An industrial intelligent full-stack domestication multivariate time series prediction and intelligent diagnosis system

Abstract

The application relates to the technical field of industrial control and artificial intelligence, and discloses an industrial intelligent full-stack domestication multivariate time sequence prediction and intelligent diagnosis system, which comprises the following steps: accessing heterogeneous operation data of equipment in a data source layer; shunting the operation data through a measuring point prefix routing mechanism in a backend service layer, and triggering a degradation fallback logic to return a simulation sequence when an interface is abnormal, and constructing a sliding window sample; extracting features from the sliding window sample through a multivariate model in an intelligent agent layer, outputting a prediction deviation vector, generating a dynamic abnormal threshold value based on historical residual errors; performing online abnormal detection on the dynamic abnormal threshold value in the backend service layer, injecting a prompt word into the prediction deviation vector when an abnormality occurs, outputting a diagnosis result by combining knowledge retrieval and fine-tuning of a large language model in the intelligent agent layer; and performing elastic computing scheduling and rendering in a front-end display layer. The application improves the abnormality capturing accuracy and the fault diagnosis professionalism.

Description

Technical Field

[0001] This invention relates to the fields of industrial control and artificial intelligence technology, specifically to a domestically produced, full-stack industrial intelligent multivariate time series prediction and intelligent diagnostic system. Background Technology

[0002] In industrial production and equipment operation and maintenance, real-time monitoring of equipment operating status and fault diagnosis are core aspects of ensuring production safety. With the popularization of sensing devices and artificial intelligence technologies, status prediction models based on multivariate time-series data are beginning to be introduced into industrial centralized control systems.

[0003] In actual industrial implementations, data acquisition from heterogeneous devices typically shares a common basic query interface. This centralized request model is prone to response congestion and data stream pollution during multi-tenant concurrent access. Furthermore, if the underlying network channel experiences jitter or disconnection, the training and online inference links of the upper-layer models, which rely on continuous time-series samples, will come to a standstill.

[0004] After data acquisition, most existing condition monitoring systems rely on manually set fixed physical thresholds to determine anomalies. However, during long-term operation, industrial equipment is affected by changes in ambient temperature, natural aging of components, or fluctuations in scheduled loads, causing the operating baseline of its physical parameters to dynamically drift. Fixed judgment boundaries cannot capture subtle, hidden fault characteristics when the equipment is under low load conditions, while under high load and full load conditions, normal, regular fluctuations are easily misjudged as faults, resulting in a large number of invalid alarms.

[0005] When equipment triggers an alarm, conventional monitoring systems typically only output the name of the abnormal measurement point or the alarm code, failing to provide in-depth mechanistic analysis of the underlying causes of the data anomalies. Some systems attempt to directly connect to general-purpose large language models to generate fault analysis reports; however, due to the lack of support from maintenance manuals and procedures specific to particular equipment models, the conclusions output by these general models are often broad and lack engineering guidance value. Furthermore, industrial equipment faults often exhibit cascading effects; if the large amount of abnormal data generated by concurrent alarms from multiple measurement points is directly input into a large language model, it is highly likely to exceed the model's input length limit, leading to inference truncation. In addition, control centers typically need to run a large number of data monitoring panels concurrently, and existing systems generally adopt a continuous, indiscriminate background polling calculation strategy. This strategy continues to call the underlying hardware model for inference calculations even when the monitoring panels are not visible or idle, resulting in significant consumption of computing resources and congestion of network request channels. Summary of the Invention

[0006] The technical problems solved by this invention are as follows: In industrial production environments, the acquisition of multi-source heterogeneous data is prone to causing interface response blockage and data confusion; traditional fixed alarm thresholds cannot adapt to the baseline drift caused by the dynamic operating conditions of equipment, which can easily lead to alarm logic failure; general large language models lack the underlying mechanism correlation support for industrial specific equipment, which can easily generate irrelevant answers and exceed the input length limit when processing concurrent chain alarms; at the same time, continuous background time-series prediction and diagnostic polling will lead to the ineffective consumption of computing power resources of the underlying computing unit and network request channel congestion.

[0007] To address the above problems, the present invention provides the following technical solution:

[0008] This invention provides an industrial intelligent full-stack domestic multivariate time series prediction and intelligent diagnosis system, comprising: a data source layer, a backend service layer, an intelligent agent layer, and a frontend display layer; The data source layer is used to access heterogeneous operational data from external devices; The backend service layer, connected to the data source layer, is used to perform traffic splitting and time-series preprocessing on the running data through the measurement point prefix routing mechanism, and to construct sliding window samples. The intelligent agent layer, connected to the backend service layer, deploys a multivariate Transformer model and a large language model, which are used to extract features and model the temporal coupling relationship of the sliding window samples through the multivariate Transformer model, and output the predicted value and prediction deviation vector of each measurement point. The agent layer is also used to extract historical prediction residuals from the training phase of the multivariate Transformer model to generate dynamic anomaly thresholds. The backend service layer is also used to perform online anomaly detection by comparing the dynamic anomaly threshold with the real-time prediction deviation vector. When an abnormal measurement point is detected, the backend service layer injects the prediction deviation vector into the preset diagnostic prompt words, and the intelligent agent layer calls the large language model that integrates domain knowledge retrieval and fine-tuning to output intelligent diagnostic results; The front-end presentation layer connects to the back-end service layer and is used to perform elastic computing scheduling based on page state, and to implement multi-dimensional security protection strategies in each level of the interaction link.

[0009] Furthermore, when the backend service layer uses the test point prefix routing mechanism to distribute runtime data, it is specifically used for: Extract the measurement point prefix of each measurement point identifier in the set of measurement point identifiers corresponding to the input running data, call the internally stored routing mapping table, and determine the target data interface terminal point corresponding to the measurement point prefix; Based on the same target data interface terminal point, the corresponding measurement point identifiers are aggregated into independent request subsets, and each request subset is sent to the corresponding physical node or logical service area in the data source layer. When initiating a data acquisition request, if the status monitoring component built into the backend service layer identifies that the interface response time exceeds the request timeout threshold set by combining the historical average response time of the underlying database and the latency tolerance, or if the error rate reaches the circuit breaker limit set according to the system's concurrent load capacity, the backend service layer triggers a degradation fallback logic, redirecting the data stream acquisition instruction to the internally configured simulated data generation module, and generating a simulated time series sequence with superimposed Gaussian white noise based on the mean and variance of the corresponding measurement points under historical normal operating conditions, and returning it.

[0010] Furthermore, when constructing the sliding window sample, the backend service layer is specifically used for: Construct a set of reference timestamp sequences with equal step sizes, and map the original non-fixed-interval running data onto the reference timestamp sequences; When the data loss period at a certain measuring point is less than the preset tolerance period, the most recent valid historical observation value is used for forward filling. Linear interpolation is performed when the missing duration exceeds the tolerance period and there are valid first and last observation nodes. Extract continuous multivariate observation vectors of fixed length, and construct a two-dimensional sliding window matrix by vertically splicing them in chronological order. Stack multiple two-dimensional sliding window matrices along the new dimension to reconstruct a three-dimensional input tensor containing three dimensions: batch size, sliding window length, and total number of physical measurement points, as the sliding window sample.

[0011] Furthermore, when the intelligent agent layer extracts features from the sliding window samples using a multivariate Transformer model, it is specifically used for: The fully connected mapping network built inside the multivariate Transformer model performs temporal feature linear mapping on the input three-dimensional input tensor, and generates an initial feature representation matrix using a linear projection weight matrix initialized with uniform distribution. A position encoding matrix is ​​constructed based on the sliding window length of the sliding window sample and the preset hidden layer dimension. For the even-numbered dimension index position and the odd-numbered dimension index position in the position encoding matrix, a sine function and a cosine function based on the same frequency scaling factor are used for mapping and assignment, respectively. The initial feature representation matrix and the location encoding matrix are subjected to element-wise matrix addition to generate a hidden state matrix that incorporates location information.

[0012] Furthermore, the multivariate Transformer model contains multiple stacked encoder layers. When the agent layer models the temporal coupling relationship in the current encoder layer, it is specifically used for: The input hidden state matrix is ​​divided into multiple independent attention heads along the feature dimension, and the hidden state matrix is ​​mapped into a query matrix, a key matrix, and a value matrix respectively using the corresponding learnable projection weight matrix; Perform a matrix dot product operation between the query matrix and the transposed key matrix to generate the original attention score matrix; After adjusting the original attention score matrix by division using a scaling factor based on the square root of the single-head dimension, a normalized exponential function is applied to transform it into an attention weight matrix. Perform matrix multiplication between the attention weight matrix and the corresponding value matrix to output the subspace feature matrix; After concatenating the subspace feature matrices of all attention heads and mapping them back to the original dimension, the updated deep feature representation matrix is ​​output through residual connections, layer normalization processing, and a feedforward neural network submodule.

[0013] Furthermore, when the intelligent agent layer extracts historical prediction residuals from the model training phase to generate dynamic anomaly thresholds, it is specifically used for: Using the complete training dataset of the multivariate Transformer model, a global forward inference test is performed. The actual observed value of each measurement point at any historical moment is subtracted from the corresponding theoretical predicted value to obtain the residual value of the corresponding measurement point at the corresponding moment, thus constructing a historical residual time series. After truncating and filtering the historical residual time series by upper and lower boundaries, all residual values ​​of each measurement point in the historical residual time series are summed and divided by the total number of samples to obtain the benchmark mean. The sum of squares of the differences between each residual value and the benchmark mean is calculated, divided by the total number of samples, and then the square root is taken to obtain the benchmark standard deviation. The identifier of each measurement point is structurally bound to its corresponding benchmark mean and benchmark standard deviation.

