A digital transformation maturity assessment method and system based on deep learning

Abstract

The application relates to the technical field of digital transformation evaluation, and discloses a digital transformation maturity evaluation method and system based on deep learning, which comprises the following steps: obtaining to-be-evaluated time series data, historical steady-state benchmark data and historical period reference data; extracting time series change characteristics of the to-be-evaluated time series data relative to the historical steady-state benchmark data and the historical period reference data; determining the deviation property of the to-be-evaluated time series data according to the time series change characteristics, wherein the deviation property comprises a small stable deviation, a transient fluctuation, a periodic fluctuation and a trend drift; if the deviation property is determined to be the trend drift, performing a preset period of continuous review on the time series change characteristics; and generating a classification control strategy for a preset evaluation model according to the determination result of the deviation property and the review result. The application significantly improves the accuracy and robustness of digital transformation maturity evaluation in a complex dynamic business environment.

Description

Technical Field

[0001] This application relates to the field of digital transformation assessment technology, and more specifically, to a method and system for assessing the maturity of digital transformation based on deep learning. Background Technology

[0002] In the field of enterprise digital transformation maturity assessment, using deep learning models to process multimodal data and conduct time-series evaluations has become an important method. To adapt to the dynamic changes in the business environment, existing systems typically embed feature drift detection logic, which calculates the statistical distance between the current data distribution and the historical baseline distribution through a sliding window. When the distance exceeds a set threshold, it automatically triggers incremental training or fine-tuning of the deep learning model.

[0003] However, during the deepening of digital transformation in manufacturing enterprises, seasonal production fluctuations (such as order surges in the fourth quarter) often occur, leading to significant but short-lived pulse-like fluctuations in equipment utilization and employee overtime hours. Existing feature drift detection logic mainly relies on a single distribution distance calculation, lacking the ability to distinguish between temporary business shocks and persistent feature drifts driven by transformation. When seasonal rush work data enters the sliding window, the system may trigger unnecessary incremental training due to statistical differences exceeding the threshold, internalizing short-term noise as model features. When production returns to normal in the first quarter of the following year, the model evaluates based on the learned high-load pattern, misjudging normal production rhythms as efficiency declines, resulting in a significant and unrealistic drop in maturity scores. This oversensitivity to short-term noise means that closed-loop evolution not only fails to achieve self-correction but also introduces new evaluation biases, resulting in a systematically low maturity assessment result after production returns to normal.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application discloses a digital transformation maturity assessment method and system based on deep learning. It aims to solve the technical problem in existing technologies where the inability to distinguish between temporary business shocks and persistent feature drift leads to excessive sensitivity of the model to short-term noise and introduces assessment bias.

[0006] The technical solution of this application is as follows: Firstly, this application discloses a method for assessing the maturity of digital transformation based on deep learning, the steps of which include: Acquire time-series data to be evaluated that characterizes the current level of digital operation of an enterprise, historical steady-state benchmark data that reflects the stable operating baseline status of an enterprise, and historical cycle reference data that reflects the annual operating patterns of the same period; Extract the time series variation characteristics of the time series data to be evaluated relative to historical steady-state baseline data and historical periodic reference data; Based on the characteristics of time series changes, determine the nature of the deviation of the time series data to be evaluated. The deviation nature includes stable small deviation, transient fluctuation, periodic fluctuation and trend drift. If the deviation is determined to be a trend drift, then a continuous review of the time series change characteristics is performed at a preset period. Based on the determination and verification results of the deviation nature, a classification control strategy is generated for the preset evaluation model. The classification control strategy includes: if the deviation nature is determined to be a stable small deviation, the time series data to be evaluated is directly input into the evaluation model, and the maturity evaluation result is calculated and output; if the deviation nature is determined to be a transient fluctuation or a periodic fluctuation, the global weights of the evaluation model are frozen, the time series data to be evaluated is input into the evaluation model after the global weights are frozen, and the maturity evaluation result with environmental interpretation labels is calculated and output; if the verification confirms that the deviation nature is a trend drift, the affected feature subset is located, and based on the affected feature subset, the directional parameter update is performed on the local network branch in the evaluation model corresponding to the feature subset, keeping the mapping weight of the evaluation model in a frozen state, the time series data to be evaluated is input into the evaluation model after the local network branch update is completed, and the maturity evaluation result is calculated and output.

[0007] Secondly, this application discloses a digital transformation maturity assessment system based on deep learning, comprising: The data acquisition module is used to acquire time-series data that characterizes the current level of digital operation of an enterprise, historical steady-state benchmark data that reflects the stable operating benchmark status of an enterprise, and historical cycle reference data that reflects the annual operating patterns of the same period. The feature extraction module is used to extract the time series variation features of the time series data to be evaluated relative to historical steady-state benchmark data and historical periodic reference data; The property determination module is used to determine the deviation properties of the time series data to be evaluated based on the characteristics of time series changes. The deviation properties include stable small deviation, transient fluctuation, periodic fluctuation and trend drift. The continuous review module is used to perform a continuous review of the time-series change characteristics at a preset period if the deviation is determined to be a trend drift. The strategy generation module generates classification control strategies for a preset evaluation model based on the determination and review results of the deviation nature. These strategies include: if the deviation nature is determined to be a stable, small deviation, the time series data to be evaluated is directly input into the evaluation model, and the maturity evaluation result is calculated and output; if the deviation nature is determined to be a transient or periodic fluctuation, the global weights of the evaluation model are frozen, the time series data to be evaluated is input into the evaluation model with the global weights frozen, and the maturity evaluation result with environmental interpretation labels is calculated and output; if the review confirms that the deviation nature is a trend drift, the affected feature subset is located, and based on the affected feature subset, directional parameter updates are performed on the local network branches in the evaluation model corresponding to the feature subset, keeping the mapping weights of the evaluation model frozen, the time series data to be evaluated is input into the evaluation model after the local network branches are updated, and the maturity evaluation result is calculated and output.

[0008] In summary, this application provides a deep learning-based method and system for assessing the maturity of digital transformation. By constructing a multi-dimensional time-series benchmark state and introducing a deep quantification and property determination mechanism for time-series change characteristics, it achieves accurate identification of feature drift in digital transformation maturity assessment. This scheme can effectively distinguish between short-term business shocks, cyclical fluctuations, and persistent transformation changes. Through a hierarchical control strategy, it freezes model weights and provides environmental explanations when facing short-term noise, and performs targeted local parameter updates when facing real transformation. This mechanism not only avoids model mistraining and assessment bias caused by seasonal fluctuations, but also ensures that the deep learning model can maintain the consistency and depth of historical assessment benchmarks while absorbing new transformation characteristics, thereby significantly improving the accuracy and robustness of digital transformation maturity assessment in complex and dynamic business environments. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a deep learning-based digital transformation maturity assessment method provided in an embodiment of this application.

[0010] Figure 2 This is a flowchart illustrating the classification control strategy provided in an embodiment of this application.

[0011] Figure 3 This is a schematic diagram of the structure of a deep learning-based digital transformation maturity assessment system provided in an embodiment of this application.

[0012] Figure 4 This is a schematic diagram of the continuous review process provided in the embodiments of this application.

