A Deep Learning-Based Information Processing Method and System

Abstract

This invention discloses a deep learning-based information processing method and system, relating to the field of industrial production monitoring technology. The method includes: constructing a data tensor for the pressing process on an industrial production line; calculating the feature map drift based on the data tensor; fine-tuning the deep learning model structure based on the feature map drift; identifying redundant structures in the fine-tuned deep learning model, whereby the redundant structures represent attention heads that incur additional computational time due to adjustment without improving model recognition accuracy; correcting the fine-tuned deep learning model based on the redundant structures; and processing the pressing process data based on the corrected deep learning model. This invention combines feature map drift and redundant structures to adjust the deep learning model and process pressing process data, thereby achieving pressing information processing and improving processing efficiency and reliability.

Description

Technical Field

[0001] This invention relates to the field of industrial production monitoring technology, and in particular to an information processing method and system based on deep learning. Background Technology

[0002] In multimodal monitoring of industrial production lines, deep learning technology is typically used to fuse data from high-frequency vibration sensors and infrared thermal imaging cameras to detect abnormal conditions during assembly in real time. The cross-modal attention module within the deep learning model is responsible for establishing a feature space mapping between the temporal patterns of vibration signals and the local high-temperature morphology of thermal images. However, under the long-term continuous operation of the production line, minor wear of the pressing molds and seasonal changes in ambient temperature can cause non-rigid drift in the originally rigidly aligned association patterns between modes. To maintain high recognition accuracy on the drifted data, existing deep learning models, during periodic incremental fine-tuning, often spontaneously allocate more attention heads to resolve the blurred associations and broaden the intermediate layer dimensions to fit more complex decision boundaries. This compensatory deepening phenomenon caused by data drift implicitly increases the computational cost of forward inference in the model. Existing monitoring systems often only focus on the model's accuracy on the validation set, making it difficult to detect the spontaneous expansion of the model's internal structural complexity during fine-tuning. This leads to the gradual accumulation of inference delays, eventually exceeding the strict time limit for real-time decision-making on the production line. This results in the risk of occasional timeouts or the forced use of expired results, leading to low information processing efficiency and low reliability.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0004] The main objective of this invention is to propose a deep learning-based information processing method and system that can adjust the deep learning model by combining feature mapping drift and redundant structure, and perform data processing of the pressing process to achieve pressing information processing, thereby improving processing efficiency and reliability.

[0005] On one hand, embodiments of the present invention provide an information processing method based on deep learning, comprising the following steps: Construct a data tensor for the pressing process on an industrial production line, wherein the data tensor is used to represent a two-dimensional array storing data; Based on the data tensor, the feature mapping drift is calculated, which represents the degree to which the mapping relationship between the vibration feature sequence and the heatmap feature sequence in the deep learning model has drifted. Based on the feature map drift, the deep learning model is fine-tuned, including adjusting the attention head in the cross-modal attention module. Based on the deep learning model after fine-tuning the model structure, redundant structures are identified. These redundant structures represent attention heads that incur additional computational time due to the adjustment but do not improve the model's recognition accuracy. Based on the aforementioned redundant structure, the deep learning model after fine-tuning the model structure is corrected. Data processing for the pressing process is performed based on the revised deep learning model.

[0006] On the other hand, embodiments of the present invention provide an information processing system based on deep learning, comprising: The data acquisition module is used to construct a data tensor for the pressing process on the industrial production line, wherein the data tensor is used to represent a two-dimensional array storing data; The feature mapping drift calculation module is used to calculate the feature mapping drift based on the data tensor. The feature mapping drift represents the degree of drift in the mapping relationship between the vibration feature sequence and the heatmap feature sequence in the deep learning model. The model structure fine-tuning module is used to perform model structure fine-tuning on the deep learning model based on the feature map drift. The model structure fine-tuning includes adjusting the attention head in the cross-modal attention module. The redundant structure identification module is used to identify redundant structures in the deep learning model after fine-tuning the model structure. The redundant structure is used to represent the attention head that causes additional computation time due to the adjustment but does not improve the model's recognition accuracy. The model correction module is used to correct the deep learning model after the model structure fine-tuning process based on the redundant structure. The pressing process data processing module is used to process the pressing process data based on the corrected deep learning model.

[0007] The embodiments of this application include at least the following beneficial effects: First, a data tensor of the pressing process on the industrial production line is constructed. Based on the data tensor, the feature mapping drift is calculated. Then, based on the feature mapping drift, the deep learning model is fine-tuned. Based on the fine-tuned deep learning model, redundant structures are identified. Based on the redundant structures, the fine-tuned deep learning model is corrected. Finally, based on the corrected deep learning model, the pressing process data is processed. Thus, the deep learning model can be adjusted by combining the feature mapping drift and the redundant structures, and the pressing process data can be processed to achieve pressing information processing, thereby improving processing efficiency and reliability.

[0008] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the description and the drawings. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0010] Figure 1 This is a flowchart illustrating an information processing method based on deep learning, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the steps of calculating the feature map drift based on the data tensor in an embodiment of the present invention. Figure 3 This is a schematic diagram of the overall architecture of an information processing method based on deep learning according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an information processing system based on deep learning according to an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0012] In related technologies, multimodal monitoring in industrial production lines typically employs deep learning to fuse data from high-frequency vibration sensors and infrared thermal imaging cameras to detect abnormal states during assembly in real time. The cross-modal attention module within the deep learning model is responsible for establishing a feature space mapping between the temporal patterns of vibration signals and the local high-temperature morphology of thermal images, thus outputting a decision within a millisecond-level time window. However, under the actual conditions of long-term continuous operation on production lines, minor wear of pressing molds and seasonal changes in workshop ambient temperature can cause a slow, non-uniform shift in the noise spectrum of vibration signals and the background temperature distribution of thermal images. This equipment aging and gradual environmental change cause a non-rigid drift in the originally rigidly aligned high-order correlation patterns between the two modes. To maintain high recognition accuracy on the drifted data, when the model performs periodic incremental fine-tuning using new samples, its cross-modal attention module spontaneously allocates more attention heads to resolve the blurred indirect correlations, while the intermediate layer dimensions are widened by fitting a more tortuous decision boundary. This compensatory deepening phenomenon caused by data drift implicitly increases the computational load of the model's forward inference. Existing monitoring systems often focus only on whether the model's accuracy on the validation set meets the standards, failing to detect the spontaneous expansion of the model's internal structural complexity during fine-tuning. As inference latency gradually increases and eventually exceeds the strict time limit for real-time decision-making on the production line, the system faces the risk of occasional timeouts or being forced to use outdated results. This raises a specific scenario-based technical problem: in multimodal monitoring of industrial production lines, when the feature mapping relationship between vibration and thermal imaging modes undergoes non-rigid drift due to equipment wear and changes in ambient temperature, the deep learning model spontaneously deepens the cross-modal fusion structure during fine-tuning to maintain recognition accuracy. This leads to a gradual accumulation of inference latency, compromising the real-time performance and reliability of online anomaly detection.

[0013] The embodiments of this application will be explained in detail below with reference to the accompanying drawings: Figure 1 This is an optional flowchart of an information processing method based on deep learning provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S101 to S106.

[0014] Step S101: Construct a data tensor for the pressing process on the industrial production line. The data tensor is used to represent a two-dimensional array storing data. Step S102: Calculate the feature mapping drift degree based on the data tensor. The feature mapping drift degree is used to represent the degree of drift in the mapping relationship between the vibration feature sequence and the heat map feature sequence in the deep learning model. Step S103: Based on the feature mapping drift, perform model structure fine-tuning on the deep learning model. The model structure fine-tuning includes adjusting the attention head in the cross-modal attention module. Step S104: Fine-tune the deep learning model according to the model structure, identify redundant structures, and use redundant structures to represent attention heads that generate additional computation time due to adjustment without improving the model's recognition accuracy. Step S105: Based on the redundant structure, correct the deep learning model after fine-tuning the model structure; Step S106: Perform data processing for the pressing process based on the corrected deep learning model.

[0015] Steps S101 to S106 as shown in the embodiments of this application can adjust the deep learning model by combining feature mapping drift degree and redundant structure, and perform data processing of the pressing process to realize pressing information processing, thereby improving processing efficiency and reliability.

