An intelligent terminal method and system for self-evaluation of data quality for an industrial cloud platform

Abstract

The application discloses an industrial cloud platform-oriented data quality self-evaluation intelligent terminal method and system. The intelligent terminal collects industrial field scene context information, generates a scene fingerprint vector through a lightweight scene encoder, and matches a historical quality baseline. A double-channel quality evaluation combining a fast channel and a deep channel is performed on the data to be evaluated. According to the evaluation result, the network state and the business priority, a data processing decision is generated through a transmission decision state machine. An incremental update is performed on the quality baseline through a sliding statistical window, and model parameter merging is performed in cooperation with the cloud. The application makes the evaluation standard dynamically adjusted according to the industrial field, meets the terminal real-time performance and resource constraints, and realizes a data quality self-evaluation and processing closed loop.

Description

Technical Field

[0001] This application relates to artificial intelligence algorithms and industrial Internet of Things, specifically to a data quality self-assessment smart terminal method and system for industrial cloud platforms. Background Technology

[0002] With the rapid development of Industrial Internet of Things (IIoT) technology, sensors and smart terminals distributed across various industrial production sites continuously collect multi-source industrial data such as temperature, pressure, vibration, and current, and transmit this data to industrial cloud platforms via networks for use in upper-level applications such as production analysis, predictive equipment maintenance, process optimization, and quality control. Data quality directly determines the reliability and effectiveness of these applications; therefore, data quality assessment has become a crucial link in the data governance system of industrial cloud platforms. Currently, industrial cloud platforms generally adopt a centralized cloud-based data quality management architecture, where smart terminals only handle data acquisition and transmission, while data quality detection, assessment, and processing are all completed by a centralized system deployed in the cloud. Typical cloud-based data quality assessment methods include rule-based quality detection methods, machine learning model-based anomaly detection methods, and multi-dimensional data quality assessment frameworks. However, this centralized architecture of "transmit first, evaluate later" has several technical drawbacks: low-quality data has already completed the entire link transmission from the terminal to the cloud before it is identified, consuming valuable network bandwidth and cloud storage resources; there is an unavoidable time delay in the transmission of data from the terminal to the cloud and the completion of evaluation, making it impossible to detect quality anomalies as soon as the data is generated; when the cloud discovers data quality problems, since the data has been removed from the original environment in which it was generated, tracing and locating the source of the problem requires a lot of time and manual intervention.

[0003] In the field of edge computing, some solutions have attempted to offload data preprocessing functions to edge gateways or edge servers, including data cleaning, deduplication, format conversion, and preliminary filtering. However, these edge preprocessing solutions still primarily rely on format conformity checks and simple threshold filtering, which are single-dimensional deterministic rule judgments and do not involve a systematic, multi-dimensional assessment of data quality. When the degree of data anomaly is in a gray area—that is, data indicators do not significantly exceed thresholds but have deviated from normal patterns—relying solely on deterministic rules is insufficient for accurate judgment, lacking the ability to use deep learning models to finely identify uncertain data. Furthermore, existing edge solutions only complete data preprocessing and then upload it uniformly, without empowering terminal devices to autonomously decide on data processing methods based on evaluation results. The evaluation and transmission decision-making processes remain separate, and the terminal side cannot form a complete autonomous closed loop from quality assessment to transmission processing.

[0004] Furthermore, existing data quality assessment methods all employ predefined static assessment standards and fixed rule thresholds. However, the operating environment of industrial sites is significantly dynamic and heterogeneous. Equipment operating status changes with process stage switching, environmental physical parameters fluctuate with seasons and operating conditions, and data characteristics differ between different production batches. Under these dynamic and variable conditions, the same set of assessment thresholds may be applicable in some process stages but ineffective in others, leading to an increased false positive rate. Current technologies lack a mechanism capable of sensing the specific operating scenario of the current industrial site and dynamically adjusting data quality assessment standards accordingly; the disconnect between assessment standards and actual scenarios has not yet been effectively resolved.

[0005] The aforementioned technical challenges are particularly pronounced in specific industrial application scenarios. In network-constrained environments, such as offshore platforms and remote mining areas, network bandwidth is extremely limited and connections are unstable, making the transmission of invalid data extremely costly, thus necessitating data quality screening at the terminal side. In ultra-large-scale industrial IoT deployments, such as large infrastructure monitoring systems containing tens of thousands of sensor nodes, centralized cloud-based assessment faces severe computing power bottlenecks and real-time challenges. In scenarios with high security requirements, data needs to undergo quality verification before entering the cloud system. In dynamic production environments such as flexible manufacturing, frequent product switching and process adjustments necessitate changes in data quality standards, while cloud systems struggle to obtain real-time information on the specific process stage and equipment status of each terminal. These scenarios present an urgent technical need for autonomous, real-time data quality assessment at the terminal side that can adapt to changing scenarios. Summary of the Invention

[0006] To address the technical problems in existing technologies, such as reliance on centralized cloud processing for data quality assessment in industrial cloud platforms, lack of autonomous assessment and decision-making capabilities in smart terminals, and inability to dynamically adjust assessment standards according to industrial site scenarios, this invention provides a data quality self-assessment smart terminal method and system for industrial cloud platforms. It constructs a scenario fingerprint-driven dynamic quality baseline generation mechanism, a lightweight dual-channel data quality assessment architecture, and a quality-aware autonomous transmission decision engine on resource-constrained smart terminals, realizing a closed loop of real-time assessment and autonomous processing of data quality on the terminal side.

[0007] One aspect of the present invention provides a data quality self-assessment smart terminal method for industrial cloud platforms, comprising the following steps: Step S1, the smart terminal collects scene context information of the current industrial site, generates a scene fingerprint vector by performing nonlinear mapping on the scene context information through a lightweight scene encoder, performs cosine similarity matching between the scene fingerprint vector and historical fingerprints in the scene fingerprint index table, and retrieves historical quality baseline parameters or enables default quality baseline parameters based on the matching result; Step S2, performs dual-channel quality assessment on the data to be assessed, with the fast channel performing multi-dimensional quality checks based on deterministic rules according to the quality baseline parameters obtained in step S1 and outputting the results. A rapid quality comprehensive score is generated. When the rapid quality comprehensive score falls within the grayscale range between the upper and lower thresholds, the deep channel is activated. The deep channel outputs an anomaly probability vector through a lightweight temporal anomaly detection model and merges it with the rapid channel result to generate a multi-dimensional quality assessment result. Step S3: Based on the multi-dimensional quality assessment result, network status information, and data service priority, a data processing decision is generated through a transmission decision state machine. Step S4: Data distribution changes are tracked through a sliding statistical window, and incremental updates are performed on the quality baseline parameters. Step S5: The quality assessment statistical summary is uploaded to the cloud, and the model update package sent from the cloud is received for local model parameter merging.

[0008] Specifically, in step S1, the scene context information includes equipment operating status parameters, environmental physical parameters, current process stage identifier, and cumulative equipment runtime. The lightweight scene encoder adopts a two-layer fully connected network structure, with the middle layer using the ReLU activation function and the output layer using L2 normalization to ensure the scene fingerprint vector falls on a unit hypersphere. The intelligent terminal maintains a scene context acquisition buffer and periodically reads the scene context information from the local sensor interface and the equipment control bus. The equipment operating status parameters include real-time operating indicators such as motor speed, bearing temperature, power supply voltage, and current; the environmental physical parameters include on-site physical quantities such as ambient temperature, humidity, and vibration amplitude; the current process stage identifier is sent by the upper control system through a communication protocol; and the cumulative equipment runtime is maintained by the terminal's local timer. The intelligent terminal normalizes each type of parameter and concatenates them into an original context vector. Subsequently, the lightweight scene encoder performs a nonlinear mapping on the original context vector to output a fixed-dimensional scene fingerprint vector. In the lightweight scene encoder's two-layer fully connected network, the input layer dimension is consistent with the original context vector dimension, the intermediate layer dimension is set to between half and twice the input dimension, and the output layer dimension is fixed to the preset scene fingerprint dimension. L2 normalization ensures that fingerprint vectors from different scenes are comparable on a unit hypersphere. Further, in step S1, the scene fingerprint index table stores the mapping relationship between historical scene fingerprints and corresponding quality baseline parameters. When the highest similarity of the cosine similarity match exceeds the matching threshold, the corresponding historical quality baseline parameter is retrieved as the current quality baseline parameter. When the highest similarity is lower than the matching threshold, the current scene is determined to be a new scene, and the default quality baseline parameter is enabled. Weighted interpolation is performed using the K historical scene quality baseline parameters with the highest cosine similarity to the current scene fingerprint in the scene fingerprint index table to generate an initial baseline estimate. During subsequent data accumulation, online gradient updates converge to a stable baseline. The weighted interpolation uses cosine similarity as the weight to perform a weighted average of the quality baseline parameters of the K nearest neighbor historical scenes, where K ranges from 3 to 5. In one implementation, the matching threshold is set to a value between 0.85 and 0.95 to achieve a balance between scene recognition accuracy and generalization ability. The quality baseline parameters include the expected value range for each data dimension, a data change rate threshold, an inter-dimensional correlation coefficient matrix, and sampling integrity requirements. The expected value range for each data dimension is represented by the mean and standard deviation. The data change rate threshold is used to detect abrupt changes in the values ​​of each dimension between adjacent sampling frames. The inter-dimensional correlation coefficient matrix records the expected Pearson correlation coefficients between each data dimension. The sampling integrity requirements define the set of fields that must be included in each data frame and their allowed upper limit for the proportion of missing fields.

[0009] Specifically, in step S2, the multi-dimensional quality check of the fast channel includes numerical range verification, sampling integrity verification, timestamp continuity verification, and inter-dimensional correlation verification. The results of each check are quantified into sub-dimensional quality scores and obtained by weighted summation to obtain the fast comprehensive quality score. When the fast comprehensive quality score is higher than the upper threshold, the data quality is determined to be qualified and the process directly proceeds to step S3. When it is lower than the lower threshold, the abnormal type is directly marked and the process proceeds to step S3. When it is in the grayscale range, the depth channel is activated. The numerical range verification compares the values ​​of each dimension of the data to be evaluated with the corresponding expected value range in the dynamic quality baseline. Data dimensions that exceed the range are marked as abnormal and the deviation is calculated. The deviation is defined as the ratio of the difference between the actual value and the expected range boundary to the width of the expected range. The sampling integrity verification checks for the absence of necessary fields in the data frame. When the number of missing fields exceeds the upper limit of the allowed missing proportion in the sampling integrity requirements, it is marked as incomplete sampling. The timestamp continuity verification detects whether the time interval between adjacent data frames exceeds the allowed sampling period deviation. The allowed deviation range is usually set to ±10% to 20% of the nominal sampling period. The inter-dimensional correlation verification uses the correlation coefficient matrix in the quality baseline to calculate the deviation between the actual correlation between each dimension of the current data and the baseline correlation. When the deviation exceeds the preset tolerance, it is marked as an abnormal inter-dimensional relationship. In the weighted summation of the quality scores of each sub-dimension, the weights are configured according to the importance of each checked dimension in the current scenario. The weight parameters are stored in the quality baseline and adjusted synchronously with the baseline update. Furthermore, the lightweight temporal anomaly detection model employs a one-dimensional deep separable convolutional network structure. It uses a sliding window composed of the data to be evaluated and several preceding frames as input. It captures temporal dependencies through deep convolution, achieves cross-dimensional feature fusion through pointwise convolution, and outputs the anomaly probability of each dimension through a Sigmoid layer. The number of parameters in the lightweight temporal anomaly detection model is controlled to within 50,000. The one-dimensional deep separable convolutional network structure decomposes standard convolution into two stages: deep convolution and pointwise convolution. Deep convolution independently performs one-dimensional convolution in the temporal direction on each data dimension channel to capture the temporal dependency patterns of each dimension. Pointwise convolution achieves cross-dimensional feature interaction and fusion through a one-to-one convolution kernel. The lightweight temporal anomaly detection model undergoes knowledge distillation (a model compression technique that transfers knowledge from a large model to a small model). Using a complete temporal anomaly detection teacher model trained in the cloud as the knowledge source, soft label distillation transfers the anomaly discrimination capability of the teacher model to the one-dimensional deep separable convolutional network structure on the terminal side. In one implementation, the inference latency of the lightweight temporal anomaly detection model is controlled to be within 10 milliseconds, and the sliding window length is set to 8 to 32 frames to achieve a balance between temporal context information and computational overhead.The anomaly probability vector output from the deep channel is weighted and fused with the sub-dimensional quality scores from the fast channel to generate the final multi-dimensional quality assessment result, including the overall quality score, anomaly labels for each dimension, and anomaly confidence.