[0013] Labeling Explanation: 210, Data Acquisition Module; 220, Feature Extraction Module; 230, Property Determination Module; 240, Continuous Verification Module; 250, Strategy Generation Module. Detailed Implementation

[0014] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0015] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0016] Firstly, referring to Figure 1 This application provides a method for assessing the maturity of digital transformation based on deep learning. The steps of this method include: S1. Obtain time-series data to be evaluated that characterizes the current level of digital operation of the enterprise, historical steady-state benchmark data that reflects the stable operating benchmark status of the enterprise, and historical cycle reference data that reflects the annual operating pattern of the same period. S2. Extract the time series variation characteristics of the time series data to be evaluated relative to historical steady-state baseline data and historical periodic reference data; S3. Based on the characteristics of time series changes, determine the nature of the deviation of the time series data to be evaluated. The nature of the deviation includes stable small deviation, transient fluctuation, periodic fluctuation and trend drift. S4. If the deviation is determined to be a trend drift, then perform a continuous review of the time series change characteristics at a preset period. S5. Based on the determination and verification results of the deviation nature, generate a classification control strategy for the preset evaluation model. The classification control strategy includes: if the deviation nature is determined to be a stable small deviation, directly input the time series data to be evaluated into the evaluation model, calculate and output the maturity evaluation result; if the deviation nature is determined to be a transient fluctuation or periodic fluctuation, freeze the global weights of the evaluation model, input the time series data to be evaluated into the evaluation model after the global weights are frozen, calculate and output the maturity evaluation result with environmental interpretation labels; if the verification confirms that the deviation nature is a trend drift, locate the affected feature subset, perform directional parameter updates on the local network branches in the evaluation model corresponding to the feature subset based on the affected feature subset, keep the mapping weights of the evaluation model in a frozen state, input the time series data to be evaluated into the evaluation model after the local network branches are updated, calculate and output the maturity evaluation result.

[0017] The time-series data to be evaluated refers to the various operational indicators recorded in chronological order that are currently occurring within an enterprise, reflecting its current actual operating status. This data is not limited to single values ​​from a single business system, but can include equipment operation records from the Manufacturing Execution System (MES), order fulfillment records from the Enterprise Resource Planning (ERP) System, material turnover records from the Warehousing System, shift scheduling and attendance records from the Human Resources System, inspection result records from the Quality System, and energy consumption records from the Energy Management System. This data typically includes explicit timestamps and can be arranged in minute, hour, shift, day, or week order order. For example, in a manufacturing workshop, within the same time period, equipment operating status, work order completion quantity, abnormal downtime, overtime hours, and energy consumption per unit output can be collected simultaneously; these data collectively constitute the time-series data to be evaluated for that time period.

[0018] Historical steady-state baseline data refers to data accumulated by an enterprise during a period of stable operation, representing its normalized level of digital transformation when unaffected by short-term shocks. A stable operating period does not simply refer to earlier data points, but rather to a data range within which the enterprise's maturity assessment results, key business indicators, and cross-system collaboration relationships do not show significant abrupt changes. For example, if an enterprise's order structure, equipment load, staffing, and delivery rhythm remain relatively stable over several consecutive assessment periods, and its maturity score does not experience sudden increases or decreases, then this range is more suitable as historical steady-state baseline data. The purpose of this type of data is to provide a reference for what kind of operating model the enterprise should exhibit under normal circumstances.

[0019] Historical cycle reference data refers to data recorded by a company in the same historical period (e.g., the same season or month of the previous year). It is primarily used to reflect seasonal or cyclical patterns in the company's operations. This type of data emphasizes the correspondence between time positions; its purpose is not to reflect the company's normal operating level, but rather to reflect the recurring business rhythms at specific calendar positions. For example, if a company experiences a concentrated order delivery period, high equipment utilization, and increased overtime work at the end of each year, then data from the same calendar period of the previous year can serve as an important reference for identifying whether the current year's performance reflects normal seasonal demand.

[0020] Deviation characteristics refer to the specific reasons and manifestations of current operational data deviating from the normal baseline. These include stable, small deviations (normal fluctuations), transient fluctuations (sudden and brief anomalies), cyclical fluctuations (seasonal changes consistent with historical patterns), and trend drifts (representing persistent changes in genuine transformation and upgrading). Stable, small deviations typically manifest as limited fluctuations in current data relative to historical steady-state baseline data, but the overall structural relationship remains unchanged. Transient fluctuations typically manifest as sudden shifts within a short period, lacking historical repetition and failing to persist in subsequent cycles. Cyclical fluctuations typically manifest as relatively significant current deviations, but with high consistency in the direction, rhythm, and duration of change with historical reference data. Trend drifts typically manifest as current data not only deviating from historical steady-state baseline data, but also as this deviation persisting across multiple subsequent evaluation cycles, gradually forming a new stable structure, such as continuously increasing equipment connectivity, continuously shortening production scheduling response time, and continuously improving cross-departmental collaboration efficiency.

[0021] In the operational logic of the above overall technical solution, it is first necessary to simultaneously construct baseline data states for three different time dimensions. In one possible example, these three types of data can be obtained by manually exporting reports from various business systems of the enterprise (such as manufacturing execution systems and human resources systems) periodically, and then having data analysts manually extract data fragments for the corresponding time periods based on experience. For example, analysts can export the data for the current month at the end of each month as the time series data to be evaluated, and review last year's reports to find the data for months considered stable as historical steady-state baseline data. At the same time, they can retrieve the full-month operation ledgers, work order records, personnel scheduling, and energy consumption data from the previous year that completely correspond to the current evaluation window calendar, and compile them into a comparative sample for the same period as historical period reference data.

[0022] After obtaining multi-dimensional benchmark data, it is necessary to extract the changes in the current data relative to historical benchmarks. In one possible example, extracting time-series variation features can be achieved by calculating the statistical differences between the current data and the benchmark data on various specific indicators. Specifically, the mean difference and variance rate of change between the time-series data to be evaluated and the historical steady-state benchmark data on individual indicators such as equipment utilization rate and employee overtime hours can be calculated separately. The statistical differences of these surface values ​​are then summarized to form the time-series variation features used for subsequent judgment. In the same way, the time-series variation features of the current data relative to historical periodic reference data can be extracted.

[0023] Subsequently, based on the extracted time-series change characteristics, the deviation nature of the current data is classified and determined. If the determination result is a trend drift, it means that the enterprise has initially been considered to have undergone genuine transformation and upgrading. To prevent false positives caused by single data fluctuations, the model is not immediately updated, but instead enters a continuous review phase. Over several subsequent evaluation periods, the stability of this drift trend is continuously observed. To more clearly illustrate this determination process, it can be understood as a tiered screening process. First, it is determined whether the current deviation is significant only relative to the historical steady-state benchmark data, but not significant relative to the historical periodic reference data; if so, it is more likely to be identified as a periodic fluctuation. If the current deviation deviates from both the historical steady-state benchmark data and the historical periodic reference data, but the deviation only occurs in a local short period of time and then quickly falls back, it is more likely to be identified as a transient fluctuation. If the current deviation has a limited amplitude, a short duration, and does not cause changes in cross-modal structural relationships, it is more likely to be identified as a stable, small deviation. A current offset is preliminarily identified as a trend drift only when it simultaneously meets the following characteristics: First, the offset is not a recurring pattern from the same historical period; second, the offset is not a single sudden spike, but rather persists throughout the continuous evaluation period; third, the offset is reflected not only in a single indicator but also in the coordinated changes of multiple related indicators; and fourth, the direction of the offset is consistent with the enterprise's digital transformation goals, such as increased automation, shorter process response times, and faster closed-loop anomaly handling. Upon entering the continuous review phase, the same feature extraction and deviation determination are repeated in subsequent continuous evaluation periods, with a focus on checking whether the aforementioned offset persists, gradually stabilizes, or expands into a new normal structure. If the offset disappears, weakens, or reverts to a historical cycle pattern during the review period, the trend drift determination is revoked; if the offset persists and forms a stable pattern during the review period, the trend drift is confirmed.

[0024] Finally, based on the determined nature of the deviation, a differentiated classification control strategy is implemented for the deep learning model. If the deviation is only stable and small, it indicates that the business is normal, and the evaluation result is output directly. If it is a transient or cyclical fluctuation (such as year-end rush work), all weights of the model are forcibly frozen to prevent the model from learning such short-term noise. An environmental explanation label is added to the output result to indicate to management that the current score fluctuation is caused by short-term environmental factors. Only when a real trend drift is confirmed will the specific business features that have changed (i.e., the affected feature subset) be located. The parameters are updated only for the local network branches in the model that process these specific features. At the same time, the mapping weights of the model's core are strictly frozen. The time series data to be evaluated is input into the evaluation model after the local network branches have been updated, and the maturity evaluation result is calculated and output.

