condition decoding network, as well as the untrained diagnostic feature generation network, an untrained stress state analysis model is constructed.

[0027] In one possible implementation, the step of using the signal characterization extraction network to extract the pressure distribution features characterized by the mold pressure time-series signal includes:

[0028] Using the signal characterization extraction network, the local pressure distribution features corresponding to multiple time segments in the mold pressure time series signal are extracted, and the pressure distribution coupling features between each local pressure distribution feature are extracted.

[0029] By integrating the obtained local pressure distribution features and pressure distribution coupling features, the pressure distribution features representing the mold pressure time series signal are obtained; the local pressure distribution features are used to indicate at least one of the following: single-segment frequency features, energy distribution pattern, acquisition conditions, time regularity, physical source category, and comprehensive high-level characterization of the corresponding time series segment; the pressure distribution coupling features are used to indicate at least one of the following: consistency, time series relationship, and frequency change between each time series segment.

[0030] In one possible implementation, the step of using the diagnostic feature generation network to convert the obtained stress distribution features into diagnostic semantic features includes:

[0031] Based on the network parameters of the diagnostic feature generation network in the current training cycle, a first association rule is obtained between each single-dimensional stress distribution feature and each single-dimensional diagnostic semantic feature learned by the diagnostic feature generation network; wherein, each single-dimensional diagnostic semantic feature represents one of the multi-dimensional benchmark stress condition indicators; the multi-dimensional benchmark stress condition indicators include at least the complete set of the multi-dimensional sample stress condition indicators and the multi-dimensional inferred stress condition indicators.

[0032] According to the first association rule, the pressure distribution features are converted into multiple single-dimensional diagnostic semantic features; wherein, the multiple single-dimensional diagnostic semantic features characterize the multi-dimensional inferred pressure condition index.

[0033] Multiple single-dimensional diagnostic semantic features are integrated to generate diagnostic semantic features.

[0034] In one possible implementation, before converting the obtained pressure distribution features into diagnostic semantic features using the diagnostic feature generation network, the method further includes:

[0035] Obtain the analysis and processing load during the current training cycle;

[0036] When it is determined that the analysis and processing load exceeds the load limit, according to the second association rule between each preset feature encoding form and each feature resource consumption, a target feature encoding form whose corresponding feature resource consumption is less than the benchmark resource ratio is determined from each feature encoding form; wherein, the benchmark resource ratio is: the feature resource consumption corresponding to the feature encoding form used by the network parameters of the diagnostic feature generation network;

[0037] Real-time monitoring of the computational resource consumption required for the current training cycle;

[0038] Based on the correlation rules between the computing resource consumption, the analysis and processing load, and the resource consumption of the feature, the compression level of the target feature encoding format is dynamically selected.

[0039] According to the selected compression level, the pressure distribution feature is converted into the target feature encoding form to obtain the encoded pressure distribution feature; wherein, the higher the compression level, the less resources the generated encoded pressure distribution feature occupies.

[0040] In one possible implementation, optimizing the network parameters of the diagnostic feature generation network based on the error between the obtained inferred pressure state diagnostic data and the corresponding sample pressure state diagnostic data includes:

[0041] Based on the error between the multidimensional inferred pressure condition indicators contained in the obtained inferred pressure state diagnostic data and the multidimensional sample pressure condition indicators contained in the corresponding sample pressure state diagnostic data, the diagnostic difference is determined.

[0042] The state matching error is determined based on the difference between the obtained inferred pressure state diagnostic data and the corresponding sample pressure state diagnostic data.

[0043] Based on the diagnostic difference and the state matching error, when the diagnostic accuracy criterion is not met, the network parameters of the diagnostic feature generation network are optimized.

[0044] In one possible implementation, the diagnostic feature generation network includes a primary diagnostic feature generation subnetwork and a deep diagnostic feature fusion subnetwork.

[0045] The process of using the diagnostic feature generation network to convert the obtained stress distribution features into diagnostic semantic features includes:

[0046] The primary diagnostic features are used to generate a sub-network, which converts the obtained pressure distribution features into multiple primary diagnostic semantic feature vectors; wherein each primary diagnostic semantic feature vector represents a preliminary prediction of at least one single-dimensional inferred pressure condition index.

[0047] The deep diagnostic feature fusion subnetwork is used to fuse multiple primary diagnostic semantic feature vectors and learn the high-order correlation between the multiple primary diagnostic semantic feature vectors to generate the diagnostic semantic features represented by the fused multidimensional inferred pressure condition index.

[0048] In one possible implementation, after at least one complete training cycle has been completed, the method further includes:

[0049] Obtain an external pressure monitoring verification dataset; wherein, the external pressure monitoring verification dataset includes mold pressure time series signals and their actual pressure state diagnostic data that were not involved in the construction of the pressure time series sample set;

[0050] Using the pressure state analysis model obtained through current training, inference is made on the external pressure monitoring verification dataset to obtain the corresponding inferred verification pressure state diagnostic data.

[0051] Calculate the generalization error between the inferred and verified pressure state diagnostic data and the actual pressure state diagnostic data;

[0052] Based on the generalization error, the hyperparameters of subsequent training cycles are dynamically adjusted; wherein the hyperparameters include at least one of the following: learning rate, batch size, and training data sample weight distribution.

[0053] In a second aspect, embodiments of the present invention provide a server system, including a server, the server being used to execute the method described in the first aspect.

[0054] Compared to existing technologies, the beneficial effects of this invention include: The injection mold pressure monitoring and processing method and system disclosed in this invention acquires a pressure time-series sample set, where each sample contains a mold pressure time-series signal and corresponding sample pressure state diagnostic data; a pressure state analysis model is trained based on the sample set, the model including a pre-trained signal representation extraction network, a working condition decoding network, and an untrained diagnostic feature generation network. During training, pressure distribution features are extracted, converted into diagnostic semantic features, decoded to generate inferred diagnostic data, and the parameters of the diagnostic feature generation network are optimized based on the error between the inference and the sample data; the current mold pressure time-series signal is acquired in real time, input into the trained model, and the current pressure state diagnostic data is output. This invention improves the accuracy and real-time performance of pressure state diagnosis by accurately extracting the local and coupling features of the pressure signal and dynamically optimizing the parameters of the diagnostic feature generation network, effectively supporting injection molding process optimization and mold maintenance early warning. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating the steps of the injection mold pressure monitoring and processing method provided in an embodiment of the present invention.

[0057] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0059] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0060] In order to solve the technical problems mentioned in the background art Figure 1 This is a schematic flowchart of the injection mold pressure monitoring and processing method provided in the embodiments of this disclosure. The injection mold pressure monitoring and processing method will be described in detail below.

[0061] Step S201: Obtain a pressure time series sample set; wherein each pressure time series sample includes a corresponding mold pressure time series signal and sample pressure state diagnostic data containing multi-dimensional sample pressure condition indicators of the mold pressure time series signal.

[0062] Step S202: Based on the pressure time-series sample set, perform multiple batch parameter tuning on the untrained pressure state analysis model; wherein, the pressure state analysis model includes: a pre-trained signal representation extraction network and a working condition decoding network, and an untrained diagnostic feature generation network; each training cycle includes:

[0063] Step S203: Using the signal characterization extraction network, extract the pressure distribution characteristics represented by the mold pressure time-series signal;

[0064] Step S204: Using the diagnostic feature generation network, the obtained pressure distribution features are converted into diagnostic semantic features; wherein, the diagnostic semantic features are used to characterize the multidimensional inferred pressure condition index of the mold pressure time series signal.

[0065] Step S205: Using the operating condition decoding network, decode the diagnostic semantic features to generate inferred pressure state diagnostic data;

[0066] Step S206: Optimize the network parameters of the diagnostic feature generation network based on the error between the obtained inferred pressure state diagnostic data and the corresponding sample pressure state diagnostic data;

[0067] Step S207: Obtain the current mold pressure timing signal of the current injection mold during the injection process, and output the corresponding current pressure state diagnostic data based on the trained pressure state analysis model.

[0068] In this embodiment of the invention, for example, a server system in a precision injection molding workshop is used as the execution entity. This workshop is equipped with 10 injection molding machines (corresponding to molds M1 to M10), primarily producing casings for consumer electronics products. The server needs to perform status diagnosis of the injection molding process through pressure monitoring data. The specific execution flow is as follows:

[0069] The server first constructs a pressure timing sample set. Pressure monitoring data from molds M1 to M10 during the injection molding cycle is collected using pressure sensors (sampling frequency 1kHz) deployed on the injection molding machine. (Each data point for mold M1 is a 10-second time-domain signal, corresponding to the complete cycle of filling → holding pressure → cooling → mold opening.) Simultaneously, process parameters (e.g., injection temperature 230℃, injection speed 80mm / s, mold temperature 60℃) are recorded from the injection molding machine control system. Signal pattern analysis is performed on the pressure monitoring data: Short-Time Fourier Transform (STFT) is used to identify changes in signal energy and frequency components. For example, during the filling stage (0-3 seconds), the signal energy is concentrated at 50-100Hz; during the holding pressure stage (3-6 seconds), it drops to 10-30Hz; and during the cooling stage (6-10 seconds), it further drops to 1-5Hz. Based on the energy transition moments, the transition points between "filling-holding pressure" and "holding pressure-cooling" are determined, and the 10-second signal is divided into three inherently consistent segments (with stable signal frequency and energy distribution within the same segment). Subsequently, a pre-trained pressure feature extraction model is invoked to extract multi-dimensional sample pressure condition indicators from each segment. For example, during the filling stage, the indicators include "filling time 3.0 seconds" and "peak pressure 85 MPa"; during the holding stage, the indicators include "average holding pressure 70 MPa" and "fluctuation range ±2 MPa"; and during the cooling stage, the indicators include "attenuation rate 5 MPa / s" and "residual pressure 15 MPa". Combined with process parameter records, a pre-trained diagnostic generation model is used to generate sample pressure state diagnostic data (such as "uniform pressure distribution during the filling stage, stable holding pressure, and no overall abnormalities"). This results in a pressure time-series sample set containing 100,000 samples, with each sample consisting of "mold pressure time-series signal" and "sample pressure state diagnostic data".