[0016] In some embodiments, steps S101-S106 can first construct a data tensor for the pressing process on the industrial production line. The data tensor is used to represent a two-dimensional array storing data. Historical operation log files can be manually exported from the production line control system periodically, and a script program can be used to parse the text-type status records in the log files into numerical matrices, which are then concatenated to form a data tensor in the form of a two-dimensional array. To make this implementation feasible, the log file preferably contains at least a portion of the following information: pressing start time, pressing end time, equipment station number, pressing result marker, vibration-related record fields, and temperature-related record fields. If the vibration and temperature information in the log is not the original waveform or original heat map, but rather a summary value that has been preprocessed by the control system, the log can be segmented according to the pressing cycle, and then the fields within the same pressing cycle can be written into the matrix rows in a fixed order. For example, one row can correspond to one pressing cycle, with the first part of the columns recording the vibration-related statistical values ​​within that cycle, and the second part recording the temperature-related statistical values ​​within that cycle. If multiple sub-records exist within the same period, they can be concatenated in chronological order first, and then their length can be standardized by padding, truncation, or segmentation. Although the resulting data tensor is not as detailed as the original sensor signal, it can still serve as a substitute input for offline verification, initial model training, or when online data acquisition is unavailable.

[0017] Then, based on the data tensor, the feature mapping drift is calculated. The feature mapping drift represents the degree of drift in the mapping relationship between the vibration feature sequence and the heatmap feature sequence in the deep learning model. An auxiliary classifier network can be pre-trained. The current data tensor is input into this classifier network, and the degree of drift in the mapping relationship between the vibration feature sequence and the heatmap feature sequence is indirectly evaluated by the fluctuation of the confidence score output by the classifier network. To more clearly illustrate how this auxiliary classifier network is obtained, a period of normal pressing data during the baseline stage and data from subsequent operation stages where drift characteristics have been confirmed can be collected before system deployment. These two types of data are labeled as "baseline association state" and "offset association state," respectively. The auxiliary classifier network is then trained using this labeled data. After training, when a new data tensor is input, if the classifier's confidence in the "baseline association state" continuously decreases, or shows significant fluctuations over multiple consecutive pressing cycles, it indicates a difference between the current input and the association pattern of the baseline stage. To avoid misjudgments caused by a single abnormal fluctuation, the confidence level trend over several consecutive pressing cycles can be smoothed, and then an indirect judgment of the degree of drift can be given based on this. The advantage of this method is that the implementation path is relatively direct, making it suitable for quickly establishing drift monitoring capabilities in the early stages of a system. However, since its output is usually a holistic judgment result, it is more suitable as a preliminary screening method.

[0018] Next, based on the feature map drift, the deep learning model undergoes fine-tuning, including adjusting the attention heads in the cross-modal attention module. A global parameter update strategy can be employed: when the drift exceeds a preset threshold, the entire deep learning model is retrained using a dataset containing the latest data tensor through backpropagation. During this process, the weights of the attention heads in the cross-modal attention module are globally updated. This preset threshold can be determined using historical operating data. Specifically, during a stable operating phase, the range of feature map drift changes over a period of time can be statistically analyzed, and the value representing the upper bound of normal fluctuations can be used as the trigger threshold. When the drift exceeds this threshold for multiple consecutive pressing cycles in subsequent operations, a global parameter update is triggered to reduce the probability of false triggers caused by occasional noise. During model retraining, the latest data tensor can be combined with a portion of baseline stage samples to form a fine-tuning dataset. This prevents the model from becoming overly biased towards the current short-term operating conditions and forgetting the original stable features. During training, the attention heads in the cross-modal attention module redistribute their focus on vibration feature sequences and thermal image feature sequences. For example, some attention heads may be more inclined to capture the correspondence between vibration impact during pressing and local temperature rise, while others may be more inclined to capture the correspondence between low-frequency vibration changes caused by continuous friction and thermal background rise. Although global parameter updates have a large computational cost, they can more completely restore the model's adaptability to new working conditions when drift is significant and the impact has extended to the feature extraction and fusion layers.

[0019] Based on the fine-tuned deep learning model, redundant structures are identified. Redundant structures represent attention heads that incur additional computational time due to adjustments without improving model recognition accuracy. This can be achieved by setting a fixed weight threshold. Specifically, all attention heads in the cross-modal attention module of the fine-tuned deep learning model are iterated through, and the sum of the absolute values ​​of the attention weight matrix of each attention head is extracted. If the sum of the absolute values ​​of the weights of an attention head is lower than the fixed weight threshold, it is directly identified as a redundant structure that incurs additional computational time due to adjustments without improving model recognition accuracy. The fixed weight threshold is preferably derived from statistical results of the model in the baseline stage or historical effective models. Specifically, the weight activity level of each attention head under normal and effective working conditions can be statistically analyzed first, and then a boundary value that can distinguish between "continuously effective participation in computation" and "long-term low contribution to computation" can be selected as the threshold reference. In practice, we can not only observe the sum of the absolute values ​​of the weights after a single fine-tuning, but also repeatedly count the weight activity of the same attention head under several consecutive batches of input. If an attention head is below the threshold in multiple batches and its output variation is consistently small, it is more reliable to classify it as a redundant structure. The reason for this approach is that a single instance of low weight may only be due to the special distribution of a particular batch of samples, while a continuous low-activity state better indicates that the attention head has lost its practical function.

[0020] Based on the redundant structure, the deep learning model after fine-tuning the model structure is corrected. After identifying the redundant structure, the deep learning model can be corrected by forcibly setting the output tensor of the corresponding attention head to zero at the code level, or by directly pruning the computational branch of the attention head in the computation graph. To avoid the correction operation causing sudden impact on the model stability, the "forced zeroing of output tensors" method can be used for verification first. That is, without changing the overall topology of the model, the output of the identified attention head is temporarily hidden, and the recognition results, inference time, and false positives and false negatives of the model on the verification data are observed. If the verification results show that the model performance does not decrease significantly after the attention head is hidden, then computation graph pruning is further performed to remove the corresponding branch from the deployed model. The advantage of this two-step process is that the accuracy of the redundancy judgment is first verified by soft suppression, and then the inference computation is truly reduced by hard pruning, thereby reducing the risk of false pruning.

[0021] Finally, the pressing process data is processed based on the corrected deep learning model. After the correction, the model still maintains the original input and output interfaces when processing real-time pressing process data. Therefore, it can be directly connected to the existing production line judgment process. Real-time pressing process data can be input into the corrected deep learning model, and the pressing quality judgment results can be output, such as pressing qualified, pressing misalignment, pressing incomplete, local overheating, etc.

[0022] It is understandable that a data tensor represents a two-dimensional array storing data. In this embodiment, it is essentially a structured data carrier that organizes multi-source observation results from the same pressing process in a unified order. One dimension of this two-dimensional array can represent a time segment, a sampling segment, or an aligned event sequence, while the other dimension can represent vibration characteristic values, temperature distribution characteristic values, or statistics further extracted from the original signal. For example, within a pressing cycle, a vibration sampling sequence corresponding to the pressing action can be extracted first, and then several thermal image frames corresponding to that time segment can be selected. Several sampling points or statistical features in the vibration sampling sequence and several pixel region features in the thermal image frames are arranged in a predetermined order to form a unified two-dimensional array. The data tensor constructed in this way retains both the dynamic change information of the vibration mode and the spatial temperature distribution information of the thermal mode, facilitating joint analysis by subsequent deep learning models.

[0023] Feature mapping drift represents the degree of shift in the mapping relationship between vibration feature sequences and heatmap feature sequences in a deep learning model. This mapping relationship includes not only the simple temporal correspondence between the two types of data, but also the stable correlation pattern learned by the model after long-term training, where "a certain type of vibration change is usually accompanied by a certain type of heat distribution change." When mold wear, lubrication conditions change, or ambient temperature rises or falls slowly, even if the pressing action itself does not fundamentally change, this stable correlation may shift. For example, the local temperature rise area corresponding to a certain vibration peak may shift in position, weaken in intensity, or experience a time lag. In this case, the correspondence relied upon by the cross-modal fusion layer inside the model is no longer consistent with the baseline state, and feature mapping drift is used to characterize the strength of this shift.