[0025] Specifically, environmental explanation labels can be attached to maturity assessment results in the form of structured tags. The tags should include at least the deviation category, a summary of the triggering basis, and a suggested explanation. For example, when periodic fluctuations are identified, a periodic fluctuation label can be attached to the result record, along with information indicating that the current window's change rhythm is consistent with historical periods and that the model has not performed learning updates. This allows for a clear distinction between changes in the score stemming from changes in the company's capabilities and changes in external business rhythms when reviewing assessment results later. For local network branch updates in trend-drift scenarios, the entire model is not retrained. Instead, feature attribution and attention weight backtracking are performed based on vector space offsets. The contribution ratio of each input indicator and modal branch to the representation shift is quantified, and indicators and modal dimensions with contributions exceeding a preset threshold are selected. This accurately locates the affected feature subset, and then the corresponding data input channels and feature processing branches are determined based on the affected feature subset. Finally, only the trainable parameters in these branches are updated. The mapping weights are frozen, meaning that the core mapping part of the model responsible for maintaining the consistency between the overall maturity score scale and the historical assessment criteria does not participate in the update, thereby avoiding the model from disrupting the continuity of the original scoring system when learning new business models.

[0026] In some preferred embodiments, the specific solutions disclosed in this embodiment for the steps of obtaining the enterprise's time-series data to be evaluated, historical steady-state benchmark data, and historical periodic reference data are as follows: Real-time acquisition of enterprise multimodal operation data, and extraction of current sliding window data from the enterprise multimodal operation data as time series data to be evaluated; Data from consecutive periods in which the volatility of maturity assessment scores is lower than a preset volatility threshold are selected from the historical assessment results of enterprises and used as historical steady-state benchmark data. Extract the calendar interval data corresponding to the enterprise's previous year and the current sliding window data, and use it as historical period reference data.

[0027] In this specific implementation plan, data acquisition probes deployed at the edge of the enterprise network are used to pull equipment data from the manufacturing execution system and business flow data from the enterprise resource planning system in real time. A fixed-length time window (such as two weeks) is set, and the data within the window is captured in real time as time progresses.

[0028] In a more concrete and easily implementable example, data acquisition probes can access various business systems via API calls, message subscriptions, or read-only database synchronization. For manufacturing execution systems, they can acquire records such as equipment start / stop status, work order execution progress, and downtime reason codes; for enterprise resource planning systems, they can acquire records such as order placement time, delivery commitment time, and procurement arrival status; for human resources systems, they can acquire information such as shift schedules, attendance records, and overtime application records; and for quality systems, they can acquire information such as inspection batches, pass / fail determination, and rework / repair records. After data acquisition, a unified data access list is first established, marking each type of data with its source system, acquisition time, business meaning, and modality, and then splicing them together according to a unified timeline.

[0029] The current sliding window data refers to a data segment formed by truncating a continuous period of time backward from the current assessment point. The window length, L, can be pre-set according to the company's assessment pace, and can be configured for daily, weekly, or monthly assessment needs. When a company's production pace is fast and it wants to identify short-term shocks more promptly, L can be a shorter interval; when a company's business changes are slower and it wants to observe more stable structural changes, L can be a longer interval. As a new time slice arrives, the window's start and end points move forward synchronously, ensuring that the time-series data to be assessed always reflects the latest operational status.

[0030] When constructing historical steady-state benchmark data, the system automatically iterates through historical assessment records from the past year, calculates the volatility of maturity scores within each time period, and automatically selects three consecutive months of data with volatility below a preset volatility threshold (e.g., volatility below a specific percentage) as the steady-state benchmark. Simultaneously, it automatically aligns the timeline, extracting calendar interval data from the previous year that perfectly corresponds to the current sliding window (e.g., data from the same week last year) as historical cycle reference data. The preset volatility threshold can be determined by first selecting a reference period from the company's historical data that has been confirmed to be free of major organizational changes, major equipment upgrades, major order shocks, and major external anomalies. The maturity assessment results from multiple consecutive assessment windows within this reference period are then arranged chronologically. The natural fluctuation range of these results under normal and stable operating conditions is observed, and a threshold range that covers normal fluctuations but excludes obvious abnormal fluctuations is set accordingly. This threshold range is then fixed as the preset volatility threshold for subsequent automatic selection of steady-state intervals. If the volatility characteristics of different business segments within the company differ significantly, different preset volatility thresholds can be configured for different scenarios such as discrete manufacturing, process manufacturing, and assembly manufacturing.

[0031] For filtering data over continuous periods, the preferred approach not only checks the stability of the maturity assessment score itself but also simultaneously checks for obvious abnormal peaks, long-term missing data, or data collection failures in key business indicators. This is to avoid mistakenly selecting seemingly stable intervals with abnormal underlying data as historical steady-state benchmark data. For extracting historical periodic reference data, the system preferably uses calendar mapping rules for strict alignment. That is, the calendar interval covered by the current sliding window is used to extract the corresponding calendar interval from the previous year. If the enterprise has special rhythms such as shutdowns during Spring Festival, long holidays, or annual maintenance, business event alignment can be superimposed on top of calendar alignment, making the historical periodic reference data more accurately reflect repeatable periodic patterns.

[0032] Optionally, in order to more accurately capture the complex intrinsic correlations between multimodal data, the specific scheme disclosed in this embodiment for the step of extracting the time series variation characteristics of the time series data to be evaluated relative to historical steady-state benchmark data and historical periodic reference data is as follows: The time series data to be evaluated is input into a pre-set deep learning model to extract implicit representation vectors that characterize the enterprise's digital operation mode and maturity status. Calculate the offset of the implicit representation vector relative to the historical implicit representation vector of the historical steady-state baseline data and the historical periodic reference data in the vector space; Temporal variation features are extracted based on offset.

[0033] In this specific implementation plan, a pre-trained deep learning model is invoked (the model is in inference mode, and the weights are temporarily frozen). The time-series data to be evaluated is input into the model, and the network layers inside the model automatically fuse data features from different modalities and map them into a low-dimensional mathematical space, generating a latent representation vector. This vector highly condenses the enterprise's current digital operation model. Similarly, historical steady-state baseline data and historical periodic reference data are also converted into corresponding historical latent representation vectors.

[0034] Subsequently, in this high-dimensional vector space, the spatial distance (such as Euclidean distance or cosine similarity) between the current latent representation vector and two historical latent representation vectors is calculated, and these spatial offsets are transformed into specific temporal change features.

[0035] The pre-defined deep learning model can be implemented as follows: First, a training sample set is formed from the enterprise's historical operational data. Each sample in the training sample set contains multimodal operational data within a time window and a maturity assessment label or phased assessment result corresponding to that time window. Preferably, the multimodal operational data covers at least two categories of data: equipment, business processes, personnel, quality, and energy consumption, to ensure the model can learn cross-modal relationships. Then, the training samples undergo unified preprocessing, including time alignment, outlier labeling, categorical field encoding, continuous field normalization, and missing value imputation. Next, a deep learning model for maturity assessment is constructed. This model can include a multimodal input layer, an intramodal feature extraction layer, a cross-modal fusion layer, and a maturity output layer. The multimodal input layer receives data from different sources; the intramodal feature extraction layer extracts temporal features within each modality; the cross-modal fusion layer learns the collaborative relationships between modalities such as equipment, business processes, and personnel; and the maturity output layer outputs the maturity assessment results. During the training phase, historical samples are continuously input into the model until the model's evaluation results on the validation data stabilize and the output results can well reflect the changing trend of historical maturity labels. At this point, the trained model is solidified as a pre-defined deep learning model. In this implementation scheme, this model is not directly used for updating during the temporal change feature extraction phase, but is first used as a stable representation extractor.