[0014] Furthermore, when the backend service layer performs online anomaly detection, it specifically performs the following: based on the 3-Sigma principle, multiply the baseline standard deviation of each measurement point by 3, compare the calculation result with the preset minimum channel constraint parameter, and take the maximum value of the two; add the baseline mean and subtract the maximum value respectively to obtain the upper limit of dynamic anomaly judgment and the lower limit of dynamic anomaly judgment as the dynamic anomaly threshold. Obtain the real-time prediction residual value of each measuring point at the current moment from the real-time prediction deviation vector; When the real-time prediction residual value is strictly greater than the upper limit of the dynamic anomaly judgment or strictly less than the lower limit of the dynamic anomaly judgment, it is determined that the current data of the corresponding measurement point has jumped out of the safe channel, and the corresponding anomaly flag variable is assigned a value of 1.

[0015] Furthermore, when the backend service layer injects the prediction deviation vector into the preset diagnostic prompts, it is specifically used for: Extract the absolute value of the real-time prediction residual of each abnormal measurement point in the prediction deviation vector at the current moment, and divide the absolute value by the benchmark standard deviation of the corresponding measurement point under historical healthy working conditions to calculate the normalized deviation of the corresponding abnormal measurement point. The abnormal measurement points are sorted in descending order based on the normalized deviation value, and the top K core abnormal measurement point data are extracted. A preset basic prompt word template matching the current device category is loaded, and the extracted real-time prediction residual and the dynamic abnormal threshold are converted into quantified text descriptions. The abnormal measurement point identifiers and the converted text descriptions are concatenated into an abnormal feature statement and injected into the specified empty placeholders of the basic prompt word template to dynamically construct a complete diagnostic request prompt word vector.

[0016] Furthermore, when the intelligent agent layer invokes the large language model that integrates domain knowledge retrieval and fine-tuning, it is specifically used for: The internal embedding model is invoked to convert the diagnostic request prompt word vector into a quantized query vector. The cosine similarity score is obtained by calculating the inner product value between the query vector and each knowledge base text block vector in the pre-built vector index library, and dividing it by the product of the Euclidean norms of the query vector and the knowledge base text block vector. After removing irrelevant text blocks whose scores are lower than the similarity threshold set based on the dispersion of the domain corpus, several top-ranked items are extracted, and their physical troubleshooting procedure texts are extracted and structurally concatenated to form a retrieval knowledge context. The retrieval knowledge context is then concatenated with the original diagnostic request prompt word vector at the text level to generate enhanced prompt words. The enhanced prompts are distributed to the corresponding hardware computing units via the MINDIE inference gateway. Based on the original fixed parameter set of the large language model, and combined with the LoRA fine-tuning weight parameter matrix set consisting of the product of the dimension reduction matrix and the dimension increase matrix injected by the bypass, the intelligent diagnostic results are output by performing autoregressive forward inference.

[0017] Furthermore, when the front-end presentation layer performs elastic computation scheduling based on page state, it is specifically used for: Capture the state change event of the current monitoring tab in the front-end display interface using the page visibility application interface; When the event of the tab being hidden is captured, a global suspension command is triggered, pausing all built-in polling timers that are active on the current page, and cutting off the timed request data stream sent to the backend service layer, so that the corresponding computing node enters the computing power idle and memory release state. When the event of the tab becoming visible is captured, a wake-up command is triggered to reactivate and initialize the corresponding polling timer. Before resuming polling, the time difference is calculated based on the suspended timestamp and the current wake-up timestamp, and a compensation range query request is automatically sent to the backend service layer to draw the missing historical actual data and model prediction data.

[0018] This invention provides a domestically developed, full-stack industrial intelligent multivariate time series prediction and intelligent diagnostic system. It offers the following advantages: 1. This invention uses a measurement point prefix routing mechanism to divert heterogeneous operational data, establishing isolated processing channels for different business data. Combined with a status monitoring component, when an interface response times out or the error rate reaches a preset circuit breaker threshold, the system triggers a degradation fallback logic and generates a simulated time-series sequence with superimposed Gaussian white noise for return. This mechanism maintains the structural continuity of the time-series sliding window matrix when the underlying network channel is abnormal, ensuring the stability of the downstream deep learning model training process.

[0019] 2. This invention utilizes a multivariate Transformer model to extract temporal coupling features and shifts the anomaly detection benchmark to the dimension of real-time prediction residuals. The system calculates the benchmark mean and benchmark standard deviation of historical residuals and introduces a minimum channel constraint parameter to construct dynamic anomaly judgment upper and lower limits. This calculation logic enables the judgment boundary to adapt to the benchmark drift caused by equipment operating conditions and avoids the problem of excessively narrow judgment channels caused by the low variance of historical data, thereby improving the system's accuracy in capturing anomalies.

[0020] 3. In the intelligent diagnosis stage, this invention performs descending truncation of abnormal measurement points based on normalized deviation, avoiding the exceeding of the large language model context input length caused by concurrent cascading alarms. Simultaneously, the system generates enhanced prompt words by calculating cosine similarity to retrieve text blocks from the knowledge base, and performs inference based on LoRA fine-tuning weights formed by the product of the reduced-dimensional matrix and the increased-dimensional matrix, while keeping the original parameters of the large language model fixed. This process effectively integrates external procedure texts and industry-specific features, improving the engineering troubleshooting guidance value of the final output diagnostic results. Attached Figure Description

[0021] Figure 1 This is an architecture diagram of the domestically produced industrial intelligent full-stack multivariate time series prediction and intelligent diagnosis system of the present invention; Figure 2This is a flowchart of the industrial intelligent multivariate time series prediction and intelligent diagnosis method of the present invention; Figure 3 This is a flowchart of the time series model training system of the present invention; Figure 4 This is a flowchart of the real-time prediction and intelligent verification process of the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1 This invention provides an industrial intelligent full-stack domestic multivariate time series prediction and intelligent diagnosis system, which may include: a data source layer, a backend service layer, an intelligent agent layer, and a frontend display layer.

[0024] The data source layer connects to external devices to acquire operational data. It accesses real-time databases, time-series databases, and business databases. Real-time databases include OPC and SK databases. The data source layer supports access to distributed photovoltaic data sources, conventional thermal power data sources, and new energy data sources. The data source layer is configured with a measurement point prefix routing module. This module enables the data source layer to perform data routing and distribution.

[0025] The backend service layer connects to the data source layer to obtain raw time-series data. The backend service layer is built on the Spring Cloud microservice framework. It configures a service registration mechanism and a Feign invocation mechanism. The backend service layer externally encapsulates and provides time-series model training APIs and time-series model prediction APIs. The backend service layer configures a MINDIE inference gateway. Communication between service nodes within the backend service layer uses HTTPS encryption and Feign's internal authentication mechanism.

[0026] The agent layer connects to the backend service layer. It runs on a Python environment and the FastAPI framework. The agent layer is configured with multivariate Transformer models, LoRA model management services, and RAG retrieval services. It is deployed on the Ascend 310PNPU hardware platform and uses a Docker containerized deployment environment. The agent layer isolates and schedules the underlying NPU computing resources through a container device mapping mechanism.

[0027] The front-end presentation layer connects to the back-end service layer for data interaction. It is built on the Vue.js framework, combined with the ElementUI component library and the ECharts chart library. The front-end presentation layer includes a device tree browsing interface, a multi-panel monitoring interface, an automatic refresh module, and an intelligent Q&A interface.

[0028] See attached document Figure 2 The system executes a macro-level workflow during runtime, which may include the following main steps: S100, the data source layer accesses heterogeneous operational data from external devices, and the backend service layer performs traffic splitting and time-series preprocessing on the operational data through a measurement point prefix routing mechanism to construct sliding window samples.

[0029] The S200 and intelligent agent layers use a multivariate Transformer model to extract features from sliding window samples and model the temporal coupling relationship, outputting the predicted values ​​and prediction bias vectors for each measurement point.

[0030] The S300 intelligent agent layer extracts historical prediction residuals from the model training phase to generate dynamic anomaly thresholds. The backend service layer compares these dynamic anomaly thresholds with the real-time prediction deviation vectors to perform online anomaly detection.

[0031] S400. When an abnormal measurement point is detected, the backend service layer injects the predicted deviation vector into the diagnostic prompt words, and the intelligent agent layer calls the large language model that integrates domain knowledge retrieval and fine-tuning to output the intelligent diagnostic results.

[0032] The S500 system performs elastic computing scheduling based on page state at the front-end presentation layer and implements multi-dimensional security protection strategies in each level of the interaction link to ensure the safe and stable operation of the system's computing resources and data.

[0033] In the data access and preprocessing process of step S100 above, in an industrial environment, the naming of measurement points for the same business system or similar equipment usually follows specific specifications, and their headers often contain business characters that can identify the equipment's affiliation. Based on this data characteristic, to solve the response blocking and data confusion problems caused by directly querying the underlying heterogeneous database through a unified interface, this system executes intelligent distribution and isolation logic based on measurement point prefixes during the data acquisition phase. In this embodiment, the data routing mechanism built between the backend service layer and the data source layer specifically includes the following processing steps: S111. The backend service layer receives a measurement point data acquisition request sent by the frontend presentation layer. This request contains a set of multivariable measurement point identifiers selected by the operator. These identifiers typically correspond to the actual physical operating parameters of the generator set or photovoltaic equipment, such as operating temperature, system pressure, or output voltage. Assume the total number of measurement points requiring joint modeling in this request is... The set of input measurement point identifiers is defined to include Each measurement point is identified by a unique string.

[0034] S112. The backend service layer parses the string of each measurement point identifier in the measurement point identifier set and extracts the corresponding measurement point prefix. Specifically, the system extracts a corresponding number of characters from the beginning of each measurement point identifier string based on a preset prefix truncation length parameter as the measurement point prefix for that measurement point. As a preferred method, the prefix truncation length parameter is determined based on the aforementioned industrial field business coding specifications, and is typically set to 4 characters. After obtaining the measurement point prefix, the backend service layer calls the internally stored routing mapping table for matching queries. This routing mapping table records the mapping relationship between preset business prefixes and underlying data interface terminal points. The system determines the target data interface terminal point corresponding to the current measurement point prefix through table lookup operations. To ensure the completeness of the data distribution algorithm logic, when no matching item is found in the routing mapping table for the extracted measurement point prefix, the system assigns the measurement point identifier to a preset public interface terminal point by default or throws an identifier exception prompt to the front-end display layer.