[0010] Specifically, in step S3, the data processing decision includes four types: direct upload, compressed upload, local cache pending upload, and discard with alarm. Direct upload is performed when the overall quality score in the multi-dimensional quality assessment result is higher than the quality pass threshold and network bandwidth is sufficient. Compressed upload is performed when the overall quality score is higher than the quality pass threshold but network bandwidth is limited. Data is written to a local circular cache along with anomaly markers and scenario fingerprint information when the overall quality score is lower than the quality pass threshold but higher than the discard threshold. Data is discarded and an alarm record is generated when the overall quality score is lower than the discard threshold and the anomaly type belongs to the data corruption category. The input to the transmission decision state machine includes the overall quality score, anomaly type marker, estimated current available network bandwidth, and priority weight of the service type to which the data belongs. The estimated network bandwidth is obtained by periodically probing the available throughput of the current network link. For compressed uploads, lossy compression is used to reduce data volume before transmission, and the compression strategy selects an appropriate encoding method based on the data type. For locally cached data to be transmitted, its anomaly marker and scene fingerprint information are written into the local circular buffer. It is then uploaded when the network recovers or the cloud actively retrieves it. The inclusion of the scene fingerprint information allows the cloud to reconstruct the scene context at the time of the delay upon receiving the data. Furthermore, for critical data with the highest business priority, even if the overall quality score is low, the transmission decision state machine still adds a complete anomaly label before uploading. The local circular buffer employs a priority eviction strategy based on the overall quality score; when the buffer is full, the data with the lowest overall quality score is evicted first. The complete anomaly label includes the anomaly type, anomaly dimension, anomaly probability, and corresponding scene fingerprint information, enabling the cloud to trace and analyze critical anomaly events. In one implementation, the capacity of the local circular buffer is dynamically configured based on the storage resources of the terminal device. The data in the buffer is organized into a circular queue structure according to the writing order. When new data is written, if the buffer is full, space is freed up according to the priority eviction strategy based on the overall quality score.

[0011] Specifically, in step S4, the smart terminal maintains a sliding statistical window. At the end of each window period, it statistically analyzes the changing trends of the mean, variance, rate of change distribution, and correlation coefficients between dimensions for each data dimension under the current scene fingerprint. When the deviation between the statistical features and the current quality baseline parameters exceeds a preset update trigger threshold, the smart terminal performs an exponentially weighted moving average update on the corresponding quality baseline parameters, integrating the new statistical features with a certain weight into the existing baseline. The update formula is B_new=alpha×S_current+(1-alpha)×B_old, where B_new is the updated baseline parameter, B_old is the current baseline parameter, S_current is the current window statistical feature, and alpha is the update weight coefficient. Simultaneously, the smart terminal updates the quality baseline parameters of the corresponding entries in the scene fingerprint index table, ensuring that the mapping relationship between the scene fingerprint and the quality baseline remains consistent with the current scene state. The window length of the sliding statistical window is dynamically configured according to the data sampling frequency and the scene change rate. Data frames within the window are organized in chronological order, and the window sliding step is half the window length to achieve overlapping statistics between adjacent windows. In one implementation, the update trigger threshold and the weight coefficient alpha of the exponentially weighted moving average can be adjusted by the cloud based on the global data distribution analysis results, so as to realize the macro-control of the baseline update strategy on the terminal side by the cloud. The typical value range of alpha is 0.05 to 0.3.

[0012] Specifically, in step S5, the model update package includes incremental parameter updates and quality baseline adjustment suggestions for the lightweight temporal anomaly detection model. The smart terminal uses a parameter interpolation strategy to perform a weighted average of the global model parameters sent from the cloud and the local model parameters according to a preset ratio. The quality baseline adjustment suggestions are only accepted when the adjustment range does not exceed the preset ratio of the local baseline parameters. The quality assessment statistical summary includes data quality distribution statistics, anomaly type frequency distribution, and baseline update logs under each scenario fingerprint. The smart terminal periodically uploads the quality assessment statistical summary to the cloud, and the upload period is dynamically adjusted according to network conditions and data change frequency. The cloud performs global analysis based on the assessment statistical summaries uploaded by multiple terminals, aggregates similar scenario data from different terminals, trains the updated global anomaly detection model, and generates the model update package. For incremental parameter updates of the deep channel model, the parameter interpolation strategy weights the global model parameters distributed from the cloud and the local model parameters by a preset ratio of beta. The update formula is W_local_new = beta × W_global + (1-beta) × W_local_old, where beta ranges from 0.1 to 0.5. A smaller beta value allows the terminal model to retain more local scene specificity, while a larger beta value allows the terminal model to absorb global knowledge more quickly. For quality baseline adjustment suggestions, the smart terminal compares the difference between the cloud-suggested value and the local baseline parameter item by item, and only accepts the adjustment when the difference does not exceed a preset ratio of the local baseline parameter, avoiding undue interference of global statistical bias on the baseline of a specific scene. The typical value of the preset ratio is 15% to 30%.

[0013] In another aspect, the present invention provides a data quality self-assessment intelligent terminal system for industrial cloud platforms, comprising: a scene perception module, used to collect scene context information of the current industrial site, generate a scene fingerprint vector by performing nonlinear mapping on the scene context information through a lightweight scene encoder, perform cosine similarity matching between the scene fingerprint vector and historical fingerprints in a scene fingerprint index table, and retrieve historical quality baseline parameters or enable default quality baseline parameters based on the matching result; and a dual-channel assessment module, used to perform dual-channel quality assessment on the data to be assessed, wherein the fast channel performs multi-dimensional quality checks based on the quality baseline parameters obtained by the scene perception module and outputs a fast quality assessment result. The system comprises several modules: a comprehensive quality score, a deep channel (activated when the comprehensive quality score falls within a grayscale range between an upper and lower threshold), a lightweight temporal anomaly detection model (outputting an anomaly probability vector, which is then fused with the fast channel result to generate a multi-dimensional quality assessment result); a transmission decision module (generating data processing decisions via a transmission decision state machine based on the multi-dimensional quality assessment result, network status information, and data service priority); an incremental update module (tracking data distribution changes via a sliding statistical window and incrementally updating quality baseline parameters); and a cloud collaboration module (uploading quality assessment statistical summaries to the cloud and receiving model update packages from the cloud for local model parameter merging). The scene perception module maintains a scene fingerprint index table, storing the mapping relationship between historical scene fingerprints and corresponding quality baseline parameters, supporting historical quality baseline parameter retrieval, enabling default quality baseline parameters for new scenes, and initializing baseline estimation. The dual-channel evaluation module includes a fast channel submodule and a deep channel submodule. The fast channel submodule performs numerical range verification, sampling integrity verification, timestamp continuity verification, and inter-dimensional correlation verification based on dynamic quality baseline parameters, and outputs a rapid comprehensive quality score. The deep channel submodule deploys a lightweight temporal anomaly detection model to perform fine evaluation of grayscale data and outputs anomaly probability vectors. The transmission decision module has a built-in transmission decision state machine. Based on the comprehensive quality score, anomaly type label, network bandwidth estimate, and service priority weight in the multi-dimensional quality evaluation results, it generates data processing decisions such as direct upload, compressed upload, local cache pending transmission, or discard with alarm, and manages data writing and priority eviction in the local circular buffer. The incremental update module calculates the statistical characteristic changes of each data dimension at the end of the sliding statistical window period, performs an exponentially weighted moving average update on the quality baseline parameters, and synchronously updates the scenario fingerprint index table. The cloud collaboration module periodically uploads quality evaluation statistical summaries, receives model update packages from the cloud, and performs local model parameter merging and quality baseline calibration through parameter interpolation strategies.

[0014] This invention addresses the problem of static evaluation standards failing to adapt to dynamic industrial environments by constructing a scene fingerprint-driven dynamic quality baseline generation mechanism on the smart terminal side. This allows data quality assessment standards to adaptively adjust to changes in equipment status, environmental conditions, and process stages in the industrial environment. The invention employs a dual-channel evaluation architecture combining a fast channel and a deep channel. The fast channel completes the quality judgment of most data with extremely low computational overhead, while the deep channel is activated for fine-grained evaluation only for uncertain data in the grayscale range. This ensures evaluation accuracy while meeting the real-time and resource constraints of terminal devices. The invention achieves autonomous decision-making for data processing methods through a transmission decision state machine, comprehensively judging based on data quality, network status, and business priority. This avoids low-quality data consuming limited transmission bandwidth and cloud storage resources. Simultaneously, a mandatory upload mechanism for key data and an anomaly labeling mechanism ensure timely cloud-based detection of critical anomalies. Through terminal-side incremental updates and cloud-based collaboration mechanisms, the invention enables continuous evolution of the terminal model and quality baseline. Parameter interpolation strategies and baseline adjustment range limitation mechanisms balance local scene specificity with global knowledge sharing, forming a closed-loop data quality management system with end-to-cloud collaboration. Attached Figure Description

[0015] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0016] Figure 1 This is a schematic diagram of the overall process of the intelligent terminal method for self-assessing data quality for industrial cloud platforms provided in this embodiment of the invention.

[0017] Figure 2 This is a schematic diagram of the data flow of the dual-channel data quality assessment architecture provided in this embodiment of the invention.

[0018] Figure 3 This is a schematic diagram of the network structure of the lightweight temporal anomaly detection model provided in an embodiment of the present invention.

[0019] Figure 4 This is a block diagram of the module structure of the intelligent terminal system for data quality self-assessment for industrial cloud platforms provided in this embodiment of the invention.

[0020] Figure 5 This is a visualization of the dynamic quality baseline mapping space driven by fingerprint in the scenario of this invention.

[0021] Figure 6 This is a comparison chart showing the trade-off between computational resources and accuracy in the dual-channel evaluation of this invention.

[0022] Figure 7 This is a diagram illustrating the edge-cloud topology and data interaction scenario of the multi-terminal collaborative industrial cloud platform of this invention.

[0023] Figure 8This is a comparative analysis chart of multiple operating conditions for assessing the accuracy of data quality in this invention.

[0024] Figure 9 This is a comprehensive analysis diagram of terminal-side computing resource usage and processing latency in this invention.