[0036] Furthermore, the latent representation vector can be understood as a set of internal feature expressions formed by the model after completing multimodal fusion. While this result does not directly correspond to a single business indicator, it comprehensively reflects the overall state of the enterprise's operating model within the current time window. For example, when equipment utilization increases, order response speed improves, abnormal downtime decreases, and overtime work does not increase synchronously, the latent representation vector generated by the model will exhibit an internal expression corresponding to efficiency-improvement transformation. Conversely, when equipment utilization increases, overtime work significantly increases, quality fluctuations intensify, and this pattern is similar to historical year-end sales surges, the latent representation vector generated by the model will be closer to the internal expression corresponding to seasonal high loads. In other words, the latent representation vector is not a simple compression of a single indicator, but a comprehensive encoding of the synergistic relationships between different modalities.

[0037] In practice, the time-series data to be evaluated, historical steady-state baseline data, and historical periodic reference data can be input into the same pre-defined deep learning model to obtain three comparable latent representation vectors. Then, the offsets of the current latent representation vector from the historical steady-state baseline latent representation vector and the historical periodic reference latent representation vector in the vector space are calculated. Based on these two sets of offsets, all time-series change features, such as the magnitude of change, duration, periodic similarity, and cross-modal feature correlation structure changes, are quantified and extracted.

[0038] The magnitude of change can be determined by comparing the difference between the implicit representation vector of the current assessment period and the historical steady-state benchmark vector, supplemented by the synchronous changes of key business indicators; the duration can be accumulated according to the number of assessment periods that continuously exceed the steady-state range, and once an intermediate period returns to the steady-state range, the count is restarted; the trend of numerical decline can be judged by observing whether the magnitude of change continues to decrease and whether the rate of decrease is significant in the subsequent assessment periods after the deviation occurs; the change in cross-modal feature correlation structure can be obtained by comparing the correlation patterns between equipment local feature vectors, business local feature vectors, and personnel local feature vectors before and after. For example, if the increase in equipment utilization rate originally led to a synchronous increase in energy consumption, and after the transformation it becomes an increase in equipment utilization rate but a decrease in energy consumption per unit output, then it can be regarded as a structural change has occurred; the periodic similarity can be compared with the reference segment of the same month of the previous year, the same week of the previous quarter, or the historical promotion period. If the two are similar in terms of fluctuation direction, peak and trough position, and continuous rhythm, it indicates that the periodic similarity is high.

[0039] Furthermore, by combining the feature contribution analysis results within the model, it is possible to identify whether equipment modality, business process modality, personnel modality, or quality modality contributes more to the current offset, thus providing a direct basis for locating the affected feature subsets in the future. For example, if the offset mainly comes from features related to equipment network status, automatic work reporting coverage, and production scheduling response timeliness, these features can be grouped into the same affected feature subset; if the offset mainly comes from features related to overtime hours, temporary employment ratio, and short-term output surges, and is consistent with historical patterns, it is more suitable as a basis for determining periodic fluctuations without triggering model learning and updates.

[0040] In some preferred embodiments, the step of inputting the time series data to be evaluated into a preset deep learning model to extract latent representation vectors includes: The time series data to be evaluated is input into multiple parallel feature extraction branches in the deep learning model, and local time series features of the time series data to be evaluated in different modal data are extracted respectively. The local temporal features are fused, and the fused local temporal features are mapped to the latent representation space to generate latent representation vectors.

[0041] Furthermore, the pre-defined deep learning model can be obtained through joint training with historical steady-state benchmark data and historically labeled deviation data. In a feasible example, historical steady-state benchmark data can be collected from the manufacturing enterprise during a period of continuous stable operation. This data can then be combined with records of the enterprise's digital transformation, major process switching, holiday sales surges, and sudden anomalies to form a training sample set labeled with deviation characteristics. Subsequently, multiple parallel feature extraction branches are constructed according to modality type, and a mapping network is connected after the fusion layer. Iterative training is performed with the goal of ensuring that the output latent representation vector can stably distinguish between steady-state samples, transient fluctuation samples, periodic fluctuation samples, and trend drift samples. After training, independent validation data can be used to test the model's representation stability under different months, different production lines, and different order densities. When the model maintains high consistency across multiple consecutive validation batches, and the output latent representation vector is clustered under similar business conditions and separable under dissimilar business conditions, the model can be solidified as a pre-defined deep learning model and put into use. The purpose of this setup is to enable the model to learn not only the changes in a single indicator but also the internal structural characteristics of the manufacturing enterprise's operational status under multi-modal joint conditions.

[0042] In this specific implementation plan, multiple parallel feature extraction branches refer to designing independent network channels at the input of the deep learning model for data of different modalities. Each channel is specifically responsible for processing data of a particular modality, thereby avoiding feature interference between different modalities in shallow networks. Local temporal features refer to the variation patterns of single-modal data along its own time axis. After extracting the local temporal features of each modality, these local features are interacted and integrated through a cross-modal fusion layer (such as an attention mechanism or feature concatenation network). Finally, they are projected into a unified low-dimensional mathematical space through a fully connected layer or mapping network to form an implicit representation vector that comprehensively reflects the overall operational status of the enterprise.

[0043] Furthermore, to make the above concepts easier to implement, local temporal features can be understood as local patterns that can be identified when each modality is observed individually. For example, for the equipment operation modality, local temporal features may manifest as a continuous surge in current within a short period, periodic fluctuations in rotational speed, and rapid amplification of vibration within a few consecutive windows; for the business order modality, local temporal features may manifest as concentrated order placement, continuous backlog of certain processes within multiple windows, and delayed appearance of completion records; for the personnel energy consumption modality, local temporal features may manifest as a continuous increase in the number of night shift workers, higher electricity consumption during a certain shift, and a stable difference between weekend and weekday shifts. These local temporal features are not directly obtained by manually writing rules, but are automatically learned and formed by the corresponding branches during the training process. The role of the cross-modal fusion layer is to establish the correspondence between these local patterns, such as identifying whether an increase in equipment vibration occurs simultaneously with an emergency order insertion, or whether an increase in the number of workers on a shift occurs simultaneously with a decrease in energy consumption per unit output. The implicit representation vector obtained after mapping to the implicit representation space is no longer a single sensor value or a single business field, but a comprehensive compressed representation of the current operating status of the enterprise, which can be directly used for subsequent spatial offset calculation and temporal change feature extraction.

[0044] In a specific embodiment, the time-series data to be evaluated by the manufacturing enterprise includes equipment operation modal data (such as equipment current and speed at the second level), business order modal data (such as order placement and completion records at the hour level), and personnel energy consumption modal data (such as attendance and electricity consumption at the shift level). The deep learning model internally sets up three parallel feature extraction branches: for equipment operation modal data, a one-dimensional convolutional neural network branch is used to capture local abrupt changes in high-frequency signals; for business order modal data, a long short-term memory network branch is used to extract long-term dependencies of discrete events over time; and for personnel energy consumption modal data, a multilayer perceptron combined with time window statistics is used to extract periodic aggregated features. These three branches output local feature vectors for equipment, business, and personnel, respectively. Subsequently, the model uses a multi-head attention mechanism to weight and fuse these three vectors, identifying, for example, the correlation weights between high-frequency equipment fluctuations and order delays, and finally outputs a 512-dimensional latent representation vector through a nonlinear mapping layer.