[0024] Redundant structures refer to attention heads that incur additional computational time when adjusted without improving model recognition accuracy. Specifically, some attention heads, although still participating in forward computation after fine-tuning, consistently contribute low output, or their attention patterns highly overlap with other attention heads, or their presence does not improve abnormal recognition results while continuously consuming computational resources and storage bandwidth. These attention heads can be considered redundant structures. Identifying redundant structures does not rely on a single phenomenon but rather on a comprehensive judgment combining weight states, output activity levels, impact on the final recognition result, and degree of overlap with other attention heads.

[0025] This embodiment first constructs a data tensor reflecting the physical state of the current pressing process as the basic input for subsequent analysis. The data tensor is preferably derived from online acquisition signals during the actual operation of the pressing equipment, as online acquisition signals can more completely preserve the transient changes during stages such as pressing, holding, and return. Next, the system does not blindly retrain the model directly, but rather quantitatively assesses the degree of deviation of the current data distribution from the model's original understanding, i.e., calculating the feature mapping drift. This drift reflects both whether the single-modal input itself deviates from the baseline and whether the cross-modal fusion relationship has decoupled or reorganized. When a drift in the mapping relationship is confirmed, model structure fine-tuning is triggered, allowing the model to adaptively adjust its cross-modal attention modules to a certain extent. This adjustment manifests as a reallocation of parameters for existing attention heads, a reassessment of the participation level of different attention heads, or a reorganization of the functional division between attention heads, without requiring a change to the model's external interface. Finally, the model is corrected by eliminating or suppressing these redundant structures, and the corrected lightweight, high-precision model is used to perform actual pressing process data processing. This forms a processing chain of "first detecting drift, then fine-tuning the orientation, then identifying redundancy, and then performing corrections," enabling the model to adapt to gradually changing operating conditions without continuously accumulating inference burden during the adaptation process.

[0026] Through the above technical solution, this embodiment quantifies the feature mapping drift degree and triggers targeted model structure fine-tuning, enabling the model to learn the new mapping relationship after drift and maintain high recognition accuracy. This embodiment introduces a redundant structure identification and correction mechanism after fine-tuning. By accurately locating and processing attention heads generated during fine-tuning that only increase computation time without improving recognition accuracy, the disorderly expansion of the model structure is effectively curbed. This logic of first fine-tuning adaptation and then pruning correction not only ensures the model's adaptability to gradual changes in operating conditions but also fundamentally eliminates the additional computational overhead caused by compensatory deepening. This ensures that the corrected model can strictly meet the millisecond-level real-time decision time limit of the production line when performing data processing during the pressing process, significantly improving the real-time performance and reliability of online anomaly detection. This is because the drift degree calculation step determines whether the model truly needs adjustment, avoiding indiscriminate and frequent fine-tuning; the model structure fine-tuning step determines how to allow the model to readjust to the current operating conditions; and the redundant structure identification and correction step further determines which adaptive changes are merely computational burdens rather than effective capabilities. The three components are interconnected, ensuring that the model will not become inaccurate due to refusing adjustments, nor will it slow down its inference due to excessive adjustments. Therefore, it can maintain stable online judgment capabilities on press-fitting production lines that operate continuously for a long time.

[0027] In some embodiments, the step S101 of constructing the data tensor of the pressing process on the industrial production line may include, but is not limited to, the following steps: Vibration data during the pressing process is collected using vibration sensors; Temperature distribution data during the pressing process was collected using an infrared thermal imager. Time-aligned vibration data and temperature distribution data; A data tensor is constructed based on the time-aligned vibration data and temperature distribution data.

[0028] In some embodiments, vibration data during the pressing process can be collected first using vibration sensors, and temperature distribution data during the pressing process can be collected using an infrared thermal imager. This embodiment mainly focuses on the real-time acquisition and precise synchronization of multimodal physical signals. Vibration data at the moment of pressing can be collected in real time using a high-frequency vibration sensor deployed on the pressing head; this vibration data can be represented as a one-dimensional time-series voltage signal. Simultaneously, temperature distribution data of the pressing area is collected using an infrared thermal imager installed on the side; this temperature distribution data can be represented as a two-dimensional thermal radiation image frame sequence. To enable the acquisition results to be directly used for subsequent modeling, the vibration sensor is preferably fixedly installed at a location that can stably reflect the pressing impact and mechanical transmission vibration, such as the pressing head, mold connection parts, or the surface of structural components closely related to the pressing force transmission path; the infrared thermal imager is preferably arranged facing the pressing contact area or its adjacent heat-affected zone to continuously acquire local temperature rise, heat diffusion, and background temperature changes during the pressing process. Before data acquisition begins, a no-load or standard component press-fit test can be performed to confirm that there is no obvious saturation distortion in the vibration channel, that the thermal image field of view completely covers the target area, and that the timestamp recording function of both types of equipment is normal.

[0029] Then, the vibration data and temperature distribution data are time-aligned. A synchronization mechanism based on hardware trigger signals can be used. For example, when the programmable logic controller of the press machine issues a pressing command, a rising edge trigger level is simultaneously sent to the data acquisition card of the vibration sensor and the infrared thermal imager. Using the moment of receiving this trigger level as the absolute time zero point, and based on the high sampling rate of the vibration sensor and the low frame rate of the infrared thermal imager, high-frequency one-dimensional vibration data segments and low-frequency two-dimensional temperature distribution data frames are bound at the same timestamp through linear interpolation or time window aggregation. To illustrate this process more clearly, the following approach can be used: First, extract a segment of vibration data covering the period before, during, and after pressing, centered on the trigger moment; then, select several frames from the thermal imaging frame sequence that temporally cover the same pressing cycle; subsequently, using the time point of the thermal image frame as an alignment anchor, extract sampling segments within the corresponding time window from the vibration data near each anchor point, and summarize and represent these segments, for example, by extracting peak values, mean values, energy characterization values, or segmented sampling values; finally, bind each thermal image frame and its corresponding vibration segment representation as a multimodal sample at the same time. If the vibration sampling time resolution is much higher than the thermal image frame time resolution, a time window aggregation method is preferred because this method can reduce the weakening of high-frequency impact characteristics by single-point interpolation; if a smoother description of vibration change trends is required, a linear interpolation method can be used to map the vibration characteristics onto the time axis of the thermal image frame.

[0030] Then, based on the time-aligned vibration data and temperature distribution data, a data tensor is constructed. The time-aligned vibration data segments can be flattened and concatenated with the corresponding temperature distribution data pixel matrix to construct a two-dimensional array-like data tensor containing multimodal features. During the flattening process, all original sampling points can be retained, or the vibration segments can be segmented for statistical analysis or feature compression before being concatenated with the target region pixel values, regional mean, or regional temperature difference features in the thermal image, thus balancing information preservation and computational efficiency.

[0031] Through the above technical solution, this embodiment acquires underlying vibration and temperature signals from physical sensors, avoiding information loss from intermediate log records. Simultaneously, strict time alignment eliminates time misalignment caused by different sampling frequencies and transmission delays. This strict multimodal synchronization achieved during the data input stage ensures that the constructed data tensor accurately reflects the physical concurrency relationship between the vibration and thermal modes during the pressing process. This provides high-quality aligned input for the cross-modal attention module in the subsequent deep learning model, significantly reducing the probability of misjudgment due to time misalignment during feature fusion, thereby improving the accuracy of multimodal monitoring. This embodiment not only improves the quality of input data but also directly enhances the reliability of subsequent drift calculations and redundant structure identification.

[0032] In some embodiments, such as Figure 2 As shown, in step S102, the feature map drift is calculated based on the data tensor, which may include, but is not limited to, the following steps: Step S201: Obtain baseline information and state information of network layers in the deep learning model. The state information includes the vibration feature sequence output by the vibration feature extraction network and the heat map feature sequence output by the infrared thermal feature extraction network. The baseline information includes the vibration statistical distribution baseline, the temperature statistical distribution baseline, and the attention weight matrix baseline entropy value. Step S202: Calculate the single-mode dispersion based on the data tensor, vibration statistical distribution baseline, and temperature statistical distribution baseline. The single-mode dispersion includes vibration dispersion and temperature dispersion. Step S203: Calculate the feature map drift based on the single-mode discreteness, vibration feature sequence, heatmap feature sequence, and baseline entropy value of the attention weight matrix.

