Multi-mode printer state prediction method and system

Abstract

The invention discloses a multi-mode printer state prediction method and system, and relates to the technical field of state prediction. According to the method, a state evolution trend analysis mechanism is introduced, state fragments with risks can be recognized in advance before obvious faults occur, early warning and trend guiding of key parameter changes in the printing task process are achieved, and the risk of printing quality reduction or equipment damage caused by sudden state change is effectively reduced. Meanwhile, a condition modeling and state clustering mode driven by multi-modal data is adopted, the adaptability and generalization ability of the prediction model to complex operation working conditions are improved, and the method is particularly suitable for actual industrial environments such as frequent switching of printing tasks and changeable equipment operation states.

Description

Technical Field

[0001] The present invention relates to the technical field of state prediction, and in particular to a method and system for predicting the state of a multi-modal printer. Background Art

[0002] With the continuous evolution of smart manufacturing and industrial Internet of Things technologies, printing equipment has gradually transformed from traditional single functions to multimodal intelligent systems. Accurate prediction of equipment operating status and abnormal warning have become key links in improving the stability of printing tasks and printing quality. Especially in the context of increasingly complex printer structures and dynamically changing working environments, it is difficult to fully control the operating status of the equipment by relying solely on a single data dimension (such as temperature, voltage, number of printed pages, etc.). Therefore, the need to integrate multimodal data (structural sensor signals, logs, printing loads, etc.) and realize state trend prediction based on intelligent algorithms is becoming increasingly prominent. However, existing technologies mostly focus on the evaluation of the current state of the equipment or the correction of abnormalities based on static models. There is still a lack of a systematic method that can identify trends, identify potential abnormal fragments, and predict state evolution for dynamic evolution processes at different stages. Especially in scenarios such as cold starts and frequent mode switching, there are still great challenges in accurately modeling the laws of equipment state evolution.

[0003] CN118963684B discloses a printer state correction method and system. This method establishes a theoretical aging curve model between aging parameters and theoretical data curves, grading the actual printing effect level. Based on the curve change level and aging level, an attenuation coefficient is determined to correct the current printer state and trigger the state correction process. Although this method can compensate for the aging status of multiple sensors to a certain extent, thereby improving the accuracy of the correction coefficient, its core is still based on the fitting and correction of a single indicator based on a static aging model. It fails to capture the complex dynamic correlations between multimodal data. In particular, it lacks a response mechanism for data trend changes and the ability to predict future state evolution trends. This makes it difficult to adapt to complex operating scenarios where the device frequently switches between modes or where early potential anomalies gradually accumulate.

[0004] CN119766019A proposes a model predictive control method and system for an industrial 3D printer. By constructing a model predictive controller (MPC), it combines motor state information to output an optimal voltage vector in real time to control the motor's operating state, improving motor response speed and printing accuracy. This method provides an optimized path control strategy for issues such as path tracking and torque control during 3D printing. However, its core focus is on motor-level motion control optimization, primarily addressing hardware response issues caused by the complexity of the printing path. It does not involve modeling and predicting the evolution of the printer's overall system state, nor does it address how to mine potential abnormal states or trend changes from multimodal data. Its support for the full-cycle state transition and prediction of the device in complex scenarios is relatively limited.

[0005] In summary, existing printer state correction or control methods mostly focus on single modal data modeling or local control optimization of unit hardware, and have not yet formed a complete framework for fusing multimodal data and establishing an evolution trend prediction model. In particular, there is a lack of accurate state identification and prediction mechanisms under cold start and sudden modal fluctuation conditions. Summary of the Invention

[0006] The present invention is proposed in view of the problems of correction or control methods of existing printers under cold start and sudden modal fluctuation conditions.

[0007] Therefore, the problem to be solved by the present invention is how to realize the dynamic evolution prediction of the printing device status and early identification of potential faults based on the conditional segmentation modeling and trend prediction mechanism of the feature vector, combined with the abnormal fragment time series clustering and state-level transition sequence generation technology.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0009] In a first aspect, the present invention provides a multimodal printer state prediction method, which includes obtaining a multimodal data stream set, determining whether it is in a cold start stage, and outputting a stage label; conditionally slicing the data stream set according to the stage label, and calculating a feature vector for each slice; constructing a local polynomial regression model for each slice, and mapping the trend vector of key indicators in the next 3 to 5 cycles through each feature vector; if the trend mutation and modal fluctuation conditions are met, marking the current slice as a potential abnormal fragment; clustering and sorting the potential abnormal fragment set in the time dimension, generating a state-level transition sequence through inter-segment transfer association, and outputting a state prediction result based on trend consistency; based on the comparison of the state prediction result with the subsequent true state label, feedback is used to adjust the weights of each regression model.