[0045] Furthermore, in this embodiment, the three types of modal data can be organized as model input samples according to a unified evaluation period. When the evaluation period is daily, the equipment operation modal data can be divided into a high-frequency window sequence continuously collected within that natural day; the business order modal data can be organized into the sequence of occurrence of various events and their time intervals within that natural day; and the personnel energy consumption modal data can be organized into the number of attendees, team composition, and electricity consumption summary for the corresponding shifts within that natural day. After the equipment operation modal data enters the one-dimensional convolutional neural network branch, it can first undergo normalization processing and then be input into a multi-layer convolutional and pooling structure in chronological order to highlight the sharp fluctuations in short periods; after the business order modal data enters the long short-term memory network branch, the sequence of events and the intervals between events can be preserved, thereby identifying the lag pattern of completion delays occurring several hours after a sudden increase in orders; after the personnel energy consumption modal data enters the multi-layer perceptron combined with time window statistics branch, statistical features such as the proportion of day shifts / night shifts and the changes in electricity consumption across multiple consecutive shifts can first be generated, and then sent to the network to extract higher-level periodic patterns. The local feature vectors output by the three branches can be unified to the same dimension through linear transformation, and then input into a multi-head attention mechanism for fusion.

[0046] Optionally, the temporal variation features include the magnitude of change, duration, numerical decline trend, cross-modal feature association structure changes, and periodic similarity; The steps for determining the deviation nature of the time series data to be evaluated based on the characteristics of time series changes include: When the duration is less than the preset duration threshold and the downward trend of the value is greater than the preset downward trend threshold, the deviation is determined to be a transient fluctuation. When the periodic similarity is greater than the preset similarity threshold, the deviation is determined to be a periodic fluctuation. When the duration exceeds a preset duration threshold and the cross-modal feature association structure change exceeds a preset structural mutation threshold, the deviation is determined to be a trend drift. Otherwise, the deviation is determined to be a stable, small-amplitude deviation.

[0047] The aforementioned preset duration threshold, preset fallback threshold, preset similarity threshold, and preset structural mutation threshold are not arbitrarily given, but are determined through a combination of historical data statistics and manual review. In an implementable example, events with clearly attributable causes can be first collected from the company's past data, such as short-term order surges, annual seasonal volume boosts, process upgrades, and continuous equipment anomalies. Then, the performance range of each event in the five types of time-series change characteristics can be calculated. Subsequently, a boundary value that can better distinguish different deviation types is selected as the initial threshold, and business experts review several boundary samples to eliminate obviously unreasonable threshold settings. After completion, the consistency between the automatic judgment results and actual business results can be continuously recorded during a trial operation period. The thresholds can be fine-tuned based on misjudgments until a threshold set suitable for the target company is formed. The preset thresholds obtained in this way can match the company's own business rhythm, order fluctuation intensity, and production organization methods, avoiding the deviations caused by using a uniform fixed standard.

[0048] In this specific implementation plan, the temporal variation characteristics are refined into five core dimensions. The magnitude of change reflects the absolute size of the current data's deviation from the benchmark; the duration records the time span during which the deviation persists; the numerical decline trend measures the speed and tendency of the data to revert to the historical benchmark after the deviation occurs; the cross-modal feature correlation structure change characterizes whether the inherent linkage between different business indicators has been broken and reorganized (e.g., two originally positively correlated indicators become negatively correlated); and the periodic similarity quantifies the degree of matching between the current deviation pattern and historical periodic reference data in terms of form and rhythm. By setting a series of thresholds (preset duration threshold, preset decline threshold, preset similarity threshold, preset structural mutation threshold), these continuous feature indicators are transformed into discrete logical judgment conditions, thereby achieving accurate classification of the nature of the deviation. The magnitude of change, duration, numerical decline trend, cross-modal feature correlation structure change, and periodic similarity can be calculated using the methods described above.

[0049] In one specific implementation, a manufacturing company experienced a sharp increase in equipment utilization and energy consumption in mid-November. Extracting the time-series change characteristics revealed that while the fluctuation was significant, it only lasted for 3 days (less than the preset duration threshold of 7 days), and the values ​​showed an 80% decline on the 4th day (greater than the preset decline threshold of 50%). At this point, the system determined the deviation to be a transient fluctuation (possibly due to a short-term rush of production caused by an urgent order). In another scenario, the company experienced a comprehensive increase in all indicators in December. The system calculated that the current change pattern had a periodic similarity of 0.85 with the historical cycle reference data for December of the previous year (greater than the preset similarity threshold of 0.8), thus classifying it as a periodic fluctuation (a typical year-end production surge). In the third scenario, the company introduced a new intelligent scheduling system, which detected a decrease in equipment idle rate and an increase in on-time order delivery rate. This deviation lasted for 21 days (greater than the preset duration threshold of 7 days), and the cross-modal characteristic correlation structure change index between equipment idle rate and order delivery reached 0.6 (exceeding the preset structural change threshold of 0.4), indicating a substantial change in the underlying business logic. In this case, the deviation is determined to be a trend drift. If none of the above conditions are met, for example, if the indicators only fluctuate slightly around the baseline, it is determined to be a stable, small deviation.

[0050] Furthermore, in the actual deployment of the above embodiments, the judgment process can be solidified into the execution flow of the rule engine. Taking the scenario in mid-November as an example, firstly, continuous daily data on equipment utilization rate and energy consumption are read from the equipment acquisition gateway, and the order execution status of the same period is obtained from the MES system. After feature extraction is completed, the rule engine first checks the duration and then checks the downward trend of the values. When both conditions are met simultaneously, the deviation duration is short and the decline is obvious, and transient fluctuations are immediately output. For the scenario in December, the multimodal data segment of December of the previous year in the historical cycle reference data warehouse is automatically retrieved and matched with the current segment in terms of rhythm. When the matching result is higher than the preset similarity threshold, even if the change is large, it is preferentially classified as a periodic fluctuation. For scenarios that introduce intelligent production scheduling systems, in addition to continuously observing equipment idle rate and order on-time delivery rate, related indicators such as workshop changeover loss, work-in-process dwell time, and unit output energy consumption can also be observed simultaneously. If these related indicators also change in the same direction of improvement, and this relationship exists stably within multiple continuous windows, the judgment of trend drift will be more reliable.

[0051] In practical applications, the steps for performing continuous review at preset intervals on deviation events determined to be trend-driven drift include: Track the evolution trajectory of temporal variation characteristics within a preset number of sliding windows; If the change amplitude remains within the preset stable range within the sliding window and the cross-modal feature association structure changes show a consistent diffusion trend, then the determination result of trend drift is confirmed as true. If the downward trend of the value within the sliding window exceeds the preset downward threshold or the cross-modal feature correlation structure changes return to the historical steady state, then the determination result of trend drift is confirmed as false, and the determination result is corrected to transient fluctuation.

[0052] The aforementioned preset number of sliding windows can be set according to the company's assessment frequency. For example, when conducting maturity assessments weekly, several consecutive weeks can be set as a continuous review and observation period; when conducting maturity assessments daily, several working days or calendar days can also be used as continuous sliding windows. Tracking the evolution trajectory of time-series changes can be specifically understood as follows: at the arrival of each new window, not only should the magnitude of change, the trend of numerical decline, and the changes in cross-modal feature correlation structures be recalculated, but these features should also be compared with the previous window to see if they tend to stabilize, continue to spread, or begin to decay. Keeping the magnitude of change within the preset stable range does not require the change value to remain completely unchanged, but rather requires that the change persist within an acceptable fluctuation range without significant back-and-forth oscillations; if the changes in cross-modal feature correlation structures show a consistent diffusion trend, it emphasizes that this structural reshaping is not limited to one or two indicators, but gradually affects more related business links, and the direction of change is consistent. Conversely, if a significant decline occurs in subsequent windows, or if the previously changed correlation structure returns to the historical steady-state relationship, it indicates that the initial deviation lacks the stability and diffusion that a true transformation should have, and it should be corrected as a transient fluctuation.