[0070] Based on the sample set, the server trains the pressure state analysis model. This model includes a pre-trained signal representation extraction network (bidirectional LSTM structure), a pre-trained condition decoding network (BERT model), and an untrained diagnostic feature generation network (fully connected layer + attention mechanism). Before training, the signal representation extraction network is pre-trained: local data in signal segments are randomly masked (e.g., signals during the 1.5-2.0 second filling phase), and the network infers the masked segments and calculates the error (e.g., original signal pressure 60-75 MPa, inferred value 62-73 MPa, MSE = 1.8 MPa²). After 5000 rounds of optimization, the inference error is reduced to MSE < 0.5 MPa². The training phase employs multi-batch parameter tuning (1000 samples per batch, 100 batches in total). The workflow for a single training cycle is as follows: First, the signal characterization extraction network extracts the pressure distribution features of the mold pressure time-series signal, including local pressure distribution features (e.g., during the filling stage, "92% of the frequency is in the 50-100Hz range" and "energy is concentrated in the 200-300Hz band") and pressure distribution coupling features (e.g., "the ratio of filling peak value to average holding pressure is 1.21" and "the product of average holding pressure and cooling attenuation rate is 350MPa² / s"). After integration, a 512-dimensional feature vector is output. Next, the diagnostic feature generation network converts the pressure distribution features into diagnostic semantic features: the primary sub-network generates multiple primary vectors (e.g., "filling time ≈ 3.0 seconds" and "average holding pressure ≈ 70MPa"), and the deep sub-network fuses higher-order correlations (e.g., "filling time is positively correlated with holding pressure") through an attention mechanism, ultimately outputting 256-dimensional diagnostic semantic features (representing multi-dimensional inferred pressure condition indicators). If the analysis and processing load is too high during training (CPU utilization 85% > 70%), the server dynamically switches the feature encoding format (e.g., from one-hot encoding (5MB / sample) to binary encoding (2MB / sample)) and selects compression level 2 (compression rate 40%) based on computational resource consumption (12GB memory usage), compressing the feature vector to 307 dimensions to reduce resource consumption. Subsequently, the condition decoding network decodes the diagnostic semantic features into natural language inference diagnostic data (e.g., "peak filling pressure 85MPa, holding pressure fluctuation ±2MPa, overall no abnormalities"). By comparing the error between the inferred data and the sample data (e.g., diagnostic difference = 0, state matching error = 0.02 < 0.1), the gradient descent method is used to optimize the parameters of the diagnostic feature generation network. During the training phase, an external validation dataset (5000 untrained data points from mold M11) is introduced. If the generalization error is high (e.g., the inferred holding pressure is 65 MPa, the actual value is 70 MPa, and the difference is 5 MPa), the hyperparameters are dynamically adjusted (the learning rate is reduced from 0.001 to 0.0005, and the batch size is increased from 1000 to 1500). Ultimately, the generalization error is reduced to a diagnostic difference of <1 MPa and a state matching error of <0.03.

[0071] After model deployment, the server monitors the current injection mold in real time. It receives the real-time pressure signal (1kHz sampling, 10-second cycle) from mold M1 via industrial Ethernet and inputs it into the trained model: the signal characterization extraction network extracts the current pressure distribution features (e.g., 90% frequency share during filling, 345MPa² / s product of holding pressure-cooling coupling features); the diagnostic feature generation network converts this into diagnostic semantic features ("filling time 3.0 seconds, average holding pressure 69MPa, fluctuation ±2.5MPa, attenuation rate 5.2MPa / s"); and the working condition decoding network outputs the current pressure state diagnostic data (e.g., "peak filling pressure 83MPa slightly low, holding pressure stable, fine-tuning injection speed recommended"). This enables real-time status monitoring and process optimization guidance during the injection molding process.

[0072] In this embodiment of the invention, the acquisition of the pressure time series sample set can be performed through the following example.

[0073] Collect pressure monitoring datasets for multiple injection molds corresponding to the injection process, and collect the process parameter records corresponding to each pressure monitoring dataset;

[0074] According to the preset sampling period, each pressure monitoring dataset is time-series segmented to obtain multiple pressure signal segments corresponding to the corresponding pressure monitoring dataset, and each obtained pressure signal segment is used as the corresponding mold pressure time-series signal.

[0075] Using a pre-trained pressure feature extraction model, pressure condition indicators are identified for each mold pressure time series signal, resulting in multi-dimensional sample pressure condition indicators of the corresponding mold pressure time series signal.

[0076] Using the pre-trained diagnostic generation model, based on each obtained process parameter record and each sample pressure condition index, sample pressure state diagnostic data corresponding to each mold pressure time sequence signal is generated.

[0077] Based on the pressure timing signal of each mold and its corresponding sample pressure state diagnostic data, a pressure timing sample set is obtained.

[0078] In this embodiment of the invention, for example, the server first performs a data collection operation. For molds M1 to M10 in the workshop (used for injection molding consumer electronic components such as mobile phone casings and laptop casings, respectively), pressure monitoring data during the injection molding process is continuously collected by pressure sensors (sampling frequency 1kHz) deployed at the cavities and gates of each injection molding machine. For example, when mold M1 is producing a certain model of mobile phone casing, a single pressure monitoring dataset is a time-series signal lasting 24 hours, covering 500 complete injection cycles (each cycle is 10 seconds, including filling, holding pressure, and cooling stages). At the same time, the server synchronously retrieves the process parameter records corresponding to each pressure monitoring dataset from the injection molding machine control system via industrial Ethernet, including key parameters such as injection temperature (220-240℃), injection speed (60-100mm / s), mold temperature (50-70℃), and holding pressure time (2-4 seconds), ensuring that the pressure data corresponds one-to-one with the process conditions.

[0079] Next, the server performs time-series segmentation on the pressure monitoring dataset according to the preset sampling period (10 seconds / cycle, corresponding to a complete injection molding process): for example, for 24 hours of data for mold M1, a continuous signal is extracted every 10 seconds to obtain 500 pressure signal segments. Each segment contains the complete pressure change process within the injection molding cycle (such as the pressure time-domain waveform of the 0-3 second filling stage, the 3-6 second holding stage, and the 6-10 second cooling stage), and each segment is marked as "mold pressure time-series signal".

[0080] Subsequently, the server calls the pre-trained pressure feature extraction model (trained based on 500,000 historical injection pressure data points, using a CNN-LSTM hybrid architecture) to identify multi-dimensional sample pressure condition indicators for each mold pressure time-series signal. Taking a pressure signal segment (10-second loop) of mold M1 as an example, the model extracts local pressure features (such as the pressure rise slope in the filling stage from 0 to 3 seconds and the pressure fluctuation variance in the holding stage from 3 to 6 seconds) through convolutional layers. After capturing the temporal correlation through LSTM layers, it outputs the multi-dimensional sample pressure condition indicators for this segment, specifically including "filling time 2.9 seconds", "peak filling pressure 86 MPa", "average holding pressure 71 MPa", "pressure fluctuation amplitude in the holding stage ±1.8 MPa", "pressure decay rate in the cooling stage 4.9 MPa / s", and "final residual pressure 14 MPa". These indicators cover the core pressure characteristics of each stage of the injection molding process.

[0081] Based on the extracted multidimensional sample pressure condition indicators, the server further calls a pre-trained diagnostic generation model (based on the BERT-base architecture, with training data including 50,000 pressure condition diagnostic texts annotated by process experts), and combines the corresponding process parameters to generate sample pressure condition diagnostic data. For example, the process parameters corresponding to the pressure signal segment of mold M1 are "injection temperature 230℃ (standard 225-235℃), injection speed 80mm / s (standard 75-85mm / s), mold temperature 60℃ (standard 55-65℃)". The model generates sample pressure condition diagnostic data by combining the pressure condition indicators (fill time 2...). The process parameters were matched (the filling time was 2.9 seconds, within the standard range of 2.7-3.1 seconds, and the average holding pressure was 71 MPa, within the standard range of 68-73 MPa). The sample pressure status diagnostic data was generated as follows: "The pressure distribution during the filling stage was uniform (filling time 2.9 seconds, pressure peak 86 MPa meets the process requirements), the pressure stability during the holding stage was good (average 71 MPa, fluctuation range 1.8 MPa ≤ ±3 MPa), the pressure decay during the cooling stage was normal (decay rate 4.9 MPa / s, residual pressure 14 MPa), the current process parameters matched the pressure status 98%, and there were no abnormalities in the injection molding process."

[0082] Finally, the server associates and stores each mold pressure time-series signal (such as the 10-second pressure waveform data of mold M1) with its corresponding sample pressure state diagnostic data (including the above-mentioned multi-dimensional sample pressure condition indicators and diagnostic text) to form a pressure time-series sample set: by processing the historical data of molds M1 to M10, a dataset containing 100,000 samples is constructed. Each sample covers the complete mapping relationship of "mold pressure time-series signal - sample pressure state diagnostic data", providing basic data support for subsequent model training.

[0083] In this embodiment of the invention, before performing multiple batch parameter tuning on the untrained pressure state analysis model based on the pressure time series sample set, the following implementation method is also provided.

[0084] Based on the model architecture configuration scheme of the pressure state analysis model, an untrained signal representation extraction network and a diagnostic feature generation network are constructed, as well as a pre-trained working condition decoding network is obtained.

[0085] Based on the collected time-series data of multiple pressure monitoring, the parameters of the untrained signal representation extraction network are fined in multiple batches, and the pre-trained signal representation extraction network is output.

[0086] Based on the pre-trained signal representation extraction network and working condition decoding network, as well as the untrained diagnostic feature generation network, an untrained stress state analysis model is constructed.

[0087] In this embodiment of the invention, for example, the server first constructs an untrained signal representation extraction network and a diagnostic feature generation network according to the architecture configuration scheme of the pressure state analysis model, and obtains a pre-trained working condition decoding network. The signal representation extraction network adopts a "bidirectional LSTM + attention mechanism" architecture: the input layer receives a 10-second pressure time-series signal (10,000 sampling points), extracts temporal correlation features through a 3-layer bidirectional LSTM (256 hidden units per layer, dropout rate 0.2), and then focuses on key pressure segments (such as peak filling moments and pressure holding fluctuation segments) through a multi-head attention layer (8 heads, feature dimension 512). The diagnostic feature generation network adopts a "3-layer fully connected + residual connection" structure, with the input being the 512-dimensional pressure distribution features output by the signal representation extraction network, which are then mapped to 256-dimensional diagnostic semantic features after ReLU activation. Meanwhile, the server retrieves the pre-trained working condition decoding network from the company's model library. This network is based on the BERT-base architecture (12-layer Transformer, 768-dimensional hidden states) and has been trained and converged using 50,000 expert-annotated stress diagnosis texts (including descriptions of normal / abnormal states). It can decode diagnostic semantic features into natural language diagnostic data.