[0010] As a preferred embodiment of the multimodal printer state prediction method described in the present invention, the judgment of the cold start phase includes: constructing a window for the first k time slices in the multimodal data stream set, extracting a set of modal feature components, and calculating the heteroscedasticity for each modal feature to form a characteristic heteroscedasticity vector, wherein k is a constant; calculating the fluctuation density in the current window based on the characteristic heteroscedasticity vector, and calculating the ratio of the fluctuation density mean stored in the historical stable phase as a cold start discrimination factor; if the ratio is greater than or equal to the density offset threshold, the output phase label is a cold start; otherwise, the output phase label is a stable phase.

[0011] As a preferred solution of the multimodal printer state prediction method described in the present invention, the division of the conditional slices includes: matching the stage labels with the multimodal data stream sets, and for each stage label, extracting the data stream sequence generated in the corresponding stage; based on the time sequence between data points, dividing the data stream sequence into several continuous data segments, each segment length is a fixed period; in each segment, according to whether the print head temperature exceeds the upper and lower floating range of the corresponding stage mean, the data points in the segment are re-sliced ​​to generate a conditional slice set.

[0012] As a preferred solution of the multimodal printer state prediction method of the present invention, the construction of the local polynomial regression model includes: for any feature vector, retrieving the last T in the corresponding condition slice e A sequence of key indicators per unit time; wherein the key indicators include but are not limited to the modal values ​​of the print head temperature and feed rate; a weight sequence is introduced on the key indicator sequence, wherein each weight is dynamically calculated based on the modal variance and historical mean of the data point, so that high fluctuation points account for a higher proportion in the fitting, and a weight vectorized trend sequence s is constructed. k ; Based on s k The constructed key indicator sequence adopts a polynomial fitting modeling method of no more than 3rd order, and independently constructs a local prediction expression for each modal indicator.

[0013] As a preferred solution of the multi-modal printer state prediction method of the present invention, the confirmation of the abnormal segment includes: calculating the continuous difference Δ of the trend vector k , and construct a trend mutation factor group, where any trend mutation factor is |Δ j+1 -Δ j |; Combine all trend mutation factor groups in the stage history shards to calculate the stability reference interval in, and are the mean and standard deviation of the mutation factors, respectively, and θ is the adjustment threshold coefficient; determine whether any mutation factor in the trend mutation factor group exceeds the stability reference interval Λ. If so, identify the current fragment as a deviation-type fragment, and record the trigger point and the corresponding time point; if it is a deviation-type fragment, extract the multimodal data group contained in the feature vector of the fragment, calculate the variation amplitude within the current window for each mode, and construct a modal fluctuation sequence; if the variation amplitude of at least two items in the modal fluctuation sequence is higher than the upper limit of the historical modal interval, then confirm that the fragment is an abnormal fragment.

[0014] As a preferred embodiment of the multimodal printer state prediction method of the present invention, the generation of the state level transition sequence includes: constructing an index chain list of monotonically increasing segments according to the timestamp information of the abnormal segment set, and embedding the interval parameter γ between segments in the index chain structure. t, establish a sequential constraint structure; place adjacent potential abnormal fragments in γ t The fragments in the same time range are grouped into clusters, which are recorded as TQ, and each TQ is given a fragment sequence label to form a fragment cluster sequence aggregated by time dimension; based on the consistency ratio η of the internal fragments of each TQ in the characteristic trend direction h Construct a cross-cluster transfer criterion that satisfies η h When it is greater than the preset direction consistency threshold, a state transition path from TQ to TQ+1 is established; based on the state transition path, a state-level transition sequence is generated, and each pair of continuous cluster association pairs in the transition sequence is marked as a first-level state transition pair.

[0015] As a preferred embodiment of the multimodal printer state prediction method described in the present invention, the state prediction result output based on trend consistency includes: for each pair of continuous segment clusters, comparing whether the overall trend direction of the corresponding segments in the original sequence remains continuous; when trend consistency is established, it is considered that the state marked by the previous segment cluster remains unchanged in the next segment cluster, and the state label of the previous segment cluster is continued to be assigned to the next segment cluster, which is regarded as a continuous evolution process of the state; if it is judged that the trend is inconsistent, it is regarded as a potential transition of the state, and the state label of the previous segment cluster is not directly continued, and the current segment cluster is independently state-labeled, and the position of the segment cluster is marked as a state change node in the state-level transition sequence; after completing the trend consistency judgment and state label processing for all continuous segment clusters, the final state label of each segment cluster is rearranged according to the original time order, and a state prediction result sequence covering all segment clusters is output as the final trend-guided state prediction output

[0016] In a second aspect, the present invention provides a multimodal printer state prediction system, comprising: a cold start determination module for acquiring a multimodal data stream set, determining whether it is in a cold start phase, and outputting a phase label;

[0017] The conditional sharding module is used to conditionally shard the data stream set according to the stage label and calculate the feature vector for each shard;

[0018] The anomaly marking module is used to build a local polynomial regression model for each shard. It uses each eigenvector to map the trend vector of key indicators in the next 3 to 5 cycles. If the trend mutation and modal fluctuation conditions are met, the current shard is marked as a potential anomaly segment.