[0033] In some embodiments, baseline information and state information of network layers in the deep learning model can be obtained first. State information includes the vibration feature sequence output by the vibration feature extraction network and the heatmap feature sequence output by the infrared thermal feature extraction network. Baseline information includes the vibration statistical distribution baseline, temperature statistical distribution baseline, and the baseline entropy value of the attention weight matrix. When obtaining state information, after the current data tensor is input into the deep learning model, the vibration feature sequence output by the forward propagation of the vibration feature extraction network and the heatmap feature sequence output by the infrared thermal feature extraction network are intercepted in real time through hook functions attached to intermediate layers of the model. To ensure that the intercepted results can be used for subsequent comparisons, the hook functions are preferably attached to network layer output positions that are structurally stable, semantically clear, and directly related to cross-modal fusion. For example, the vibration branch can select the output of the last temporal feature encoding layer, and the heatmap branch can select the output of the spatial feature encoding layer before entering the cross-modal attention module. After interception, this state information can be temporarily stored in a cache and associated with the current pressing cycle's time identifier, workstation identifier, and model version information for subsequent traceability and analysis. Understandably, state information refers to the intermediate feature representations activated or output by specific network layers within a deep learning model when processing current input data. For example, the vibration feature sequence output by the vibration feature extraction network can be understood as a temporal representation extracted by the model from the original vibration data; the heat map feature sequence output by the infrared thermal feature extraction network can be understood as a spatial or spatiotemporal representation extracted by the model from the thermal image; and the intermediate attention allocation results in the cross-modal attention module reflect how the model currently establishes the correspondence between the two types of features.

[0034] Then, based on the data tensor, vibration statistical distribution baseline, and temperature statistical distribution baseline, the single-mode dispersion is calculated, which includes vibration dispersion and temperature dispersion. The divergence between the vibration data distribution in the current data tensor and the vibration statistical distribution baseline can be calculated separately as the vibration dispersion; similarly, the divergence between the current temperature data distribution and the temperature statistical distribution baseline can be calculated as the temperature dispersion. This can be implemented as follows: First, separate the vibration-related and temperature-related components from the current data tensor; for the vibration-related component, its amplitude distribution, fluctuation range, local energy concentration, or several frequency band response characteristics can be statistically analyzed, and then compared with the corresponding statistical results at the baseline stage to obtain the vibration dispersion; for the temperature-related component, the background area temperature level, target area temperature rise distribution, hot spot location stability, or regional temperature difference change can be statistically analyzed, and then compared with the temperature statistical distribution baseline to obtain the temperature dispersion. If the current vibration data shows an increase in background noise, a change in the shape of impact peaks, or a shift in frequency band energy compared to the baseline, the vibration dispersion will increase; if the current thermal image shows an overall shift in background temperature, local hot spot diffusion, or a shift in the center of heat distribution compared to the baseline, the temperature dispersion will increase. Through this modal comparison, it is possible to first determine whether the drift mainly comes from the vibration side, the temperature side, or both sides changing simultaneously.

[0035] Then, based on the single-mode dispersion, vibration feature sequence, heatmap feature sequence, and the baseline entropy of the attention weight matrix, the feature mapping drift is calculated. Specifically, the entropy of the real-time attention weight matrix generated between the current vibration feature sequence and the heatmap feature sequence can be calculated, and its difference can be compared with the baseline entropy of the attention weight matrix to obtain the dispersion of the higher-level correlation. The dispersion of the vibration and temperature at the lower level and the dispersion of the higher-level correlation are then weighted and fused to finally output a comprehensive feature mapping drift value. A hierarchical judgment approach can be adopted: if the vibration dispersion is high and the temperature dispersion is low, and the higher-level correlation dispersion is also high, it indicates that the cross-modal mapping shift is mainly caused by changes in the vibration mode; if the temperature dispersion is high and the vibration dispersion is low, and the higher-level correlation dispersion increases, it indicates that the original correspondence is mainly disrupted by changes in the thermal mode background; if the dispersion of both single modes is not high, but the higher-level correlation dispersion increases significantly, it indicates that although the surface statistical features of the single modes have not changed much, the coupling mode between the two has been reorganized, which is an important manifestation of non-rigid drift. During weighted fusion, the weights of each component are preferably set based on historical verification results. That is, whichever type of indicator is more sensitive to the subsequent decline in recognition performance is given a higher fusion weight. If the precise weights cannot be determined temporarily, a relatively balanced fusion method can be used, and then gradually adjusted based on the running results.

[0036] Through the above technical solution, this embodiment delves into the feature extraction and fusion mechanism of deep learning models. First, by calculating the single-modal discreteness, it accurately captures the changes in vibration noise caused by mold wear and the temperature background shift caused by gradual environmental changes. Then, combining deep feature sequences and baseline entropy values, it further quantifies the degree of decoupling of the correlation between these two low-level changes at the cross-modal fusion layer. This hierarchical quantification logic, from the low-level single-modal data distribution to the high-level cross-modal feature mapping, can comprehensively and accurately characterize the true state of non-rigid drift under complex working conditions. Its unique technical effect lies in providing highly interpretable and targeted quantitative indicators for subsequent model structure fine-tuning, avoiding over- or under-tuning due to misjudgment of drift degree, thus ensuring the effectiveness of model correction from the source.

[0037] In some embodiments, step S203, calculating the feature map drift based on the single-mode discreteness, vibration feature sequence, heatmap feature sequence, and attention weight matrix baseline entropy, may include, but is not limited to, the following steps: Step S301: Calculate the attention dispersion increment based on the baseline entropy value of the attention weight matrix. The attention dispersion increment is used to represent the amount of discrete increase in the correlation between the vibration feature sequence and the heat map feature sequence in the cross-modal attention weight matrix. Step S302: Identify the degree of feature change between the vibration feature sequence and the heat map feature sequence in the cross-modal attention weight matrix; Step S303: Weighted summation of single-modal discreteness, attention discreteness increment, and feature change degree to obtain feature mapping drift.

[0038] In some embodiments, the attention dispersion increment can be calculated first based on the baseline entropy value of the attention weight matrix. The attention dispersion increment represents the discrete increase in the correlation between the vibration feature sequence and the heat map feature sequence within the cross-modal attention weight matrix. Specifically, the attention dispersion increment can be calculated by comparing the entropy difference between the current state and the baseline state. This current state can be obtained in real time during the online operation of the equipment. Specifically, vibration signals can be collected by vibration sensors deployed on the pressing equipment, and heat map sequences can be collected by infrared thermal imaging devices deployed above or to the side of the pressing station. The data acquisition module then organizes the two types of data into a data tensor based on a unified timestamp. During the forward inference process of the model, the cross-modal attention weight matrix corresponding to the current sample is directly read from the cross-modal attention module and compared with the baseline entropy value of the attention weight matrix saved in the baseline stage to obtain the attention dispersion increment.

[0039] Then, the degree of feature change between the vibration feature sequence and the heatmap feature sequence in the cross-modal attention weight matrix is ​​identified. Matrix analysis methods can be used to extract the feature vectors of the vibration feature sequence and the heatmap feature sequence before and after the interaction, comparing the changes in their direction and magnitude to quantify the degree of feature change. The feature vectors before and after the interaction can be understood as follows: one set consists of the vibration feature sequence and the heatmap feature sequence before input to the cross-modal attention module, and the other set consists of the corresponding feature representations after cross-modal attention alignment, reweighting, or fusion. By comparing the magnitude of change in these two sets of representations in the dominant correlation direction, the stability of deep feature coupling can be determined.