[0088] Next, the server uses the collected pressure monitoring time-series data (historical pressure monitoring dataset of molds M1-M10, totaling 2 million 10-second pressure time-series signals) to pre-train the untrained signal representation extraction network. The specific process is as follows: the server performs random masking on the time-series data in batches (256 signals per batch)—randomly selecting 15% of the time-series segments of each signal (e.g., 0.5 seconds for the filling phase and 0.3 seconds for the holding phase) and replacing them with zero-value masks to generate masked pressure time-series signals; the masked signals are input into the untrained signal representation extraction network, and the network outputs the predicted pressure values ​​of the masked segments (inferred from the features of the unmasked portions); the server calculates the mean squared error (MSE) between the predicted values ​​and the original masked segments, and updates the network parameters through backpropagation using the Adam optimizer (learning rate 0.001, weight decay 1e-5). After 10,000 training rounds, when the validation set MSE drops to 0.8 MPa² (initial MSE 12.5 MPa²) and does not decrease for 500 consecutive rounds, the server stops pre-training and outputs the pre-trained signal representation extraction network (at this point, the network can accurately capture the temporal dependence and local features of the stress signal).

[0089] Finally, the server connects the pre-trained signal representation extraction network, the pre-trained working condition decoding network, and the untrained diagnostic feature generation network in the order of "signal input → signal representation extraction → diagnostic feature generation → working condition decoding" to construct an untrained stress state analysis model. The output of the signal representation extraction network is connected to the input of the diagnostic feature generation network through a 512-dimensional feature vector, and the output of the diagnostic feature generation network (256-dimensional diagnostic semantic features) is connected to the input of the working condition decoding network, forming a complete end-to-end model architecture, which is ready for subsequent batches of parameter tuning.

[0090] In this embodiment of the invention, the step of performing multiple batches of parameter tuning on the untrained signal characterization extraction network based on the collected multiple pressure monitoring time series data can be implemented through the following example.

[0091] In each training cycle, local signal suppression is performed on the pressure monitoring time series data to obtain locally distorted data of locally shielded random signal segments;

[0092] Feature extraction is performed on the locally distorted data to obtain signal characterization results;

[0093] Based on the signal characterization results, the locally masked random signal segments in the locally distorted data are inferred to obtain the inferred signal segments;

[0094] Based on the error between the obtained inferred signal segment and the random signal segment, the network parameters of the signal characterization extraction network are optimized.

[0095] In an embodiment of the invention, for example, the server, for constructing the pressure state analysis model, first constructs an untrained signal representation extraction network (bidirectional LSTM + attention mechanism, inputting a 10-second pressure signal (10,000 sampling points) and an output 512-dimensional feature network) and a diagnostic feature generation network (3-layer fully connected + residual connection) based on a preset architecture configuration scheme, and calls a pre-trained working condition decoding network (BERT-based architecture, supporting the conversion of diagnostic semantic features into natural language). Subsequently, the server uses 2 million historical pressure monitoring time-series data points (each a 10-second injection cycle signal) from molds M1-M10 to pre-train the signal representation extraction network.

[0096] In the specific training, the server processes 256 signals per batch: for each 10-second signal (10,000 sampling points), a 15% random mask is performed—randomly selecting a 0.5-second segment (500 sampling points) from the filling phase (0-3 seconds) and a 0.3-second segment (300 sampling points) from the holding phase (3-6 seconds), replacing them with zero values ​​to generate locally distorted data (e.g., the original filling phase pressure of a certain signal in mold M1 is 50-86 MPa, after masking the corresponding segment becomes 0 MPa). The locally distorted data is input into the untrained signal representation extraction network. The network captures the temporal correlation of the unmasked segments through a bidirectional LSTM layer (e.g., the pressure trend 0.2 seconds before filling, the average pressure before and after holding), focuses on the key features of the unmasked segments through an attention layer (e.g., the pressure slope 0.1 seconds before the filling peak), and outputs the predicted pressure value of the masked segment (e.g., the predicted pressure of the masked segment in the filling phase is 52-84 MPa, and in the holding phase it is 68-72 MPa). The server calculates the mean squared error (MSE) between the predicted value and the original mask segment. For example, for mold M1, the original mask segment has a mean pressure of 72 MPa, and the predicted mean is 71 MPa, resulting in an MSE of 1.2 MPa². The network updates the LSTM layer weights and attention layer parameters through backpropagation using the Adam optimizer (learning rate 0.001, weight decay 1e-5). After 10,000 training rounds (processing 7812 batches of data per round), when the validation set MSE decreases from the initial 12.5 MPa² to 0.8 MPa² and converges for 500 consecutive rounds, the server stops pre-training, obtaining the pre-trained signal representation extraction network. Finally, the server concatenates this pre-trained network, the pre-trained working condition decoding network, and the untrained diagnostic feature generation network to construct the untrained pressure state analysis model, laying the foundation for subsequent optimization of the diagnostic feature generation network.

[0097] In this embodiment of the invention, the step of using the signal characterization extraction network to extract the pressure distribution features represented by the mold pressure time-series signal can be implemented through the following example.

[0098] Using the signal characterization extraction network, the local pressure distribution features corresponding to multiple time segments in the mold pressure time series signal are extracted, and the pressure distribution coupling features between each local pressure distribution feature are extracted.

[0099] By integrating the obtained local pressure distribution features and pressure distribution coupling features, the pressure distribution features characterized by the mold pressure timing signal are obtained.

[0100] In this embodiment of the invention, for example, the server invokes a pre-trained signal representation extraction network (bidirectional LSTM + attention mechanism architecture) to extract pressure distribution features from the mold pressure time-series signal. Taking a 10-second pressure time-series signal of mold M1 (divided into three time segments: filling stage 0-3 seconds, holding stage 3-6 seconds, and cooling stage 6-10 seconds) as an example:

[0101] First, the network extracts the local pressure distribution characteristics of each time segment: For the filling stage segment (0-3 seconds), Fourier transform is used to extract frequency characteristics (92% of the signal in the 50-100Hz band, corresponding to the high-speed flow characteristics of the melt) and energy distribution pattern (65% of the energy in the 200-300Hz band, indicating flow stability). Combined with time domain analysis, the time law is obtained (pressure rises linearly from 50MPa to 86MPa, slope 28.3MPa / s); For the holding stage segment (3-6 seconds), frequency characteristics (88% of the frequency in the 10-30Hz band, corresponding to the slowdown of melt flow), energy distribution (energy in the 50-100Hz band decreases to 20%) and pressure fluctuation characteristics (variance 1.8MPa², fluctuation amplitude ±1.8MPa); For the cooling stage segment (6-10 seconds), frequency characteristics (95% of the frequency in the 1-5Hz band, corresponding to static pressure decay) and physical source characteristics (pressure decay rate 4.9MPa / s, residual pressure 14MPa) are extracted.

[0102] Subsequently, the network extracts the pressure distribution coupling characteristics between various local pressure distribution features: calculates the consistency between the filling and holding pressure stages (the ratio of the peak filling pressure of 86MPa to the average holding pressure of 71MPa is 1.21, within the range of process standard 1.1-1.3), the timing relationship (the time delay from the end of the filling stage to the start of the holding pressure stage is 0.02 seconds, ≤ standard 0.05 seconds, with no lag in the transition); calculates the frequency change characteristics between the holding and cooling stages (the 10-30Hz frequency band accounts for 88% in the holding pressure stage, and the 1-5Hz frequency band accounts for 95% in the cooling stage, with a smooth frequency band transition without abrupt changes).

[0103] Finally, the server integrates all local pressure distribution features (frequency, energy, and temporal patterns of the filling / holding / cooling stages) and pressure distribution coupling features (segment consistency, temporal relationships, frequency changes, etc.), and maps the multidimensional features into a 512-dimensional vector through a fully connected layer of the network. This vector serves as the pressure distribution feature representing the pressure timing signal of the mold (e.g., bits 1-128 of the vector correspond to local filling features, bits 129-256 correspond to local holding features, bits 257-384 correspond to local cooling features, and bits 385-512 correspond to coupling features).

[0104] In this embodiment of the invention, the local pressure distribution feature is used to indicate at least one of the following: single-segment frequency characteristics, energy distribution pattern, acquisition conditions, time regularity, physical source category, and comprehensive high-level characterization of the corresponding time segment; the pressure distribution coupling feature is used to indicate at least one of the following: consistency, timing relationship, and frequency change between each time segment.

[0105] In an embodiment of the invention, for example, when the server extracts features from the 10-second pressure timing signal (filling 0-3 seconds, holding pressure 3-6 seconds, cooling 6-10 seconds) of mold M1, it first obtains the local pressure distribution features of each timing segment: the single-segment frequency feature of the filling stage segment is "92% of the 50-100Hz frequency band" (corresponding to the dynamic pressure fluctuation of high-speed melt flow), the energy distribution pattern is "65% of the energy in the 200-300Hz frequency band" (indicating the intensity of flow turbulence), the time pattern is "pressure linearly increases from 50MPa to 86MPa, with a slope of 28.3MPa / s", and the physical source category is... "Dynamic impact pressure of melt filling the cavity"; the single-segment frequency characteristics of the holding stage segment are "88% of the frequency band in the 10-30Hz range" (corresponding to the static pressure of melt feeding), the energy distribution pattern is "the energy proportion in the 50-100Hz range decreases to 20%" (turbulence weakens), and the overall high-level characterization is "pressure fluctuation variance of 1.8MPa²" (reflecting the stability of holding pressure); the single-segment frequency characteristics of the cooling stage segment are "95% of the frequency band in the 1-5Hz range" (corresponding to the static attenuation of cavity pressure), the time law is "pressure attenuation rate of 4.9MPa / s", and the physical source category is "residual pressure release from melt cooling and contraction".

[0106] Simultaneously, the server extracts the pressure distribution coupling characteristics between the three segments: the consistency characteristic is "the ratio of the filling peak of 86MPa to the average holding pressure of 71MPa is 1.21" (compliant with process standards 1.1-1.3); the timing relationship characteristic is "the transition delay from filling to holding pressure stage is 0.02 seconds" (≤ standard 0.05 seconds, no hysteresis); and the frequency change characteristic is "the 10-30Hz frequency band accounts for 88% in the holding pressure stage, and drops to 95% in the 1-5Hz frequency band in the cooling stage" (the frequency band proportion transitions smoothly without abrupt changes). These local features accurately describe the pressure characteristics of each stage, and the coupling characteristics reflect the correlation between stages, together constituting the complete pressure distribution characteristics of the mold pressure timing signal.