[0019] The state transition module is used to cluster and sort the set of potential abnormal fragments in the time dimension, generate a state-level transition sequence through the transition association between fragments, and output the state prediction results based on trend consistency;

[0020] The model adjustment module is used to adjust the weights of each regression model based on the comparison of the state prediction results and the subsequent true state labels.

[0021] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the multimodal printer state prediction method as described in the first aspect of the present invention are implemented.

[0022] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, the steps of the multimodal printer state prediction method as described in the first aspect of the present invention are implemented.

[0023] The beneficial effects of the present invention are as follows: by introducing a state evolution trend analysis mechanism, the present invention can identify state segments with risks in advance before obvious faults occur, achieve early warning and trend guidance of key parameter changes during the printing task, and effectively reduce the risk of print quality degradation or equipment damage caused by sudden state changes. At the same time, the present invention adopts a multimodal data-driven conditional modeling and state clustering method to improve the adaptability and generalization ability of the prediction model to complex operating conditions, and is particularly suitable for actual industrial environments such as frequent switching of printing tasks and changeable equipment operating states. Through trend consistency judgment and state label update mechanism, a state prediction sequence with evolutionary significance can be output to provide a decision-making basis for subsequent task scheduling and maintenance intervention. In summary, the present invention can realize dynamic evolution modeling and trend prediction of the operating state of the printing equipment, thereby effectively improving the system's recognition and response capabilities to potential anomalies, cold start instability and modal fluctuations. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0025] Figure 1 A flowchart of a method for predicting a state of a multimodal printer;

[0026] Figure 2 This is the structural diagram of the multimodal printer state prediction system. DETAILED DESCRIPTION

[0027] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0028] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0029] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0030] As mentioned in the background technology section above, existing printer state correction or control methods often focus on single-modal data modeling or local control optimization of unit hardware. A comprehensive framework for fusing multimodal data and establishing evolution trend prediction models has yet to be established. In particular, there is a lack of accurate state identification and prediction mechanisms under cold start and sudden modal fluctuation conditions. Therefore, a multimodal printer state prediction solution is needed.

[0031] Figure 1 FIG. 1 is a flow chart of a method for predicting a multi-modal printer state according to an embodiment of the present invention. Figure 1 As shown, the multimodal printer state prediction method includes the following steps:

[0032] S1: Obtain a set of multimodal data streams, determine whether it is in the cold start phase, and output the phase label.

[0033] First, the multimodal data stream includes but is not limited to print head temperature, feed rate, power consumption characteristics, and electronic control signals. In this embodiment of the present invention, to determine whether the collected multimodal data stream is in the cold start phase, a comprehensive judgment is made based on the volatility of data characteristics within the initial time window.

[0034] Preferably, the cold start phase judgment includes the following steps:

[0035] 1) Build a window for the first k time slices in the multimodal data stream set, extract the modal feature component set, and calculate the heteroscedasticity of each modal feature to form a feature heteroscedasticity vector, where k is a constant.

[0036] It should be noted that heteroscedasticity refers to an index of the unevenness of the fluctuation amplitude between the characteristic values ​​of each time slice, which can effectively reflect the level of instability of the feature in the initial stage. Traditional state judgment often relies on basic statistical indicators such as mean and variance, while the present invention adopts heteroscedasticity indicators to more sensitively capture the violent fluctuations or local instability that may exist in the early stage of cold start, thereby improving the accuracy of cold start discrimination. In specific operations, for the feature sequence of each modal channel, the local variance of each time slice in the window is calculated in turn, and weighted according to the time slice position weight, and finally the heteroscedasticity value under a single mode is output. The set of heteroscedasticity values ​​of all modes constitutes the characteristic heteroscedasticity vector in the multimodal case.

[0037] 2) The volatility density in the current window is calculated based on the characteristic heteroskedasticity vector, and the ratio of the volatility density means stored in the historical stable phase is calculated as the cold start discriminant factor.

[0038] Specifically, the fluctuation density is the weighted result of the values ​​of each dimension of the feature heteroskedasticity vector, which can intuitively reflect the overall level of feature fluctuation in the current window. This step is designed to uniformly measure the fluctuation trend of system state from the perspective of multimodal features, avoiding the one-sided judgment caused by a single feature anomaly, thereby improving the consistency and reliability of state judgment.