[0040] The feature mapping drift is obtained by weighting and summing the single-modal dispersion, attention dispersion increment, and feature change degree. The single-modal dispersion (reflecting changes in underlying data), attention dispersion increment (reflecting the degree of association defocusing), and feature change degree (reflecting the shift in feature structure) are weighted and summed according to preset weight coefficients. These weight coefficients can be set based on historical experience. For example, if the production line is more sensitive to fluctuations in underlying signals, the weight of the single-modal dispersion can be appropriately increased; if the production line focuses more on the stability of deep semantic associations, the weights of the latter two can be increased. Furthermore, the preset weight coefficients can be determined through historical samples: first, collect data from a period of stable production and a period of data where gradual drift has been confirmed; compare the feature mapping drift obtained under different weight combinations with the actual defect warning effect; select the weight combination that can issue warnings earlier and has a lower false alarm rate as the parameter configuration for online use, and then solidify this configuration in the edge computing device or industrial control server.

[0041] This embodiment refines the calculation dimension of high-level correlation dispersion. When the pressing condition changes gradually, the originally clear physical correspondence becomes blurred. When allocating attention weights, the model no longer decisively focuses on a few key feature pairs as in the baseline state, but instead distributes the weights evenly across more feature pairs. This dispersion of weights manifests as an increase in dispersion. The degree of feature change focuses on measuring the structural shift of the feature sequence itself during cross-modal interaction. Even if the attention weights are not significantly dispersed, if the content of the features involved in the interaction changes fundamentally, it also means that the mapping relationship has drifted.

[0042] Through the above technical solution, this embodiment deconstructs the complex feature mapping drift phenomenon into three independent yet complementary dimensions: changes in the underlying statistical distribution, loss of focus in the upper layers, and shifts in the structure of the upper layers' features. By quantifying these three dimensions separately and performing weighted fusion, a comprehensive drift evaluation system is constructed. This multi-dimensional evaluation method effectively solves the problem of missed or misjudged results by single indicators. Its unique technical effect lies in its ability to extremely sensitively and accurately capture early, subtle non-rigid drifts caused by equipment aging or gradual environmental changes on industrial production lines. In the above application scenario, when the mold experiences minor wear, the change in the underlying vibration signal may not be obvious, but the structure of the upper-layer features has already begun to shift. At this time, by comprehensively weighting and summing, an accurate drift warning can still be output in a timely manner, thus providing an extremely reliable trigger basis for subsequent model structure fine-tuning and avoiding a precipitous drop in recognition accuracy caused by the model continuously running on incorrect data distributions. Further analysis reveals the following reasons: the degree of single-modal dispersion can quickly reflect the slow shift of visible physical quantities on the sensor side; the increment of attention dispersion can reflect whether the model begins to lose its original focusing ability when establishing cross-modal correspondence; and the degree of feature change can reveal more hidden changes in deep coupling structures. When any one of the three causes an anomaly, the overall result will respond, thus this embodiment has stronger foresight regarding gradual anomalies; when all three change simultaneously, the overall result will be further enhanced, thus possessing better confirmatory capability.

[0043] In some embodiments, in step S202, calculating the single-mode dispersion based on the data tensor, vibration statistical distribution baseline, and temperature statistical distribution baseline may include, but is not limited to, the following steps: Extract time-aligned vibration data and temperature distribution data from the data tensor; Based on the time-aligned vibration data, vibration statistical characteristics are calculated, including the vibration mean and vibration variance. Based on the time-aligned temperature distribution data, calculate the temperature statistical characteristics, which include the temperature mean and temperature variance. The degree of vibration dispersion is calculated based on the vibration statistical characteristics and the vibration statistical distribution baseline. The degree of temperature dispersion is calculated based on the statistical characteristics of temperature and the baseline of temperature statistical distribution.

[0044] In some embodiments, time-aligned vibration data and temperature distribution data can be extracted from the data tensor first. In specific implementations, when a data tensor containing time-aligned vibration data and temperature distribution data is received, the data is first decomposed by mode. Time alignment can be achieved by unifying the sampling clock, timestamp interpolation, or using the start of the pressing stroke as a common reference time. For example, if a vibration sensor continuously outputs vibration sampling points and an infrared thermal imaging device continuously outputs thermal images, the data acquisition controller can merge the two types of data into the same data tensor according to the same pressing cycle number, and perform time interpolation on the thermal images or window resampling on the vibration data to make them correspond on the same time axis.

[0045] Then, based on the time-aligned vibration data, vibration statistical characteristics are calculated, including the vibration mean and vibration variance. Similarly, based on the time-aligned temperature distribution data, temperature statistical characteristics are calculated, including the temperature mean and temperature variance. For vibration data, the arithmetic mean and variance of all sampling points throughout the entire pressing cycle are calculated; for temperature distribution data, the average temperature and temperature variance across all time frames and spatial pixels are calculated.

[0046] Then, based on the vibration statistical characteristics and the vibration statistical distribution baseline, the vibration dispersion is calculated. The calculated vibration average and variance can be compared with the standard mean and standard deviation recorded in the vibration statistical distribution baseline. For example, the vibration dispersion can be obtained by calculating the relative error. If the relative error method is used, the deviation of the current vibration average from the baseline mean and the deviation of the current vibration variance from the baseline variance can be determined, and then the two can be combined according to a preset ratio to obtain the vibration dispersion. Finally, the degree of temperature dispersion is calculated based on the temperature statistical characteristics and the temperature statistical distribution baseline. Using the same calculation method as for vibration dispersion, the temperature statistical characteristics are compared with the temperature statistical distribution baseline to obtain the degree of temperature dispersion.

[0047] Through the above technical solution, this embodiment utilizes the two most basic statistical quantities, mean and variance, to construct a lightweight and robust underlying data distribution monitoring mechanism. The computational complexity of mean and variance is extremely low, making them ideal for high-frequency real-time calculations on resource-constrained industrial edge computing devices. Its unique technical advantage lies in its ability to quickly and stably identify temperature background shifts caused by seasonal changes in workshop ambient temperature, and vibration noise increases caused by minor wear of pressing molds, with extremely low computational overhead. This method of calculating dispersion based on fundamental statistical characteristics not only avoids the computational delays caused by complex feature extraction but also greatly improves the tracking stability of gradual physical anomalies, providing solid and reliable underlying data support for subsequent comprehensive drift calculations. This is because seasonal environmental changes and slow equipment wear typically manifest first as slow shifts in long-term statistics, rather than sudden waveform distortions. Therefore, continuous monitoring using fundamental statistical characteristics makes it easier to form smooth and interpretable trend curves. In this way, the system can make judgments based on the dispersion changes over multiple consecutive periods, rather than being misled by a single abnormal peak.

[0048] In some embodiments, step S301, calculating the attention dispersion increment based on the baseline entropy value of the attention weight matrix, may include, but is not limited to, the following steps: Obtain the cross-modal attention weight matrix of the cross-modal attention module in the network layer. The cross-modal attention weight matrix is ​​used to represent the correlation between the vibration feature sequence and the heat map feature sequence. Calculate the two-dimensional information entropy of the cross-modal attention weight matrix; The attention dispersion increment is calculated based on the two-dimensional information entropy and the baseline entropy value of the attention weight matrix.

[0049] In some embodiments, the cross-modal attention weight matrix of the cross-modal attention module in the network layer can be obtained first. The cross-modal attention weight matrix is ​​used to represent the correlation between the vibration feature sequence and the heatmap feature sequence. Specifically, during the forward inference process of the deep learning model, the current cross-modal attention weight matrix is ​​extracted from the cross-modal attention module. To ensure consistent extraction positions, the output of the cross-modal attention module of the specified network layer can be pre-read in the model deployment file, and the matrix can be written to the buffer after each loading cycle. Since the elements in this matrix have usually been normalized (e.g., through the Softmax function), and the sum of all its elements is 1, it can be directly regarded as a two-dimensional probability distribution matrix.

[0050] Then, the two-dimensional information entropy of the cross-modal attention weight matrix is ​​calculated. This can be achieved by using the formula for information entropy, multiplying each element of the matrix by its natural logarithm, summing all products, and taking the negative of the summation. Considering that complex mathematical processes should not be directly exposed in an online implementation, a pre-built information entropy calculation module can be called within the software. This module iterates through the matrix cells, reads each weight value, performs numerical stabilization on the minimum values, accumulates their information contribution, and finally outputs the two-dimensional information entropy.