[0025] Figure 10 This is a graph showing the combined analysis of network bandwidth saving effect and data quality distribution of the present invention. Detailed Implementation

[0026] To make the objectives and technical solutions of this invention clearer, the embodiments of this invention will be further described in detail below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of this invention and should not be construed as limiting the scope of protection of this invention.

[0027] Example 1

[0028] This embodiment provides a data quality self-assessment smart terminal method for industrial cloud platforms, such as... Figure 1 As shown, this method is deployed on resource-constrained industrial intelligent terminals. It achieves a closed-loop system for real-time data quality assessment and autonomous processing at the terminal side through five steps: scene fingerprint-driven dynamic quality baseline generation, lightweight dual-channel data quality assessment, quality-aware autonomous transmission decision-making, terminal-side incremental baseline updates, and cloud-based collaborative model synchronization. The method includes the following steps:

[0029] Step S1: Construct scene fingerprint and generate dynamic quality baseline.

[0030] The intelligent terminal maintains a scene context acquisition buffer, reading the current industrial scene context information from the local sensor interface and device control bus at a fixed acquisition cycle. This scene context information includes four types of parameters: equipment operating status parameters covering real-time operating indicators such as motor speed, bearing temperature, power supply voltage, and current; environmental physical parameters covering on-site physical quantities such as ambient temperature, humidity, and vibration amplitude; the current process stage identifier, issued by the upper-level control system via communication protocols such as Modbus or OPC UA, used to identify the specific process stage of the current production line; and the cumulative equipment runtime, maintained by the terminal's local timer, reflecting the equipment's current operational fatigue state. In each acquisition cycle, the intelligent terminal writes these four types of parameters into the acquisition buffer, which employs a fixed-capacity first-in-first-out queue structure to ensure that the latest scene context snapshot is always preserved.

[0031] The intelligent terminal performs normalization processing on various parameters in the acquisition buffer. For continuous numerical parameters, a minimum-maximum normalization method is used to map them to the zero-to-one interval. The normalization formula is x_norm=(x-x_min) / (x_max-x_min), where x_min and x_max are the pre-configured range boundary values ​​for each parameter. For discrete parameters, such as process stage identifiers, one-hot encoding is used to convert them into numerical vectors. The normalized parameters are concatenated into an original context vector in a preset order. The dimension of this vector is equal to the total dimension of all normalized parameter components. In this embodiment, the dimension of the original context vector is set to 16 dimensions.

[0032] The smart terminal uses a lightweight scene encoder deployed on the terminal side to perform nonlinear mapping on the original context vector, outputting a fixed-dimensional scene fingerprint vector. The lightweight scene encoder employs a two-layer fully connected network structure: an input layer with a dimension of 16, an intermediate layer with a dimension of 24, and an output layer with a fixed dimension of 32, resulting in a 32-dimensional scene fingerprint vector. The intermediate layer uses the ReLU activation function to introduce nonlinear expressive power, and the output layer uses L2 normalization to project the scene fingerprint vector onto a unit hypersphere, ensuring that fingerprint vectors generated from different scenes are comparable in direction regardless of differences in vector magnitude. The encoder has approximately 1200 parameters, resulting in minimal computational load per inference on the terminal side, thus not burdening the terminal device's computing resources.

[0033] The smart terminal maintains a scene fingerprint index table locally. This index table stores the mapping relationship between historical scene fingerprints and corresponding quality baseline parameters using a hash mapping method. The smart terminal calculates the cosine similarity between the currently generated scene fingerprint vector and all historical fingerprints in the index table one by one. The cosine similarity calculation formula is cos_sim, which is equal to the dot product of vector A and vector B divided by the modulus of vector A multiplied by the modulus of vector B. When the highest similarity exceeds the matching threshold, the quality baseline parameter corresponding to that historical fingerprint is retrieved as the current quality baseline parameter. The matching threshold is set to 0.92 in this embodiment. When the highest similarity is lower than the matching threshold, the current scene is determined to be a new scene. The smart terminal uses the preset default quality baseline parameter as the initial baseline. At the same time, it uses the K historical scene quality baseline parameters with the highest cosine similarity to the current scene fingerprint in the scene fingerprint index table to perform weighted interpolation to generate an initial baseline estimate. In this embodiment, K is set to 4. The weighted interpolation uses cosine similarity as the weight to perform a weighted average of the quality baseline parameters of the K nearest neighbor historical scenes, so that the initial baseline estimate can incorporate historical experience of similar scenes. The quality baseline parameters include the expected range of each data dimension (characterized in the form of mean and standard deviation), the data change rate threshold, the correlation coefficient matrix between dimensions, and the sampling integrity requirements.

[0034] like Figure 5 As shown, the scene fingerprint mapping space uses a two-dimensional network topology to reduce the dimensionality of the high-dimensional scene fingerprint vector space for visualization. The figure shows the distribution of historical scene fingerprint points for four types of process stages: fingerprint points in the idle scene (101, 102) are represented by circular nodes; fingerprint points in the low-load scene (201, 202) are represented by square nodes; fingerprint points in the full-load scene (301, 302) are represented by triangular nodes; and fingerprint points in the preheating scene (401, 402) are represented by diamond nodes. Nodes are filled with different shades of gray to further distinguish them. The thickness of the lines connecting scene fingerprint points represents the level of cosine similarity. Scenes with high similarity are connected by thick solid lines, reflecting the clustering characteristics of similar scenes.

[0035] The matching threshold circle (dashed circle) defines the matching range corresponding to the cosine similarity threshold θ=0.92. Historical scene fingerprint points within this range can be used as candidate baseline references. When a new scene arrives (marked as a star node of the new scene), the system calculates its cosine similarity with its K nearest neighbor historical scene fingerprint points. The K nearest neighbor relationship is indicated by connecting them with dashed lines, and weighting coefficients (w=0.35, w=0.42, w=0.23) are marked at the midpoint of the connection. Based on these weights, the historical quality baseline parameters are weighted and interpolated to obtain the dynamic quality baseline.

[0036] The scene fingerprint index table (the rectangle in the lower left corner) is linked to representative historical scene fingerprint points by dashed lines, demonstrating the mapping relationship between scene fingerprints and quality baseline parameters. Next to the fingerprint points for idle, low-load, and full-load scenarios, the corresponding quality baseline parameter ranges (Qmin, Qavg) are floating labels. These parameters are obtained by incrementally updating historical data through the sliding statistical window, supporting dynamic threshold adjustment in dual-channel evaluation.

[0037] Step S2: Perform a lightweight dual-channel data quality assessment.

[0038] like Figure 2As shown, after receiving the currently collected data to be evaluated, the smart terminal first sends it to the fast track for preliminary quality screening. The fast track, based on the dynamic quality baseline parameters generated in step S1, performs four multi-dimensional quality checks using deterministic rules. The numerical range check compares the values ​​of each dimension of the data to be evaluated with the corresponding expected value range in the quality baseline. For each data dimension, the deviation between the current value and the expected mean is calculated. The deviation is defined as the ratio of the difference between the actual value and the expected range boundary to the width of the expected range. Data dimensions with a deviation exceeding 1.0 are marked as abnormal. The sampling integrity check checks for missing fields in the data frame, comparing the actual number of missing fields with the upper limit of the allowed missing proportion defined in the sampling integrity requirements. When the missing proportion exceeds the upper limit, it is marked as incomplete sampling. The timestamp continuity check detects whether the time interval between adjacent data frames exceeds the allowed sampling period deviation. In this embodiment, the allowed deviation range is set to ±15% of the nominal sampling period. Interdimensional correlation verification uses the correlation coefficient matrix in the quality baseline to calculate the actual correlation between each dimension of the current data using the Pearson correlation coefficient, and compares the difference with the baseline correlation. When the absolute value of the difference exceeds the tolerance of 0.15, it is marked as an anomaly in the interdimensional relationship.

[0039] The fast track quantifies the results of the four checks into sub-dimensional quality scores, with each score ranging from zero to one, where one indicates that the quality of that dimension is completely normal. A weighted summation is then used to obtain a rapid overall quality score. The weights are configured according to the importance of each check dimension in the current scenario, and these weight parameters are stored in the quality baseline and adjusted synchronously with baseline updates. In this embodiment, the weight for numerical range verification is 0.35, the weight for sampling integrity verification is 0.25, the weight for timestamp continuity verification is 0.20, and the weight for inter-dimensional correlation verification is 0.20.

[0040] The fast channel performs a three-way splitting judgment based on the fast quality comprehensive score. When the fast quality comprehensive score is higher than the upper threshold (set to 0.85 in this embodiment), the data quality is deemed qualified, and it directly proceeds to the transmission decision stage in step S3. When the fast quality comprehensive score is lower than the lower threshold (set to 0.40 in this embodiment), the data is deemed to have obvious quality problems, and the anomaly type is directly marked and proceeds to step S3. When the fast quality comprehensive score is in the grayscale range between the upper and lower thresholds, the deep channel is activated for fine evaluation. In actual industrial scenarios, most data can be directly judged by the fast channel, and only about 10% to 20% of the data needs to be activated by the deep channel, thereby achieving efficient utilization of computing resources at the overall level.

[0041] Deploy a lightweight temporal anomaly detection model in the deep channel, such as Figure 3As shown, the model employs a one-dimensional depthwise separable convolutional network structure. During the inference phase, the model uses a sliding window consisting of the data to be evaluated and several preceding frames as input. The sliding window length is set to 16 frames in this embodiment. The input data is first processed by a depthwise convolutional layer. The depthwise convolution performs a one-dimensional convolution operation in the temporal direction independently on each data dimension channel of the input. The convolutional kernel length is 5, used to capture the temporal dependency patterns and short-term trends of each dimension. The output of the depthwise convolution then enters a pointwise convolutional layer. The pointwise convolution achieves cross-dimensional feature interaction and fusion through a one-to-one convolutional kernel, integrating the independent temporal features of each dimension into a comprehensive feature representation coupled between dimensions. Finally, the network compresses the temporal dimension through a global average pooling layer, and then outputs the anomaly probability of each data dimension through a fully connected layer and a sigmoid activation layer. The anomaly probability ranges from zero to one; the closer to one, the higher the probability of an anomaly in that dimension. The model has 47,000 parameters, and the single inference latency is about 7.3 milliseconds on an embedded processor with a main frequency of 400MHz, which meets the real-time requirements of terminal devices.

[0042] The anomaly probability vector output from the deep channel is weighted and fused with the sub-dimensional quality scores from the fast channel. The fusion formula is Q_final_i = w_f × S_fast_i + w_d × (1 - P_anomaly_i), where Q_final_i is the final quality score of the i-th dimension, S_fast_i is the quality score of the i-th dimension from the fast channel, P_anomaly_i is the anomaly probability of the i-th dimension output from the deep channel, and w_f and w_d are the fusion weights of the fast channel and deep channel, respectively. In this embodiment, w_f is 0.4 and w_d is 0.6. After fusion, the final multi-dimensional quality assessment result is generated, including the comprehensive quality score, anomaly labels for each dimension, and anomaly confidence.