[0053] In this specific implementation plan, a preset number of sliding windows corresponds to an observation period (e.g., a sliding window of four consecutive evaluation cycles). The evolution trajectory refers to the dynamic change process of time-series change characteristics within the observation period. The preset stability interval is used to measure whether the new business model has stabilized and no longer experiences drastic fluctuations. The consistent diffusion trend means that this structural change not only persists locally but also gradually spreads and solidifies to other related business processes (e.g., the efficiency improvement brought about by intelligent scheduling spreads from a single workshop to the entire factory). If the data shows stable and continuously diffusing characteristics within the observation period, the review is passed; conversely, if the data drops sharply within the observation period, or the originally changed cross-modal correlation structure degenerates back to the historical steady state, it indicates that the previous change was not a true transformation but a long-term fluctuation. In this case, the system will overturn the initial judgment and downgrade it to a transient fluctuation.

[0054] Furthermore, the pre-defined stable range can be determined using data from the stable phases of successful historical transformation cases. For example, past improvement projects that have been confirmed and implemented by management can be collected, such as production line automation upgrades, warehouse scheduling system implementation, and process parameter optimization. After these projects enter the stable operation phase, the normal fluctuation range of their changes over multiple consecutive evaluation periods can be extracted and used as the source of the stable range for current enterprise review. The consistency diffusion trend can also be obtained through business chain tracking. For example, first observe the decrease in equipment idle rate after the implementation of intelligent scheduling, and then observe whether the shortening of work-in-process waiting time, the improvement of order on-time delivery rate, the acceleration of warehouse turnover, and the improvement of unit energy consumption occur in sequence. If these changes gradually unfold along a reasonable business chain without significant reverse fluctuations, it can be identified as a consistency diffusion trend.

[0055] In one specific implementation, during week 1, a trend drift in a company's operational data is initially determined (manifested as increased automated test coverage and decreased product defect rate). This is followed by a four-week continuous review phase. During the sliding window from week 2 to week 5, the system continuously tracks these characteristics. If the automated test coverage stabilizes in a new high-level range (with changes remaining within a preset stable range), and this efficiency improvement further drives improvements in warehouse turnover (the cross-modal feature correlation structure shows a consistent diffusion trend), the system confirms the trend drift as true at the end of week 5, subsequently triggering directional parameter updates for local network branches. Conversely, if in week 3, due to a failure of new testing equipment, the company is forced to resume manual testing, causing a rapid rebound in the product defect rate (the downward trend is greater than a preset decline threshold), and the correlation structure returns to its historical steady state before the introduction of the new equipment, then the trend drift is confirmed as false during the review period, corrected to a transient fluctuation, thereby preventing the model from making erroneous parameter updates.

[0056] Furthermore, in this embodiment, automated test coverage, product defect rate, and warehouse turnover rate can be set as the core tracking indicators for the review task, and a review weekly report can be automatically generated at the end of each week. The weekly report can record in sequence: whether the changes in this week compared to the initial judgment point in week 1 have remained the same, whether the changes have stabilized, whether the changes have begun to spread to subsequent business links, whether there has been a significant decline, and whether there are external disturbance factors. If the second and third weeks only show an increase in test coverage, but have not yet been reflected in the warehouse turnover rate, the pending confirmation status can be maintained without updating the model in advance; if the warehouse turnover rate and defect rate have continued to improve and stabilized in the fourth and fifth weeks, it indicates that the change has expanded from the local testing link to a wider range of operational links, meeting the review pass conditions. Conversely, if equipment failure occurs in the third week, causing manual testing to revert, in addition to recording the downward trend of the values, maintenance work orders, downtime records, and abnormal repair events can be read simultaneously to help confirm that the change is due to equipment failure rather than a real transformation.

[0057] In a preferred embodiment of this application, before determining the deviation nature of the time series data to be evaluated based on the time series variation characteristics, the method further includes: Calculate the distribution variance of historical steady-state baseline data, and calculate the preliminary deviation tolerance threshold based on the distribution variance; If the change is less than the initial deviation from the tolerance threshold, it is determined that no feature drift has occurred, and the time series data to be evaluated is directly input into the evaluation model to calculate and output the maturity evaluation result. If the change is greater than or equal to the initial deviation from the tolerance threshold, the deviation determination step is triggered.

[0058] Furthermore, the aforementioned historical steady-state benchmark data can be obtained through filtering enterprise operation records. To more clearly and completely describe the acquisition method, in an implementable example, continuous time periods without major equipment failures, large-scale order insertions, institutional adjustments, or process changes, and where production line load is within the normal range, can be selected from the enterprise's past production operation logs. These time periods are then reviewed and confirmed by business management personnel and marked as steady-state samples. Subsequently, key indicators such as daily output, equipment utilization rate, unit energy consumption, and order on-time delivery rate are extracted from these steady-state samples to form a historical steady-state benchmark dataset. The distribution variance calculated based on this dataset reflects the natural fluctuation level that the enterprise is allowed to have under normal conditions. A preliminary deviation tolerance threshold is then formed based on this natural fluctuation level, enabling the pre-filtering mechanism to have enterprise-specific adaptability.

[0059] Furthermore, the evaluation model can be a maturity assessment model that has already been trained and whose parameters have not been updated. In other words, if the current filter determines that no feature drift has occurred, there is no need to initiate the complex deviation analysis process within the deep learning model; instead, the existing maturity assessment model can be used to perform a standard scoring output on the current period's data. This approach ensures the continuity of the evaluation results while avoiding repetitive deep analysis of obviously normal daily fluctuations.

[0060] In this specific implementation plan, the distribution variance is a statistical indicator that measures the natural dispersion of historical steady-state benchmark data during stable periods. A larger variance indicates more drastic fluctuations in the company's business under normal conditions; a smaller variance indicates very stable daily operations. The initial deviation tolerance threshold is a dynamic limit calculated based on this distribution variance (e.g., set at 3 times the standard deviation of the distribution variance). This mechanism performs a rapid binary classification decision by calculating whether the apparent change in the current time series data to be evaluated exceeds this tolerance limit customized based on historical characteristics. If it does not exceed the limit, it means the current fluctuation is completely within the normal range of the company's daily fluctuations and does not require further investigation; if it exceeds the limit, it indicates a potentially noteworthy anomaly or transformation, thus triggering a subsequent in-depth deviation determination.

[0061] Furthermore, to avoid interference from individual outliers in the pre-filtering process, it is preferable to perform steady-state cleaning before calculating the distribution variance of historical steady-state benchmark data. For example, data that is obviously missing, collected incorrectly, or collected during downtime and maintenance should be removed first; special dates such as holidays, the first day of trial production, and temporary major repairs can be archived separately and not included in the steady-state statistics. For rapid calculation of the magnitude of change, only a small number of key indicators most sensitive to the business can be selected, such as one or more of daily output, unit energy consumption, equipment utilization rate, and order on-time delivery rate, thus ensuring that the pre-filtering is sufficiently lightweight. Exceeding the limit can be understood as the deviation of the current period from the steady-state mean exceeding the natural fluctuation range commonly seen by the enterprise under normal operating conditions, at which point it is worthwhile to initiate subsequent complex analysis. In this way, the pre-filtering retains sensitivity to significant deviations without frequently triggering the deep analysis chain due to ordinary daily noise.

[0062] In a specific embodiment, when processing data from a process manufacturing enterprise, the distribution variance of key indicators (such as daily output and average daily energy consumption) in its historical steady-state baseline data is first extracted. Assuming the historical variance of daily output is 100, a preliminary deviation tolerance threshold of 30 is calculated (i.e., daily output is allowed to fluctuate within 30 units above or below the baseline mean). When new time-series data to be evaluated enters the system, a lightweight surface numerical comparison is first performed. If the current daily output change is only 15 (less than the preliminary deviation tolerance threshold of 30), it is directly determined that no feature drift has occurred, bypassing complex implicit representation extraction and multi-dimensional feature determination, and the data is directly input into the evaluation model to output the current maturity score. However, when the daily output change reaches 45 on a certain day (greater than the preliminary deviation tolerance threshold of 30), this fluctuation is identified as exceeding the normal tolerance range. The subsequent deep learning model is then activated to extract implicit representation vectors and perform a deep judgment strictly according to the aforementioned rules for transient fluctuations, periodic fluctuations, or trend drifts.