[0051] Next, based on the two-dimensional information entropy and the baseline entropy value of the attention weight matrix, the attention dispersion increment is calculated. The calculated two-dimensional information entropy can be subtracted from the pre-saved baseline entropy value of the attention weight matrix. If the difference is positive, it is the attention dispersion increment; if the difference is negative or close to zero, it indicates that the current concentration of the attention distribution is comparable to the baseline state, and the dispersion increment can be recorded as zero. The purpose of truncating negative values ​​to zero is to ensure that drift assessment focuses more on "whether the dispersion has increased" rather than misjudging occasional increased concentration as a reverse anomaly. For ease of implementation, a lightweight judgment logic can be set in the inference service: when the current entropy value is higher than the baseline entropy value and maintains an increasing trend for multiple consecutive periods, the attention dispersion increment is considered valid, thereby further reducing the risk of false triggering caused by single-period noise fluctuations.

[0052] This embodiment introduces two-dimensional information entropy as the core indicator for quantifying the dispersion of attention weights. Information entropy, in information theory, measures the uncertainty or disorder of a system. When the weight values ​​in the cross-modal attention weight matrix are highly concentrated on a few elements (i.e., the model is very confident that certain vibration features are strongly correlated with certain heatmap features), its two-dimensional information entropy is low; conversely, when the weight values ​​tend to be evenly distributed (i.e., the model is confused about the correlation between features and cannot accurately locate the correspondence), its two-dimensional information entropy is high. The baseline entropy value of the attention weight matrix can be obtained as follows: run the trained cross-modal model under stable operating conditions, continuously inputting data tensors from multiple normal pressing cycles; extract the cross-modal attention weight matrix at the same network layer and the same cross-modal attention module for each cycle, and calculate the corresponding two-dimensional information entropy; then statistically summarize these two-dimensional information entropies to obtain the baseline entropy value representing normal operating conditions. To improve robustness, the baseline entropy value can be the average result of multiple normal cycles, or an allowable fluctuation range can be saved simultaneously to distinguish between normal fluctuations and abnormal diffusion. The baseline entropy value obtained in this way is matched with the actual production line, specific model structure, and specific sensor deployment method, and is highly targeted.

[0053] Through the above technical solution, this embodiment abstracts the complex cross-modal association state into a two-dimensional probability distribution and uses information entropy, a rigorous mathematical tool, to quantify its disorder level. By comparing the current entropy value with the baseline entropy value, the evolution trajectory of the mapping relationship within the model can be clearly revealed. Its unique technical effect lies in the fact that when equipment aging causes the originally rigidly aligned high-order association modes to drift non-rigidly, the model often spontaneously attempts to find alternative associations in a wider feature space, resulting in attention weight diffusion. By calculating the two-dimensional information entropy and obtaining the discreteness increment, this hidden change in the model's internal state can be captured with extreme sensitivity. This quantification method is not only simple and efficient in calculation, but also has strong theoretical interpretability, making the evaluation of high-level association defocusing phenomena no longer dependent on subjective experience, but based on a solid mathematical foundation.

[0054] In some embodiments, step S302, identifying the degree of feature change between the vibration feature sequence and the heatmap feature sequence in the cross-modal attention weight matrix, may include, but is not limited to, the following steps: Calculate the cross-covariance matrix between the vibration feature sequence and the heat map feature sequence. The cross-covariance matrix is ​​used to measure the linear correlation between the vibration feature sequence and the heat map feature sequence. Analyze the distribution changes of singular values ​​in the cross-covariance matrix; Based on the changes in singular value distribution, the degree of feature change of vibration feature sequences and heatmap feature sequences in the cross-modal attention weight matrix is ​​identified.

[0055] In some embodiments, the cross-covariance matrix between the vibration feature sequence and the heatmap feature sequence can be calculated first. The cross-covariance matrix is ​​used to measure the linear correlation between the vibration feature sequence and the heatmap feature sequence. In a specific implementation, the vibration feature sequence and the heatmap feature sequence participating in the cross-modal attention calculation are first obtained, centered, and then the cross-covariance matrix of the two is calculated. Centering refers to removing the average level of each feature dimension in the current sample, so that subsequent analysis focuses more on the changing relationship rather than the absolute bias. For example, the vibration features within a pressing cycle can be arranged into a sequence by time steps, and the heatmap features within the corresponding cycle can be arranged into a sequence by time frames or regional segments. Within the same batch of samples, their respective means are removed, and the cross-covariance matrix is ​​calculated.

[0056] Then, the distribution changes of singular values ​​in the cross-covariance matrix are analyzed. Singular value decomposition (SVD) can be performed on the cross-covariance matrix to obtain a series of singular values ​​arranged from largest to smallest. Based on the changes in the singular value distribution, the degree of feature change of the vibration feature sequence and the heatmap feature sequence in the cross-modal attention weight matrix can be identified. The currently calculated singular value sequence can be compared with the standard singular value sequence under baseline conditions. The standard singular value sequence can be pre-established before model deployment: under stable operating conditions, vibration feature sequences and heatmap feature sequences from multiple normal pressing cycles are collected. The cross-covariance matrix and singular value sequence are calculated for each cycle, and then these singular value sequences are statistically summarized to obtain the baseline reference distribution. During the online phase, the current singular value sequence calculated for each new cycle can be compared from three perspectives. For example, the cosine similarity and relative entropy between two singular value sequences can be calculated, or the energy proportion changes of the top K largest singular values ​​can be directly compared. If the current singular value distribution becomes flatter than the baseline distribution, or the maximum singular value decreases significantly, it indicates that the previously strong correlation in a specific dimension is weakening. Based on this, a specific numerical value of the degree of feature change can be quantified and output.

[0057] This embodiment utilizes cross-covariance matrix (CCM) and singular value decomposition (SVD) techniques to delve into the deep structural coupling relationships between cross-modal features. The CCM comprehensively characterizes the linear correlation between two different modal feature sequences across all dimensions. SVD extracts singular vectors representing the main correlation directions and singular values ​​representing the correlation strength from the CCM. The distribution of singular values ​​(e.g., the size of the largest singular value, the energy percentage of the first few singular values) directly reflects the dominant coupling mode of the two modal features in the fusion space. Significant changes in the singular value distribution indicate a fundamental structural shift in the coupling mode between the features. To more clearly and completely describe the data sources involved in the computation, the vibration feature sequence and heatmap feature sequence can be taken from either the input or output of the cross-modal attention module, as long as consistency is maintained between the baseline and online phases. A more suitable approach is to consistently take the input features from the cross-modal attention module within the same network layer, as this location has not yet been directly affected by the current attention reweighting result, making it easier to analyze the structural relationships between the original interactive objects; if more attention is paid to the coupling stability after fusion, the input features can also be consistently taken from the output features.

[0058] Through the above technical solution, this embodiment transcends the limitation of merely focusing on the surface numerical values ​​of attention weights, delving into the linear algebraic space of feature vectors and utilizing cross-covariance and singular value decomposition to reveal the essential structure of cross-modal feature interactions. Changes in the singular value distribution can accurately reflect complex transformations such as rotation, stretching, or compression of the feature subspace. Its unique technical effect lies in the fact that when the gradual change in the pressing condition leads to a deep structural shift in the correlation pattern between vibration and heatmap (e.g., the correlation center shifts from one frequency band to another), even if the concentration of attention weights (i.e., information entropy) has not changed significantly, this embodiment can still detect this hidden feature structure shift in advance through subtle changes in the singular value distribution. This deep feature change recognition mechanism greatly enhances the comprehensiveness and foresight of drift degree calculation, ensuring that the model can still make accurate adaptive adjustment decisions when facing extremely complex non-rigid drifts. This is because the singular value distribution describes "which directions are most important and whether the dominance of these directions has changed" in the cross-modal correlation structure. As long as the dominant direction shifts, even if the surface attention remains focused, the singular value structure will undergo subtle changes first. Therefore, this embodiment can detect deep structural shifts early. The system can not only know whether the model is out of focus, but also whether the structure the model is focused on has changed, thus enabling the final output feature map drift to have both breadth and depth.