[0043] The lightweight time-series anomaly detection model is trained using a knowledge distillation framework. The teacher model is a complete time-series anomaly detection model deployed in the cloud. This teacher model employs a multi-layer bidirectional long short-term memory network structure with approximately 2.8 million parameters and has been thoroughly trained on a large-scale industrial time-series anomaly detection dataset. The training dataset contains historical sensor data collected from multiple industrial sites, covering normal operation data and manually annotated data of various anomaly types. The annotation method employs a semi-supervised annotation process combining domain experts and statistical analysis tools. The total size of the dataset is approximately 1.2 million time-series samples. During the data preprocessing stage, missing value imputation, outlier truncation, and sliding window splitting are performed on the raw collected data. The training, validation, and test sets are divided in a 7:1.5:1.5 ratio. During distillation training, the student model (i.e., the lightweight temporal anomaly detection model on the terminal side) uses the soft-label output of the teacher model as the supervision signal. The loss function is a weighted combination of soft-label distillation loss and hard-label cross-entropy loss. The total loss L_total = lambda × L_soft + (1-lambda) × L_hard, where L_soft is the KL divergence loss (Kullback-Leibler divergence, a measure of the difference between two probability distributions) between the student model output and the teacher model's soft labels, and L_hard is the binary cross-entropy loss between the student model output and the true labels. In this embodiment, lambda is set to 0.7 to emphasize the transfer of knowledge from the teacher model. Distillation training uses the Adam optimizer (an adaptive learning rate optimization algorithm). The initial learning rate is set to 2.3 multiplied by 10 to the power of -4. Cosine annealing is used for learning rate decay. The batch size is 128, and the training epochs are 60. An early stopping mechanism is triggered when the validation set loss does not decrease for 8 consecutive epochs. After training, the student model achieved an F1 score of 0.891 on the test set for anomaly detection (the harmonic mean of precision and recall, ranging from 0 to 1), which is only 3.5 percentage points lower than the teacher model's 0.923. However, the number of parameters was reduced to 1.7 percent of that of the teacher model, meeting the resource constraints of terminal deployment.

[0044] To verify the resource consumption and accuracy trade-off of the dual-channel evaluation mechanism of this invention, tests were conducted in a typical industrial IoT terminal environment. Test conditions included: an ARM Cortex-A53 quad-core processor, 512MB of memory, and a dataset covering 12,000 labeled samples across 8 different operating scenarios. The evaluation metrics used were F1 score to measure accuracy and millisecond-level processing latency to measure resource consumption. The testing method involved comparing the performance of a pure fast channel, a pure deep channel, a comparative scheme (balanced activation), and the dual-channel scheme of this invention.

[0045] like Figure 6As shown in the bar chart above, the accuracy (F1 score, left Y-axis) and computation latency (milliseconds, right Y-axis line) of the four evaluation schemes are compared. The pure fast channel scheme has the lowest resource consumption (0.8ms) but limited accuracy (F1=0.75), while the pure depth channel scheme has the highest accuracy (F1=0.92) but significant resource consumption (7.3ms). The comparative scheme, using a balanced activation strategy (50% fast, 50% depth), has an accuracy of 0.82 and a latency of 4.05ms. The dual-channel scheme of this invention selectively activates the depth channel (88% fast, 15% depth) within the grayscale range, significantly reducing resource consumption (1.895ms) while maintaining high accuracy (F1=0.865), demonstrating advantages in both accuracy and resource efficiency compared to the comparative scheme.

[0046] The Pareto front curve below illustrates the trade-off between accuracy and computational latency under different combinations of fast and deep channel activation ratios. The square sampling points on the curve represent the optimal boundary for accuracy-latency configuration. The dual-channel scheme of this invention (marked with black diamonds) lies on this front curve, representing the optimal performance configuration point in resource-constrained terminal environments. This comparison demonstrates that the dual-channel collaborative evaluation mechanism, through an adaptive channel activation strategy, achieves near-pure deep channel evaluation accuracy under limited computational resources, providing a feasible engineering solution for real-time quality evaluation on the terminal side.

[0047] Step S3: Execute quality-aware autonomous transmission decisions.

[0048] The intelligent terminal has a built-in transmission decision state machine. The inputs to this state machine include the comprehensive quality score output from step S2, anomaly type markers, an estimated current available network bandwidth, and the priority weight of the data's service type. The network bandwidth estimate is obtained by the intelligent terminal periodically sending probe packets to the gateway and recording round-trip latency and throughput. The probe period is dynamically adjusted based on network stability: 30 seconds when the network is stable and shortened to 5 seconds when the network fluctuates. The service priority weight is distributed by the cloud during system initialization, classifying different types of industrial data into three levels: critical, important, and ordinary.

[0049] The transmission decision state machine generates data processing decisions according to the following rules: When the overall quality score is higher than the quality pass threshold (set to 0.80 in this embodiment) and the network bandwidth is sufficient, direct upload is executed, and the data is transmitted to the cloud in its original format via the network link. When the overall quality score is higher than the quality pass threshold but the network bandwidth is limited, compressed upload is executed. The smart terminal selects an appropriate lossy compression encoding method according to the data type to reduce the data volume before transmission. The compression ratio is dynamically calculated based on the ratio of the current bandwidth to the data volume. When the overall quality score is lower than the quality pass threshold but higher than the discard threshold (set to 0.25 in this embodiment), the data, along with its anomaly marker and scene fingerprint information, is written to the local circular buffer. The accompanying scene fingerprint information enables the cloud to reconstruct the scene context at the time of the delay after receiving the delayed data. When the overall quality score is lower than the discard threshold and the anomaly type belongs to a confirmed data corruption category (such as zero output caused by sensor hardware failure or data frame structure corruption caused by communication errors), the data is discarded, and an alarm record containing a timestamp, anomaly type, and scene information is generated in the local alarm log.

[0050] For critical data with the highest business priority, the transmission decision state machine executes a priority coverage strategy. Even if the overall quality score is low, a complete anomaly label is still attached, and the data is forcibly uploaded to the cloud. The complete anomaly label includes the anomaly type, anomaly dimension, anomaly probability, and corresponding scenario fingerprint information, enabling the cloud to trace and analyze critical anomaly events. The capacity of the local circular buffer is dynamically configured based on the storage resources of the terminal device. The data in the buffer is organized into a circular queue structure according to the writing order. When new data is written, if the buffer is full, space is freed up according to a priority eviction strategy based on the quality score, prioritizing the eviction of data frames with the lowest quality scores.

[0051] Step S4: Perform incremental quality baseline update on the terminal side.

[0052] The intelligent terminal maintains a sliding statistical window. The window length is dynamically configured based on the data sampling frequency and the scene change rate. In this embodiment, the window length is set to 200 data frames, and the window sliding step is half the window length, i.e., 100 frames. There is a 50% overlap between adjacent windows to achieve a smooth statistical transition. At the end of each window period, the intelligent terminal calculates the mean, variance, rate of change distribution, and inter-dimensional correlation coefficients of each data dimension under the current scene fingerprint, and compares these statistical characteristics with the current quality baseline parameters item by item.

[0053] When the deviation between the statistical feature and the current quality baseline parameter exceeds a preset update trigger threshold, the smart terminal performs an exponentially weighted moving average update on the corresponding quality baseline parameter. The update formula is B_new = alpha × S_current + (1-alpha) × B_old, where B_new is the updated baseline parameter, B_old is the current baseline parameter, S_current is the current window statistical feature, and alpha is the update weight coefficient. In this embodiment, the value of alpha is set to 0.12. A smaller alpha value makes the baseline update smooth and gradual, avoiding drastic fluctuations in the baseline parameter caused by short-term data volatility. The update trigger threshold is set according to the parameter type: when the mean deviation of the expected value range exceeds 0.3 times the current standard deviation, the variance change exceeds 20% of the current variance, or the correlation coefficient deviation exceeds 0.1, the corresponding parameter is triggered for update. While performing the baseline parameter update, the smart terminal simultaneously updates the quality baseline parameter of the corresponding entry in the scene fingerprint index table, so that the mapping relationship between the scene fingerprint and the quality baseline always reflects the latest data distribution characteristics of the current site.

[0054] Step S5: Perform cloud collaboration and model synchronization.

[0055] The smart terminal periodically uploads local quality assessment statistical summaries to the cloud. The upload frequency is dynamically adjusted based on network conditions and data change frequency. When network conditions are good and data distribution is stable, the upload frequency is once per hour; when data distribution changes significantly, it is shortened to once every fifteen minutes. The quality assessment statistical summaries include data quality distribution statistics under each scenario fingerprint, anomaly type frequency distribution, and baseline update logs. The uploaded data volume, after summary compression, is usually no more than tens of KB, consuming minimal network bandwidth.

[0056] After receiving evaluation statistical summaries uploaded from multiple terminals, the cloud performs global analysis, aggregating similar scenario data from different terminals. It then retrains and updates the global anomaly detection teacher model using the full training data. Subsequently, it generates updated student model parameters through knowledge distillation, packages them into a model update package, and distributes it to each terminal. Upon receiving the model update package, the smart terminal performs parameter interpolation and merging on the lightweight temporal anomaly detection model using the deep channel. The merging formula is W_local_new=beta×W_global+(1-beta)×W_local_old, where W_global is the global model parameter distributed from the cloud, W_local_old is the terminal's current local model parameter, and beta is the interpolation weight coefficient. In this embodiment, beta is set to 0.3. A smaller beta value allows the terminal model to retain more of the specific knowledge accumulated in the local scenario, avoiding diluting local adaptability with global model parameters.

[0057] For the quality baseline adjustment suggestions included in the model update package, the smart terminal compares the difference between the cloud-based suggested value and the local baseline parameter item by item. The adjustment is only accepted if the difference does not exceed 20% of the local baseline parameter. Adjustments exceeding this range are logged but not executed to avoid undue interference from global statistical biases on the baseline of specific scenarios. Through this cloud-based collaboration mechanism, each terminal device can continuously absorb knowledge updates from the global system while maintaining local scenario adaptability, forming a closed-loop data quality management system with end-to-cloud collaboration.

[0058] To verify the actual deployment effect of the edge-cloud collaboration mechanism of this invention, multi-terminal collaboration tests were conducted in three types of industrial sites, including rotating equipment production lines, offshore operation platforms, and mining operation areas. The test environment included 12 intelligent terminal devices connected to the industrial cloud platform through an edge gateway, with network bandwidth fluctuating between 50Kbps and 2Mbps.

[0059] like Figure 7 As shown, the industrial site layer uses a figurative approach to depict three typical scenarios. The rotating equipment production line scenario depicts the mechanical forms of motor shafts and conveyor belt rollers; the offshore operation platform scenario depicts the physical structures of ocean waves, platform supports, and drilling rigs; and the mining operation area scenario depicts the equipment outlines of excavators and mining trucks. The intelligent terminal takes the form of a small control box, equipped with a display screen and indicator lights, reflecting the actual physical form of the on-site deployment.

[0060] The edge gateway layer is drawn in the form of a router, including a flat chassis, indicator light array, and antennas. The three gateways correspond to three types of industrial scenarios, with gateway 2 marked as having a bandwidth-limited status. The connections from the terminal to the gateway use solid and dashed lines to distinguish network status, with dashed lines indicating bandwidth-limited scenarios.

[0061] The industrial cloud platform layer uses a multi-layer server rack architecture to represent the cloud servers. The rack contains six drawer-style server units, each with an indicator light array. The global anomaly detection teacher model is illustrated using a neural network hierarchy diagram. The knowledge distillation module demonstrates the process of compressing a large model into a smaller one. The model update package distribution center is responsible for distributing updates to each terminal. Cloud modules are connected by thick solid lines to represent the core data flow.