[0039] 3) If the ratio is greater than or equal to the density offset threshold, the output stage label is cold start; otherwise, the output stage label is stable stage.

[0040] Specifically, the present invention calculates the ratio of the fluctuation density value calculated in the current window to the average fluctuation density value stored during the historical stable phase to form a cold start discrimination factor. If this ratio is greater than or equal to a preset density offset threshold, the current multimodal data stream is determined to be in the cold start phase; otherwise, it is determined to be in the stable phase. The density offset threshold is a system-preset parameter, and its value is set based on the tolerance range of fluctuations and the need to control the risk of misjudgment in the specific application scenario.

[0041] It is particularly worth noting that, through the above-mentioned cold start stage discrimination step, the present invention can quickly and accurately divide the state stages at the initial stage of multimodal data collection, that is, to provide a phased basis for subsequent data processing, state prediction and anomaly detection. Compared with the traditional method of simple judgment that only relies on time periods, the method described in the present invention can realize stage division at the level of operating data characteristics based on the dynamic comparison of characteristic volatility and historical volatility levels, and has stronger adaptability and higher accuracy. In addition, the design of the cold start discrimination factor can also be expanded to the detection of other state transition nodes, such as equipment startup, load mutation or system reset, etc., to provide a universal discrimination framework for state management.

[0042] S2: Conditionally shard the data stream set according to the stage label and calculate the feature vector for each shard.

[0043] It should be noted that in the conventional data processing process, the stage attributes are often ignored and only the time window is relied upon for slicing, which easily leads to the mixing of data in different operating states, affecting the accuracy of feature extraction. The present invention uses the stage label as the basis for slicing, so that each slicing only corresponds to the data in the same operating state (such as cold start or stable stage) of the system, thereby ensuring the stage consistency of subsequent feature vector calculations. Through the above operations, the precise binding of data and operating state can be effectively achieved, making feature extraction more discriminative and interpretable. The specific operations are as follows:

[0044] The optimal conditional sharding process includes the following steps:

[0045] A) Match the stage labels with the multimodal data stream set, and for each stage label, extract the data stream sequence generated in the corresponding stage.

[0046] The above operation is based on the joint indexing of the timestamps and stage labels in the multimodal data stream set, and the extraction of the corresponding multimodal data stream sequence for the time segment labeled as cold start. Similarly, the same data screening is performed for the time segment labeled as stable stage, avoiding the confusion of features from different stages, thereby ensuring the independence and purity of cold start features and stable operation features in subsequent calculations. This avoids the risk of the window spanning two operating states by using the sliding window interception method used in conventional methods. The present invention strictly demarcates the boundaries through stage labels, ensuring that each segment of data belongs to a single state, thus achieving high reliability and state purity in the data sharding process.

[0047] B) Based on the temporal order of the data points, the data stream sequence is divided into several continuous data segments, each of which has a fixed period length. Specifically, the fixed period can be preset based on the device's operating characteristics and the data sampling frequency, such as 10 seconds, 30 seconds, or 1 minute. This is not a strict limitation in the present embodiment. This ensures that each data segment is of sufficient length for statistical feature extraction, while avoiding smoothing out state fluctuations due to excessive length.

[0048] C) In each segment, the data points in the segment are further segmented based on whether the print head temperature exceeds the upper and lower floating range of the corresponding stage mean to generate a conditional segment set.

[0049] Specifically, the mean μ and standard deviation σ of the print head temperature are calculated for historical data during the current phase (e.g., cold start or stabilization phase), and a floating range threshold (e.g., μ±1.5σ) is set. During operation, the data points within each time period are traversed to determine whether the print head temperature falls within the set range. If it exceeds, it is determined to be a temperature fluctuation point, and based on this, the original data segment is segmented into multiple conditional segments.

[0050] This operation differs from existing methods that simply partition data by fixed time or number of data points. It is highly adaptive and state-aware, accurately capturing the implicit correlation between printhead temperature fluctuations and device operating status. This conditional slicing not only preserves the temporal continuity of the original data, but also effectively responds to dynamic changes in key device parameters and adjusts data slices, providing more representative fragments for subsequent feature vector calculations.

[0051] Furthermore, calculating the feature vector for each shard includes the following steps:

[0052] For each conditional slice, data normalization is first performed to eliminate the dimensional effects between different modes and different sampling scales. For each conditional slice, statistical indicators of all modes in the slice are collected to construct its segment feature vector, including but not limited to the mean and variance. The above multiple statistical features are spliced ​​into a feature vector of uniform length as the representative expression of the current conditional slice.