[0107] In this embodiment of the invention, the step of using the diagnostic feature generation network to convert the obtained pressure distribution features into diagnostic semantic features can be implemented through the following example.

[0108] Based on the network parameters of the diagnostic feature generation network in the current training cycle, a first association rule is obtained between each single-dimensional stress distribution feature and each single-dimensional diagnostic semantic feature learned by the diagnostic feature generation network; wherein, each single-dimensional diagnostic semantic feature represents one of the multi-dimensional benchmark stress condition indicators; the multi-dimensional benchmark stress condition indicators include at least the complete set of the multi-dimensional sample stress condition indicators and the multi-dimensional inferred stress condition indicators.

[0109] According to the first association rule, the pressure distribution features are converted into multiple single-dimensional diagnostic semantic features; wherein, the multiple single-dimensional diagnostic semantic features characterize the multi-dimensional inferred pressure condition index.

[0110] Multiple single-dimensional diagnostic semantic features are integrated to generate diagnostic semantic features.

[0111] In an embodiment of the invention, for example, in the current training cycle (the 50th batch of training), the server calls the diagnostic feature generation network (a 3-layer fully connected + residual connection structure, inputting 512-dimensional stress distribution features and outputting 256-dimensional diagnostic semantic features) to perform the conversion from stress distribution features to diagnostic semantic features. First, the server reads the network parameters of the current training cycle (including the weight matrices of each layer, bias terms, and residual connection coefficients), and obtains the first association rule between the single-dimensional stress distribution features and the single-dimensional diagnostic semantic features by parsing the weight matrix of the second fully connected layer (dimension 512×128) and the weight matrix of the third layer (128×256). For example, regarding the pressure distribution characteristics of mold M1 (a 512-dimensional vector, where bits 1-10 correspond to the proportion of the 50-100Hz frequency band during the filling stage, bits 11-20 correspond to the pressure rise slope during the filling stage, bits 21-30 correspond to the proportion of the 10-30Hz frequency band during the holding stage, bits 31-40 correspond to the pressure fluctuation variance during the holding stage, and bits 41-50 correspond to the pressure decay rate during the cooling stage), the server identified:

[0112] The correlation weight between the first 10 pressure distribution features (the proportion of the 50-100Hz frequency band during the filling stage) and the single-dimensional diagnostic semantic feature "filling time" is 0.85 (higher than the average weight of 0.3 for other feature dimensions), with a bias term of -0.02, forming the first association rule: when the proportion of this frequency band x (normalized 0-1) satisfies x≥0.9, the inferred value of filling time = 2.8+0.1×(x-0.9);

[0113] The correlation weight between the pressure distribution features of the 11th to 20th positions (pressure rise slope during the filling stage) and the single-dimensional diagnostic semantic feature "peak filling pressure" is 0.92, with a bias term of 0.05. The rule is: when the slope k (MPa / s) satisfies k≥25, the peak filling pressure = 75 + 0.3 × (k-25).

[0114] The 21st-30th positions (the proportion of the 10-30Hz frequency band during the pressure holding stage) are associated with the "average pressure holding pressure" by a weight of 0.88. The rule is: when the frequency band proportion y ≥ 0.85, the average pressure holding pressure = 68 + 0.2 × (y - 0.85).

[0115] The 31st-40th positions (variance of pressure fluctuation during the pressure holding stage) are associated with "pressure holding fluctuation amplitude" with a weight of 0.79. The rule is: when the variance σ² ≤ 2.0 MPa², the fluctuation amplitude = ±√σ².

[0116] The weight of the 41st-50th position (pressure decay rate during cooling stage) is 0.83, and the rule is: when the decay rate d (MPa / s) satisfies 4.5≤d≤5.5, the inferred value = d.

[0117] Based on the first association rule mentioned above, the server transforms the 512-dimensional pressure distribution characteristics of mold M1:

[0118] The feature value of the first 10th feature is 0.92 (the 50-100Hz frequency band accounts for 92% of the filling stage). According to the rules, the filling time is calculated as 2.8 + 0.1 × (0.92 - 0.9) = 2.82 seconds. We take 2.8 seconds (single-dimensional diagnostic semantic feature 1).

[0119] The slope corresponding to the 11th-20th feature values ​​is 28.3 MPa / s. The peak filling pressure is calculated as 75 + 0.3 × (28.3 - 25) = 75 + 0.99 = 75.99 MPa. We take 76 MPa (single-dimensional diagnostic semantic feature 2).

[0120] The 21st-30th feature value is 0.89 (the 10-30Hz frequency band accounts for 89% during the pressure holding stage). The average pressure holding pressure is calculated as 68 + 0.2 × (0.89 - 0.85) = 68 + 0.008 = 68.008 MPa. We take 68 MPa (single-dimensional diagnostic semantic feature 3).

[0121] The variance corresponding to the 31st-40th eigenvalues ​​is 1.8MPa², and the pressure holding fluctuation range is calculated to be ±√1.8≈±1.34MPa (single-dimensional diagnostic semantic feature 4).

[0122] The 41st to 50th feature values ​​correspond to an attenuation rate of 4.9 MPa / s, which is directly taken as 4.9 MPa / s (single-dimensional diagnostic semantic feature 5).

[0123] Finally, the server integrates the above-mentioned single-dimensional diagnostic semantic features (fill time 2.8 seconds, peak filling pressure 76 MPa, average holding pressure 68 MPa, holding pressure fluctuation ±1.34 MPa, cooling attenuation rate 4.9 MPa / s, etc., a total of 256 items) into a 256-dimensional vector according to a preset dimensional order (e.g., first position: fill time, second position: peak filling pressure... 256th position: residual pressure) to generate diagnostic semantic features. Each value of this vector corresponds to the inferred value of a single-dimensional diagnostic semantic feature, which is used by the subsequent working condition decoding network to generate natural language diagnostic data.

[0124] In this embodiment of the invention, before converting the obtained pressure distribution features into diagnostic semantic features using the diagnostic feature generation network, the following implementation method is also provided.

[0125] Based on the second association rule between each preset feature encoding form and each feature resource consumption amount, a target feature encoding form with a corresponding feature resource consumption amount less than the benchmark resource ratio is determined from each feature encoding form; wherein, the benchmark resource ratio is: the feature resource consumption amount corresponding to the feature encoding form used by the network parameters of the diagnostic feature generation network;

[0126] The pressure distribution features are converted into the target feature encoding form to obtain the encoded pressure distribution features.

[0127] In this embodiment of the invention, for example, before the server calls the diagnostic feature generation network to process the pressure distribution features of mold M1 (a 512-dimensional vector covering local and coupled features of the filling, holding, and cooling stages), it needs to perform feature encoding format conversion to reduce resource consumption. First, the server reads a preset second association rule table, which records the mapping relationship between feature encoding formats and feature resource consumption, specifically including: 32-bit floating-point encoding (4 bytes per feature value, precision ±1e-7), 16-bit integer encoding (2 bytes, precision ±1e-3), and 8-bit integer encoding (1 byte, precision ±1e-2). The network parameters of the current diagnostic feature generation network default to 32-bit floating-point encoding, and its corresponding feature resource consumption is 4 bytes / feature value (i.e., the baseline resource ratio).

[0128] By comparing the resource consumption of each encoding format in the second association rule table, the server determines the target feature encoding format: the resource consumption of both 16-bit integer encoding (2 bytes / feature value) and 8-bit integer encoding (1 byte / feature value) is less than the baseline resource consumption (4 bytes / feature value). Considering the accuracy requirements of the pressure distribution characteristics (the process allows a pressure index error ≤ ±0.5MPa, corresponding to an encoding accuracy ≥ ±0.1MPa), the server excludes 8-bit integer encoding (accuracy ±1e-2, which may lead to errors exceeding 1MPa) and selects 16-bit integer encoding as the target feature encoding format (2 bytes / feature value, accuracy ±1e-3, meeting the error requirements).

[0129] Subsequently, the server performs encoding conversion on the 512-dimensional pressure distribution characteristics of mold M1: the original feature value is a 32-bit floating-point number (range [-1, 1], which, after normalization, corresponds to the actual pressure range of 0-100 MPa). The server converts it to a 16-bit integer (range [-32768, 32767]) using a linear mapping formula (integer encoding value = round(float-point feature value × 32767)). For example, the floating-point feature value corresponding to the peak pressure during the filling stage is 0.86 (corresponding to an actual pressure of 86 MPa), and the converted 16-bit integer encoding value is round(0.86 × 32767) = 28180; the floating-point feature value corresponding to the average holding pressure is 0.71 (actual pressure of 71 MPa), and the converted value is round(0.71 × 32767) = 23264. After the conversion, the total resource consumption of the 512-dimensional pressure distribution feature decreased from 512×4=2048 bytes to 512×2=1024 bytes, a 50% reduction. Furthermore, verification showed that the actual pressure error after decoding each feature value was ≤±0.3MPa (≤ process allowable error ±0.5MPa), meeting the diagnostic accuracy requirements. The server inputs the encoded pressure distribution feature (512-dimensional 16-bit integer vector) into the diagnostic feature generation network to continue performing subsequent semantic feature conversion.

[0130] In this embodiment of the invention, the step of determining the target feature encoding form whose corresponding feature resource usage is less than the baseline resource ratio from each feature encoding form according to the second association rule between each preset feature encoding form and each feature resource usage can be implemented through the following example.

[0131] Obtain the analysis and processing load during the current training cycle;

[0132] When it is determined that the analysis and processing load exceeds the load limit, according to the second association rule between each feature encoding form and the resource consumption of each feature, a target feature encoding form whose resource consumption is less than the baseline resource ratio is determined from each feature encoding form.

[0133] In this embodiment of the invention, for example, during the 50th batch training of the diagnostic feature generation network (each batch processes 512 stress time series samples, each sample containing 512-dimensional stress distribution features), the server needs to monitor the analysis and processing load of the current training cycle in real time to dynamically adjust the feature encoding form. The server collects hardware resource usage data through the system monitoring module (a resource monitoring tool deployed in the Linux kernel layer): the current CPU is an Intel Xeon Gold 6248 (20 cores and 40 threads), and the monitoring shows that its utilization rate is 85% (exceeding the preset load limit of 80%, which is set according to the hardware heat dissipation threshold and model training latency requirements to avoid the CPU frequency from dropping or training from stalling due to continuous high load); the memory utilization rate is 72% (23GB of 32GB memory is used), of which the stress distribution feature storage occupies 12GB (512 samples × 512 features × 4 bytes / feature = 1,048,576 bytes / batch, accounting for 40% after caching multiple batches).