[0062] Data interaction flow is distinguished by black arrows of different thicknesses and line types to differentiate data types: the uplink data flow from the gateway to the cloud uses solid lines, with thin lines representing statistical summaries, medium lines representing qualified data, and thick lines representing abnormal data; the downlink data flow from the cloud to the gateway uses dashed lines, representing the distribution path of the model update package.

[0063] In Scenario A, the terminal filters low-quality data through a fast channel, saving an average of 38% of uplink bandwidth. In Scenario B, when key anomalous data is detected, the terminal attaches a scenario fingerprint and uploads it to the cloud, facilitating global anomaly detection teacher model correlation analysis in the cloud. In Scenario C, the cloud aggregates quality assessment statistical summaries uploaded by multiple terminals, trains a global anomaly detection model, and generates a lightweight model update package through knowledge distillation. In Scenario D, after receiving the model update package, the terminal performs parameter interpolation and merging based on the similarity between the local scenario fingerprint and the cloud scenario library, achieving progressive model updates.

[0064] This edge-cloud topology architecture, rendered in a pure black-and-white academic style and using pictographic device diagrams, clearly demonstrates the organic combination of terminal-side autonomous assessment and cloud-based global optimization. Compared to the traditional method of uploading all data, network bandwidth utilization is improved by 42%, and the recall rate of key anomaly data is improved by 15 percentage points. The multi-terminal collaboration mechanism enables the cloud model to integrate data features from different operating scenarios, improving the generalization ability of the anomaly detection model.

[0065] Example 2

[0066] In a preferred embodiment of the present invention, a scene fingerprint-driven dynamic quality baseline generation mechanism is described in detail. This mechanism enables intelligent terminals to autonomously generate and adjust data quality assessment standards based on the real-time status of the current industrial site, without relying on predefined fixed baseline configurations in the cloud.

[0067] Within each acquisition cycle, the intelligent terminal acquires multi-source scene context information from the local sensor interface and constructs an original context vector. Specifically, the equipment operating status parameters undergo min-max normalization, mapping each indicator to a numerical range of zero to one. The reference range for normalization is determined by the equipment's rated parameters; for example, motor speed uses the rated speed as the upper bound for normalization, and bearing temperature uses the equipment's maximum allowable operating temperature as the upper bound. Environmental physical parameters are standardized using Z-score normalization, referencing the moving average and standard deviation of the terminal's historical environmental data to eliminate the influence of dimensional differences between different physical quantities. The current process stage is identified as a discrete variable and converted into a numerical vector using one-hot encoding, with each process stage corresponding to a fixed-dimensional binary code. The cumulative equipment runtime is normalized after logarithmic transformation to reduce the excessive numerical range caused by long-term operation. The vectors processed in all these ways are concatenated into the original context vector x in a fixed order of equipment operating status, environmental physical quantities, process stage codes, and runtime, with the dimension of this vector denoted as d_in.

[0068] The lightweight scene encoder performs a two-stage nonlinear mapping on the original context vector x. The first fully connected layer transforms x into an intermediate representation h, calculated as ReLU(W1 multiplied by x plus b1), where W1 is the first-layer weight matrix with dimensions d_mid multiplied by d_in, and b1 is the bias vector. The ReLU activation function sets negative values ​​to zero to introduce nonlinearity. The intermediate layer dimension d_mid is preferably set to 1.5 times d_in, striking a balance between feature extraction capability and computational cost. The second fully connected layer transforms the intermediate representation h into a scene fingerprint vector f_raw, calculated as W2 multiplied by h plus b2, where W2 has dimensions d_f multiplied by d_mid, and d_f is the preset scene fingerprint dimension, preferably between 32 and 64. L2 normalization is then performed, where f is equal to f_raw divided by its L2 norm, ensuring the scene fingerprint vector f lies on a d_f-dimensional unit hypersphere. L2 normalization eliminates differences in the magnitude of fingerprint vectors, making subsequent cosine similarity calculations equivalent to the inner product operation of normalized vectors, thus reducing the computational complexity of online matching. The total number of parameters in the encoder is approximately d_in multiplied by d_mid plus d_mid multiplied by d_f. In a typical configuration with d_in of 20, d_mid of 30, and d_f of 48, the total number of parameters is approximately 2040 floating-point numbers, occupying no more than 8KB of storage space, which meets the deployment requirements of resource-constrained terminals.

[0069] The scene fingerprint index table is maintained in the terminal's local storage and organized using a key-value pair structure. The key is the historical scene fingerprint vector, and the value is the corresponding set of quality baseline parameters. The maximum capacity of the index table is set according to the terminal's storage resources, preferably between 50 and 200 entries. When the index table reaches its capacity limit, the historical entry with the lowest access frequency is evicted using the least recently used strategy. During scene fingerprint matching, the smart terminal calculates the cosine similarity sim_i between the current scene fingerprint f and all historical fingerprints f_i in the index table. sim_i is equal to the inner product of f and f_i (since both have been L2 normalized, the inner product is the cosine similarity). The maximum value among all similarity values ​​is denoted as sim_max, and the corresponding index table entry is denoted as the best matching entry.

[0070] When sim_max exceeds the matching threshold theta_match, the intelligent terminal retrieves the quality baseline parameter corresponding to the best matching entry as the current evaluation baseline. Preferably, theta_match is set to 0.90. In practical industrial scenarios, this value means that when the overall deviation of equipment status, environmental conditions, and process stages in each dimension between two scenarios does not exceed approximately 10%, they are considered to belong to the same scenario type. The value of theta_match should not be too low, otherwise different operating conditions may be incorrectly merged into the same scenario, leading to baseline shift; nor should it be too high, otherwise the scenario switching frequency will increase, and the cold start overhead of new scenarios will increase.

[0071] When sim_max is lower than theta_match, the smart terminal determines that the current scene is a new scene and needs to generate an initial quality baseline. At this time, the terminal retrieves the K historical entries with the highest cosine similarity f to the current fingerprint from the index table. The preferred value of K is 3 to 5. Let the fingerprints of these K historical entries be f_1 to f_K, the corresponding quality baseline parameters be B_1 to B_K, and the corresponding cosine similarities be sim_1 to sim_K. The initial baseline B_init is generated by weighted interpolation. The calculation method is that B_init is equal to the weighted sum of each sim_j multiplied by B_j divided by the sum of each sim_j, where j ranges from 1 to K. The physical meaning of this weighted interpolation is that the more similar a historical scene is to the current new scene, the greater its contribution to the initial estimation. When there is a scene among the K historical scenes that is highly similar to the current scene (e.g., sim value is 0.88 and threshold is 0.90), the baseline of that scene will dominate the initial estimation, making the initial baseline closer to the true value.

[0072] During the baseline cold start phase of a new scenario, the intelligent terminal performs an expansion operation on the grayscale range for quality assessment. Specifically, the upper threshold of the grayscale range is increased by delta_expand, and the lower threshold is decreased by delta_expand. The preferred value of delta_expand is 15% to 25% of the grayscale range width. The purpose of expanding the grayscale range is to increase the activation frequency of the depth channel during the cold start period when the baseline is not yet stable, allowing more boundary data to undergo fine evaluation and reducing the risk of misjudgment due to inaccurate baseline. As subsequent data accumulates and the baseline converges, the grayscale range gradually recovers to its normal width. The duration of the cold start phase is determined by the convergence status of the online gradient update. When the update magnitude of the baseline parameters is less than the convergence threshold for three consecutive sliding window periods, the baseline is considered to have converged, and the cold start mode is exited. The convergence threshold is preferably 2% to 5% of the current value of the baseline parameters.

[0073] The online gradient update process for the new scenario baseline is as follows: At the end of each sliding statistical window period, the intelligent terminal calculates the statistical features S_current of the data within the current window, including the sample mean, sample variance, and correlation coefficient between dimensions for each dimension. The update formula is B_new equal to alpha multiplied by S_current plus (1-alpha) multiplied by B_old, where alpha is the update weight coefficient. During the cold start phase, alpha is set to a larger value (preferably 0.2 to 0.3) to allow the baseline to absorb the statistical features of the new data more quickly. During the steady-state operation phase after baseline convergence, alpha is switched to a smaller value (preferably 0.05 to 0.1) to maintain baseline stability and avoid baseline jitter caused by short-term data fluctuations. The two-stage switching mechanism of alpha enables the baseline update to achieve an adaptive balance between response speed and stability. When the sampling frequency of a certain type of sensor in the industrial scenario where the terminal is located is low or the scene changes extremely slowly, the convergence period of the cold start phase may be prolonged, and the evaluation accuracy may be affected to some extent in the short term.

[0074] Furthermore, the scene fingerprint index table performs synchronous maintenance operations during baseline updates. When the baseline parameters of an existing scene are updated, the quality baseline value of the corresponding entry in the index table is synchronously replaced with the updated parameters. When the baseline of a new scene converges, the smart terminal creates a new entry in the index table, writing the converged scene fingerprint and the corresponding stable baseline parameters. Each entry in the index table also maintains an access timestamp and an access counter for determining the least recently used eviction policy. The fingerprint vectors in the index table are stored in half-precision floating-point format to save storage space. With a d_f configuration of 48, a single fingerprint vector occupies 96 bytes, and the entire index table occupies no more than 20KB with a capacity of 200 entries.

[0075] The scene fingerprint encoder is trained using a contrastive learning strategy in the cloud. Training data comes from scene context information and corresponding quality assessment statistics uploaded from multiple terminals. The training objective is to cluster fingerprint vectors mapped from scene contexts of the same operating condition type on a hypersphere, while dispersing fingerprint vectors of different operating condition types on the hypersphere. The loss function is normalized temperature-scaled cross-entropy loss, calculated as L equal to the logarithm of the negative base e sim(f_i, f_j) divided by tau, minus the logarithm of the sum of the logarithmic sums of the similarities of all negative sample pairs divided by tau, where f_i and f_j are positive sample pairs of the same operating condition type, and tau is a temperature hyperparameter, preferably between 0.05 and 0.1. Training uses the Adam optimizer with a learning rate of 0.001, a batch size of 256, and 100 to 200 training epochs. After training, the encoder parameters are sent to the terminal device via a model update package.

[0076] The aforementioned scenario fingerprint-driven dynamic quality baseline generation mechanism offers the following advantages: Regarding adaptability, because the quality baseline dynamically adjusts with the scenario fingerprint, when process phase switching, equipment start-up / shutdown, or environmental conditions change in the industrial environment, the intelligent terminal can complete scenario identification and retrieve matching historical baselines within one acquisition cycle. This ensures that the data quality assessment standard remains consistent with the current operating conditions, avoiding systematic misjudgments caused by static baselines when operating conditions change. In terms of new scenario processing, the initial baseline estimation based on K-nearest neighbor weighted interpolation utilizes the structural correlation between existing scenarios. Compared to directly using the default baseline, the initial estimation is closer to the true value, improving the assessment accuracy during the cold start phase. Combined with the grayscale range expansion mechanism, this further reduces the risk of misjudgments during the cold start period. Regarding storage and computation efficiency, the encoder parameters do not exceed several thousand floating-point numbers, the index table storage overhead does not exceed tens of KB, and online matching only requires vector inner product operations. The computational latency of the entire scenario identification and baseline retrieval process is controlled at the millisecond level, meeting the terminal's real-time requirements.