[0053] It can be seen that the present invention can dynamically reflect the subtle fluctuation behavior in the operation of the device. For example, in the cold start phase, the fluctuation range of the print head temperature itself is relatively large. At this time, the conditional slicing will generate more and shorter segments, thereby carefully depicting the gradual characteristics of the temperature rise during the cold start process. In the stable stage, the temperature fluctuation range is smaller, and the conditional segments generated by the system are also longer and more uniform, thus reflecting the stability of the device operation. Compared with the limitation of the existing slicing method that cannot perceive the change of the print head state, the present invention introduces the key parameter of print head temperature as a slicing condition, thereby realizing the linkage adjustment of data segment division and the actual state of device operation, thereby significantly improving the characterization ability of data features and the model training effect.

[0054] S3: Construct a local polynomial regression model for each shard, and map the trend vectors of key indicators in the next 3 to 5 cycles through each eigenvector; if the trend mutation and modal fluctuation conditions are met, mark the current shard as a potential abnormal fragment.

[0055] S3.1: The construction of the local polynomial regression model includes:

[0056] S3.1.1: For any feature vector, retrieve the last T in the corresponding conditional slice eA sequence of key indicators per unit time; wherein, the key indicators include but are not limited to the modal values ​​of the print head temperature and feed rate. It should be noted that, in terms of the selection of key indicators, the present invention gives priority to modal values ​​that are sensitive to state changes during equipment operation, such as print head temperature and feed rate. The print head temperature indicator can reflect the dynamic change trend of the equipment heating system, while the feed rate indicator is directly related to the stability of the material supply system. These two types of indicators provide basic data guarantees for the prediction of equipment state trends. Therefore, for each slice, the present invention automatically extracts the above-mentioned key indicator value sequence of the tail time period as the data basis for the subsequent polynomial fitting model. Different from the modeling method of the overall mean or full-segment static data in the prior art, it realizes the rapid capture and processing of the dynamic trend of the local time period.

[0057] S3.1.2: Introduce a weight sequence on the key indicator sequence, where each weight is dynamically calculated based on the modal variance and historical mean of the data point, so that high volatility points account for a higher proportion in the fitting, and construct a weight vectorized trend sequence s k .

[0058] The weight calculation logic is as follows: If a data point exhibits high volatility (i.e., large modal variance) within the current window, or its value deviates significantly from the historical mean, the system assigns a higher weight to that point. Through this mechanism, the present invention allows data points with high volatility or large deviations to dominate the fitting process, effectively amplifying minor anomalies and improving trend detection accuracy.

[0059] S3.1.3: Based on s k The constructed key indicator sequence adopts a polynomial fitting modeling method of no more than 3rd order, and independently constructs a local prediction expression for each modal indicator.

[0060] Specifically, based on the above-mentioned weighted key indicator sequence, the present invention uses a polynomial fitting method of no more than 3rd order to independently construct a local prediction model for each modal indicator. The polynomial order is controlled within 3rd order to avoid overfitting while ensuring that the model has the ability to describe nonlinear trend changes. For example, for the print head temperature indicator, the fitting model form can be expressed as: T(t) = a0 + a1t e +a2t e 2 +a3t e 3 , where T(t) is the prediction time t eThe temperature value at the moment, a0~a3 are polynomial coefficients, obtained by solving based on the weighted least squares method. Similarly, an independent prediction expression is also constructed for the feed rate index. In this process, based on the fragmentation characteristics described by the eigenvector, the adaptive fitting of the polynomial model parameters is realized, so that the model is accurately mapped to the local dynamic characteristics of the current fragment, thereby improving the pertinence and accuracy of the prediction. Through the polynomial regression model, the present invention can predict the changing trend of key indicators in the next 3 to 5 cycles, providing a trend basis for the subsequent identification of abnormal fragments.

[0061] After completing the trend vector prediction, the present invention further implements an abnormal segment confirmation mechanism to determine whether the current segment has potential abnormalities by performing a difference analysis on the trend vector.

[0062] S3.2: Confirmation of abnormal fragments includes the following steps:

[0063] First, calculate the successive differences Δ for the trend vector k , and construct a trend mutation factor group, where any trend mutation factor is |Δ j+1 -Δ j |; Combine all trend mutation factor groups in the stage history shards to calculate the stability reference interval in, and The method uses the mean and standard deviation of the mutation factors, respectively, and θ as the adjustment threshold coefficient. It determines whether any mutation factor in the trend mutation factor group exceeds the stability reference interval Λ. If so, the current segment is marked as a deviating segment, and the trigger point and corresponding time point are recorded. This mechanism differs from existing methods that only detect single features such as mean deviation or volatility anomaly. It accurately captures trend mutation characteristics, effectively improving the accuracy and reliability of anomaly detection. If the segment is a deviating segment, the multimodal data set contained in the feature vector of the segment is extracted, the variation amplitude within the current window is calculated for each mode, and a modal fluctuation sequence is constructed. If the variation amplitude of at least two items in the modal fluctuation sequence exceeds the upper bound of the historical modal interval, the segment is confirmed to be an anomaly segment. This dual verification mechanism—trend mutation detection and modal fluctuation verification—effectively avoids misjudgments caused by single-indicator anomalies, ensuring that the marked segments are truly anomalous segments and achieving high detection robustness.