[0134] The server reads the preset second association rule table, which specifies: 32-bit floating-point encoding (baseline encoding form) occupies 4 bytes / feature value (precision ±1e-7, supports pressure error ≤ ±0.01MPa), 16-bit integer encoding occupies 2 bytes / feature value (precision ±1e-3, supports error ≤ ±0.3MPa), and 8-bit integer encoding occupies 1 byte / feature value (precision ±1e-2, error ≤ ±1MPa). The baseline resource proportion is 4 bytes / feature value for 32-bit floating-point encoding. The server needs to filter encoding forms (16-bit and 8-bit integer encoding) with resource occupancy less than this value. Considering the pressure diagnosis accuracy requirements (process requirements pressure condition index error ≤ ±0.5MPa), 8-bit integer encoding may cause errors in fine indicators such as pressure fluctuation amplitude exceeding 1MPa (not meeting the requirements), so it is excluded; 16-bit integer encoding has a precision of ±1e-3 (corresponding to actual pressure error ≤ ±0.3MPa, such as a filling pressure peak of 86MPa with an error ≤ 0.2MPa after encoding conversion), which meets the process requirements.

[0135] Therefore, the server determined that 16-bit integer encoding was the target feature encoding format, as its resource consumption (2 bytes / feature value) was less than the baseline resource consumption (4 bytes / feature value). Subsequently, the server performed an encoding conversion on the 512-dimensional pressure distribution features of the current batch: the original 32-bit floating-point feature values ​​(range [-1, 1], corresponding to actual pressure of 0-100 MPa after normalization) were converted to 16-bit integers (range [-32768, 32767]) through a linear mapping (integer encoding value = round(floating-point feature value × 32767)). After the conversion, the memory consumption of a single batch of pressure distribution features decreased from 1,048,576 bytes (32-bit encoding) to 524,288 bytes (16-bit encoding), and the CPU utilization rate decreased to 75% (below the load limit), ensuring a smooth training process and that the feature accuracy met diagnostic requirements.

[0136] In this embodiment of the invention, the step of optimizing the network parameters of the diagnostic feature generation network based on the error between the obtained inferred pressure state diagnostic data and the corresponding sample pressure state diagnostic data can be implemented through the following example.

[0137] Based on the error between the multidimensional inferred pressure condition indicators contained in the obtained inferred pressure state diagnostic data and the multidimensional sample pressure condition indicators contained in the corresponding sample pressure state diagnostic data, the diagnostic difference is determined.

[0138] Based on the diagnostic difference, optimize the network parameters of the diagnostic feature generation network.

[0139] In an embodiment of the present invention, for example, in the 50th batch (including 100 pressure timing samples of mold M1) of training the diagnostic feature generation network, after performing inference on a certain sample, the server needs to optimize the network parameters based on the error. The mold pressure timing signal of this sample corresponds to the sample pressure state diagnostic data, and its multi-dimensional sample pressure condition indicators are: filling time 2.9 seconds (process standard 2.7-3.1 seconds), peak filling pressure 86MPa (standard 83-89MPa), average holding pressure 71MPa (standard 68-73MPa), holding pressure fluctuation ±1.8MPa (standard ≤±3MPa), and cooling attenuation rate 4.9MPa / s (standard 4.5-5.5MPa / s).

[0140] The server outputs inferred pressure status diagnostic data through the diagnostic feature generation network and the operating condition decoding network. The multidimensional inferred pressure operating condition indicators are: filling time 3.1 seconds, peak filling pressure 88 MPa, average holding pressure 69 MPa, holding pressure fluctuation ±2.2 MPa, and cooling attenuation rate 5.1 MPa / s.

[0141] The server calculates the errors of each indicator: filling time error = 3.1 - 2.9 = 0.2 seconds (exceeding the standard upper limit by 0.2 seconds), peak filling pressure error = 88 - 86 = 2MPa (standard upper limit + 2MPa), average holding pressure error = 69 - 71 = -2MPa (standard lower limit - 2MPa), holding pressure fluctuation error = 2.2 - 1.8 = 0.4MPa, and cooling attenuation rate error = 5.1 - 4.9 = 0.2MPa / s.

[0142] Based on the process importance weights (fill time 0.2, peak filling time 0.3, average holding pressure 0.3, fluctuation range 0.1, attenuation rate 0.1), the diagnostic difference is calculated as follows: 0.2×0.2+0.3×2+0.3×(-2)+0.1×0.4+0.1×0.2=0.04+0.6-0.6+0.04+0.02=0.1 (threshold 0.05, needs optimization).

[0143] The server uses the diagnostic variance as the loss function and calculates the gradient using the Adam optimizer (learning rate 0.001): the derivative of the weights of the second fully connected layer of the network (connection pressure distribution features and primary diagnostic semantic features) is calculated, the fill time-related weight gradient is +0.03 (the weights need to be reduced to decrease the fill time inference value), the fill peak-related weight gradient is +0.02 (the weights need to be reduced), and the pressure holding mean-related weight gradient is -0.02 (the weights need to be increased).

[0144] After updating the weights according to the gradient direction, the sample is re-inferred: the filling time is 3.0 seconds (error 0.1 seconds), the filling peak is 87 MPa (error 1 MPa), the holding pressure average is 70 MPa (error -1 MPa), the diagnostic difference is reduced to 0.04 (< threshold), and the network parameter optimization of the sample is completed.

[0145] In this embodiment of the invention, optimizing the network parameters of the diagnostic feature generation network based on the diagnostic difference can be implemented through the following example.

[0146] The state matching error is determined based on the difference between the obtained inferred pressure state diagnostic data and the corresponding sample pressure state diagnostic data.

[0147] Based on the diagnostic difference and the state matching error, when the diagnostic accuracy criterion is not met, the network parameters of the diagnostic feature generation network are optimized.

[0148] In an embodiment of the present invention, for example, for a certain pressure timing sample (10-second injection cycle) of mold M1, after the server completes the calculation of multi-dimensional inferred pressure condition indicators (filling time 3.1 seconds, peak filling pressure 88MPa, etc.), it further generates inferred pressure state diagnostic data text through the condition decoding network: "The pressure distribution during the filling stage is basically uniform (filling time 3.1 seconds, peak pressure 88MPa), the pressure stability during the holding stage is average (average 69MPa, fluctuation range 2.2MPa), the pressure decay during the cooling stage is slightly faster (decay rate 5.1MPa / s), and the current process parameters match the pressure state 92%."

[0149] The server retrieves the sample pressure status diagnostic data text corresponding to this sample: "Uniform pressure distribution during the filling stage (filling time 2.9 seconds, peak pressure 86 MPa), stable pressure during the holding stage (average 71 MPa, fluctuation range 1.8 MPa), normal pressure decay during the cooling stage (decay rate 4.9 MPa / s), current process parameters match the pressure status 98%." The difference between the two is calculated using a cosine similarity algorithm: the cosine similarity of the word vectors for the text keywords ("uniform filling", "stable holding pressure", "normal decay", etc.) is 0.85. Therefore, the status matching error is 1 - 0.85 = 0.15 (the process setting threshold for status matching error is 0.1).

[0150] Based on the previously calculated diagnostic discrepancy of 0.1 (threshold 0.05), the server determines that the current sample does not meet the diagnostic accuracy criterion (diagnostic discrepancy 0.1 > 0.05, state matching error 0.15 > 0.1). Then, using the weighted sum of the diagnostic discrepancy (0.1) and the state matching error (0.15) (weight 0.5:0.5) as the comprehensive loss function (0.1 × 0.5 + 0.15 × 0.5 = 0.125), the parameters of the diagnostic feature generation network are optimized through backpropagation using the Adam optimizer (learning rate 0.001): The focus is on adjusting the attention weights of the deep diagnostic feature fusion sub-network (the attention weights during the filling stage were originally 0.3, now reduced to 0.25; the attention weights during the holding stage were increased from 0.25 to 0.3), and fine-tuning the bias terms of the fully connected layers in the primary diagnostic feature generation sub-network (the filling time-related bias term was adjusted from -0.02 to -0.03, and the holding mean-related bias term was adjusted from 0.01 to 0.02).

[0151] After optimization, the sample was re-inferred: the multidimensional inferred pressure condition indicators were adjusted to a filling time of 3.0 seconds (error of 0.1 seconds), a peak filling pressure of 87 MPa (error of 1 MPa), and an average holding pressure of 70 MPa (error of -1 MPa). The similarity of the state diagnosis text was improved to 0.93 (state matching error of 0.07), and the diagnostic difference was reduced to 0.04 (both meet the threshold requirements). The network parameter optimization was completed.

[0152] In this embodiment of the invention, the diagnostic feature generation network includes a primary diagnostic feature generation subnetwork and a deep diagnostic feature fusion subnetwork.

[0153] The process of using the diagnostic feature generation network to convert the obtained pressure distribution features into diagnostic semantic features can be implemented through the following example.

[0154] The primary diagnostic features are used to generate a sub-network, which converts the obtained pressure distribution features into multiple primary diagnostic semantic feature vectors; wherein each primary diagnostic semantic feature vector represents a preliminary prediction of at least one single-dimensional inferred pressure condition index.

[0155] The deep diagnostic feature fusion subnetwork is used to fuse multiple primary diagnostic semantic feature vectors and learn the high-order correlation between the multiple primary diagnostic semantic feature vectors to generate the diagnostic semantic features represented by the fused multidimensional inferred pressure condition index.

[0156] In an embodiment of the present invention, for example, the server calls the diagnostic feature generation network (including the primary diagnostic feature generation sub-network and the deep diagnostic feature fusion sub-network) to perform transformation on the 512-dimensional pressure distribution features of mold M1 (covering local and coupled features of the filling, holding, and cooling stages).

[0157] The primary diagnostic feature generation subnetwork generates primary diagnostic semantic feature vectors:

[0158] The primary subnetwork adopts a "3-branch fully connected structure," with each branch corresponding to a stage of the injection molding process. The input is the local and coupling features of that stage's pressure distribution characteristics, and the output is a primary vector containing a preliminary inference of the single-dimensional inferred pressure condition indicators for that stage. Specifically:

[0159] The filling stage branch takes the first 160 bits of the input pressure distribution features (local features of the filling stage: proportion of the 50-100Hz frequency band, pressure rise slope, etc.; coupled features: filling-holding consistency, temporal relationship), and processes them through two fully connected layers (first layer 160→64, ReLU activation; second layer 64→32) to generate a 32-dimensional primary diagnostic semantic feature vector for the filling stage, including "preliminary estimate of filling time 3.0 seconds" (based on the frequency band proportion of 0.92, error ±0.2 seconds), "preliminary estimate of peak filling pressure 87MPa" (based on the pressure rise slope of 28.3MPa / s, error ±2MPa), and "filling stage pressure uniformity score 0.85" (based on the filling-holding transition delay of 0.02 seconds in the coupled features).