[0059] In some embodiments, step S103, which involves fine-tuning the model structure of the deep learning model based on the feature map drift, may include, but is not limited to, the following steps: Based on the feature mapping drift degree and the preset mapping drift threshold, a structural fine-tuning signal is generated; In response to the structural fine-tuning signal, drift sample data is extracted from the data tensor. The drift sample data is used to represent the data that causes the mapping relationship between the vibration feature sequence and the heat map feature sequence in the deep learning model to drift. Obtain the product label sequence generated by the production line quality inspection module. The product label sequence is used to reflect the product quality of the press-fitted workpiece corresponding to the drift sample data. Based on the drift sample data and product label sequences, a training dataset and a validation dataset are constructed. Based on the training and validation datasets, the gradient descent optimization algorithm is used to fine-tune the model structure of the deep learning model.

[0060] In some embodiments, a structural fine-tuning signal can be generated first based on the feature map drift degree and a preset map drift threshold. The feature map drift degree calculated in real time can be compared with the preset map drift threshold. This preset map drift threshold can be set based on historical operating experience, for example, selecting the upper limit of drift degree fluctuation under normal operating conditions. When the feature map drift degree exceeds this threshold for multiple consecutive pressing cycles, the system determines that a substantial deviation has occurred in the current operating condition and then generates a structural fine-tuning signal. This threshold-based triggering mechanism effectively avoids invalid fine-tuning caused by single data noise or occasional interference.

[0061] Then, in response to the structural fine-tuning signal, drift sample data is extracted from the data tensor. This drift sample data represents the data that causes a shift in the mapping relationship between the vibration feature sequence and the heat map feature sequence in the deep learning model. After generating the structural fine-tuning signal, the system automatically extracts the drift sample data that triggered the signal from the recently cached data tensor. These drift sample data accurately record the vibration and temperature distribution characteristics that are currently shifting. Simultaneously, the system acquires the product label sequence generated by the production line quality inspection module. This product label sequence reflects the product quality of the press-fitted workpiece corresponding to the drift sample data. To ensure the accuracy of the model output after fine-tuning, input data alone is insufficient; corresponding real physical results are also required. Therefore, the system synchronously interfaces with the production line quality inspection module to obtain the product label sequence corresponding to these drift sample data. The production line quality inspection module can be an automated dimensional measuring instrument, a pull-out force testing machine, or a manual quality inspection station. The product label sequence it generates includes clear quality judgment results such as press-fitting qualified, press-fitting misaligned, and press-fitting incomplete.

[0062] Then, based on the drift sample data and product label sequences, training and validation datasets are constructed. The extracted drift sample data can be paired with corresponding product label sequences and divided into training and validation datasets according to a preset ratio (e.g., 8:2). To prevent significant forgetting during model fine-tuning, normal historical samples from the baseline phase can be mixed into the training dataset at a certain proportion.

[0063] Finally, based on the training and validation datasets, the gradient descent optimization algorithm is used to fine-tune the deep learning model structure. The constructed dataset can be used to train the deep learning model through backpropagation using gradient descent optimization algorithms (such as Adam or SGD optimizers). During fine-tuning, the attention head in the cross-modal attention module adaptively adjusts its weight parameters according to the new data distribution and the actual quality inspection labels, re-establishing an effective mapping relationship between the vibration feature sequence and the heatmap feature sequence.

[0064] Through the above technical solution, this embodiment constructs a closed-loop fine-tuning mechanism based on real quality inspection feedback. By setting a threshold to generate structural fine-tuning signals, on-demand triggering of model updates is achieved, saving computational resources. By introducing product label sequences generated by the production line quality inspection module, reliable supervision signals are provided for model fine-tuning, ensuring that the model does not deviate from actual physical quality standards when adapting to new working conditions. This supervised, targeted fine-tuning method enables the model to quickly and accurately learn the new mapping relationship after drift, thereby maintaining a high level of anomaly identification accuracy in complex and ever-changing industrial environments.

[0065] In some embodiments, step S104, identifying redundant structures based on the deep learning model after model structure fine-tuning, may include, but is not limited to, the following steps: Perform a masking operation on the adjusted attention head in the cross-modal attention module. The masking operation is used to indicate that the output of the attention head is set to zero. After performing the masking operation, the reduction in anomaly detection of the deep learning model is obtained. The reduction in anomaly detection is used to represent the reduction in the number of products whose outputs in the deep learning model are inconsistent with the product labels in the verification dataset. The increase in computation time of the deep learning model after fine-tuning the model structure; Calculate the ratio of the reduction in anomaly detection to the increase in computation time; Redundant structures are identified based on the ratio.

[0066] In some embodiments, a masking operation can be performed on the adjusted attention heads in the cross-modal attention module. The masking operation represents setting the output of the attention head to zero. Specifically, the system iterates through each fine-tuned attention head in the cross-modal attention module and performs the masking operation on each one. This masking operation does not change the physical network structure of the model; rather, at the code level of forward inference, it forces the output tensor of the currently examined attention head to be set to zero by multiplying it by an all-zero matrix or by directly blocking the data flow.

[0067] After performing the masking operation, the reduction in anomaly detection by the deep learning model is obtained. This reduction represents the decrease in the number of instances where the deep learning model's output in the validation dataset is inconsistent with the product labels. After masking a specific attention head, the system inputs the validation dataset into the model, counts the number of instances where the model's output is inconsistent with the actual product labels (i.e., the number of incorrectly identified samples), and compares this number with the number of incorrect samples before masking to obtain the anomaly detection reduction. It should be noted that if masking a specific attention head results in a decrease or no change in the number of incorrectly identified samples (i.e., a reduction greater than or equal to zero), it indicates that the attention head does not contribute positively to the model's correct identification and may even introduce interference noise.

[0068] The increase in computation time of the deep learning model after fine-tuning the model structure is then collected. Performance analysis tools can be used to collect the increase in computation time of the attention head after fine-tuning compared to the baseline state. This increase in computation time can be reflected in the additional milliseconds or floating-point operations consumed by the attention head during forward inference.

[0069] Finally, the ratio of the reduction in anomaly detection to the increase in computation time is calculated. Based on this ratio, redundant structures are identified. This ratio essentially quantifies the "cost-effectiveness" or "performance gain" of the attention head. If this ratio is extremely low (e.g., below a preset performance threshold), or even negative, it means that the attention head consumes a lot of computational resources without improving detection accuracy, and may even drag down overall performance. Based on this ratio evaluation result, the system can accurately identify these inefficient attention heads as redundant structures.

[0070] Through the above technical solution, this embodiment provides an extremely safe and quantitatively accurate method for locating redundant structures. By using masking operations, the system can verify the true contribution of each attention head without disrupting the overall model topology, avoiding irreversible damage that might result from direct physical pruning. Simultaneously, by comprehensively considering the reduction in anomaly detection and the increase in computation time, this embodiment combines the two mutually constraining indicators of model accuracy and inference latency, achieving a two-dimensional evaluation of attention head performance. This allows the system to accurately distinguish which attention heads are truly necessary structures for adapting to drift and which are merely ineffective expansions generated during fine-tuning. This provides a solid and reliable quantitative basis for subsequent model correction, ensuring that the corrected model maintains high accuracy while strictly meeting the millisecond-level real-time requirements of industrial production lines.

[0071] Figure 3This is a schematic diagram of the overall architecture of an information processing method based on deep learning provided in this application embodiment. First, data from the pressing process on an industrial production line is acquired as input. Then, by constructing a data tensor and calculating the feature map drift, the deep learning model undergoes fine-tuning of its structure. Next, redundant structures in the fine-tuned model are identified, and the deep learning model is corrected. Finally, the pressing process data is processed, and the data processing results are output, thereby realizing pressing information processing and improving processing efficiency and reliability.

[0072] The beneficial effects of implementing the embodiments of the present invention include: First, the embodiments of this application construct a data tensor of the pressing process on an industrial production line. Based on the data tensor, the feature mapping drift is calculated. Then, based on the feature mapping drift, the deep learning model is fine-tuned. Based on the fine-tuned deep learning model, redundant structures are identified. Based on the redundant structures, the fine-tuned deep learning model is corrected. Finally, based on the corrected deep learning model, the pressing process data is processed. Thus, the deep learning model can be adjusted by combining the feature mapping drift and the redundant structures, and the pressing process data can be processed to achieve pressing information processing, thereby improving processing efficiency and reliability.