[0063] Furthermore, in this process manufacturing enterprise scenario, historical steady-state benchmark data can be selected from a continuous period where equipment operation is stable, raw material supply is normal, and there are no concentrated maintenance shutdowns. For example, several natural weeks with stable operating conditions can be selected from the previous few months, and indicators such as daily output and average daily energy consumption can be extracted to form a steady-state sample library. After the market closes each day, the data for that day is read and quickly compared with the corresponding indicators in the steady-state sample library. If only daily output fluctuates slightly, while other key indicators such as energy consumption, equipment utilization rate, and order completion rate remain within the normal range, the system can directly allow it to enter the routine maturity assessment; if daily output deviates significantly, and average daily energy consumption or equipment utilization rate also shows synchronous anomalies, the priority for triggering subsequent in-depth analysis can be increased. The advantage of this setting is that the pre-filter does not mechanically look at only a single value, but rather takes into account the basic consistency between key business indicators while maintaining a lightweight approach, thus better reflecting the actual operating characteristics of process manufacturing enterprises.

[0064] After clarifying the judgment logic for deviation and the continuous review mechanism, the system needs to implement the corresponding classification control strategy based on the finally determined deviation. For scenarios judged as transient or periodic fluctuations, although the system freezes the model's global weights to prevent erroneous learning, such fluctuations often objectively cause fluctuations in the current maturity assessment score. To avoid misunderstandings by enterprise management regarding these normal score fluctuations, it is necessary to provide clear attribution explanations in the output results.

[0065] In a preferred embodiment of this application, the step of outputting maturity assessment results with environmental interpretation tags includes: If the fluctuation is determined to be transient or periodic, the abnormal fluctuation characteristics in the time series data to be evaluated are extracted to generate environmental interpretation labels. Environmental interpretation markers are attached to the maturity assessment results to explain the environmental causes of fluctuations in assessment scores.

[0066] Abnormal fluctuation characteristics refer to specific business indicators and their changing patterns in the time series data to be evaluated that directly lead to the deviation being classified as transient or cyclical fluctuations. In practice, feature attribution algorithms or attention weight backtracking mechanisms can be used to locate the input dimension that contributes the most to the current implicit representation vector shift. For example, if the current fluctuation is determined to be cyclical, backtracking might reveal that the values ​​of order volume and employee overtime hours significantly exceed historical steady-state benchmarks and are highly consistent with historical data from the same period. In this case, the names, current values, deviation magnitudes, and historical comparison data of these two indicators are extracted as abnormal fluctuation characteristics.

[0067] Subsequently, using pre-set natural language generation templates or rule engines, the extracted abnormal fluctuation features are transformed into structured environmental explanation tags. For example, explanatory text is generated such as: "Due to the seasonal order surge at the end of the year, the order volume in this period increased by 40% month-on-month, and the overtime hours of employees increased by 30%, which is a normal cyclical business fluctuation and not a decline in the enterprise's core digital capabilities."

[0068] Finally, the environmental explanation tag is added as an additional field and output to the front-end display interface or assessment report along with the maturity assessment score.

[0069] On the other hand, when continuous review confirms that the deviation is a genuine trend drift, it indicates that the enterprise's digital operation model has undergone a substantial upgrade. At this point, the evaluation model must learn and adapt to this new model. However, as mentioned earlier, global retraining is not only time-consuming but also disrupts the consistency of historical evaluation benchmarks. Therefore, this application further proposes a step of performing directional parameter updates on the local network branches corresponding to the affected feature subsets in the evaluation model, including: Localize the neuronal connections in the evaluation model that are affected by feature subsets to determine local network branches; Unlock the parameter update permission of local network branches, and use the time series data to be evaluated to perform incremental training on local network branches in order to achieve local weight updates.

[0070] In practical implementation, the neural connections in the localization evaluation model affected by feature subsets can be achieved through network topology mapping. Since the deep learning model described above employs an architecture with multiple parallel feature extraction branches, the data input channels for different modalities are independent. When the affected feature subset is located (e.g., only involving device connectivity rate and automatic reporting coverage in the device operation modality), it is possible to directly trace back to the specific one-dimensional convolutional neural network branch or specific attention head that processes these features. These network layers that directly process the affected features and their internal neural connections are marked as local network branches to be updated.

[0071] After identifying the local network branch, the gradient calculation permissions for the parameters within that branch are unlocked via the gradient control interface of the deep learning framework (e.g., setting the parameter attributes of the corresponding layer to a differentiable state). Simultaneously, the parameters of other unaffected feature extraction branches and the core cross-modal fusion layer and maturity output layer are strictly kept frozen (e.g., setting the parameter attributes of the corresponding layers to a non-differentiable state). Then, the time-series data currently confirmed to have experienced trend drift is used as training samples, and a small number of incremental iterations are performed on the unlocked local network branch using a small learning rate. Training stops when the loss function of the local network branch converges or reaches the preset number of fine-tuning steps, thus completing the update of the local weights.

[0072] Secondly, referring to Figure 2 This application also proposes a deep learning-based digital transformation maturity assessment system, including: Data acquisition module 210 is used to acquire time-series data to be evaluated that characterizes the current level of digital operation of an enterprise, historical steady-state benchmark data that reflects the stable operating benchmark status of an enterprise, and historical cycle reference data that reflects the annual operating patterns of the same period. Feature extraction module 220 is used to extract the time series variation features of the time series data to be evaluated relative to historical steady-state benchmark data and historical periodic reference data; The property determination module 230 is used to determine the deviation properties of the time series data to be evaluated based on the time series change characteristics. The deviation properties include stable small deviation, transient fluctuation, periodic fluctuation and trend drift. The continuous review module 240 is used to perform a continuous review of the time series change characteristics at a preset period if the deviation is determined to be a trend drift. The strategy generation module 250 is used to generate a classification control strategy for a preset evaluation model based on the determination and review results of the deviation nature. The classification control strategy includes: if the deviation nature is determined to be a stable small deviation, the time series data to be evaluated is directly input into the evaluation model, and the maturity evaluation result is calculated and output; if the deviation nature is determined to be a transient fluctuation or a periodic fluctuation, the global weights of the evaluation model are frozen, the time series data to be evaluated is input into the evaluation model after the global weights are frozen, and the maturity evaluation result with environmental interpretation labels is calculated and output; if the review confirms that the deviation nature is a trend drift, the affected feature subset is located, and based on the affected feature subset, the directional parameter update is performed on the local network branch in the evaluation model corresponding to the feature subset, keeping the mapping weights of the evaluation model in a frozen state, the time series data to be evaluated is input into the evaluation model after the local network branch update is completed, and the maturity evaluation result is calculated and output.

[0073] Each functional module in this system corresponds one-to-one with the steps in the aforementioned method embodiments, forming a complete closed-loop processing chain. The data acquisition module 210 ensures accurate alignment of multi-dimensional benchmark data; the feature extraction module 220 and the property determination module 230 achieve deep insight and accurate classification of complex business fluctuations; the continuous review module 240 effectively identifies and intercepts long-term abnormal fluctuations; and the strategy generation module 250 achieves a perfect balance between the stability of the evaluation benchmark and the adaptability of model evolution through differentiated model control methods. This system can be directly deployed on the industrial internet platform or private cloud environment of manufacturing enterprises, providing enterprises with continuous, objective, and interpretable dynamic assessment services for the maturity of digital transformation through software and hardware collaboration.