[0035] S113. After obtaining the target data interface endpoint, the backend service layer groups and categorizes the original set of measurement point identifiers to achieve targeted distribution of external requests and logical isolation of underlying system resources. Specifically, the system iterates through the parsed measurement point identifiers, aggregating those mapping to the same target data interface endpoint into the same request subset, thus constructing multiple independent request subsets corresponding to different underlying interfaces. After grouping, the backend service layer distributes each request subset to the corresponding physical node or logical service area in the data source layer.

[0036] In practical implementation, the data source layer is physically divided into different security isolation zones. When the parsed measurement point prefix is ​​a conventional thermal power identifier (HNJS) or a new energy identifier, the backend service layer routes the subset of requests containing that prefix to the SkRtDB real-time historical interface deployed in the first physical zone. Correspondingly, when the parsed measurement point prefix is ​​a distributed photovoltaic identifier (HNGF), the system routes the corresponding subset of requests to the PVDataService historical interface deployed in the second physical zone. Through the above-mentioned grouping and distribution mechanism, the system not only has the ability to automatically map heterogeneous data source request addresses, but also helps to establish a protective barrier between distributed photovoltaic business data and conventional thermal power business data in the access channel and memory processing space. As for the specific communication and acquisition mechanism of the OPC protocol in the underlying database, those skilled in the art can implement it using existing industrial control standards and specifications, the content of which is well-known technology in this field and will not be elaborated here.

[0037] Furthermore, regarding the data acquisition step in step S100, in the actual operating environment of an industrial site, cross-physical area data communication is limited by network conditions or the load status of the underlying database, posing a risk of response delay and connection interruption. To ensure the continuity of the upper-layer model training and online diagnostic process, this embodiment constructs an interface high availability and disaster recovery mechanism in the interaction link between the backend service layer and the data source layer. This mechanism reduces the probability of system-wide downtime caused by underlying failures by intercepting abnormal requests and injecting simulated data. Specifically, it includes the following processing steps: S121. The backend service layer initiates a data retrieval request to the data source layer through its internally encapsulated service call component. During this process, the backend service layer configures a preset request timeout threshold and enables a status monitoring component to monitor the availability of the underlying data interface in real time. In this embodiment, the request timeout threshold is set based on the historical average response time of the underlying database and the latency tolerance of the upper-layer business, for example, it can be set to between 3 and 5 seconds. As a preferred approach, the status monitoring component operates in circuit breaker mode to statistically analyze the interface request error rate and response timeout frequency within a time window in real time.

[0038] S122. The backend service layer executes degradation trigger judgment based on the statistical results of the status monitoring component. When the underlying data interface returns a valid payload within the timeout threshold, the system normally parses and flows the actual runtime sequence data. Conversely, when the status monitoring component identifies that the interface response time exceeds the timeout threshold, or the error rate caused by network disconnection and service rejection reaches the preset circuit breaker threshold, the backend service layer determines that the current physical data channel has failed. The above-mentioned circuit breaker threshold is usually set by the operations and maintenance personnel according to the system's concurrent load capacity, such as an error rate of 50% within one minute. After determining the failure, the system immediately cuts off further calls to the abnormal interface and triggers degradation and fallback logic.

[0039] S123. After triggering the fallback logic, the backend service layer redirects the data stream acquisition command to the simulated data generation module configured within the system. The simulated data generation module parses the set of measurement point identifiers and time interval parameters in the original request, and dynamically constructs a mock data stream accordingly. To ensure downstream business nodes can process this data compatiblely, the constructed mock data stream maintains strict isomorphism with the actual return value in terms of data structure. Specifically, the mock data stream not only has a timestamp sequence corresponding to the request interval, but also generates floating-point values ​​within the normal operating range for each measurement point based on preset benchmark parameters. As a specific implementation, these benchmark parameters include the mean and variance of each measurement point under historical normal operating conditions. The mock data generation module generates a simulated time series sequence with normal equipment fluctuation characteristics by superimposing Gaussian white noise conforming to the variance onto the mean.

[0040] S124. The backend service layer encapsulates the constructed simulated data stream into a standard data transmission object and returns it to the upper-layer business caller. With the supplement of the simulated data stream, the system maintains the operational conditions for subsequent sliding window sample construction and agent layer model training. This disaster recovery method enables the system to still perform basic system demonstration verification and offline training functions during underlying network anomalies.

[0041] S125. To prevent the system from being permanently degraded due to a single failure, the status monitoring component starts a preset sleep timer after the circuit breaker is triggered. When the sleep time expires, the status monitoring component enters a half-open state, allowing a small number of test requests to be sent to the data source layer. If the test request returns successfully, the status monitoring component closes the circuit breaker and resumes the real data acquisition process; if the test request still times out or reports an error, the sleep timer is reset, and the system continues to remain in a degraded state.

[0042] For the internal state machine transitions and specific code-level configuration methods of circuit breaker components in microservice architectures, those skilled in the art can refer to existing distributed service fault tolerance frameworks for deployment. The implementation principles are well-known technologies in this field and will not be elaborated here.

[0043] After data acquisition, regarding the sliding window construction step in step S100, after acquiring the underlying multi-source heterogeneous data or the simulation data generated by the disaster recovery mechanism, various physical sensors in the industrial field typically operate using independent clocks, and some data is uploaded based on an event-driven mode of numerical change amplitude. This results in the collected multi-dimensional time-series data exhibiting physical characteristics of inconsistent sampling frequencies and misaligned timestamps. To eliminate the interference of this characteristic on subsequent multivariate model joint operations, the backend service layer configures and executes time-series alignment and multi-dimensional sliding window construction logic before transmitting data to the agent layer. In this embodiment, the preprocessing process specifically includes the following steps: S131. The backend service layer extracts time information from various raw time series and constructs a unified reference timeline for timestamp alignment. Assume the total number of measurement points participating in the joint modeling in the industrial scenario is... The system reads the time interval and sampling resolution set by the user in the front-end request and generates a set of reference timestamp sequences with equal step sizes. As a preferred method, this sampling resolution is set based on the real-time requirements of the system monitoring, typically configured at the second or minute level. For each measurement point, the back-end service layer maps its original non-uniform observation data onto this reference timestamp sequence. At any aligned time... The system's multivariate observation vector is represented as containing A one-dimensional sequence of elements, where each element corresponds to a physical measurement point at time t. The actual observed value.

[0044] S132. Due to differences in the sampling mechanisms of various underlying physical sensors, the above mapping process inevitably leads to some measurement points lacking effective observations at specific reference times. To avoid null values ​​causing anomalies in feature extraction of subsequent neural network models, the backend service layer performs missing value imputation on the mapped observation vectors. In this embodiment, the system adopts a strategy combining forward imputation and linear interpolation. The system pre-sets the tolerance period for the corresponding physical measurement points, which is set according to the inertial time constant of the changes in the physical state of the equipment. When the data missing duration of a certain measurement point is less than the preset tolerance period, the system performs forward imputation using the most recent effective historical observation; when the missing duration exceeds the tolerance period and there are effective first and last observation nodes, linear interpolation is performed based on the adjacent effective values. Furthermore, to avoid long-span data loss causing interpolation distortion and triggering algorithm logic dead zones, if the missing duration exceeds the maximum interpolation limit or the missing value is at the edge of the sequence, the system will remove the observation vector segment containing the invalid span, or replace it with the historical average value of the same period, thereby maintaining the data integrity of the multivariate sequence on a unified time axis.

[0045] S133. After completing the temporal alignment and numerical imputation of the multivariate sequence, the continuous data stream with temporal attributes needs to be transformed into a discrete sample structure suitable for deep learning model processing. The backend service layer extracts historical data blocks in the forward direction of the time series to construct a sliding window sample for model input. Let the predefined sliding window size parameter be... In order to predict the next moment... The system's operational status at any given time is determined by the backend service layer, which extracts a fixed-length continuous observation vector based on historical data sequences. Specifically, the system extracts observations from time [time]. At that time Continuous A multivariate observation vector. The system will use this... The observation vectors are concatenated vertically in chronological order to construct a two-dimensional sliding window matrix that includes both time and variable dimensions. The row dimensions of this matrix correspond to the size of the sliding window. The column dimension corresponds to the total number of physical measurement points. In this embodiment, the system defaults to setting the sliding window size. This means that the system will use the multivariate state characteristics of the past 60 reference time steps to extrapolate the time. Expected performance.

[0046] Meanwhile, to support supervised training of subsequent neural network models, the system extracts the aligned time points. The real multivariate observation vectors are used as training labels corresponding to the sliding window matrix. The system generates multiple independent sliding window matrix samples and their corresponding labels by gradually shifting the window position along the time axis by setting a preset sliding step size. This sliding step size is usually set according to the expansion requirements of the training sample size, and is conventionally configured as a step size of 1 in single-step prediction scenarios.

[0047] S134. To meet the batch training specifications of the deep learning model in the agent layer and the hardware-level tensor acceleration computation requirements, the backend service layer aggregates and transforms the generated multiple sliding window matrices. Assume that within the parameter update cycle of a single network training session, [the following parameters are extracted]. The system takes continuously or randomly sampled sliding window matrix samples, stacks these independent matrices along a new dimension, and reconstructs a 3D input tensor for the network to read. The final dimensional form of this input tensor is (B, W, N), where... This represents the batch size, or BatchSize. In a real-world scenario, this input tensor specifically carries... In 1 independent training sample At each time step, The historical fluctuation characteristics of each physical parameter. After reconstruction, the backend service layer serializes and encapsulates this 3D tensor and its corresponding label data, and sends it to the agent layer for the multivariate Transformer model to read through an internal interface (such as calling the agent layer's / train interface).

[0048] For basic filtering and cleaning algorithms for outliers in time series and common data standardization and scaling methods, those skilled in the art can implement them using existing processing strategies such as moving average filtering and maximum-minimum normalization. These are well-known technologies in the field and will not be elaborated here.