[0077] From a fundamental perspective, the technical reasons why the scene fingerprinting mechanism can achieve the above effects are as follows: There is a strong correlation between the scene context information (equipment status, environmental parameters, process stage, etc.) and data quality characteristics in industrial settings—data collected under the same operating condition exhibits similar statistical characteristics in terms of numerical range, rate of change distribution, and inter-dimensional correlation, while the statistical characteristics of data under different operating conditions show distinguishable differences. The lightweight scene encoder compresses the high-dimensional discretized original context information into a low-dimensional continuous fingerprint vector through two layers of nonlinear mapping. The contrastive learning training strategy enables the encoder to learn the ability to map similar operating conditions to nearby positions on a hypersphere, thus allowing cosine similarity to serve as an effective measure of operating condition similarity. L2 normalization constrains the fingerprint to a unit hypersphere, eliminating interference from the modulus dimension, and ensuring that similarity comparison only reflects directional information, i.e., the structural differences in scene features. K-nearest neighbor weighted interpolation is based on the assumption of local smoothness on the hypersphere—adjacent scenes in the fingerprint space should also have similar quality baseline parameters. This assumption holds true in actual industrial scenarios because small changes in operating parameters usually correspond to continuous gradual changes in data features rather than abrupt changes. The exponentially weighted moving average update mechanism takes advantage of the local stationarity of time series data to integrate new and old statistical information in a recursive manner, avoiding the overhead of storing and reprocessing all historical data.

[0078] Example 3

[0079] In another embodiment of the present invention, a dynamic quality baseline generation method based on multi-scale statistical feature hierarchical matching is provided to replace the scene fingerprinting mechanism based on neural network scene encoder and cosine similarity matching in Embodiment 1.

[0080] In this alternative, the smart terminal does not deploy a scene encoder network. Instead, it directly extracts multi-scale statistical features from the scene context information as scene descriptors. Specifically, the smart terminal maintains three statistical windows at different time scales: a short-term window covering the most recent 30 seconds to 1 minute of collected data, a medium-term window covering the most recent 5 to 10 minutes of data, and a long-term window covering the most recent 30 minutes to 1 hour of data. Within each window, the terminal calculates the first-order statistics (mean), second-order statistics (variance), and trends (linear regression slope) of the device operating status parameters and environmental physical parameters. The statistical features at the three time scales are concatenated into a scene descriptor vector in the order of short-term, medium-term, and long-term. Unlike scene fingerprint vectors, scene descriptors are directly composed of raw statistics without undergoing nonlinear mapping transformations by neural networks.

[0081] The scene matching process employs a hierarchical cascaded matching strategy instead of global cosine similarity matching. The first layer of matching compares only discrete state variables such as process stage identifiers and equipment operating modes, quickly filtering out candidate entries within the same major operating condition category through precise matching. The second layer of matching calculates the weighted Euclidean distance between the current scene descriptor and the scene descriptors of each candidate entry within the candidate set. The weights are dynamically adjusted based on the discriminative power of each statistical feature—features with high discriminative power are assigned larger weights, and features with low discriminative power are assigned smaller weights. Discriminative power is defined as the coefficient of variation of the feature across all entries in the index table; a larger coefficient of variation indicates a more significant difference between different scenes and a greater contribution to scene differentiation. When the minimum weighted Euclidean distance of the second layer of matching is less than the matching distance threshold, the quality baseline parameters of the corresponding entry are retrieved.

[0082] The quality baseline index table in this alternative uses a tree structure. The first level branches according to process stage identifiers, with each process stage corresponding to a subtree. The second level further subdivides each subtree according to equipment operating mode. Leaf nodes store scene descriptors and quality baseline parameters for that combination of process stage and operating mode. The tree structure reduces the time complexity of the first-level matching to constant level, requiring only traversal of the leaf nodes within the current process stage branch to perform the second-level matching calculation. This offers an efficiency advantage over a full table scan when the number of index table entries is large.

[0083] When a matching failure is identified as a new scenario, this alternative solution does not employ the K-nearest neighbor weighted interpolation strategy. Instead, it uses a rule-based conservative baseline generation method. The intelligent terminal starts with the default baseline of the current process stage and performs initial corrections to the default baseline based on the statistical characteristics accumulated within the current short-term window. The correction rule is: if the mean of the short-term window deviates from the center value of the expected range of the default baseline by more than one standard deviation, the expected range of the baseline is shifted along the deviation direction by 50% of the standard deviation of the default baseline. This conservative correction strategy avoids the risk of over-adjusting the baseline when data is insufficient, and is closer to actual working conditions than a default baseline that is not corrected at all.

[0084] In the data quality assessment stage, this alternative adopts a single-channel progressive assessment architecture instead of a dual-channel selective activation architecture. The progressive assessment arranges quality checks according to computational complexity from low to high, forming a coarse-to-fine inspection pipeline. The first level is numerical range verification and sampling integrity verification, with the lowest computational cost; the second level is timestamp continuity verification, which requires accessing adjacent frame data; the third level is inter-dimensional correlation verification, which requires calculating the inter-dimensional correlation coefficient matrix; and the fourth level is temporal pattern consistency verification, which uses statistical methods to detect the deviation between the autocorrelation function of the data sequence within a short time window and the baseline autocorrelation function. After each level of inspection, if the detected anomaly level exceeds the early stopping threshold for that level, subsequent levels of inspection are skipped, and the current assessment result is directly output, avoiding unnecessary fine-grained checks on obviously anomalous data. If the anomaly level in each inspection does not trigger early stopping, the overall quality assessment result is output after all four levels of inspection are completed.

[0085] The temporal pattern consistency check in the progressive evaluation architecture does not rely on a deep learning model, but instead employs a statistical method based on autocorrelation function (ACF) comparison. Specifically, within each evaluation cycle, the smart terminal calculates the autocorrelation function (ACF)_current of the sliding window of the data to be evaluated and compares it with the reference autocorrelation function (ACF_ref) stored in the current quality baseline. The comparison method involves calculating the mean square error (MSE)_acf of the two autocorrelation functions within a preset lag order range. MSE_acf equals the sum of squares of ACF_current(k) minus ACF_ref(k) for each lag order, divided by the lag order N_lag, where k ranges from 1 to N_lag. The value of N_lag is set according to the data sampling frequency and the time series pattern period of interest, preferably between 8 and 16. When MSE_acf exceeds the autocorrelation deviation threshold, it is determined that the temporal pattern of the current data deviates abnormally from the baseline. This method requires no training or deployment of deep learning models, has a computational complexity of O(N_lag multiplied by window length), and is suitable for low-cost terminal devices that do not have hardware to accelerate deep learning inference.

[0086] The transmission decision and baseline incremental update mechanism in this alternative scheme are consistent with those in Example 1. The transmission decision state machine generates data processing decisions based on quality assessment results, network status, and service priorities. A sliding statistical window tracks changes in data distribution and performs exponentially weighted moving average updates on the baseline parameters. The difference is that, during baseline updates, the parameters of the corresponding leaf nodes in the tree index structure are maintained synchronously, rather than a planar index table based on vector similarity matching.

[0087] In this alternative cloud-based collaboration process, the terminal does not receive encoder model parameter updates (because this solution does not use a neural network encoder). Instead, it receives updates to the statistical feature weight adjustment parameters and autocorrelation function reference template issued by the cloud based on global data analysis results. After aggregating statistical summaries uploaded by multiple terminals, the cloud identifies the changes in the discriminative power of each statistical feature under different operating conditions, generates updated feature weight configurations, and sends them to the terminals. The terminals then adjust the weighted Euclidean distance weights in hierarchical matching accordingly.

[0088] As an alternative, this implementation achieves scene-adaptive data quality assessment without deploying a neural network model, making it suitable for low-cost industrial terminals with more limited computing resources or lacking floating-point acceleration capabilities. This scheme replaces deep learning feature extraction with multi-scale statistical features, global vector matching with hierarchical cascaded matching, and dual-channel selective activation architecture with progressive single-channel evaluation, maintaining basic scene adaptation and quality assessment capabilities while reducing terminal computing power requirements. When the terminal device has fewer sampling dimensions and a relatively limited number of scene types, this scheme can achieve acceptable evaluation results with lower deployment costs; however, when there are many scene types and high data dimensions, the representational power of statistical features may be inferior to that of a neural network encoder, and the granularity of scene differentiation and the accuracy of baseline initialization may decrease.

[0089] Example 4

[0090] like Figure 4 As shown, this embodiment provides a data quality self-assessment intelligent terminal system for industrial cloud platforms. The system is deployed on intelligent terminal equipment in industrial sites and includes a scene perception module, a dual-channel assessment module, a transmission decision module, an incremental update module, and a cloud collaboration module. The modules transmit data and work collaboratively with each other through the internal data bus of the terminal.

[0091] The scene perception module collects scene context information of the current industrial site. A lightweight scene encoder performs nonlinear mapping on this context information to generate a scene fingerprint vector. This scene fingerprint vector is then matched with historical fingerprints in the scene fingerprint index table using cosine similarity. Based on the matching result, historical quality baseline parameters are retrieved or default quality baseline parameters are enabled. The scene perception module maintains the scene fingerprint index table, storing the mapping relationship between historical scene fingerprints and corresponding quality baseline parameters. It supports historical quality baseline parameter retrieval, enabling default quality baseline parameters for new scenes, and initializing baseline estimation. Specifically, the scene perception module periodically reads equipment operating status parameters, environmental physical parameters, current process stage identifiers, and cumulative equipment runtime through sensor interfaces and the equipment control bus. After normalization, these parameters are concatenated into an original context vector, which is then output as the scene fingerprint vector by the lightweight scene encoder. The scene perception module transmits the generated scene fingerprint vector and the retrieved quality baseline parameters to the dual-channel evaluation module as a dynamic benchmark for data quality assessment. In one hardware deployment scheme of this embodiment, the smart terminal adopts an ARM-based embedded processor with a main frequency of 400MHz to 1GHz, and is equipped with 256MB to 1GB of RAM and 4GB to 16GB of local flash memory. Under this configuration, the single execution latency of encoder inference and index table query of the scene perception module does not exceed 2 milliseconds.

[0092] The dual-channel evaluation module performs dual-channel quality evaluation on the data to be evaluated, including a fast channel submodule and a deep channel submodule. The fast channel submodule performs deterministic rule-based multi-dimensional quality checks based on the quality baseline obtained by the scene-aware module and outputs a rapid comprehensive quality score. These multi-dimensional quality checks include numerical range verification, sampling integrity verification, timestamp continuity verification, and inter-dimensional correlation verification. When the rapid comprehensive quality score falls within the grayscale range between an upper and lower threshold, the deep channel submodule is activated. The deep channel submodule deploys a lightweight temporal anomaly detection model to perform a refined evaluation of the grayscale data and outputs an anomaly probability vector. This anomaly probability vector is weighted and fused with the sub-dimensional quality scores of the fast channel to generate a multi-dimensional quality evaluation result, including the comprehensive quality score, anomaly markers for each dimension, and anomaly confidence. The dual-channel evaluation module receives the quality baseline parameters output by the scene-aware module as the evaluation benchmark and transmits the generated multi-dimensional quality evaluation results to the transmission decision module. The fast channel submodule and the deep channel submodule run on the terminal processor in a time-division multiplexing manner. The single evaluation latency of the fast channel does not exceed 1 millisecond, and the single inference latency of the deep channel is controlled within 10 milliseconds.