[0064] The variation range is defined as the difference between the maximum and minimum values.

[0065] S4: Cluster and sort the set of potential abnormal fragments in the time dimension, generate a state-level transition sequence through the transition association between fragments, and output the state prediction result based on trend consistency.

[0066] S4.1: The generation of the state-level transition sequence includes the following steps:

[0067] The abnormal fragment set is constructed into an index chain list of monotonically increasing fragments according to the timestamp information, and the interval parameter γ between fragments is embedded in the index chain structure. t This parameter is used to subsequently determine the temporal correlation and attribution relationship between segments and to establish a temporal sequence constraint structure. In this way, a segment organization structure with temporal sequence constraints is formed, laying the foundation for clustering in the time dimension.

[0068] Furthermore, adjacent potential abnormal fragments are placed in γ t The segments within the range are grouped into the same time aggregation segment cluster, denoted as TQ. The same rule is applied to all continuous segments that meet the conditions, and they are unified into one TQ. A segment sequence label is assigned to each TQ to form a segment cluster sequence aggregated by time dimension.

[0069] Furthermore, after the TQ sequence is constructed, the state transition path determination phase is entered. The core of this phase is to determine whether there is a valid state transition path between different TQs based on the consistency ratio of each TQ internal segment in the characteristic trend direction. Specifically, based on the consistency ratio η of each TQ internal segment in the characteristic trend direction h Construct a cross-cluster transfer criterion that satisfies η h When the direction consistency threshold is greater than the preset direction consistency threshold, a state transition path from TQ to TQ+1 is established. According to the establishment of the state transition path one by one, a state level transition sequence is generated, and each pair of continuous cluster association pairs in the transition sequence is marked as a first-level state transition pair. Among them, the consistency ratio η h It is defined as the ratio of the number of segments with consistent directions to the total number of segments. It can be seen that this sequence fully describes the dynamic evolution chain of potential abnormal segments formed by the combination of time dimension aggregation, feature trend association, and state transition relationship.

[0070] S4.2: Outputting state prediction results based on trend consistency includes the following steps:

[0071] For each pair of consecutive segment clusters, we compare whether the overall trend direction of the corresponding segments in the original sequence remains consistent. The specific calculation method is to calculate the mean of the trend vectors for TQ and TQ+1 respectively, and calculate the angle between the two. If the angle is less than a set threshold (such as 10 degrees), the trend consistency is considered to be established. When trend consistency is established, it is believed that the state marked by the previous fragment cluster remains unchanged in the next fragment cluster, and the state label of the previous fragment cluster is continued to be assigned to the next fragment cluster. The two are regarded as a continuous evolution process of the state, and are represented in the state-level transition sequence as a smooth continuation of the state; if it is judged that the trend is inconsistent, it is regarded as a potential state transition, and the state label of the previous fragment cluster is not directly continued. The current fragment cluster is independently labeled and the position of the fragment cluster is marked as a state change node in the state-level transition sequence; after completing the trend consistency judgment and state label processing of all continuous fragment clusters, the final state label of each fragment cluster is rearranged in the original time order, and a state prediction result sequence covering all fragment clusters is output. This sequence not only retains the transition relationship in the time dimension, but also clearly reveals the continuity and transition nodes of the state, as the final trend-guided state prediction output.

[0072] S5: Based on the comparison between the state prediction results and the subsequent true state labels, feedback is provided to adjust the weights of each regression model.

[0073] First, the state prediction result sequence output in step S4 is used as input and compared with the subsequent true state labels. Specifically, the predicted state label for each segment cluster in the state prediction result sequence is extracted, and the true state label corresponding to the segment cluster in time is also extracted. This is then matched across all segment clusters to determine the state label error between them.

[0074] For example, if the state label is categorical, a classification difference function can be used. For example, if the predicted state label = the true state label, the state deviation is 1, otherwise it is 0; if the state label is numerical, Euclidean distance or absolute difference is used for quantification.