[0049] Regarding the feature extraction and modeling process in step S200 above, after receiving the 3D input tensor encapsulated and transmitted by the backend service layer, the multivariate Transformer model deployed inside the agent layer initiates the forward propagation logic for feature extraction. Combined with the specific application of industrial multivariate prediction scenarios, this model is structurally configured as a sequence feature extraction network based on a self-attention mechanism. Its internal data flow sequentially passes through a temporal feature mapping module, a positional encoding fusion node, and a multi-layered stacked encoder module. Considering the high physical dimensionality and heterogeneous dimensions of industrial multivariate time-series data, directly using it as input to a deep network often increases the difficulty of network gradient updates and convergence. To construct a unified feature representation space that facilitates gradient propagation, this embodiment configures a temporal feature mapping and positional encoding mechanism at the input of the multivariate Transformer model. This mechanism specifically includes the following processing steps: S211. The intelligent agent layer parses the received input tensor and defines the structural hyperparameters of the underlying neural network. As described in the aforementioned data preprocessing process, the input tensor is represented as a (B, W, N) dimensional form at the reading end. Considering the actual physical state of the business, where... The batch size corresponding to a single model calculation, The length of the sliding window corresponding to the time dimension, The total number of measurement points participating in the joint modeling should correspond to the total number of measurement points participating in the joint modeling. Within the multivariate Transformer model, the system initializes the hidden layer dimension parameters of the lead oxide. As a preferred approach, considering both the hardware computing power limitations of the Ascend 310PNPU and its ability to represent industrial time-series features, this embodiment fixes the hidden layer dimension parameter to [value missing]. .

[0050] S212. After establishing the basic dimensions of the network, the agent layer performs a linear mapping of temporal features on the input tensor through an internally constructed fully connected mapping network. This step aims to transform the original... 3D physical observation features projected to dimension 1 In a continuous high-dimensional latent variable space. Taking any independent time series sample matrix in a single batch processing process as an example, the calculation process of time series feature mapping is as follows: The system constructs a mapping network containing a learnable linear projection weight matrix and a matching feature bias vector. To ensure stable propagation in the initial stage of the network, this embodiment uses a Xavier uniform distribution to initialize the weights of the linear projection weight matrix; subsequently, the system performs matrix multiplication on the input time series sample matrix and the linear projection weight matrix, and adds the result of the operation to the feature bias vector element-wise in the corresponding dimension. By performing this linear transformation, the system enables discrete observations representing different physical parameters (such as temperature, pressure, and rotational speed) to achieve preliminary feature alignment in a unified high-dimensional latent space, generating an initial feature representation matrix.

[0051] S213. Although the feature space has been unified, the multivariate Transformer model, limited by its parallel computing underlying structure, often struggles to directly extract the absolute time position and relative temporal evolution order of the input time series. To enable the network to capture the decay or fluctuation patterns of industrial equipment operating conditions over time, the agent layer uses the sliding window length of the input sequence... With the set hidden layer dimension A position encoding matrix is ​​constructed. In this embodiment, the elements of the position encoding matrix are calculated alternately based on sine and cosine functions of different frequencies. Specifically, the system constructs a frequency scaling factor based on the time step position index within the matrix and the dimension index of the hidden layer. For even-numbered dimension index positions in the matrix, the system uses a sine function based on the scaling factor for mapping and assignment; for odd-numbered dimension index positions in the matrix, the system uses a cosine function based on the same scaling factor for mapping and assignment. This construction method, which applies specific frequency waveforms to different dimensions, can assign unique and continuous vectorized position identifiers to different historical time steps within the sliding window.

[0052] S214. After obtaining the location identifier, the agent layer fuses the initial feature representation matrix obtained by linear mapping with the generated location encoding matrix at the feature level. On the corresponding network computing nodes, the system performs element-wise matrix addition on the aforementioned initial feature representation matrix and location encoding matrix, injecting temporal sequence information into the multivariate high-dimensional latent features. After completing the fusion calculation, the agent layer uses the hidden state matrix fused with location information as the standard input data stream of the first-layer encoder of the multivariate Transformer model to the next layer, thereby supporting the parallel computation of multivariate coupling relationships by the self-attention mechanism.

[0053] After completing temporal feature mapping and location encoding, complex nonlinear coupling and time delay effects often exist between multiple variables such as temperature, pressure, and load within industrial equipment. Traditional linear models struggle to effectively separate these coupling features. To address this issue, this embodiment constructs a multi-layer encoder network within the intelligent agent layer and utilizes a multi-head self-attention mechanism to extract the nonlinear coupling relationships of different industrial measurement points over time. The specific calculation process includes the following steps: S221, the construction of the intelligent agent layer is... A feature extraction network consisting of stacked encoder layers with the same structure. As a preferred approach, considering the dynamic response complexity and inference latency constraints of industrial equipment, the system sets the number of encoder layers. Each encoder layer contains a multi-head self-attention submodule and a feedforward neural network submodule, with residual connections and layer normalization operations configured around each submodule. Let the hidden state matrix passed from the previous network node contain... Time step length and The hidden layer dimension.

[0054] S222. To capture multidimensional correlation features of multivariate sequences in parallel across different subspaces, the system divides the input hidden state matrix along the feature dimension in the current layer encoder. Each attention head is independent. In this embodiment, the system sets the number of heads for multi-head attention. In specific industrial monitoring scenarios, these eight independent attention heads help the model focus on the physical characteristics of different types of equipment, such as separating and extracting the slowly varying delayed coupling features of thermal systems and the transient abrupt changes of electrical systems. For each attention head, the system configures a corresponding learnable projection weight matrix. The model uses these weight matrices and the input hidden state matrix to perform linear transformations, projecting the high-dimensional feature space onto the head's dedicated query matrix, key matrix, and value matrix. During this mapping process, the hidden dimension of a single attention head is reduced to the original dimension. of One-third is used to ensure that the overall computational complexity of multi-head parallelism remains consistent with that of the single-head model.

[0055] S223. After obtaining the mapping matrix, the agent layer performs scaled dot product attention calculation within each attention head. The specific calculation logic is as follows: First, the system performs a matrix dot product operation on the query matrix and the transposed key matrix to generate an original attention score matrix that quantifies the correlation strength between time steps and physical measurement points within the sliding window. To prevent the dot product result from being too large and causing the gradient of the subsequent activation function to fall into the saturation region, the system introduces a scaling factor based on the square root of the single-head dimension to adjust the original attention score matrix by division scaling. Subsequently, the system applies a normalized exponential function on the feature dimension to transform the scaled values ​​into an attention weight matrix in probability distribution form. Finally, the system performs matrix multiplication on this attention weight matrix and the corresponding value matrix to complete the weighted aggregation of temporal features and output the subspace feature matrix extracted by this attention head. Through this calculation logic, the system has the ability to adaptively focus on historical key state features that have a significant impact on the current operating condition across local time steps.

[0056] S224. For the output of multi-head attention, the system concatenates the output subspace feature matrices of all attention heads along the feature dimension and maps them back to the original matrix through the output projection matrix. The system first calculates the global fusion feature of the multi-head self-attention submodule from the feature dimension. Then, it adds this global fusion feature to the initial hidden state matrix input to this layer using residuals and performs layer normalization. The processed data continues to flow through the feedforward neural network submodule. As a defined network hierarchy, the feedforward neural network submodule in this embodiment consists of two linear transformation layers and a nonlinear activation function placed between them. The system first calculates the feature dimension from the initial hidden state matrix input to the current layer. Mapping to extended dimensions Apply the ReLU activation function and then map it back. Dimensions. In this process, the expanded dimension is typically set as follows: This mechanism introduces a richer nonlinear expression space to the model. After computation by the feedforward network, the system again superimposes residual connections and layer normalization at the output.

[0057] S225. After the complete computation of the above link, the current layer encoder outputs the updated deep feature representation matrix. The system sequentially performs iterative calculations within the agent layer. The output of the bottom layer encoder serves as the input of the next layer encoder, extracting more abstract physical measurement point operation rules layer by layer. Finally, the system obtains the top-level deep feature matrix containing global multivariate coupling information and temporal dependencies. For the variance calculation mechanism of the normalization function within the multivariate Transformer model, those skilled in the art can refer to existing deep learning basic operators for configuration; this content is well-known in the field and will not be elaborated upon here.

[0058] Regarding the dynamic anomaly threshold generation and online detection process in step S300 above, after completing the time-series joint prediction of the multivariate Transformer model, the system needs to perform anomaly determination of the equipment operating status based on the prediction results. Those skilled in the art will understand that traditional industrial monitoring systems typically use manually set fixed physical thresholds as alarm judgment boundaries. However, during the actual operating cycle of generator sets or photovoltaic equipment, their internal physical parameters are affected by a combination of factors such as seasonal ambient temperature, natural aging of equipment components, and external power grid dispatch load, causing the overall baseline of the physical parameters to easily drift.

[0059] If a static, fixed judgment threshold is maintained, the system is prone to logic failure when facing dynamic operating conditions. Specifically, when the equipment is under low load or low power consumption, abnormal fluctuations in physical parameters often fail to reach the higher fixed threshold upper limit, thus increasing the risk of missed hidden faults. Conversely, when the equipment is under high load and full load, normal regular fluctuations in physical parameters often exceed the fixed safety boundary, causing invalid false alarms and interfering with the judgment of maintenance personnel.

[0060] To overcome the limitations of static boundary settings, the agent layer incorporates dynamic threshold generation logic based on adaptive operating conditions. In this embodiment, the dynamic determination mechanism includes the following processing logic: S311. The system shifts the anomaly judgment benchmark from the dimension of absolute physical observations to the level of prediction residuals output by the model. Prediction residuals characterize the degree of deviation between the current actual operating state of industrial equipment and the health state evolution law extracted by the multivariate Transformer model. At this stage, the system acquires the real physical observation vectors uploaded by the underlying sensors and performs difference calculations with the theoretical prediction vectors synchronously output by the multivariate Transformer model. Since the deep learning model implicitly learns the normal coupling fluctuation laws between multiple variables during the early training phase, the prediction residuals generated under normal operating conditions should remain within a certain range of random noise and be unaffected by the macroscopic baseline drift of the equipment.