[0160] The pressure holding stage branch: Input the features from bits 161 to 320 (local features of the pressure holding stage: proportion of the 10-30Hz frequency band, pressure fluctuation variance, etc.; coupled features: pressure holding-cooling frequency change), and generate a 32-dimensional primary vector for the pressure holding stage through the same structure, including "preliminary estimate of the average pressure holding is 70MPa" (based on the frequency band proportion of 0.89, with an error of ±2MPa), "preliminary estimate of the pressure holding fluctuation amplitude is ±2.0MPa" (based on the fluctuation variance of 1.8MPa², with an error of ±0.3MPa), and "stability score of the pressure holding stage is 0.88" (based on the smoothness of frequency change).

[0161] Cooling stage branch: Input features from bits 321 to 512 (local features of the cooling stage: proportion of 1-5Hz frequency band, pressure attenuation rate, etc.; coupling features: coupling strength of overall pressure distribution) to generate a 32-dimensional primary vector of the cooling stage, including "preliminary estimate of cooling attenuation rate 5.0MPa / s" (inferred based on attenuation rate features, with an error of ±0.2MPa / s) and "preliminary estimate of residual pressure 15MPa" (inferred based on local features, with an error of ±1MPa).

[0162] Deep diagnostic feature fusion subnetwork generates diagnostic semantic features:

[0163] The deep sub-network employs an "attention mechanism + residual connection" structure. It takes the three primary vectors (32 × 3 = 96 dimensions) as input and learns higher-order relationships to optimize the inferred index values, generating 256-dimensional diagnostic semantic features. Specific process:

[0164] Learning higher-order relationships: The server calculates the association weights between primary vectors through a multi-head attention layer (4 heads, 24 dimensions per head), for example:

[0165] The positive correlation weight between filling time (3.0 seconds) and average holding pressure (70 MPa) is 0.75 (process rule: longer filling time requires higher holding pressure to compensate for shrinkage).

[0166] The positive correlation weight between the peak filling pressure (87MPa) and the pressure holding fluctuation range (±2.0MPa) is 0.6 (an excessively high peak value can easily lead to increased pressure holding fluctuation).

[0167] The positive correlation weight between the cooling attenuation rate (5.0 MPa / s) and the average holding pressure (70 MPa) is 0.65 (the higher the holding pressure, the faster the cooling attenuation rate).

[0168] Optimize initial inference values: Adjust metrics based on correlation weights, for example:

[0169] The initial estimate for the filling time is 3.0 seconds. Based on the average holding pressure of 70 MPa (lower than the sample index of 71 MPa), the time is adjusted upward to 2.9 seconds (closer to the sample 2.9 seconds) with a positive correlation weight of 0.75.

[0170] The peak pressure was 87 MPa. Combined with the holding pressure fluctuation of ±2.0 MPa (higher than the sample pressure of 1.8 MPa), it was adjusted down to 86 MPa (close to the sample pressure of 86 MPa) with a positive correlation weight of 0.6.

[0171] The average holding pressure is 70 MPa. Combined with the filling time of 2.9 seconds (close to the standard) and the cooling decay rate of 5.0 MPa / s (sample 4.9 MPa / s), it is adjusted to 71 MPa (matching the sample).

[0172] Generate diagnostic semantic features: Integrate optimized multidimensional inferred pressure condition indicators (filling time 2.9 seconds, filling peak 86 MPa, average holding pressure 71 MPa, holding pressure fluctuation ±1.8 MPa, cooling attenuation rate 4.9 MPa / s, etc.), supplement auxiliary features such as stability score in the primary vector through residual connection, and finally output a 256-dimensional diagnostic semantic feature vector (the first 50 correspond to the core pressure indicators, the 51-200 correspond to the correlation weights, and the 201-256 correspond to the stability score), which fully represents the multidimensional inferred working condition of the current pressure state of mold M1.

[0173] In this embodiment of the invention, after at least one complete training cycle has been completed, the following implementation method is also provided.

[0174] Obtain an external pressure monitoring verification dataset; wherein, the external pressure monitoring verification dataset includes mold pressure time series signals and their actual pressure state diagnostic data that were not involved in the construction of the pressure time series sample set;

[0175] Using the pressure state analysis model obtained through current training, inference is made on the external pressure monitoring verification dataset to obtain the corresponding inferred verification pressure state diagnostic data.

[0176] Calculate the generalization error between the inferred and verified pressure state diagnostic data and the actual pressure state diagnostic data;

[0177] Based on the generalization error, the hyperparameters of subsequent training cycles are dynamically adjusted; wherein the hyperparameters include at least one of the following: learning rate, batch size, and training data sample weight distribution.

[0178] In an embodiment of the present invention, for example, after the server completes the 100th batch training of the pressure state analysis model (1000 samples per batch, covering the pressure time series samples of molds M1-M10), it needs to evaluate the model's generalization ability and adjust the hyperparameters through external validation.

[0179] External pressure monitoring verification dataset acquisition: The server retrieves the external pressure monitoring verification dataset from the historical database of the newly commissioned mold M11 in the workshop (producing the same series of mobile phone shells, with a cavity structure similar to M1 but a gate diameter increased by 0.2mm, and slightly different process parameters: injection temperature 235℃±5℃, injection speed 85mm / s±5mm / s). This dataset contains 5000 pressure timing signals of M11 (each signal cycles for 10 seconds, sampling frequency 1kHz) and corresponding real pressure state diagnostic data. The real data is annotated by process experts and includes multi-dimensional real pressure condition indicators (such as filling time 3.2 seconds±0.1 seconds, peak filling pressure 89MPa±2MPa, average holding pressure 73MPa±1MPa) and diagnostic text (such as "pressure distribution is slightly uneven during the filling stage, and pressure stability is good during the holding stage").

[0180] Model inference and generalization error calculation: The server loads the currently trained pressure state analysis model (the diagnostic feature generation network has been fine-tuned through 100 batches, and the diagnostic difference in the validation set is 0.04 ≤ threshold 0.05). Inference is performed on 5000 signals in the validation set in batches (500 signals per batch). Taking a pressure signal segment of M11 (filling phase 0-3.2 seconds, pressure 55-89MPa; holding phase 3.2-6.2 seconds, pressure 72-74MPa) as an example, the model outputs the following inferred and validated pressure state diagnostic data: multidimensional inferred pressure condition indicators "filling time 3.4 seconds, peak filling pressure 92MPa, average holding pressure 71MPa, holding pressure fluctuation ±2.5MPa", and diagnostic text "uneven pressure distribution during the filling phase (filling time 3.4 seconds > standard 3.3 seconds), and average pressure stability during the holding phase (fluctuation 2.5MPa > standard 2MPa)". The server calculates the generalization error: Diagnostic difference: Comparing the multidimensional inferred index with the real index (real filling time 3.2 seconds, peak pressure 89 MPa, average holding pressure 73 MPa, fluctuation range ±1.5 MPa), the error is calculated according to the process weights (fill time 0.3, peak pressure 0.3, average holding pressure 0.3, fluctuation range 0.1): 0.3×(3.4-3.2)+0.3×(92-89)+0.3×(71-73)+0.1×(2.5-1.5)=0.06+0.9-0.6+0.1=0.46 (the diagnostic difference in the training set is 0.04, and the generalization error is significantly increased); State matching error: The cosine similarity between the inferred text and the real text ("The pressure distribution in the filling stage is slightly uneven, and the pressure stability in the holding stage is good") is 0.72, and the state matching error is 1-0.72=0.28 (the state matching error in the training set is 0.08≤threshold 0.1). Overall generalization error = diagnostic difference × 0.6 + state matching error × 0.4 = 0.46 × 0.6 + 0.28 × 0.4 = 0.276 + 0.112 = 0.388 (preset generalization error threshold 0.2, hyperparameters need to be adjusted).

[0181] Dynamically adjust subsequent training hyperparameters: Based on the generalization error analysis, the server determined the cause: the increased gate diameter of M11 led to a decrease in filling resistance (shortened filling time), but the model's learning of the higher-order correlation between "filling time and holding pressure" was insufficient (there were no large gate cases in the training samples). Therefore, the hyperparameters were dynamically adjusted: Learning rate: decreased from 0.001 to 0.0005 (reducing parameter update amplitude to avoid overfitting); Batch size: increased from 1000 samples / batch to 1500 samples / batch (increasing data diversity and strengthening the learning of correlation features); Training data sample weight distribution: samples with "filling time < 3.0 seconds" and "mean holding pressure > 73 MPa" (similar to the large gate characteristics of M11) were assigned a weight of 1.5 (original weight 1.0), while the weight of ordinary samples remained at 1.0. After adjustment, the server started the 101st batch of training: the new batch contained 1500 samples, of which 30% were samples from M1-M10 similar to the M11 process (such as the low-resistance cavity case of M3). After 20 rounds of training adjustments, the server was evaluated again using the M11 validation set: the diagnostic discrepancy decreased to 0.18, the state matching error was 0.12, and the overall generalization error was 0.15 (< threshold 0.2), indicating a significant improvement in the model's generalization ability.

[0182] In this embodiment of the invention, the step of performing time-series segmentation on each pressure monitoring dataset according to a preset sampling period can be implemented through the following example.

[0183] Analyze the signal energy changes or frequency component changes in the pressure monitoring dataset to determine multiple signal mode transition points;

[0184] Based on the multiple signal mode conversion points, the segmentation points are determined, and the pressure monitoring dataset is divided into multiple pressure signal segments with inherent consistency, which serve as the corresponding mold pressure timing signals; wherein, the inherent consistency indicates that the signal modes within the same pressure signal segment remain consistent.