[0075] By comparing the state labels of all segment clusters in the sequence, a prediction error vector is constructed. This error vector fully reflects the local deviation distribution characteristics of the prediction results across different segment clusters. To further quantify the balance of the error distribution, the error distribution balance index G is calculated. This index is defined as the ratio of the standard deviation of the error vector to the mean. This index is used to characterize the amplitude of error fluctuation in the sequence. When G is large, it indicates that the local error distribution is uneven, with some segment clusters having high errors. Conversely, it indicates that the error distribution is relatively smooth and consistent. When the error distribution balance index G exceeds the preset range, the model weight adjustment mechanism is activated.

[0076] Weight adjustment involves dynamically adjusting the local weight factors of each model based on the error vector. The adjustment strategy begins by analyzing the locations of error peaks in the error vector to identify clusters of high-value segments within the error concentration. Then, based on their historical performance, the weights of models with high errors are appropriately lowered, while the weights of models with low errors are increased. Weight adjustment utilizes a proportional scaling method to ensure that the model combination automatically improves its fit to error-sensitive areas in subsequent forecasts.

[0077] Furthermore, this embodiment also provides a multimodal printer state prediction system, comprising: a cold start determination module 100 for acquiring a multimodal data stream set, determining whether it is in a cold start phase, and outputting a phase label;

[0078] The conditional sharding module 200 is used to conditionally shard the data stream set according to the stage label and calculate the feature vector for each shard;

[0079] The anomaly marking module 300 is used to build a local polynomial regression model for each shard, and map the trend vector of the key indicators in the next 3 to 5 cycles through each feature vector. If the trend mutation and modal fluctuation conditions are met, the current shard is marked as a potential anomaly segment.

[0080] The state transition module 400 is used to cluster and sort the set of potential abnormal segments in the time dimension, generate a state-level transition sequence through the transition association between segments, and output the state prediction result based on the trend consistency;

[0081] The model adjustment module 500 is used to adjust the weights of each regression model based on the comparison between the state prediction result and the subsequent true state label.

[0082] This embodiment also provides a computer device suitable for the multimodal printer state prediction method, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the multimodal printer state prediction method proposed in the above embodiment.

[0083] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0084] This embodiment also provides a storage medium storing a computer program. When the program is executed by a processor, the method for predicting the state of a multi-modal printer as proposed in the above embodiment is implemented.

[0085] In summary, by introducing the state evolution trend analysis mechanism, the present invention can identify risky state segments in advance before obvious faults occur, achieve early warning and trend guidance of key parameter changes during the printing task, and effectively reduce the risk of print quality degradation or equipment damage caused by sudden state changes. At the same time, the present invention adopts multimodal data-driven conditional modeling and state clustering methods to improve the adaptability and generalization ability of the prediction model to complex operating conditions, and is particularly suitable for actual industrial environments such as frequent switching of printing tasks and changeable equipment operating states. Through trend consistency judgment and state label update mechanism, a state prediction sequence with evolutionary significance can be output to provide a decision-making basis for subsequent task scheduling and maintenance intervention. In summary, the present invention can realize dynamic evolution modeling and trend prediction of the operating state of the printing equipment, thereby effectively improving the system's recognition and response capabilities to potential anomalies, cold start instability and modal fluctuations.

[0086] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A multimodal printer state prediction method, characterized by: include: Obtain a set of multimodal data streams, determine whether it is in the cold start phase, and output the phase label; Conditionally shard the data stream set according to the stage label and calculate the feature vector for each shard; A local polynomial regression model is constructed for each shard. The trend vectors of key indicators in the next 3 to 5 cycles are mapped through each eigenvector. If the trend mutation and modal fluctuation conditions are met, the current shard is marked as a potential abnormal segment. The set of potential abnormal fragments is clustered and sorted in the time dimension, and a state-level transition sequence is generated through the transition association between fragments. The state prediction result is output based on the trend consistency. Based on the comparison between the state prediction results and the subsequent true state labels, feedback is used to adjust the weights of each regression model.

2. The multimodal printer state prediction method according to claim 1, wherein: The judgment in the cold start phase includes: Build a window for the first k time slices in the multimodal data stream set, extract the modal feature component set, and calculate the heteroscedasticity of each modal feature to form a feature heteroscedasticity vector, where k is a constant; The volatility density in the current window is calculated based on the characteristic heteroskedasticity vector, and the ratio of the volatility density mean stored in the historical stable phase is calculated as the cold start discriminant factor; If the ratio is greater than or equal to the density deviation threshold, the output stage label is cold start; otherwise, the output stage label is stable stage.

3. The multimodal printer state prediction method according to claim 1, wherein: The division of the conditional shards includes: Match the stage labels with the multimodal data stream set, and for each stage label, extract the data stream sequence generated in the corresponding stage; Based on the time sequence between data points, the data stream sequence is divided into several continuous data segments, each of which has a fixed period length; In each segment, the data points in the segment are segmented again based on whether the print head temperature exceeds the upper and lower floating range of the corresponding stage mean to generate a conditional segmentation set.