[0061] S312. The intelligent agent layer constructs a statistical distribution model based on the extracted prediction residuals. By calculating the central tendency and dispersion of the historical residual sequence, the system generates a dynamic threshold for each independent physical measurement point that fits its current operating characteristics. This data-driven threshold generation method reduces the subjective experience bias of human intervention.

[0062] S313. In the real-time monitoring link, the agent layer compares the real-time predicted residual with the generated dynamic threshold. Through this processing method, the system achieves a mechanism for automatically adjusting the threshold boundary as the equipment operating conditions evolve. This mechanism improves the system's ability to capture minor fault characteristics while suppressing false alarm events caused by normal fluctuations in the external environment or load, providing reliable abnormal input characteristics for subsequent intelligent diagnostic processes. For the relay contact action logic and hardware-level limiting circuit principle in traditional fixed threshold alarm systems, those skilled in the art can refer to existing industrial control procedures for understanding; their content is well-known in the field and will not be elaborated upon here.

[0063] See attached document Figure 3Specifically, regarding the generation stage of threshold parameters, after clarifying the basic principle of dynamic anomaly detection based on predicted residuals, the agent layer needs to extract historical residual distribution features during the model training stage and generate baseline statistical parameters of the equipment under healthy operating conditions accordingly. To meet the conditions for model parameter generation, the system first needs to construct and train a multivariate Transformer model with accurate representation capabilities. In this embodiment, the training mechanism of the model and the subsequent residual statistical parameter generation process specifically include the following steps: S321, the system performs supervised joint training of the model based on the sliding window samples and corresponding labels constructed in the aforementioned steps. The agent layer loads the initialized multivariate Transformer model, inputs the input tensor containing the time dimension into the network, and obtains the theoretical predicted value vector of each physical measurement point in the next time step through feature extraction and output mapping of the encoder module. The system extracts the real multivariate observation vector of the corresponding time step as training labels and uses the mean squared error (MSE) as the loss function to calculate the overall numerical difference between the predicted value and the real label. Based on the calculated loss gradient, the system calls a preset optimizer (such as the Adam optimizer) to perform backpropagation, iteratively updating the projection weights and attention parameters inside the network until the network's loss value on the validation set converges to a stable state.

[0064] S322. After the multivariate Transformer model completes training and reaches convergence, the agent layer performs a global forward inference test using the complete training dataset to extract historical benchmark data. Let the total number of time-step samples in the training set be... At any historical moment The system obtains the current physical measurement point (let the measurement point index be...). The actual observed value at that moment Simultaneously, the theoretical predictions of the multivariate Transformer model output at the same time point are extracted. .

[0065] S323. The system calculates the prediction residual for each sample point in the training set across different physical measurement point dimensions. This prediction residual quantifies the deviation between the ideal state curve fitted by the model and the actual physical signal fluctuations. For each independent physical measurement point, the system subtracts its actual observed value at a specific historical moment from its corresponding theoretical prediction value to obtain the residual value for that measurement point at that moment. (Regarding the measurement point index...) At a historical moment Its residual value The calculation formula is: ; In the formula: The value is the residual. These are actual observations; This is a theoretical prediction. For measurement point index; This represents a historical moment. By traversing the entire training dataset, the system constructs a timeline for each independent measurement point, containing the total number of samples at the aforementioned time steps. The historical residual time series of each element.

[0066] S324. The intelligent agent layer performs statistical feature extraction on the historical residual time series of each measurement point. Normally, under healthy and fault-free conditions, due to the randomness of sensor environmental noise and model fitting errors, the predicted residuals generated by the measurement points approximately follow a Gaussian normal distribution. To eliminate the interference of a very small number of abnormal extreme values ​​caused by network jitter or sensor glitches on the overall distribution, the system pre-filters the residual series by truncating the upper and lower boundaries before calculation. Based on the processed residual data, the system calculates the central tendency and dispersion of the residual series for each measurement point. (The last sentence appears to be incomplete and possibly refers to a separate step regarding the measurement point index.) The system calculates the baseline mean of the residual sequence at the measurement point using the following formula. Standard deviation from the benchmark : ; ; In the formula: The baseline mean; The baseline standard deviation; The total number of samples at each time step; A historic moment; The value is the residual. This is the index of the measurement points. The benchmark mean is included. This represents the median of the basic deviation of the current physical parameters under normal operating conditions; the benchmark standard deviation Used to measure the spread width of the residual fluctuation of the current physical parameters under normal operating conditions.

[0067] S325. After extracting the statistical features of all physical measurement points, the agent layer identifies each measurement point and its corresponding benchmark mean. and benchmark standard deviation Structured binding is performed. The system encapsulates these bound baseline statistical parameters and converts them into a standard data exchange format. As a preferred approach, this format is configured as an extensible .json configuration file. The agent layer persistently stores the generated parameter configuration file, along with the weight file of the multivariate Transformer model after training and convergence, in the system's designated mount directory. After storage, the backend service layer receives the file readiness status feedback from the agent layer and prepares to call these baseline parameters in subsequent online real-time prediction stages. Regarding the specific learning rate decay strategy used by the multivariate Transformer model during training and convergence, those skilled in the art can refer to the basic training mechanisms of deep learning frameworks for implementation; this content is well-known in the field and will not be elaborated upon here.

[0068] See attached document Figure 4 After completing model training and extracting baseline statistical parameters, the system enters the online monitoring phase. At this stage, a comparison mechanism between the real-time predicted residuals and the healthy baseline needs to be established to automatically capture abnormal states. In this embodiment, the online real-time early warning mechanism based on the 3-Sigma principle specifically includes the following steps: S331. The front-end presentation layer responds to the built-in scheduled task module or operator commands by sending a real-time verification request to the back-end service layer. Upon receiving this request, the back-end service layer schedules the agent layer to load the pre-trained multivariate Transformer model weights into the NPU's computing memory. The system synchronously reads the parameter configuration file in the specified directory and performs deserialization parsing on this file. Through this parsing operation, the agent layer extracts the baseline mean value bound to each physical measurement point identifier in the current monitoring scenario. and benchmark standard deviation .

[0069] S332. The agent layer dynamically generates adaptive anomaly detection thresholds for each physical measurement point based on the baseline statistical parameters obtained through analysis. According to the 3-Sigma principle in engineering statistics, the probability that the predicted residuals of equipment physical parameters under normal operating conditions fall within three standard deviations of the baseline mean typically exceeds 99%. As a preferred approach, the system utilizes this principle to construct a safety channel that includes upper and lower boundaries. For a specific physical measurement point, the agent layer calculates its dynamic anomaly detection upper and lower limits. To avoid the generated detection channel being too narrow due to the variance of historical training data approaching zero, the system introduces a preset minimum channel constraint parameter. In specific calculations, the measurement point index is considered. The system calculates the upper limit for dynamic anomaly detection based on the following formula. Lower limit for anomaly detection : ; ; In the formula: This sets the upper limit for dynamic anomaly detection; This is the lower limit for anomaly detection; The baseline mean; The baseline standard deviation; The minimum channel constraint parameter; This serves as the index for the measurement points. Through this calculation logic, the system assigns a discrimination interval to each independent physical parameter that matches the intensity of its current fluctuation.

[0070] S333. During real-time data flow, the backend service layer collects the current physical observation vector from the data source layer. The multivariate Transformer model simultaneously completes forward inference and outputs the theoretical prediction vector for the current moment. The agent layer extracts the actual observation value and theoretical prediction value of a specific measuring point at the current moment, and calculates the real-time prediction residual for that measuring point by performing a subtraction operation. This real-time prediction residual reflects the specific magnitude by which the current transient condition deviates from the historical health benchmark.

[0071] S334. The intelligent agent layer compares the calculated real-time prediction residual with the corresponding generated upper and lower limits of dynamic anomaly judgment, and generates an anomaly flag bit for the corresponding physical measurement point accordingly. Internally, the system executes Boolean logic judgment rules: when the value of the real-time prediction residual is strictly greater than the upper limit of dynamic anomaly judgment, or strictly less than the lower limit of anomaly judgment, the system determines that the current data of the measurement point has exited the safe channel and assigns the corresponding anomaly flag variable a value of 1; conversely, when the value of the real-time prediction residual is between the lower and upper limits of anomaly judgment, i.e., within the safe channel, the system determines that the current data of the measurement point is normal and assigns the corresponding anomaly flag variable a value of 0.

[0072] S335. To prevent single-point data spikes caused by electromagnetic interference in industrial settings from falsely triggering the system's diagnostic procedures, a time window smoothing mechanism is configured at the back end of the abnormal flag variable sequence. The system counts the number of abnormal flag variables consecutively assigned a value of 1 within a set time window. When the number of consecutive limit violations does not reach the preset alarm count threshold, the system classifies it as a normal disturbance and filters it out. When the number of consecutive limit violations reaches the preset alarm count threshold, the system summarizes the abnormal flag variables corresponding to all physical measurement points participating in the joint modeling at the current moment. The backend service layer executes state branch flow based on the summary results.

[0073] S336. When all anomaly flag variables at all measurement points are equal to 0 or fail the verification of the time window smoothing mechanism, the system determines that the current equipment is in normal operation. The backend service layer pushes a regular data packet containing theoretical predictions and actual observations to the frontend display layer for chart rendering. When any measurement point has an anomaly flag variable equal to 1 and meets the continuous over-limit counting condition, the system determines that the equipment has suspected fault characteristics. At this time, the backend service layer extracts the prediction deviation vector containing the real-time prediction residuals of all measurement points and uses it as the structured input feature of the abnormal environment to trigger the subsequent intelligent diagnostic link. For the DOM node update mechanism and WebGL underlying drawing logic of the frontend chart rendering engine after receiving the regular data packet, those skilled in the art can refer to the existing frontend visualization component library documentation for development. The relevant processing procedures are well-known technologies in this field and will not be elaborated here.