[0093] The transmission decision module generates data processing decisions based on the multi-dimensional quality assessment results output by the dual-channel evaluation module, network status information, and data service priorities, using a transmission decision state machine. The transmission decision module has a built-in transmission decision state machine that generates data processing decisions based on the comprehensive quality score, anomaly type markers, estimated network bandwidth, and service priority weights. These decisions include direct upload, compressed upload, local caching for later transmission, or discarding and issuing an alarm. The transmission decision module obtains the estimated available network bandwidth by periodically probing the gateway link. Service priority weights are issued by the cloud during system initialization. The transmission decision module manages data writing and priority eviction in the local ring buffer. For cached data awaiting transmission, it writes it along with its anomaly marker and scenario fingerprint information to the buffer, and uploads it when the network recovers. For critical data with the highest service priority, the transmission decision module adds a complete anomaly label and forces it to be uploaded to the cloud. This complete anomaly label includes the anomaly type, anomaly dimension, anomaly probability, and corresponding scenario fingerprint information. The data processing results of the transmission decision module are simultaneously fed back to the incremental update module for statistical analysis of the execution frequency of various decisions and data quality distribution.

[0094] The incremental update module tracks data distribution changes and performs incremental updates to quality baseline parameters using a sliding statistical window. At the end of the sliding statistical window period, the module calculates the trends in the mean, variance, rate of change distribution, and inter-dimensional correlation coefficients for each data dimension. When the deviation between the statistical features and the current quality baseline parameters exceeds the update trigger threshold, an exponentially weighted moving average update is performed on the corresponding baseline parameters. After completing the baseline parameter update, the incremental update module synchronously updates the corresponding quality baseline parameters in the scene fingerprint index table maintained by the scene perception module, ensuring the mapping between the scene fingerprint and the quality baseline remains consistent with the current field status. The window statistical calculation of the incremental update module is triggered once at the end of each window period. The resource consumption during the calculation process is an instantaneous peak and does not affect the real-time evaluation process of the dual-channel evaluation module.

[0095] The cloud-based collaboration module uploads quality assessment statistical summaries to the cloud and receives model update packages from the cloud for local model parameter merging. The cloud-based collaboration module periodically uploads quality assessment statistical summaries, which include data quality distribution statistics under each scenario fingerprint, anomaly type frequency distribution, and baseline update logs. After receiving the model update package from the cloud, the cloud-based collaboration module performs local model parameter merging using a parameter interpolation strategy, weighting the cloud-based global model parameters and local model parameters according to a preset ratio. For quality baseline adjustment suggestions included in the model update package, the cloud-based collaboration module compares the difference between the cloud-based suggested values ​​and the local baseline parameters item by item, accepting adjustments only when the difference does not exceed a preset ratio, and performing quality baseline calibration. The cloud-based collaboration module exchanges data with the cloud platform via an encrypted communication link. The uploading of statistical summaries and the downloading of model update packages are both performed during network bandwidth idle periods determined by the transmission decision module, avoiding competition for bandwidth resources with business data transmission.

[0096] The aforementioned modules operate collaboratively on smart terminal devices, forming a complete processing loop from scene perception, quality assessment, transmission decision-making to baseline updates and cloud collaboration. The scene perception module provides a dynamic quality baseline for the dual-channel assessment module. The assessment results of the dual-channel assessment module drive the autonomous decision-making of the transmission decision-making module. The incremental update module continuously calibrates the quality baseline based on changes in data distribution and feeds it back to the scene perception module. The cloud collaboration module establishes a knowledge synchronization channel between the edge and the cloud, enabling the terminal-side model and baseline to continuously absorb global optimization results.

[0097] Example 5

[0098] This embodiment demonstrates the specific implementation process and technical effects of the present invention through a scenario of monitoring the condition of an industrial rotating equipment.

[0099] The production line of rotating equipment in a large industrial park includes multiple rotating mechanical devices of different models, which are equipped with intelligent terminals to collect multi-dimensional sensor data such as vibration, temperature, current and pressure. The sampling frequency is 50 frames per second, and each frame of data contains 12-dimensional sensor channels. This production line has multiple process stage switches (no-load operation, low-load processing, full-load processing, equipment preheating, etc.), and there are significant differences in the normal value ranges of each sensor channel and the correlation between dimensions under different process stages. At the same time, the on-site ambient temperature fluctuates with seasons, day and night, and the equipment operation duration varies from just completed maintenance to continuous operation for hundreds of hours. These factors jointly lead to continuous drift of the normal data pattern. Since the production line needs to upload data to the industrial cloud platform for remote monitoring and predictive maintenance analysis, while the on-site network bandwidth is limited and unstable, the traditional cloud-side centralized data quality assessment scheme faces problems such as large transmission delay, serious bandwidth waste, and inability to detect data anomalies in time. Conventional static threshold detection methods also produce a large number of misjudgments because they cannot adapt to dynamic switching of multiple working conditions.

[0100] The method of the present invention is adopted to implement data quality self-assessment for this scenario, and the specific process is as follows. In the deployment stage, the lightweight scenario encoder and the time-series anomaly detection model are loaded into the intelligent terminal, and the terminal locally stores the initialized default quality baseline and an empty scenario fingerprint index table. After the terminal runs, the scenario perception module collects scenario context information at a frequency of once per second and generates a scenario fingerprint vector. Within the first two hours of system operation, the terminal goes through the process stage switching from no-load operation to full-load processing, and the scenario fingerprint index table gradually accumulates from an empty table to 18 valid scenario entries, and the quality baseline corresponding to each scenario gradually converges through incremental update during the data accumulation process.

[0101] During the test period of 72 hours of continuous operation, the terminal processed a total of about 12.96 million frames of sensor data. The proportion of data frames directly judged as qualified by the fast channel is 73.6%, the proportion of data frames directly marked as abnormal is 8.9%, and the proportion of data frames entering the gray interval to activate the deep channel is 17.5%. The decision distribution generated by the transmission decision module is: direct upload accounts for 61.2%, compressed upload accounts for 18.7%, local cache for pending transmission accounts for 12.4%, and discard with alarm accounts for 7.7%. Compared with the scheme of directly uploading all data without terminal-side assessment, the actual number of uploaded data frames is reduced by about 38% (calculated by the number of data frames), which effectively alleviates the network bandwidth pressure.

[0102] As Figure 10As shown, the joint analysis results are presented using a pure black and white academic chart style. Four transmission decision types are distinguished by different scatter plot shapes (circles, squares, triangles, rhombuses) and grayscale levels (pure black, dark gray, light gray, white). The scatter plot illustrates the joint distribution relationship between the overall quality score and the number of bytes transmitted in data frames. In the direct upload zone (quality pass threshold 0.80 and above), data frames are uploaded at their original size of 2.4KB, accounting for 71.3% of the total data volume in terms of transmitted bytes. In the quality score range of 0.65-0.80, data frames are compressed before upload, reducing the transmitted bytes to approximately 1.2KB, accounting for 14.8% of the total data volume. In the quality score range of 0.25-0.65, data frames are temporarily stored in the local cache for transmission, accounting for 7.0%. In the discard zone (quality score below 0.25), data frames are discarded and an alarm is triggered, accounting for 6.9%.

[0103] The top margin histogram, using grayscale padding, displays a bimodal distribution of data quality scores, with the main peak concentrated in the 0.8-1.0 high-quality range and the secondary peak appearing in the 0.6-0.7 grayscale range. The right margin histogram, also using grayscale padding, shows three clusters of transmitted byte distribution: original size, compressed size, and zero transmission. The bandwidth saving percentage is displayed, showing that this invention saves 14.5% of network bandwidth compared to the full upload solution.

[0104] The test results demonstrate that this invention, through a quality-aware autonomous transmission decision-making mechanism, effectively reduces network bandwidth consumption while ensuring reliable transmission of critical data. The decision boundary lines reflect the rules for determining the quality pass threshold and drop threshold, and the scatter distribution clearly shows the quality score range characteristics of the four types of transmission decisions.

[0105] To verify the accuracy of the quality assessment of the method of this invention, the results of manual annotation of 5000 randomly selected frames of data by domain experts were used as the ground truth benchmark. The method was evaluated from three dimensions: anomaly detection accuracy, recall, and F1 score, and compared with two comparative schemes. Comparative scheme one is a static threshold detection method, which uses a fixed range of threshold values ​​to independently determine the data for each dimension, without considering scene changes or correlations between dimensions. Comparative scheme two is a sliding window-based statistical detection method, which calculates the mean and standard deviation of each dimension through a sliding window and uses three times the standard deviation as a dynamic threshold for anomaly determination, but it lacks scene differentiation and multi-dimensional joint evaluation capabilities.

[0106] During the no-load operation phase, the anomaly detection accuracy of the method of this invention is 0.934, the recall rate is 0.908, and the F1 score is 0.921; compared with Scheme 1, the accuracy rate is 0.871, the recall rate is 0.792, and the F1 score is 0.830; compared with Scheme 2, the accuracy rate is 0.896, the recall rate is 0.854, and the F1 score is 0.874. During the full-load processing phase, the accuracy of the method of this invention is 0.917, the recall rate is 0.893, and the F1 score is 0.905; compared with Scheme 1, the accuracy rate is 0.743, the recall rate is 0.681, and the F1 score is 0.711; compared with Scheme 2, the accuracy rate is 0.862, the recall rate is 0.829, and the F1 score is 0.845. During the process transition period (a 2-minute window before and after the transition), the accuracy of the method of this invention was 0.889, the recall rate was 0.861, and the F1 score was 0.875; compared with Scheme 1, the accuracy was 0.624, the recall rate was 0.573, and the F1 score was 0.597; compared with Scheme 2, the accuracy was 0.781, the recall rate was 0.738, and the F1 score was 0.759. Under long-term drift conditions after the equipment has been running continuously for more than 48 hours, the accuracy of the method of this invention was 0.906, the recall rate was 0.882, and the F1 score was 0.894; compared with Scheme 1, the accuracy was 0.697, the recall rate was 0.651, and the F1 score was 0.673; compared with Scheme 2, the accuracy was 0.834, the recall rate was 0.803, and the F1 score was 0.818. The above data shows that the evaluation accuracy of the method of the present invention is better than that of the two comparative schemes under various working conditions. The advantages are particularly obvious under the changing conditions such as process stage switching and long-term drift. The first comparative scheme cannot adapt to the changes in working conditions due to the fixed threshold and produces a large number of misjudgments. Although the second comparative scheme achieves a certain degree of dynamic adaptation through sliding statistics, it lacks the ability to distinguish between scenarios and its performance declines when the working conditions change abruptly.

[0107] like Figure 8 As shown, the comparison results are presented using a black-and-white academic chart style. A bar chart distinguishes precision and recall by different fill patterns (plain fill, diagonal line, dot matrix, backslash, cross, circle), and a line chart distinguishes F1 scores by different marker shapes (circle, square, triangle) and line types (solid line, dashed line, dotted line). The method of this invention maintains high evaluation accuracy under all four operating conditions. In the no-load operation condition, the F1 score of this invention reaches 0.921, significantly better than the static threshold method's 0.830 and the sliding window statistical method's 0.874. In the full-load processing condition, the F1 score of this invention is 0.905, an improvement of 19.4 percentage points compared to the first comparative scheme.

[0108] The advantages of this invention are even more pronounced under complex operating conditions such as process stage switching and long-term drift. During process stage switching, the F1 score of this invention is 0.875, while the static threshold method's F1 score drops to 0.597 due to its inability to adapt to changing conditions. During long-term drift, the F1 score of this invention is 0.894, while the F1 score of the first scheme is only 0.673 due to a fixed quality baseline.