4. The multimodal printer state prediction method according to claim 1, wherein: The construction of the local polynomial regression model includes: For any feature vector, retrieve the last T in the corresponding conditional slice e A sequence of key indicators per unit time; wherein the key indicators include but are not limited to the modal values ​​of print head temperature and feed rate; A weight sequence is introduced on the key indicator sequence, where each weight is dynamically calculated based on the modal variance and historical mean of the data point, so that high volatility points account for a higher proportion in the fitting, and a weight vectorized trend sequence s is constructed. k ; Based on s k The constructed key indicator sequence adopts a polynomial fitting modeling method of no more than 3rd order, and independently constructs a local prediction expression for each modal indicator.

5. The multimodal printer state prediction method according to claim 4, wherein: The confirmation of the abnormal fragment includes: Calculate the successive differences Δ for the trend vector k , and construct a trend mutation factor group, where any trend mutation factor is |Δ j+1 -Δ j |; Combine all trend mutation factor groups in the stage history shards to calculate the stability reference interval Λ=[μ φ -θ×σ φ ,μ φ +θ×σ φ ], where μ φ and σ φ are the mean and standard deviation of the mutation factors, respectively, and θ is the adjustment threshold coefficient; determine whether any mutation factor in the trend mutation factor group exceeds the stability reference interval Λ. If so, mark the current shard as a deviation shard, and record the trigger point and corresponding time point; If it is a deviant fragment, the multimodal data group contained in the feature vector in the fragment is extracted, the variation amplitude within the current window is calculated for each mode, and a modal fluctuation sequence is constructed; if the variation amplitude of at least two items in the modal fluctuation sequence is higher than the upper limit of the historical modal interval, the fragment is confirmed to be an abnormal fragment.

6. The multimodal printer state prediction method according to claim 1, wherein: The generation of the state level transition sequence includes: The abnormal fragment set is constructed into an index chain list of monotonically increasing fragments according to the timestamp information, and the interval parameter γ between fragments is embedded in the index chain structure. t , establish a temporal sequence constraint structure; The adjacent potential abnormal fragments are placed in γ t The fragments in the same time range are grouped into a cluster, which is recorded as TQ, and each TQ is assigned a fragment sequence label to form a fragment cluster sequence aggregated by time dimension; Based on the consistency ratio η of each TQ internal segment in the characteristic trend direction h Construct a cross-cluster transfer criterion that satisfies η h When it is greater than the preset direction consistency threshold, a state transition path from TQ to TQ+1 is established; According to the state transfer path, a state-level transition sequence is generated, and each pair of continuous cluster association pairs in the transition sequence is marked as a first-level state transition pair.

7. The multimodal printer state prediction method according to claim 6, wherein: The state prediction result output based on trend consistency includes: For each pair of consecutive segment clusters, compare whether the overall trend direction of the corresponding segments in the original sequence remains continuous; When the trend consistency is established, it is believed that the state marked by the previous fragment cluster remains unchanged in the next fragment cluster, and the state label of the previous fragment cluster is continued to be assigned to the next fragment cluster, which is regarded as a continuous evolution process of the state; If the trend is inconsistent, it is considered that there is a potential state transition. In this case, the state label of the previous fragment cluster will not be directly continued. The current fragment cluster will be independently labeled and the position of the fragment cluster will be marked as a state change node in the state level transition sequence. After completing the trend consistency judgment and state label processing of all continuous segment clusters, the final state labels of each segment cluster are rearranged according to the original time order, and a state prediction result sequence covering all segment clusters is output as the final trend-guided state prediction output.

8. A multimodal printer state prediction system, based on the multimodal printer state prediction method according to any one of claims 1 to 7, characterized in that: Also includes: The cold start discrimination module is used to obtain a set of multimodal data streams, determine whether it is in the cold start phase, and output the phase label; The conditional sharding module is used to conditionally shard the data stream set according to the stage label and calculate the feature vector for each shard; The anomaly marking module is used to build a local polynomial regression model for each shard. It uses each eigenvector to map the trend vector of key indicators in the next 3 to 5 cycles. If the trend mutation and modal fluctuation conditions are met, the current shard is marked as a potential anomaly segment. The state transition module is used to cluster and sort the set of potential abnormal fragments in the time dimension, generate a state-level transition sequence through the transition association between fragments, and output the state prediction results based on trend consistency; The model adjustment module is used to adjust the weights of each regression model based on the comparison of the state prediction results and the subsequent true state labels.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the multimodal printer state prediction method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the multimodal printer state prediction method according to any one of claims 1 to 7 are implemented.

Images