[0074] Regarding the intelligent diagnosis process in step S400 above, when the equipment operating status is determined to be abnormal by the aforementioned adaptive threshold mechanism, the system workflow automatically switches to the intelligent diagnosis stage. Traditional industrial alarm systems mostly output alarm codes or abnormal measurement point names, rarely providing direct analysis of the underlying mechanisms of abnormal data fluctuations. To establish a bridge between numerical computation and semantic understanding, this invention integrates a large language model in the intelligent agent layer and designs a context injection mechanism for abnormal features. This mechanism aims to transform the structured residual data output by the underlying numerical computation model into natural language context that the large language model can understand. In this embodiment, the process specifically includes the following steps: S411. Upon receiving an anomaly trigger signal, the backend service layer sends an anomaly feature extraction request to the agent layer. In response, the agent layer extracts all anomaly data generated at the current moment from memory. To provide structured input to the large model, the agent layer constructs a prediction bias vector sequence. This sequence contains the identifiers of all physical measurement points marked as abnormal at the current moment, the corresponding real-time prediction residual values, and the dynamic anomaly judgment boundary values ​​that trigger alarms.

[0075] S412. In complex industrial equipment failure scenarios, a single root cause often triggers a chain reaction of alarms exceeding limits at multiple related measurement points. If all abnormal measurement point data is fully injected, the generated text risks exceeding the maximum input token limit of the large language model. To avoid inference truncation due to input overload, the agent layer is configured with abnormal feature sorting and filtering logic. As a preferred approach, the system calculates the normalized deviation of each abnormal measurement point. In practice, the system extracts the absolute value of the real-time prediction residual of a specific abnormal measurement point and divides this absolute value by the baseline standard deviation statistically obtained from the measurement point under historical healthy operating conditions. The quotient is the normalized deviation of that measurement point. By calculating the normalized deviation, the system eliminates numerical differences between different physical dimensions. The agent layer sorts the abnormal measurement points in descending order based on the magnitude of the deviation and truncates the top-ranked points. Data from several core abnormal measurement points serves as the final diagnostic input source, among which... This is the preset maximum allowed number of injection measurement points.

[0076] S413, the intelligent agent layer, is equipped with a dedicated prompt word engineering module to handle the filtered core anomaly data. This module contains pre-built basic prompt word templates for different types of industrial equipment. These templates include role setting declarations, diagnostic task target instructions, and blank feature placeholders. Based on the type of equipment currently experiencing an anomaly (e.g., conventional thermal power identification HNJS or distributed photovoltaic identification HNGF), the system matches and loads the corresponding basic prompt word template from the template library.

[0077] S414. After loading the basic prompt word template, the agent layer performs structured text conversion of abnormal features. The system converts the extracted real-time prediction residuals and dynamic threshold values ​​into quantified textual descriptions, such as calculating the specific percentage or absolute deviation of the residuals from the safe channel. Subsequently, the system concatenates the abnormal measurement point identifiers with the converted numerical descriptions into an abnormal feature statement with contextual logic. As an example, this statement can be constructed as "The actual observed value of the bearing temperature measurement point of the current generator set is XX degrees Celsius higher than the model's health baseline prediction value, and exceeds the dynamic upper limit of the operating condition by XX percentage points."

[0078] S415. The agent layer injects the generated abnormal feature statements into the designated empty placeholders of the basic prompt word template, thereby dynamically constructing a complete diagnostic request prompt word vector. This constructed diagnostic request prompt word vector carries prior context clues about the current device anomaly. The system packages and encapsulates it and sends it to the downstream large language model inference link through the MINDIE inference service (such as calling the / chat / send interface). Through this context injection mechanism, the large language model obtains quantified boundary conditions before performing fault inference, which helps to generate more targeted fault cause diagnosis results. For the text tokenization processing and vector embedding methods of the underlying Transformer architecture of the large language model, those skilled in the art can refer to the standard workflow of mainstream natural language processing frameworks for understanding. The content belongs to the well-known technology in this field and will not be elaborated here.

[0079] Furthermore, to enhance the professionalism of diagnostic results, the pre-training corpus of general-purpose large language models often fails to fully cover the underlying maintenance manuals or procedural details of specific equipment models when handling tasks in vertical industrial domains. To guide the large model to output targeted fault handling guidance, the agent layer is configured with a domain knowledge retrieval module based on a retrieval enhancement generation mechanism. In this embodiment, the retrieval enhancement process specifically includes the following steps: S421. The intelligent agent layer pre-constructs an external domain knowledge base. This knowledge base includes unstructured text data such as gas turbine manuals, photovoltaic equipment operation and maintenance procedures, and historical expert diagnostic reports. The system uses a text parsing component to segment these documents into fixed-length text blocks. During this process, the system sets a segmentation threshold based on the maximum input length allowed by the underlying embedding model, for example, 512 token units. After segmentation, the system calls the embedding model to process each text block. As a specific implementation, this embedding model adopts a BERT architecture based on a bidirectional Transformer encoder. The system inputs discrete text blocks into the network layer of this embedding model, uses its internal multi-head self-attention mechanism to extract global semantic features, and takes the hidden layer representation of the [CLS] special marker node at the output, mapping it to a dense vector in a high-dimensional feature space. The system stores these dense vectors along with the corresponding text content into a vector index library, completing the initialization of the retrieval base.

[0080] S422. After entering the online diagnostic process, the agent layer obtains the diagnostic request prompt words generated in the aforementioned steps. The system calls the same embedding model used when building the knowledge base to transform the diagnostic request prompt words into a quantized query vector. This query vector represents the semantic direction of the current industrial equipment's fault anomaly in the feature space.

[0081] S423. After obtaining the query vector, the agent layer performs similarity matching calculations in the vector index. The system traverses the knowledge base text block vectors in the vector index and calculates the spatial distance between the query vector and each text block vector. As a preferred method, the system uses cosine similarity as the matching scoring standard. Specifically, the system extracts the query vector and the specific knowledge base text block vector, and calculates the inner product value between these two high-dimensional vectors. Subsequently, the system calculates the Euclidean norm of each of the query vector and the knowledge base text block vector, i.e., the corresponding vector magnitude. Finally, the system divides the calculated inner product value by the product of the two vector magnitudes, and the quotient is the similarity score between the query vector and the corresponding knowledge base text block vector. Through this calculation logic, the system quantifies the semantic relevance between the input fault features and various professional knowledge fragments.

[0082] S424. The agent layer sorts all candidate text blocks in descending order based on the calculated similarity scores. To avoid introducing noise that could interfere with the large model's judgment, the system configures a lower similarity threshold to remove irrelevant text blocks with scores below this threshold. This threshold is typically set based on the dispersion of the domain corpus, for example, 0.75. For the remaining highly relevant text blocks after filtering, the system extracts the top-ranked items and structures the physical troubleshooting procedure texts contained within them to form the retrieval knowledge context. When the scores of all text blocks in the vector index are below the set lower similarity threshold, the system determines that no valid external knowledge has been found. In this case, the retrieval knowledge context is initialized to an empty string. This processing logic ensures the execution consistency of the retrieval algorithm under extremely rare failures.

[0083] S425, the agent layer performs text-level concatenation of the generated retrieval knowledge context with the original diagnostic request prompts. The merged data stream constitutes enhanced prompts containing external expert knowledge background. This mechanism, combined with an external knowledge base, helps supplement the large language model with physical references for specific operating conditions, providing clear industrial mechanism support for its subsequent fault inference. Regarding the underlying hierarchical navigation small-world graph index construction algorithm of the vector database and the specific overlap rate setting mechanism for text segmentation, those skilled in the art can refer to existing vector retrieval frameworks for implementation; these are well-known technologies in the field and will not be elaborated upon here.

[0084] After constructing the context of abnormal features and retrieving external domain knowledge, the system workflow enters the natural language generation stage. This stage, by leveraging the computing power of the large language model at the bottom layer of the intelligent agent, transforms structured numerical deviations and retrieved text into fault handling guidance readable by operations and maintenance personnel. In this embodiment, the inference generation mechanism based on the MINDIE gateway and LoRA fine-tuning specifically includes the following steps: S431. To enable the general-purpose large language model to analyze the fault mechanisms of specific industrial equipment, the system performs LoRA-supervised fine-tuning training offline. The system collects historical fault work orders and maintenance records from the industrial site. Based on these records, the system constructs fine-tuning sample pairs, where the sample input consists of diagnostic request prompts containing abnormal measurement point deviations and equipment types, and the sample labels are texts written by human experts detailing the specific fault causes and handling instructions. On this basis, the system fixes the original fixed parameter matrix of the basic large language model (such as the DeepSeek-32B model with a multi-layer Transformer decoder structure) and injects fine-tuning weights consisting of the product of two low-rank matrices into the self-attention module of a specific Transformer layer via a bypass. Let the original hidden dimension of the injected layer be... The preset low-rank parameter is (For example, take) The weights injected via the bypass are constructed as a dimensionality-reduced matrix. With increasing dimension matrix The product of the two. The system uses cross-entropy as the loss function and utilizes the above-mentioned fine-tuning of the sample pair matrix. and Backpropagation gradient updates are performed, and after training converges, dedicated LoRA fine-tuning weights are generated.

[0085] S432. During the online inference phase, the agent layer acquires the enhanced prompt words generated in the preceding steps. These enhanced prompt words are composed of diagnostic request prompt words containing device anomaly characteristics and text-level concatenation of the retrieved knowledge context. The system uses a text segmenter configured at the lower level to segment these enhanced prompt words, mapping them into an input feature vector matrix that can be processed by a large language model. This input feature vector matrix integrates the current device's real-time quantization deviation and relevant historical procedural guidance in the data dimension.

[0086] S433, the intelligent agent layer processes the aforementioned input feature vector matrices through the internally integrated MINDIE inference gateway. As an AI computing power routing and scheduling component, the MINDIE inference gateway is responsible for sending external request tasks to the underlying Yiteng 310PNPU hardware computing unit and allocating corresponding GPU memory space. When faced with sudden diagnostic requests caused by concurrent alarms from multiple measurement points, the MINDIE inference gateway, based on a dynamic batch processing strategy, aggregates and sends multiple input feature vector matrices in parallel within a latency-allowed window. On this computing node, the system loads the basic large language model and the LoRA fine-tuned weights trained in the previous steps.