[0074] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for assessing the maturity of digital transformation based on deep learning, characterized in that, The steps of this method include: Acquire time-series data to be evaluated that characterizes the current level of digital operation of an enterprise, historical steady-state benchmark data that reflects the stable operating baseline status of an enterprise, and historical cycle reference data that reflects the annual operating patterns of the same period; Extract the time-series variation characteristics of the time-series data to be evaluated relative to the historical steady-state baseline data and the historical periodic reference data; Based on the time series change characteristics, the deviation nature of the time series data to be evaluated is determined, including stable small deviation, transient fluctuation, periodic fluctuation, and trend drift. If the deviation is determined to be a trend drift, then a continuous review of the time-series change characteristics is performed for a preset period. Based on the determination and verification results of the deviation characteristics, a classification control strategy is generated for the preset evaluation model. The classification control strategy includes: if the deviation characteristics are determined to be a stable small deviation, the time series data to be evaluated is directly input into the evaluation model, and the maturity evaluation result is calculated and output; if the deviation characteristics are determined to be transient fluctuations or periodic fluctuations, the global weights of the evaluation model are frozen, the time series data to be evaluated is input into the evaluation model with frozen global weights, and the maturity evaluation result with environmental interpretation labels is calculated and output; if the verification confirms that the deviation characteristics are trend drifts, the affected feature subset is located, and based on the affected feature subset, the directional parameter update is performed on the local network branch in the evaluation model corresponding to the feature subset, keeping the mapping weights of the evaluation model in a frozen state, the time series data to be evaluated is input into the evaluation model after the local network branch update is completed, and the maturity evaluation result is calculated and output.

2. The method for assessing the maturity of digital transformation based on deep learning according to claim 1, characterized in that, The steps for obtaining the enterprise's time-series data to be evaluated, historical steady-state baseline data, and historical periodic reference data include: Real-time acquisition of enterprise multimodal operation data, and extraction of current sliding window data from the enterprise multimodal operation data as the time series data to be evaluated; Data from consecutive periods in which the volatility of maturity assessment scores is lower than a preset volatility threshold are selected from the company's historical assessment results and used as the historical steady-state benchmark data. Extract the calendar interval data corresponding to the enterprise's previous year and the current sliding window data, and use it as the historical period reference data.

3. The method for assessing the maturity of digital transformation based on deep learning according to claim 1, characterized in that, The step of extracting the time series variation characteristics of the time series data to be evaluated relative to the historical steady-state baseline data and the historical periodic reference data includes: The time series data to be evaluated is input into a preset deep learning model to extract implicit representation vectors that characterize the enterprise's digital operation mode and maturity status. Calculate the offset of the implicit representation vector in the vector space relative to the historical implicit representation vectors of the historical steady-state baseline data and the historical periodic reference data; The temporal variation features are extracted based on the offset.

4. The method for assessing the maturity of digital transformation based on deep learning according to claim 3, characterized in that, The step of inputting the time series data to be evaluated into a preset deep learning model to extract latent representation vectors includes: The time series data to be evaluated is input into multiple parallel feature extraction branches in the deep learning model to extract local time series features of the time series data to be evaluated in different modal data. The local temporal features are fused, and the fused local temporal features are mapped to the latent representation space to generate the latent representation vector.

5. The method for assessing the maturity of digital transformation based on deep learning according to claim 1, characterized in that, The temporal variation features include the magnitude of change, duration, numerical decline trend, cross-modal feature correlation structure changes, and periodic similarity; The step of determining the deviation nature of the time series data to be evaluated based on the time series change characteristics includes: When the duration is less than a preset duration threshold and the downward trend of the value is greater than a preset downward trend threshold, the deviation is determined to be a transient fluctuation. When the periodic similarity is greater than a preset similarity threshold, the deviation is determined to be a periodic fluctuation. When the duration exceeds the preset duration threshold and the cross-modal feature association structure change exceeds the preset structural mutation threshold, the deviation is determined to be the trend drift. Otherwise, the deviation is determined to be a stable, small-amplitude deviation.

6. The method for assessing the maturity of digital transformation based on deep learning according to claim 5, characterized in that, The steps for performing continuous review at preset intervals on deviation events determined to be the aforementioned trend drift include: The evolution trajectory of the temporal change features is tracked within a preset number of sliding windows; If the change amplitude remains within a preset stable range within the sliding window and the cross-modal feature association structure change shows a consistent diffusion trend, then the determination result of the trend drift is confirmed to be true. If the downward trend of the value within the sliding window is greater than the preset downward threshold or the cross-modal feature correlation structure change returns to the historical steady state, then the determination result of the trend drift is confirmed as false, and the determination result is corrected to the transient fluctuation.

7. The method for assessing the maturity of digital transformation based on deep learning according to claim 5, characterized in that, Before determining the deviation nature of the time series data to be evaluated based on the time series variation characteristics, the method further includes: Calculate the distribution variance of the historical steady-state benchmark data, and calculate the preliminary deviation tolerance threshold based on the distribution variance; If the change is less than the initial deviation tolerance threshold, it is determined that no feature drift has occurred, and the time series data to be evaluated is directly input into the evaluation model to calculate and output the maturity evaluation result. If the magnitude of the change is greater than or equal to the initial deviation tolerance threshold, the deviation determination step is triggered.

8. The method for assessing the maturity of digital transformation based on deep learning according to claim 1, characterized in that, The step of outputting maturity assessment results with environmental interpretation tags includes: If the fluctuation is determined to be either transient or periodic, then the abnormal fluctuation features in the time series data to be evaluated are extracted to generate an environmental interpretation label. The environmental explanation markers are attached to the maturity assessment results to explain the environmental factors that cause fluctuations in the assessment scores.

9. The method for assessing the maturity of digital transformation based on deep learning according to claim 1, characterized in that, The step of performing directional parameter updates on the local network branches corresponding to the affected feature subset in the evaluation model includes: Locate the neuronal connections in the evaluation model that are affected by the feature subset to determine the local network branches; Unlock the parameter update permission of the local network branch, and use the time series data to be evaluated to perform incremental training on the local network branch to achieve local weight update.

10. A digital transformation maturity assessment system based on deep learning, characterized in that, include: The data acquisition module is used to acquire time-series data that characterizes the current level of digital operation of an enterprise, historical steady-state benchmark data that reflects the stable operating benchmark status of an enterprise, and historical cycle reference data that reflects the annual operating patterns of the same period. The feature extraction module is used to extract the time-series variation features of the time-series data to be evaluated relative to the historical steady-state benchmark data and the historical periodic reference data; The property determination module is used to determine the deviation properties of the time series data to be evaluated based on the time series change characteristics. The deviation properties include stable small deviation, transient fluctuation, periodic fluctuation, and trend drift. The continuous review module is used to perform a continuous review of the time-series change characteristics at a preset period if the deviation is determined to be a trend drift. The strategy generation module is used to generate a classification control strategy for a preset evaluation model based on the determination result and the verification result of the deviation nature. The classification control strategy includes: if the deviation nature is determined to be a stable small deviation, the time series data to be evaluated is directly input into the evaluation model, and the maturity evaluation result is calculated and output. If the deviation is determined to be either transient fluctuation or periodic fluctuation, the global weights of the evaluation model are frozen, the time series data to be evaluated is input into the evaluation model after the global weights are frozen, and the maturity evaluation result with environmental interpretation labels is calculated and output. If the deviation is confirmed to be a trend drift, the affected feature subset is located, and based on the affected feature subset, the directional parameter update is performed on the local network branch in the evaluation model corresponding to the feature subset, keeping the mapping weights of the evaluation model in a frozen state, the time series data to be evaluated is input into the evaluation model after the local network branch update is completed, and the maturity evaluation result is calculated and output.

Images