[0185] In this embodiment of the invention, for example, the server performs time-series segmentation on the pressure monitoring dataset of mold M1 (collected continuously for 24 hours, with a sampling frequency of 1kHz, containing 500 complete injection cycles, each cycle corresponding to a 10-second pressure signal, for a total of 10,000 sampling points). First, the server analyzes the frequency components and energy changes of the signal using Short-Time Fourier Transform (STFT): setting the sliding window size to 0.1 seconds (100 sampling points) and the step size to 0.05 seconds, and calculating the power spectral density of each window. The results show that the power of the signal in the 0-3 second interval (filling stage) is mainly concentrated in the 50-100Hz frequency band (accounting for 92%), corresponding to the dynamic pressure fluctuations caused by the high-speed flow of the melt; in the 3-6 second interval (holding stage), the power shifts to the 10-30Hz frequency band (accounting for 88%), as the melt flow slows down and the pressure tends to stabilize; in the 6-10 second interval (cooling stage), the power further decreases to the 1-5Hz frequency band (accounting for 95%), corresponding to the static attenuation of the cavity pressure.

[0186] Based on the abrupt changes in frequency components, the server determines the signal mode transition points: at 2.98 seconds, the proportion of the 50-100Hz band drops sharply from 92% to 30%, while the proportion of the 10-30Hz band rises from 10% to 85%, which is identified as the "fill-hold transition point"; at 5.97 seconds, the proportion of the 10-30Hz band drops from 88% to 15%, while the proportion of the 1-5Hz band rises from 5% to 90%, which is identified as the "hold-cooling transition point".

[0187] The server uses these two transition points as dividing points to divide the 10-second pressure monitoring data into three pressure signal segments: 0-2.98 seconds (filling stage segment), 2.98-5.97 seconds (holding stage segment), and 5.97-10 seconds (cooling stage segment). Each segment exhibits inherent consistency: within the filling stage, the 50-100Hz frequency band consistently accounts for ≥90%, with energy fluctuations ≤5%; within the holding stage, the 10-30Hz frequency band accounts for ≥85%, with average pressure fluctuations ≤±2MPa; and within the cooling stage, the 1-5Hz frequency band accounts for ≥95%, with a stable pressure decay rate of 4.5-5.5MPa / s. Through this segmentation, the server converts the continuous pressure monitoring dataset into multiple independent mold pressure time-series signals, each corresponding to the complete pressure change process of a single injection cycle, providing structured input for subsequent feature extraction and model training.

[0188] In this embodiment of the invention, the step of converting the pressure distribution features into the target feature encoding form to obtain the encoded pressure distribution features can be implemented through the following example.

[0189] Real-time monitoring of the computational resource consumption required for the current training cycle;

[0190] Based on the correlation rules between the computing resource consumption, the analysis and processing load, and the resource consumption of the feature, the compression level of the target feature encoding format is dynamically selected.

[0191] According to the selected compression level, the pressure distribution feature is converted into the target feature encoding form to obtain the encoded pressure distribution feature; wherein, the higher the compression level, the less resources the generated encoded pressure distribution feature occupies.

[0192] In an embodiment of the present invention, for example, when the server is training the diagnostic feature generation network in the 75th batch (each batch processes 512 mold pressure time-series signals, corresponding to 512 512-dimensional pressure distribution features), it needs to dynamically select the compression level to balance resource consumption and feature accuracy before converting the pressure distribution features into the target feature encoding form (16-bit integer encoding).

[0193] First, the server monitors the computing resource consumption of the current training cycle in real time through the system monitoring module: CPU utilization is 85% (load limit of 80%, already overloaded), memory usage is 15GB (total memory is 32GB, pressure distribution feature storage uses 9GB, accounting for 60%, of which 512 samples × 512 features × 2 bytes / feature = 524,288 bytes / batch, the proportion is high after caching multiple batches), and GPU memory usage is 8GB (mainly used for inference computing).

[0194] Next, the server reads the preset association rule table to clarify the mapping relationship between the compression level of the target feature encoding format (16-bit integer encoding) and the computational resource consumption and feature resource usage: Compression Level 1 (lossless compression, 2 bytes / feature value, accuracy ±1e-3, stress error ≤ ±0.3MPa), Level 2 (light compression, 12-bit integer encoding, 1.5 bytes / feature value, accuracy ±1e-2, error ≤ ±0.5MPa), and Level 3 (medium compression, 8-bit integer encoding, 1 byte / feature value, accuracy ±1e-1, error ≤ ±1MPa). The rules also specify that when CPU utilization is 80%-90% and memory usage is >12GB, compression level 2 is selected; when CPU utilization is >90% and memory usage is >15GB, level 3 is selected (sacrificing some accuracy).

[0195] The current monitoring data (CPU 85%, memory 15GB) matches the condition of "CPU 80%-90% and memory > 12GB". The server dynamically selects compression level 2 (12-bit integer encoding, 1.5 bytes / feature value). Its feature resource usage (1.5 bytes / feature value) is less than the 2 bytes / feature value of level 1, and the accuracy meets the process requirements (error ≤ ±0.5MPa).

[0196] Subsequently, the server converts the 512-dimensional pressure distribution features of mold M1 according to the compression rule of level 2: the original 16-bit integer encoded value (range [-32768, 32767]) is converted into a 12-bit integer (range [-2048, 2047]) by right shifting by 4 bits (discarding the lower 4 bits), realizing 1.5 bytes / feature value storage (the high 8 bits are stored in 1 byte, and the low 4 bits are stored in the lower 4 bits of the adjacent feature, and are concatenated by bit operations). For example, the 16-bit coded value 28180 (binary 110111001010100) corresponding to the peak filling pressure, shifted 4 bits to the right, yields the 12-bit coded value 1761 (binary 11011100101), corresponding to an actual pressure of 86 MPa, with a conversion error of 0.2 MPa (≤±0.5 MPa); the coded value 23264 (101101001100000) corresponding to the average holding pressure, shifted 4 bits to the right, yields 1454 (10110100110), corresponding to 71 MPa, with an error of 0.1 MPa.

[0197] After conversion, the memory usage of a single batch of stress distribution features decreased from 524,288 bytes (Level 1) to 393,216 bytes (Level 2), and the total memory usage decreased to 12GB (stress features account for 40%). CPU utilization subsequently decreased to 78% (< load limit). Simultaneously, through verification using 100 randomly selected features, the average error of the stress index was 0.3MPa (≤±0.5MPa), meeting the diagnostic accuracy requirements. The server ultimately obtained the encoded stress distribution features (512-dimensional 12-bit integer vectors) and input them into the diagnostic feature generation network for subsequent semantic feature conversion.

[0198] In this embodiment of the invention, the step of obtaining the current mold pressure timing signal of the current injection mold during the injection process and outputting the corresponding current pressure state diagnostic data based on the trained pressure state analysis model can be implemented through the following example.

[0199] Obtain the current mold pressure timing signal during the injection process;

[0200] The current mold pressure time series signal is input into the trained pressure state analysis model, and the corresponding current pressure state diagnostic data is output; wherein, the current pressure state diagnostic data includes multi-dimensional current pressure condition indicators of the current mold pressure time series signal.

[0201] Based on the multi-dimensional current pressure condition indicators included in the current pressure condition diagnostic data, the current pressure condition indicators are compared with the preset standard pressure condition indicator range to determine whether there is an abnormality in the current pressure condition of the current injection mold.

[0202] If an anomaly is found, a corresponding pressure anomaly handling instruction is generated based on the current pressure status diagnostic data; wherein, the pressure anomaly handling instruction is used to adjust the injection molding process parameters or trigger a mold maintenance warning.

[0203] In this embodiment of the invention, for example, the server performs real-time pressure status monitoring on mold M1 (currently in its 500th batch of production, having run continuously for 8 hours) that is producing a certain model of mobile phone casing. First, the server collects the pressure timing signal of the current injection cycle in real time through pressure sensors (sampling frequency 1kHz, accuracy ±0.1MPa) deployed at the mold cavity and gate: this signal is 10 seconds of continuous time-domain data (10,000 sampling points), corresponding to a complete injection process—the filling stage (0-3.2 seconds, pressure rises from 50MPa to 90MPa), the holding stage (3.2-6.2 seconds, pressure fluctuates between 72-74MPa), and the cooling stage (6.2-10 seconds, pressure decays from 72MPa to 16MPa). The server synchronously obtains the current process parameters from the injection molding machine control system: injection temperature 232℃ (standard 225-235℃), injection speed 80mm / s (standard 75-85mm / s), mold temperature 62℃ (standard 55-65℃), ensuring that the signal is correlated with the process conditions.

[0204] The server inputs the current mold pressure timing signal into the trained pressure state analysis model (the diagnostic feature generation network has been optimized in 120 batches, with a generalization error of 0.18 ≤ threshold 0.2). The model flow is as follows: The signal characterization extraction network extracts pressure distribution features (88% in the 50-100Hz band during the filling stage, 85% in the 10-30Hz band during the holding stage, and 94% in the 1-5Hz band during the cooling stage; the filling-holding transition delay in the coupling features is 0.03 seconds); the diagnostic feature generation network converts it into diagnostic semantic features (256-dimensional vector, containing multi-dimensional current pressure condition indicators); the condition decoding network outputs the current pressure state diagnostic data: the multi-dimensional current pressure condition indicators are "filling time 3.2 seconds, peak filling pressure 90MPa, average holding pressure 72MPa, holding pressure fluctuation ±2.5MPa, cooling attenuation rate 5.6MPa / s, residual pressure 16MPa", and the diagnostic text is "the pressure distribution during the filling stage is slightly uneven (the filling time of 3.2 seconds is close to the standard upper limit), the pressure fluctuation during the holding stage is relatively large (2.5MPa > standard ±2MPa), and the pressure attenuation during the cooling stage is slightly faster (5.6MPa / s > standard 5.5MPa / s)".

[0205] The server invoked preset standard pressure condition parameters (defined by the process department and based on historical qualified product data): filling time 2.7-3.1 seconds (currently 3.2 seconds, exceeding the upper limit by 0.1 seconds), peak filling pressure 83-89 MPa (currently 90 MPa, exceeding the upper limit by 1 MPa), average holding pressure 68-73 MPa (currently 72 MPa, within the range), holding pressure fluctuation ≤ ±2 MPa (currently ±2.5 MPa, exceeding the upper limit by 0.5 MPa), cooling attenuation rate 4.5-5.5 MPa / s (currently 5.6 MPa / s, exceeding the upper limit by 0.1 MPa / s), and residual pressure 12-18 MPa (currently 16 MPa, within the range). Through item-by-item comparison, the server determined that the current pressure status had three anomalies: excessively long filling time, excessively high peak filling pressure, excessive holding pressure fluctuation, and slightly rapid cooling attenuation rate (the first three being the main anomalies, and the cooling attenuation rate being a minor anomaly).