[0087] S434. With the support of hardware computing power, the basic large language model combines LoRA to fine-tune the weights and perform autoregressive forward inference on the input feature vector matrix. Let the original fixed parameter set of the basic large language model be... The set of LoRA fine-tuning weight parameter matrices for bypass injection is as follows: The input diagnostic request prompt sequence is: The retrieved knowledge context sequence is During specific reasoning and calculation, the system will use the diagnostic request prompt sequence. With retrieval knowledge context sequence The text sequence is concatenated to form fused input features. Subsequently, the large language model uses the original fixed parameter set... With LoRA fine-tuning weight parameter matrix set With the synergistic effect of [unclear], the internal autoregressive generator function is invoked to process the aforementioned fused input features. The large language model performs logical deduction based on the abnormal features of the input and performs word-by-word prediction with reference to the retrieved standard text, ultimately outputting a natural language diagnostic text vector representing the cause of the fault and suggested solutions. .

[0088] S435, the agent layer will generate diagnostic text vectors The data is decoded into a natural language string and encapsulated into a standard format diagnostic response payload, which is then returned to the backend service layer. The backend service layer pushes this diagnostic result along with the real-time time-series data stream to the frontend presentation layer. Upon receiving the data, the frontend presentation layer renders the corresponding natural language text in the diagnostic result box of the multi-panel monitoring interface. Simultaneously, the frontend presentation layer updates the monitoring chart status in conjunction with this diagnostic text, highlighting abnormal measurement points that generate alarms and their corresponding time-series curves. For the specific decoding strategies employed by the large language model during autoregressive generation, such as core sampling or temperature penalty, those skilled in the art can refer to the basic configuration of existing natural language generation tasks for implementation; these are well-known technologies in the field and will not be elaborated upon here.

[0089] Regarding the elastic computing scheduling mechanism in step S500 above, in industrial control center scenarios, the system typically needs to run a large number of data monitoring panels concurrently. Since the multivariate Transformer model and large language model deployed within the agent layer consume significant hardware computing power during inference, continuous background polling of all panels can easily lead to excessive consumption of underlying NPU computing resources and congestion of request channels. To establish a dynamic balance between computing power overhead and monitoring requirements, this embodiment designs an elastic computing and on-demand resource scheduling mechanism in the front-end presentation layer. This mechanism specifically includes the following processing steps: S511, the front-end presentation layer constructs a multi-panel monitoring interface. Operators create multiple independent chart rendering containers within the same monitoring view. For each chart rendering container, the front-end presentation layer provides an independent parameter configuration interface. Through this interface, operators bind corresponding multivariate time-series prediction model identifiers, auxiliary diagnostic large language model identifiers, and timed refresh cycle parameters to specific panels. This decoupled configuration mechanism allows the system to monitor industrial equipment tree nodes with different sampling frequencies or different operating conditions in parallel within the same physical screen.

[0090] S512: The front-end presentation layer obtains the timed refresh cycle parameters set for each panel and constructs a polling timer for the corresponding lifecycle. The system performs conditional branching on the input cycle value. When the refresh cycle parameter of a specific panel is equal to 0, the front-end presentation layer determines that panel as an offline static data view and removes the timed task associated with that panel from the event loop. When the refresh cycle parameter is strictly greater than 0, the system allocates an independent timing thread for it. This timing thread sends real-time verification requests containing the target model identifier to the back-end service layer at fixed time intervals according to the set cycle length. To prevent request backlog storms caused by network latency, the system injects request lock status check logic before initiating verification requests. If the previous verification request for the current panel has not yet received a response payload from the back-end, the timer will actively abandon sending this request until the previous response ends and the request lock is released.

[0091] S513. To reduce computational overhead in non-visual states, the front-end presentation layer injects page focus and visibility monitoring logic into the global routing node. As a preferred approach, the system utilizes the PageVisibilityAPI at the underlying level of the front-end browser environment to capture state change events of the current monitoring tab. This monitoring logic enables the system to detect in real time whether the operator's business window focus remains within the current real-time verification panel area.

[0092] S514. The front-end presentation layer dynamically takes over the state machine flow of the underlying scheduled tasks based on captured page state change events. When the system captures events such as the current tab being hidden, the browser window being minimized, or the user switching to another non-monitored route within the system, the front-end presentation layer triggers a global suspension command. This command suspends all active polling timers on the current page. During the timer suspension period, the front-end presentation layer cuts off the timed request data stream sent to the back-end service layer. Since no new verification trigger command is received, the computing nodes corresponding to the agent layer enter a computing power idle and memory release state.

[0093] S515. When the operator's view focus switches back to the real-time verification tab, and the system captures the event that the page has returned to a visible state, the front-end presentation layer triggers a wake-up command. The system extracts the configuration context cached by each panel before suspension, reactivates and initializes the corresponding polling timer. To compensate for the data timeline gap caused during view suspension, before resuming polling, the front-end presentation layer calculates the time difference based on the suspension timestamp and the current wake-up timestamp, and automatically sends a compensation range query request to the backend, plotting the missing historical actual data and model prediction data into the chart at once. After completing the gap tracing, the system resumes regular cross-layer data communication and inference scheduling with the backend service layer and the agent layer. With the above-mentioned linkage logic of starting and stopping on demand based on the front-end view state, the system avoids redundant network requests in the invisible state and achieves elastic allocation of the underlying NPU hardware computing power. For the specific mounting method of component lifecycle hook functions in the front-end single-page application (SPA) framework and the prevention and cleanup mechanism of timer memory leaks, those skilled in the art can refer to existing front-end engineering development specifications for implementation. The content belongs to the well-known technology in this field and will not be elaborated here.

[0094] Regarding the system security protection system in step S500 above, in industrial production and power control scenarios, the leakage of core data or external network attacks often leads to unexpected production accidents. To meet the data boundary protection requirements of industrial control systems, this embodiment constructs a multi-dimensional system security protection system in the interaction links of the data source layer, backend service layer, intelligent agent layer, and frontend presentation layer. The specific implementation logic of this protection system includes the following steps: S521. The system executes network compliance zone deployment logic. In accordance with the security protection regulations for power secondary systems, the system is divided into different security isolation domains at the physical and logical network levels. As a specific deployment architecture, the core API components of the backend service layer and the business database of the data source layer are deployed in the third secure intranet zone. The AI ​​inference model and external knowledge base carried by the intelligent agent layer are deployed in the fourth secure buffer zone. Forward and reverse physical isolation gateways are configured between the two zones to achieve physical boundary blocking between core production data and external access networks, preventing unauthorized network domains from penetrating the underlying industrial data without authorization.

[0095] S522. During the data acquisition phase, the system executes channel isolation logic. Based on the aforementioned measurement point prefix routing mechanism, the backend service layer performs authentication and traffic separation for distributed photovoltaic data sources and conventional thermal power data sources. The request subsets corresponding to conventional thermal power identifiers and the request subsets corresponding to distributed photovoltaic identifiers are allocated to independent memory processing spaces and physical storage slices. This mechanism establishes access isolation between heterogeneous data sources, helping to prevent cross-access of multi-tenant business data and data flow pollution.

[0096] S523. The system is configured with end-to-end transmission encryption and internal service authentication mechanisms. At cross-domain interaction nodes between the backend service layer and the frontend presentation layer, the system uses HTTPS encrypted transmission protocol to ensure the confidentiality of communication payloads. During communication between microservice nodes within the backend service layer, the system uses the Feign interceptor to inject dynamic authorization tokens into the request headers. This internal authentication mechanism ensures the legitimate verification of microservice communication links, preventing malicious nodes from impersonating and stealing data exchange messages within the cluster.

[0097] S524. System access control policy for the model interface. The multivariate Transformer model and large language model deployed within the agent layer consume significant underlying NPU computing resources during operation. To prevent direct probing and malicious calls from the public network environment, the system configures the agent layer's network ports to allow access only from designated internal network segments, cutting off its public network exposure. Any model training or prediction commands initiated by the front-end presentation layer must undergo identity authentication and parameter filtering by the back-end service layer, and be forwarded via a pre-defined internal routing component.

[0098] S525. The system is configured with traffic-aware rate limiting and circuit breaker protection logic. In industrial settings, sudden surges in concurrent monitoring requests can easily occur during centralized equipment startup / shutdown or network jitter recovery phases, potentially leading to memory overflow in the underlying computing units or task queue blockage. The backend service layer integrates a circuit breaker protection component at the request gateway. As a preferred approach, this component uses Spring Cloud Hystrix, allocating independent concurrency limits to different business interfaces based on thread pools or semaphore isolation strategies. When the number of concurrent requests routed to the intelligent agent layer reaches the set threshold, the circuit breaker protection component triggers rate limiting interception, quickly returning a preset default status message or cached data to the excessive caller, thereby maintaining the operational stability of the core scheduling system.

[0099] S526. The system performs a full audit log operation for AI inference. For the large language model question-and-answer interaction and time-series anomaly diagnosis process at the agent layer, the system constructs an independently running asynchronous logger. This logger extracts the user's operation identity, query timestamp, original diagnostic request prompts, RAG plugin knowledge base hit fragments, and the natural language diagnostic text finally generated by the model, without blocking the main business thread. The system serializes and persists the above-mentioned related data to an independent audit database, providing a traceable basis for subsequent fault review, model compliance review, and data leakage tracing. Regarding the TLS handshake key exchange process at the underlying HTTPS protocol and the specific data transfer mechanism of the physical isolation gateway, those skilled in the art can refer to existing network security standards and specifications for configuration; the content is well-known technology in this field and will not be elaborated here.