[0073] like Figure 4 As shown, this embodiment of the invention also provides an information processing system based on deep learning, comprising: The data acquisition module 401 is used to construct a data tensor for the pressing process on the industrial production line. The data tensor is used to represent a two-dimensional array that stores data. The feature mapping drift calculation module 402 is used to calculate the feature mapping drift based on the data tensor. The feature mapping drift is used to represent the degree of drift in the mapping relationship between the vibration feature sequence and the heat map feature sequence in the deep learning model. The model structure fine-tuning processing module 403 is used to perform model structure fine-tuning processing on the deep learning model based on the feature map drift degree. The model structure fine-tuning processing includes adjusting the attention head in the cross-modal attention module. The redundant structure identification module 404 is used to identify redundant structures in the deep learning model after fine-tuning the model structure. Redundant structures are used to represent attention heads that generate additional computation time due to adjustment but do not improve the model's recognition accuracy. The model correction module 405 is used to correct the deep learning model after the model structure has been fine-tuned based on the redundant structure. The pressing process data processing module 406 is used to process the pressing process data based on the corrected deep learning model.

[0074] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0075] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

Claims

1. An information processing method based on deep learning, characterized in that, Includes the following steps: Construct a data tensor for the pressing process on an industrial production line, wherein the data tensor is used to represent a two-dimensional array storing data; Based on the data tensor, the feature mapping drift is calculated, which represents the degree to which the mapping relationship between the vibration feature sequence and the heatmap feature sequence in the deep learning model has drifted. Based on the feature map drift, the deep learning model is fine-tuned, including adjusting the attention head in the cross-modal attention module. Based on the deep learning model after fine-tuning the model structure, redundant structures are identified. These redundant structures represent attention heads that incur additional computational time due to the adjustment but do not improve the model's recognition accuracy. Based on the aforementioned redundant structure, the deep learning model after fine-tuning the model structure is corrected. Data processing for the pressing process is performed based on the revised deep learning model.

2. The method according to claim 1, characterized in that, The data tensor for constructing the pressing process on the industrial production line includes: Vibration data during the pressing process is collected using vibration sensors; Temperature distribution data during the pressing process was collected using an infrared thermal imager. The vibration data and the temperature distribution data are time-aligned. The data tensor is constructed based on the time-aligned vibration data and temperature distribution data.

3. The method according to claim 1, characterized in that, The step of calculating the feature map drift based on the data tensor includes: Obtain baseline information and state information of network layers in the deep learning model. The state information includes vibration feature sequence output by the vibration feature extraction network and heat map feature sequence output by the infrared thermal feature extraction network. The baseline information includes vibration statistical distribution baseline, temperature statistical distribution baseline and attention weight matrix baseline entropy value. Based on the data tensor, the vibration statistical distribution baseline, and the temperature statistical distribution baseline, the single-mode dispersion is calculated, wherein the single-mode dispersion includes vibration dispersion and temperature dispersion. The feature map drift is calculated based on the single-mode discreteness, the vibration feature sequence, the heatmap feature sequence, and the baseline entropy of the attention weight matrix.

4. The method according to claim 3, characterized in that, The step of calculating the feature map drift based on the single-modal discreteness, the vibration feature sequence, the heatmap feature sequence, and the baseline entropy value of the attention weight matrix includes: Based on the baseline entropy value of the attention weight matrix, the attention dispersion increment is calculated. The attention dispersion increment is used to represent the amount by which the correlation between the vibration feature sequence and the heatmap feature sequence increases discretely in the cross-modal attention weight matrix. Identify the degree of feature variation between the vibration feature sequence and the heatmap feature sequence in the cross-modal attention weight matrix; The feature map drift is obtained by weighted summing of the single-modal discreteness, the attention discreteness increment, and the feature change degree.

5. The method according to claim 3, characterized in that, The step of calculating the single-mode discreteness based on the data tensor, the vibration statistical distribution baseline, and the temperature statistical distribution baseline includes: Extract time-aligned vibration data and temperature distribution data from the data tensor; Based on the time-aligned vibration data, vibration statistical characteristics are calculated, including the vibration mean and vibration variance. Based on the time-aligned temperature distribution data, temperature statistical characteristics are calculated, including the average temperature and temperature variance. The vibration dispersion is calculated based on the vibration statistical characteristics and the vibration statistical distribution baseline. The degree of temperature dispersion is calculated based on the temperature statistical characteristics and the temperature statistical distribution baseline.

6. The method according to claim 4, characterized in that, The step of calculating the attention dispersion increment based on the baseline entropy value of the attention weight matrix includes: Obtain the cross-modal attention weight matrix of the cross-modal attention module in the network layer. The cross-modal attention weight matrix is ​​used to represent the correlation between the vibration feature sequence and the heat map feature sequence. Calculate the two-dimensional information entropy of the cross-modal attention weight matrix; The attention dispersion increment is calculated based on the two-dimensional information entropy and the baseline entropy value of the attention weight matrix.

7. The method according to claim 4, characterized in that, The identification of the degree of feature change between the vibration feature sequence and the heatmap feature sequence in the cross-modal attention weight matrix includes: Calculate the cross-covariance matrix between the vibration feature sequence and the heat map feature sequence, the cross-covariance matrix being used to measure the linear correlation between the vibration feature sequence and the heat map feature sequence; Analyze the distribution variation of the singular values ​​of the cross-covariance matrix; Based on the changes in the singular value distribution, the degree of feature change between the vibration feature sequence and the heatmap feature sequence in the cross-modal attention weight matrix is ​​identified.

8. The method according to claim 1, characterized in that, The step of fine-tuning the deep learning model structure based on the feature mapping drift includes: Based on the feature mapping drift degree and the preset mapping drift threshold, a structural fine-tuning signal is generated; In response to the structural fine-tuning signal, drift sample data is extracted from the data tensor, the drift sample data being used to represent data that causes a drift in the mapping relationship between the vibration feature sequence and the heatmap feature sequence in the deep learning model; Obtain the product label sequence generated by the production line quality inspection module. The product label sequence is used to reflect the product quality of the press-fitted workpiece corresponding to the drift sample data. Based on the drift sample data and the product label sequence, a training dataset and a validation dataset are constructed; Based on the training dataset and the validation dataset, the gradient descent optimization algorithm is used to fine-tune the model structure of the deep learning model.

9. The method according to claim 8, characterized in that, The process of identifying redundant structures in the deep learning model after fine-tuning based on the model structure includes: A masking operation is performed on the adjusted attention head in the cross-modal attention module, wherein the masking operation is used to indicate that the output of the attention head is set to zero; After performing the masking operation, the reduction in anomaly detection of the deep learning model is obtained. The reduction in anomaly detection is used to represent the reduction in the number of times the output of the deep learning model in the verification dataset is inconsistent with the product label. The increase in computation time of the deep learning model after fine-tuning the model structure; Calculate the ratio of the reduction in anomaly detection to the increase in calculation time; The redundant structure is identified based on the ratio.

10. An information processing system based on deep learning, characterized in that, include: The data acquisition module is used to construct a data tensor for the pressing process on the industrial production line, wherein the data tensor is used to represent a two-dimensional array storing data; The feature mapping drift calculation module is used to calculate the feature mapping drift based on the data tensor. The feature mapping drift represents the degree of drift in the mapping relationship between the vibration feature sequence and the heatmap feature sequence in the deep learning model. The model structure fine-tuning module is used to perform model structure fine-tuning on the deep learning model based on the feature map drift. The model structure fine-tuning includes adjusting the attention head in the cross-modal attention module. The redundant structure identification module is used to identify redundant structures in the deep learning model after fine-tuning the model structure. The redundant structure is used to represent the attention head that causes additional computation time due to the adjustment but does not improve the model's recognition accuracy. The model correction module is used to correct the deep learning model after the model structure fine-tuning process based on the redundant structure. The pressing process data processing module is used to process the pressing process data based on the corrected deep learning model.

Images