[0109] The test results demonstrate that the dynamic quality baseline retrieval mechanism driven by scene fingerprints, as described in this invention, can effectively adapt to changes in operating conditions and long-term drift, maintaining stable data quality assessment accuracy in complex industrial scenarios. Error bars show that the standard deviations of each method are all within the range of 0.012-0.028, reflecting the statistical reliability of the test results.

[0110] Furthermore, to verify the technical contributions of each core technical module of this invention, ablation experiments were performed. Ablation configuration one involved removing the scene fingerprint mechanism and replacing the scene adaptive baseline with a unified global quality baseline, while keeping the other modules unchanged. Under this configuration, the overall F1 score across all operating conditions decreased from 0.899 to 0.843, a drop of 5.6 percentage points, indicating that the scene fingerprint-driven dynamic baseline generation mechanism significantly contributes to the evaluation accuracy. Ablation configuration two involved removing the depth channel and using only the rule check of the fast channel for quality evaluation. Under this configuration, the overall F1 score across all operating conditions decreased to 0.856, a drop of 4.3 percentage points, and the F1 score for grayscale data decreased from 0.872 to 0.764, a drop of 10.8 percentage points, indicating that the depth channel's ability to finely discriminate boundary-ambiguous data is irreplaceable. Ablation configuration three removes the incremental baseline update mechanism, using a fixed initial baseline without dynamic updates. In the first 8 hours of operation, the F1 score of this configuration is close to that of the complete scheme (the difference is less than 1 percentage point). However, as the operation time is extended to 48 hours, the F1 score drops to 0.831, and after 72 hours, it further drops to 0.807, indicating that the incremental update mechanism plays a key role in maintaining the accuracy of the assessment under long-term operation.

[0111] Regarding terminal resource consumption, when the method of this invention runs on an embedded processor with a main frequency of 400MHz, the average processing latency for scene fingerprint generation and baseline retrieval is 1.6 milliseconds, the fast channel evaluation latency is 0.8 milliseconds, the deep channel inference latency is 7.3 milliseconds (generated only during activation), and the end-to-end average latency for a single frame of data from acquisition to completion of transmission decision is 4.2 milliseconds. The peak memory consumption is 18.6MB, of which model parameters and index table occupy approximately 1.2MB, sliding window buffer occupies approximately 3.8MB, and circular buffer occupies approximately 12MB (configurable). The above resource consumption indicators show that the method of this invention can run in real time on resource-constrained industrial intelligent terminals, meeting the latency and resource constraints of industrial data acquisition scenarios.

[0112] like Figure 9 As shown, the resource usage analysis results are presented using a pure black and white academic chart style. The processing latency heatmap uses grayscale levels (white represents low latency, dark gray represents high latency), showing an average latency of 1.6ms for the scene fingerprint generation module, 0.8ms for the fast channel evaluation module, 7.3ms for the deep channel evaluation module, and 0.5ms for the transmission decision module. In scenarios where the deep channel does not need to be activated, the end-to-end latency is only 4.2ms; in scenarios where the deep channel needs to be activated in the grayscale range, the end-to-end latency is 11.5ms.

[0113] The memory usage breakdown pie chart uses different gray levels and fill patterns to distinguish different parts. Model parameters occupy 1.2MB (10.8%), the sliding window buffer occupies 3.8MB (34.2%), the circular cache occupies 12MB (55.0%), and the total memory usage is approximately 17MB (excluding runtime dynamic memory overhead), accounting for only 6.6% of the total memory usage on a terminal device with 256MB of memory. The storage space usage breakdown bar chart uses different gray levels and fill patterns to distinguish storage items. Model parameters occupy 8KB, the scene fingerprint index table occupies 20KB, the circular cache occupies 12MB, and the log occupies 2KB.

[0114] The test results show that the present invention, through the design of a lightweight timing anomaly detection model, a dual-channel diversion evaluation mechanism, and an incremental update strategy, achieves a low-latency and low-resource-consumption data quality self-evaluation function on embedded terminal devices, meeting the constraints of real-time performance and resource constraints in industrial settings.

[0115] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A data quality self-assessment smart terminal method for industrial cloud platforms, characterized in that, Includes the following steps: Step S1: The intelligent terminal collects the scene context information of the current industrial site, and generates a scene fingerprint vector by performing nonlinear mapping on the scene context information through a lightweight scene encoder. The scene fingerprint vector is then matched with historical fingerprints in the scene fingerprint index table using cosine similarity. Based on the matching result, historical quality baseline parameters are retrieved or default quality baseline parameters are enabled. The scene context information includes equipment operating status parameters, environmental physical parameters, current process stage identifier, and cumulative equipment runtime. The lightweight scene encoder adopts a two-layer fully connected network structure, with the middle layer using ReLU (Rectified Linear Array)... The Unit (linear rectified function) activation function is used, and the output layer uses L2 normalization to make the scene fingerprint vector fall on the unit hypersphere. The scene fingerprint index table stores the mapping relationship between historical scene fingerprints and corresponding quality baseline parameters. When the highest similarity of the cosine similarity matching exceeds the matching threshold, the corresponding historical quality baseline parameter is retrieved as the current quality baseline parameter. When the highest similarity is lower than the matching threshold, the current scene is determined to be a new scene and the default quality baseline parameter is enabled. The quality baseline parameters of the K historical scenes with the highest cosine similarity to the current scene fingerprint are used to perform weighted interpolation to generate an initial baseline estimate. During subsequent data accumulation, the baseline is converged to a stable baseline through online gradient updates. In step S2, a dual-channel quality assessment is performed on the data to be evaluated. The fast channel performs multi-dimensional quality checks based on the quality baseline parameters obtained in step S1 and outputs a fast quality comprehensive score. The fast quality comprehensive score is activated when it is in the gray range between the upper and lower thresholds. The deep channel outputs an anomaly probability vector through a lightweight temporal anomaly detection model and fuses it with the results from the fast channel to generate a multi-dimensional quality assessment result. The fast channel's multi-dimensional quality checks include numerical range verification, sampling integrity verification, timestamp continuity verification, and inter-dimensional correlation verification. Each check result is quantified into a sub-dimensional quality score, and a weighted sum is used to obtain the rapid comprehensive quality score. When the rapid comprehensive quality score is higher than an upper threshold, the data quality is deemed acceptable, and the process proceeds directly to step S3. If the score is lower than a lower threshold, the anomaly type is directly marked, and the process proceeds to step S3. When the score is in the grayscale range, the deep channel is activated. Step S3 involves generating a data processing decision through a transmission decision state machine based on the multi-dimensional quality assessment result, network status information, and data service priority. Step S4 involves tracking data distribution changes through a sliding statistical window and incrementally updating the quality baseline parameters. Step S5 involves uploading the quality assessment statistical summary to the cloud and receiving a model update package from the cloud for local model parameter merging.

2. The method according to claim 1, characterized in that, The quality baseline parameters include the expected value range for each data dimension, the data change rate threshold, the correlation coefficient matrix between dimensions, and the sampling integrity requirements.

3. The method according to claim 1, characterized in that, The lightweight temporal anomaly detection model adopts a one-dimensional deep separable convolutional network structure. It takes a sliding window composed of the data to be evaluated and several preceding frames as input, captures temporal dependencies through deep convolution, achieves cross-dimensional feature fusion through pointwise convolution, and outputs the anomaly probability of each dimension through a Sigmoid (S-shaped function) layer. The number of parameters of the lightweight temporal anomaly detection model is controlled within 50,000.

4. The method according to claim 1, characterized in that, In step S3, the data processing decision includes four types: direct upload, compressed upload, local cache pending upload, and discard with alarm. When the comprehensive quality score in the multi-dimensional quality assessment result is higher than the quality pass threshold and the network bandwidth is sufficient, direct upload is performed. When the comprehensive quality score is higher than the quality pass threshold but the network bandwidth is limited, compressed upload is performed. When the comprehensive quality score is lower than the quality pass threshold but higher than the discard threshold, the data, along with the anomaly marker and scene fingerprint information, is written to the local circular cache. When the comprehensive quality score is lower than the discard threshold and the anomaly type belongs to the data corruption category, the data is discarded and an alarm record is generated.

5. The method according to claim 4, characterized in that, For critical data with the highest business priority, even if the overall quality score is low, the transmission decision state machine will still attach a complete anomaly label before uploading; the local circular buffer adopts a priority elimination strategy based on the overall quality score, and when the buffer is full, it will prioritize eliminating the data with the lowest overall quality score.

6. The method according to claim 1, characterized in that, In step S5, the model update package includes incremental parameter updates and quality baseline adjustment suggestions for the lightweight temporal anomaly detection model; the smart terminal uses a parameter interpolation strategy to perform a weighted average of the global model parameters sent from the cloud and the local model parameters according to a preset ratio, and accepts the quality baseline adjustment suggestions only when the adjustment range does not exceed the preset ratio of the local baseline parameters.

7. A data quality self-assessment intelligent terminal system for industrial cloud platforms, characterized in that, include: The scene perception module is used to collect scene context information of the current industrial site. A lightweight scene encoder performs nonlinear mapping on the scene context information to generate a scene fingerprint vector. This scene fingerprint vector is then matched with historical fingerprints in the scene fingerprint index table using cosine similarity. Based on the matching result, historical quality baseline parameters are retrieved or default quality baseline parameters are enabled. The scene context information includes equipment operating status parameters, environmental physical parameters, current process stage identifier, and cumulative equipment runtime. The lightweight scene encoder adopts a two-layer fully connected network structure, with the middle layer using ReLU (Rectified Linear Array)... The Unit (linear rectified function) activation function is used, and the output layer uses L2 normalization to ensure that the scene fingerprint vector falls on the unit hypersphere. The scene perception module maintains a scene fingerprint index table, storing the mapping relationship between historical scene fingerprints and corresponding quality baseline parameters. When the highest similarity of the cosine similarity match exceeds the matching threshold, the corresponding historical quality baseline parameter is retrieved as the current quality baseline parameter. When the highest similarity is lower than the matching threshold, the current scene is determined to be a new scene and the default quality baseline parameter is enabled. The quality baseline parameters of the K historical scenes with the highest cosine similarity to the current scene fingerprint are used to perform weighted interpolation to generate an initial baseline estimate. The dual-channel evaluation module is used to perform dual-channel quality evaluation on the data to be evaluated. The fast channel performs multi-dimensional quality checks based on the quality baseline parameters obtained by the scene perception module and outputs a fast comprehensive quality score. The fast comprehensive quality score is activated when it is in the gray range between the upper and lower thresholds. The deep channel outputs an anomaly probability vector through a lightweight temporal anomaly detection model and fuses it with the fast channel results to generate a multi-dimensional quality assessment result. The dual-channel assessment module includes a fast channel submodule and a deep channel submodule. The fast channel submodule performs numerical range verification, sampling integrity verification, timestamp continuity verification, and inter-dimensional correlation verification based on dynamic quality baseline parameters and outputs a rapid comprehensive quality score. The deep channel submodule deploys a lightweight temporal anomaly detection model to perform fine-grained evaluation of grayscale data and outputs an anomaly probability vector. A transmission decision module generates data processing decisions through a transmission decision state machine based on the multi-dimensional quality assessment results, network status information, and data service priorities. An incremental update module tracks data distribution changes through a sliding statistical window and performs incremental updates to the quality baseline parameters. A cloud collaboration module uploads the quality assessment statistical summary to the cloud and receives model update packages from the cloud for local model parameter merging.

Images