Claims

1. A domestically developed, full-stack industrial intelligent multivariate time series prediction and intelligent diagnostic system, characterized in that, include: Data source layer, backend service layer, intelligent agent layer, and frontend presentation layer; The data source layer is used to access heterogeneous operational data from external devices; The backend service layer, connected to the data source layer, is used to perform traffic splitting and time-series preprocessing on the running data through a measurement point prefix routing mechanism, and to construct sliding window samples; specifically, when constructing sliding window samples, the backend service layer is used for: Construct a set of reference timestamp sequences with equal step sizes, and map the original non-fixed-interval running data onto the reference timestamp sequences; When the data loss period at a certain measuring point is less than the preset tolerance period, the most recent valid historical observation value is used for forward filling. Linear interpolation is performed when the missing duration exceeds the tolerance period and there are valid first and last observation nodes. Extract continuous multivariate observation vectors of fixed length, and construct a two-dimensional sliding window matrix by vertically splicing them in chronological order. Stack multiple two-dimensional sliding window matrices along the new dimension to reconstruct a three-dimensional input tensor containing three dimensions: batch size, sliding window length, and total number of physical measurement points, as the sliding window sample. The agent layer, connected to the backend service layer, deploys a multivariate Transformer model and a large language model. It is used to extract features and model temporal coupling relationships of the sliding window samples through the multivariate Transformer model, outputting predicted values ​​and prediction bias vectors for each measurement point. The multivariate Transformer model generation process includes: supervised joint training of the model based on the constructed sliding window samples and corresponding labels; the agent layer loads the initialized multivariate Transformer model, inputs the input tensor containing the time dimension into the network, performs feature extraction and output mapping through the encoder module to obtain the theoretical predicted value vectors for each physical measurement point in the next time step, extracts the real multivariate observation vectors for the corresponding time step as training labels, and uses mean squared error as the loss function to calculate the overall numerical difference between the predicted values ​​and the real labels. The intelligent agent layer is also used to extract the historical prediction residuals generated during the training phase of the multivariate Transformer model to generate dynamic anomaly thresholds, including: generating dynamic thresholds for each independent physical measurement point that fit its current working condition characteristics by calculating the central tendency and dispersion of the historical residual sequence. The backend service layer is also used to perform online anomaly detection by comparing the dynamic anomaly threshold with the real-time prediction deviation vector. When an abnormal measurement point is detected, the backend service layer injects the prediction deviation vector into the preset diagnostic prompt words, and the intelligent agent layer calls the large language model that integrates domain knowledge retrieval and fine-tuning to output intelligent diagnostic results; The front-end presentation layer connects to the back-end service layer and is used to perform elastic computing scheduling based on page state, and to implement multi-dimensional security protection strategies in each level of the interaction link.

2. The domestically produced industrial intelligent full-stack multivariate time series prediction and intelligent diagnostic system as described in claim 1, characterized in that, When the backend service layer uses the test point prefix routing mechanism to distribute runtime data, it is specifically used for: Extract the measurement point prefix of each measurement point identifier in the set of measurement point identifiers corresponding to the input running data, call the internally stored routing mapping table, and determine the target data interface terminal point corresponding to the measurement point prefix; Based on the same target data interface terminal point, the corresponding measurement point identifiers are aggregated into independent request subsets, and each request subset is sent to the corresponding physical node or logical service area in the data source layer. When initiating a data acquisition request, if the status monitoring component built into the backend service layer identifies that the interface response time exceeds the request timeout threshold set by combining the historical average response time of the underlying database and the latency tolerance, or if the error rate reaches the circuit breaker limit set according to the system's concurrent load capacity, the backend service layer triggers a degradation fallback logic, redirecting the data stream acquisition instruction to the internally configured simulated data generation module, and generating a simulated time series sequence with superimposed Gaussian white noise based on the mean and variance of the corresponding measurement points under historical normal operating conditions, and returning it.

3. The domestically produced industrial intelligent full-stack multivariate time series prediction and intelligent diagnostic system as described in claim 1, characterized in that, When the intelligent agent layer extracts features from the sliding window samples using a multivariate Transformer model, it is specifically used for: The fully connected mapping network built inside the multivariate Transformer model performs temporal feature linear mapping on the input three-dimensional input tensor, and generates an initial feature representation matrix using a linear projection weight matrix initialized with uniform distribution. A position encoding matrix is ​​constructed based on the sliding window length of the sliding window sample and the preset hidden layer dimension. For the even-numbered dimension index position and the odd-numbered dimension index position in the position encoding matrix, a sine function and a cosine function based on the same frequency scaling factor are used for mapping and assignment, respectively. The initial feature representation matrix and the location encoding matrix are subjected to element-wise matrix addition to generate a hidden state matrix that incorporates location information.

4. The domestically produced industrial intelligent full-stack multivariate time series prediction and intelligent diagnostic system as described in claim 3, characterized in that, The multivariate Transformer model contains multiple stacked encoder layers. When the agent layer models temporal coupling relationships in the current encoder layer, it is specifically used for: The input hidden state matrix is ​​divided into multiple independent attention heads along the feature dimension, and the hidden state matrix is ​​mapped into a query matrix, a key matrix, and a value matrix respectively using the corresponding learnable projection weight matrix; Perform a matrix dot product operation between the query matrix and the transposed key matrix to generate the original attention score matrix; after adjusting the original attention score matrix by division using a scaling factor based on the square root of the single-head dimension, apply a normalized exponential function to transform it into an attention weight matrix. Perform matrix multiplication between the attention weight matrix and the corresponding value matrix to output the subspace feature matrix; After concatenating the subspace feature matrices of all attention heads and mapping them back to the original dimension, the updated deep feature representation matrix is ​​output through residual connections, layer normalization processing, and a feedforward neural network submodule.

5. The domestically produced industrial intelligent full-stack multivariate time series prediction and intelligent diagnostic system as described in claim 1, characterized in that, When the intelligent agent layer extracts historical prediction residuals from the model training phase to generate dynamic anomaly thresholds, it is specifically used for: Using the complete training dataset of the multivariate Transformer model, a global forward inference test is performed. The actual observed value of each measurement point at any historical moment is subtracted from the corresponding theoretical prediction value to obtain the residual value of the corresponding measurement point at the corresponding moment, and a historical residual time series is constructed. After performing upper and lower boundary truncation filtering on the historical residual time series, the residual values ​​of each measurement point in the historical residual time series are summed and divided by the total number of samples to obtain the baseline mean. Calculate the sum of squares of the differences between each residual value and the benchmark mean, divide by the total number of samples, and then take the square root to obtain the benchmark standard deviation. The identifiers of each measuring point are structurally bound to the corresponding benchmark mean and benchmark standard deviation.

6. The domestically produced industrial intelligent full-stack multivariate time series prediction and intelligent diagnostic system as described in claim 5, characterized in that, When the backend service layer performs online anomaly detection, it is specifically used for: Based on the 3-Sigma principle, the reference standard deviation of each measurement point is multiplied by 3, and the calculation result is compared with the preset minimum channel constraint parameter, and the maximum value of the two is taken. The benchmark mean and the maximum value are added and subtracted respectively to obtain the upper limit and lower limit of dynamic anomaly judgment, which are used as the dynamic anomaly threshold. Obtain the real-time prediction residual value of each measuring point at the current moment from the real-time prediction deviation vector; When the real-time prediction residual value is strictly greater than the upper limit of the dynamic anomaly judgment, or strictly less than the lower limit of the anomaly judgment, it is determined that the current data of the corresponding measurement point has jumped out of the safe channel, and the corresponding anomaly flag variable is assigned a value of 1.

7. The domestically produced industrial intelligent full-stack multivariate time series prediction and intelligent diagnostic system as described in claim 1, characterized in that, When the backend service layer injects the prediction deviation vector into the preset diagnostic prompt words, it is specifically used for: Extract the absolute value of the real-time prediction residual of each abnormal measurement point in the prediction deviation vector at the current moment, and divide the absolute value by the benchmark standard deviation of the corresponding measurement point under historical healthy working conditions to calculate the normalized deviation of the corresponding abnormal measurement point. The abnormal measurement points are sorted in descending order according to the magnitude of the normalized deviation, and the data of the top K core abnormal measurement points are extracted. Load a preset basic prompt word template that matches the current device category, convert the extracted real-time prediction residual and the dynamic anomaly threshold into a quantified text description, and concatenate the anomaly measurement point identifier with the converted text description into an anomaly feature statement, which is then injected into the specified empty placeholder of the basic prompt word template to dynamically construct a complete diagnostic request prompt word vector.

8. The domestically produced industrial intelligent full-stack multivariate time series prediction and intelligent diagnostic system as described in claim 7, characterized in that, When the intelligent agent layer invokes the large language model that integrates domain knowledge retrieval and fine-tuning, it is specifically used for: The internal embedding model is invoked to convert the diagnostic request prompt word vector into a quantized query vector. The cosine similarity score is obtained by calculating the inner product value between the query vector and each knowledge base text block vector in the pre-built vector index library, and dividing it by the product of the Euclidean norms of the query vector and the knowledge base text block vector. After removing irrelevant text blocks whose scores are lower than the similarity threshold set based on the dispersion of the domain corpus, several top-ranked items are extracted, and their physical troubleshooting procedure texts are extracted and structurally concatenated to form a retrieval knowledge context. The retrieval knowledge context is then concatenated with the original diagnostic request prompt word vector at the text level to generate enhanced prompt words. The enhanced prompts are distributed to the corresponding hardware computing units via the MINDIE inference gateway. Based on the original fixed parameter set of the large language model, and combined with the LoRA fine-tuning weight parameter matrix set consisting of the product of the dimension reduction matrix and the dimension increase matrix injected by the bypass, the intelligent diagnostic results are output by performing autoregressive forward inference.

9. The domestically produced industrial intelligent full-stack multivariate time series prediction and intelligent diagnostic system as described in claim 1, characterized in that, When the front-end presentation layer performs elastic calculation scheduling based on page state, it is specifically used for: Capture the state change event of the current monitoring tab in the front-end display interface using the page visibility application interface; When the event of the tab being hidden is captured, a global suspension command is triggered, pausing all built-in polling timers that are active on the current page, and cutting off the timed request data stream sent to the backend service layer, so that the corresponding computing node enters the computing power idle and memory release state. When an event is captured that the tab has become visible again, a wake-up command is triggered to reactivate and initialize the corresponding polling timer; Before resuming polling, the time difference is calculated based on the suspended timestamp and the current wake-up timestamp, and a compensation range query request is automatically sent to the backend service layer to draw the missing historical actual data and model prediction data.

Images