[0206] For anomalies, the server generates pressure anomaly handling instructions based on the correlation of process parameters:

[0207] Excessive filling time and high peak pressure: Based on the current injection speed of 80mm / s (standard lower limit 75mm / s), it is inferred that the low speed may be causing prolonged filling time and excessively high peak pressure. Instruction 1 is generated: "Adjust injection speed to 85mm / s (standard upper limit), reduce filling time to within 3.0 seconds, and control peak filling pressure ≤ 89MPa".

[0208] Pressure holding fluctuation exceeds the limit: Given the current average pressure holding value of 72 MPa (close to the upper limit of 73 MPa), it is inferred that the pressure fluctuation may be due to unstable pressure holding. Instruction 2 is generated: "Fine-tune the average pressure holding value to 71 MPa, and simultaneously activate the pressure fluctuation compensation algorithm (automatically adjust the pressure holding value by ±0.5 MPa when the fluctuation exceeds ±1.5 MPa)".

[0209] Cooling decay rate is slightly fast: associated mold temperature 62℃ (standard median 60℃), it is recommended to maintain the current mold temperature and continuously monitor the decay rate for the next 3 cycles. If it exceeds 5.5MPa / s, adjust to 61℃.

[0210] The server sends the above processing instructions to the injection molding machine control system via industrial Ethernet. At the same time, the workshop monitoring platform displays an abnormal warning: "The current pressure status of mold M1 has excessive filling time (3.2 seconds), excessive peak pressure (90MPa), and excessive holding pressure fluctuation (±2.5MPa). The parameter adjustment instruction has been automatically issued. It is expected to return to normal in the next cycle." If the abnormality still exists after 3 cycles after adjustment (such as the filling time still exceeding 3.1 seconds), the server will trigger a mold maintenance warning: "Mold M1 may have local blockage in the cavity. It is recommended to stop the machine and check the surface finish of the gate and cavity."

[0211] Through this process, the server enables real-time pressure monitoring, anomaly diagnosis, and closed-loop control of the injection molding process, ensuring product quality stability.

[0212] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 performs the aforementioned injection mold pressure monitoring processing method. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0213] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A method for monitoring and processing pressure in injection molds, characterized in that, include: Obtain a pressure time series sample set; wherein each pressure time series sample includes a corresponding mold pressure time series signal, and sample pressure state diagnostic data containing multi-dimensional sample pressure condition indicators of the mold pressure time series signal; Based on the aforementioned pressure time-series sample set, multiple batches of parameter tuning are performed on the untrained pressure state analysis model; wherein, the pressure state analysis model includes: a pre-trained signal representation extraction network and a working condition decoding network, as well as an untrained diagnostic feature generation network; each training cycle includes: The signal characterization extraction network is used to extract the pressure distribution characteristics represented by the mold pressure time-series signal; The obtained pressure distribution features are converted into diagnostic semantic features using the diagnostic feature generation network; wherein, the diagnostic semantic features are used to characterize the multidimensional inferred pressure condition index of the mold pressure time series signal. Using the aforementioned working condition decoding network, the diagnostic semantic features are decoded to generate inferred stress state diagnostic data; Based on the error between the obtained inferred pressure state diagnostic data and the corresponding sample pressure state diagnostic data, the network parameters of the diagnostic feature generation network are optimized. Acquire the current mold pressure timing signal of the current injection mold during the injection process, and output the corresponding current pressure state diagnostic data based on the trained pressure state analysis model; The acquisition of the stress time series sample set includes: Collect pressure monitoring datasets for multiple injection molds corresponding to the injection process, and collect the process parameter records corresponding to each pressure monitoring dataset; Analyze the signal energy changes or frequency component changes in the pressure monitoring dataset to determine multiple signal mode transition points; Based on the multiple signal mode transition points, the segmentation points are determined, and the pressure monitoring dataset is divided into multiple pressure signal segments with inherent consistency; wherein, the inherent consistency indicates that the signal modes within the same pressure signal segment remain consistent. Multiple pressure signal segments corresponding to the corresponding pressure monitoring dataset are obtained, and each obtained pressure signal segment is used as the corresponding mold pressure time sequence signal. Using a pre-trained pressure feature extraction model, pressure condition indicators are identified for each mold pressure time series signal, resulting in multi-dimensional sample pressure condition indicators of the corresponding mold pressure time series signal. Using the pre-trained diagnostic generation model, based on each obtained process parameter record and each sample pressure condition index, sample pressure state diagnostic data corresponding to each mold pressure time sequence signal is generated. Based on the pressure timing signal of each mold and its corresponding sample pressure state diagnostic data, a pressure timing sample set is obtained. The step of using the signal characterization extraction network to extract the pressure distribution features represented by the mold pressure time-series signal includes: Using the signal characterization extraction network, the local pressure distribution features corresponding to multiple time segments in the mold pressure time series signal are extracted, and the pressure distribution coupling features between each local pressure distribution feature are extracted. By integrating the obtained local pressure distribution features and pressure distribution coupling features, the pressure distribution features representing the mold pressure time series signal are obtained; the local pressure distribution features are used to indicate at least one of the following: single-segment frequency features, energy distribution pattern, acquisition conditions, time regularity, physical source category, and comprehensive high-level characterization of the corresponding time series segment; the pressure distribution coupling features are used to indicate at least one of the following: consistency, time series relationship, and frequency change between each time series segment.

2. The method according to claim 1, characterized in that, Before performing multiple batch parameter tuning on the untrained pressure state analysis model based on the pressure time series sample set, the method further includes: Based on the model architecture configuration scheme of the pressure state analysis model, an untrained signal representation extraction network and a diagnostic feature generation network are constructed, as well as a pre-trained working condition decoding network is obtained. In each training cycle, local signal suppression is performed on the pressure monitoring time series data to obtain locally distorted data of locally shielded random signal segments; Feature extraction is performed on the locally distorted data to obtain signal characterization results; Based on the signal characterization results, the locally masked random signal segments in the locally distorted data are inferred to obtain the inferred signal segments; Based on the error between the obtained inferred signal segment and the random signal segment, the network parameters of the signal representation extraction network are optimized, and the pre-trained signal representation extraction network is output. Based on the pre-trained signal representation extraction network and working condition decoding network, as well as the untrained diagnostic feature generation network, an untrained stress state analysis model is constructed.

3. The method according to claim 1, characterized in that, The process of using the diagnostic feature generation network to convert the obtained stress distribution features into diagnostic semantic features includes: Based on the network parameters of the diagnostic feature generation network in the current training cycle, a first association rule is obtained between each single-dimensional stress distribution feature and each single-dimensional diagnostic semantic feature learned by the diagnostic feature generation network; wherein, each single-dimensional diagnostic semantic feature represents one of the multi-dimensional benchmark stress condition indicators; the multi-dimensional benchmark stress condition indicators include at least the complete set of the multi-dimensional sample stress condition indicators and the multi-dimensional inferred stress condition indicators. According to the first association rule, the pressure distribution features are converted into multiple single-dimensional diagnostic semantic features; wherein, the multiple single-dimensional diagnostic semantic features characterize the multi-dimensional inferred pressure condition index. Multiple single-dimensional diagnostic semantic features are integrated to generate diagnostic semantic features.

4. The method according to claim 1, characterized in that, Before converting the obtained stress distribution features into diagnostic semantic features using the diagnostic feature generation network, the method further includes: Obtain the analysis and processing load during the current training cycle; When it is determined that the analysis and processing load exceeds the load limit, according to the second association rule between each preset feature encoding form and each feature resource consumption, a target feature encoding form whose corresponding feature resource consumption is less than the benchmark resource ratio is determined from each feature encoding form; wherein, the benchmark resource ratio is: the feature resource consumption corresponding to the feature encoding form used by the network parameters of the diagnostic feature generation network; Real-time monitoring of the computational resource consumption required for the current training cycle; Based on the correlation rules between the computing resource consumption, the analysis and processing load, and the resource consumption of the feature, the compression level of the target feature encoding format is dynamically selected. According to the selected compression level, the pressure distribution feature is converted into the target feature encoding form to obtain the encoded pressure distribution feature; wherein, the higher the compression level, the less resources the generated encoded pressure distribution feature occupies.

5. The method according to claim 1, characterized in that, The step of optimizing the network parameters of the diagnostic feature generation network based on the error between the obtained inferred pressure state diagnostic data and the corresponding sample pressure state diagnostic data includes: Based on the error between the multidimensional inferred pressure condition indicators contained in the obtained inferred pressure state diagnostic data and the multidimensional sample pressure condition indicators contained in the corresponding sample pressure state diagnostic data, the diagnostic difference is determined. The state matching error is determined based on the difference between the obtained inferred pressure state diagnostic data and the corresponding sample pressure state diagnostic data. Based on the diagnostic difference and the state matching error, when the diagnostic accuracy criterion is not met, the network parameters of the diagnostic feature generation network are optimized.

6. The method according to claim 1, characterized in that, The diagnostic feature generation network includes a primary diagnostic feature generation subnetwork and a deep diagnostic feature fusion subnetwork. The process of using the diagnostic feature generation network to convert the obtained stress distribution features into diagnostic semantic features includes: The primary diagnostic features are used to generate a sub-network, which converts the obtained pressure distribution features into multiple primary diagnostic semantic feature vectors; wherein each primary diagnostic semantic feature vector represents a preliminary prediction of at least one single-dimensional inferred pressure condition index. The deep diagnostic feature fusion subnetwork is used to fuse multiple primary diagnostic semantic feature vectors and learn the high-order correlation between the multiple primary diagnostic semantic feature vectors to generate the diagnostic semantic features represented by the fused multidimensional inferred pressure condition index.

7. The method according to claim 1, characterized in that, After at least one complete training cycle has been completed, it also includes: Obtain an external pressure monitoring verification dataset; wherein, the external pressure monitoring verification dataset includes mold pressure time series signals and their actual pressure state diagnostic data that were not involved in the construction of the pressure time series sample set; Using the pressure state analysis model obtained through current training, inference is made on the external pressure monitoring verification dataset to obtain the corresponding inferred verification pressure state diagnostic data. Calculate the generalization error between the inferred and verified pressure state diagnostic data and the actual pressure state diagnostic data; Based on the generalization error, the hyperparameters of subsequent training cycles are dynamically adjusted; wherein the hyperparameters include at least one of the following: learning rate, batch size, and training data sample weight distribution.

8. A server system, characterized in that, Includes a server, the server being used to perform the method according to any one of claims 1-7.

Images