Online monitoring system for production process of high-purity lithium salt
By collecting and analyzing variables in the lithium salt production process in real time through an online monitoring system, and using a deep time-series prediction model, real-time quality monitoring of the high-purity lithium salt production process was achieved. This solved the problem of information lag caused by offline detection and improved the monitoring accuracy and control capability of the production line.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG STARRY PHARMA
- Filing Date
- 2026-04-08
- Publication Date
- 2026-05-12
AI Technical Summary
In the existing lithium salt production process, the quality monitoring of high-purity lithium salts relies on offline detection, which leads to delayed information feedback and the inability to monitor in real time, resulting in the production of a large number of inferior products.
A high-purity lithium salt production process online monitoring system is adopted. The system collects process variables in real time through the data acquisition module, constructs a time-series feature matrix, and uses a deep time-series prediction model to predict quality and generate alarm information, thereby realizing real-time quality monitoring.
It enables real-time quality monitoring of the lithium salt production process, improves monitoring accuracy, reduces the production of substandard products, and enhances the real-time control capability of the production line.
Smart Images
Figure CN122020064A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of online monitoring technology for lithium salt production, and more specifically, to an online monitoring system for the production process of high-purity lithium salts. Background Technology
[0002] The level of trace impurities (especially moisture content and free metal ion concentration) inside high-purity lithium hexafluorophosphate constitutes a key limiting factor for the battery's electrochemical window, cycle life, and thermal runaway critical temperature.
[0003] In current industrial practice, the verification of lithium hexafluorophosphate product quality mainly relies on offline testing in central laboratories. This offline testing includes chemical titration and mass spectrometry analysis, which involves sampling and detection delays of several hours. Due to the continuous nature of the crystallization and drying processes, this information feedback causes a lag. Furthermore, by the time the analytical instruments detect quality exceeding the limits, a large amount of irreversible substandard products have already been produced and packaged on the production line, making quality monitoring impossible in complex, nonlinearly coupled chemical scenarios. Summary of the Invention
[0004] This invention provides an online monitoring system for the production process of high-purity lithium salts, which at least solves the problem of low accuracy in lithium salt quality monitoring in related technologies.
[0005] According to one embodiment of the present invention, an online monitoring system for high-purity lithium salt production process is provided, comprising: The data acquisition module is used to collect process variables in real time and construct a time-series feature matrix based on the process variables; The quality prediction module is used to input the time series feature matrix into a pre-trained deep time series prediction model to obtain the predicted value of the target product quality index. The alarm generation module is used to compare the predicted value with the preset quality specifications, and generate and output alarm information when the predicted value meets the warning conditions.
[0006] In one exemplary embodiment, the real-time acquisition process variables include: A health assessment is performed on the real-time data of the process variables to generate health labels; When the health status label is determined to be abnormal, a sensor maintenance alarm is triggered.
[0007] In one exemplary embodiment, performing a health assessment on the real-time data of the process variable includes: A flatness check is performed on the real-time data, wherein when the change in the value of the real-time data within a preset check time is less than a preset flatness threshold, the health label is determined to be abnormal. And / or, A change rate check is performed on the real-time data, wherein when the instantaneous change rate of the real-time data is greater than a preset change rate threshold, the health label is determined to be abnormal.
[0008] In one exemplary embodiment, after generating and outputting alarm information, the following steps are also performed: Obtain the actual test values of the target product quality indicators; Based on the predicted value and the actual test value, calculate the multidimensional prediction residual vector; The multidimensional prediction residual vector is normalized based on the inverse of the historical residual covariance matrix. An exponentially weighted moving average is calculated on the first distance sequence within a preset time period to obtain a smoothing control statistic. The first distance is obtained through a normalization process. When the smoothing control statistic is greater than the preset mismatch control limit, the model mismatch condition is determined to be met, and a model update instruction is triggered.
[0009] According to another embodiment of the present invention, an online monitoring method for high-purity lithium salt production process is provided, comprising: Real-time acquisition of process variables; Construct a time-series feature matrix based on the process variables; The time-series feature matrix is input into a pre-trained deep time-series prediction model to obtain the predicted values of the target product quality indicators. The predicted value is compared with the preset quality specification, and when the predicted value meets the warning conditions, an alarm message is generated and output.
[0010] In one exemplary embodiment, the real-time acquisition process variables include: A health assessment is performed on the real-time data of the process variables to generate health labels; When the health status label is determined to be abnormal, a sensor maintenance alarm is triggered.
[0011] In one exemplary embodiment, performing a health assessment on the real-time data of the process variable includes: A flatness check is performed on the real-time data, wherein when the change in the value of the real-time data within a preset check time is less than a preset flatness threshold, the health label is determined to be abnormal. And / or, A change rate check is performed on the real-time data, wherein when the instantaneous change rate of the real-time data is greater than a preset change rate threshold, the health label is determined to be abnormal.
[0012] In one exemplary embodiment, after generating and outputting the alarm information, the method further includes: Obtain the actual test values of the target product quality indicators; Based on the predicted value and the actual test value, calculate the multidimensional prediction residual vector; The multidimensional prediction residual vector is normalized based on the inverse of the historical residual covariance matrix. An exponentially weighted moving average is calculated on the first distance sequence within a preset time period to obtain a smoothing control statistic. The first distance is obtained through a normalization process. When the smoothing control statistic is greater than the preset mismatch control limit, the model mismatch condition is determined to be met, and a model update instruction is triggered.
[0013] According to yet another embodiment of the present invention, a computer-readable storage medium is also provided, wherein a computer program is stored therein, wherein the computer program is configured to perform the steps in any of the above method embodiments when executed.
[0014] According to yet another embodiment of the present invention, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.
[0015] This invention extracts historical data of multidimensional process parameters and establishes a mapping model between process fluctuations and product impurity concentration to output quality predictions in real time, thereby achieving accurate online monitoring. Therefore, it can solve the problem of low accuracy in online monitoring of lithium salt production and improve the accuracy of online monitoring in lithium salt production. Attached Figure Description
[0016] Figure 1 This is a flowchart of an online monitoring method for a high-purity lithium salt production process according to an embodiment of the present invention; Figure 2 This is a structural block diagram of an online monitoring system for high-purity lithium salt production process according to an embodiment of the present invention. Detailed Implementation
[0017] The technical solutions of the embodiments of this application will be described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.
[0018] In the following description, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.
[0019] Furthermore, in this application, directional terms such as "upper," "lower," "left," and "right" may be defined relative to the orientation of the components shown in the accompanying drawings. It should be understood that these directional terms can be relative concepts, used for relative description and clarification, and may change accordingly depending on the orientation of the components in the accompanying drawings.
[0020] In this application, unless otherwise expressly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral part; it can be a direct connection or an indirect connection through an intermediate medium. Furthermore, the term "coupled" can refer to an electrical connection that enables signal transmission.
[0021] As used herein, “about,” “approximately,” or “approximately” includes the stated value and the average value within an acceptable range of deviation from the given value, wherein the acceptable range of deviation is determined by a person skilled in the art taking into account the measurement under discussion and the error associated with the measurement of the given quantity (i.e., the limitations of the measurement system).
[0022] This embodiment provides an online monitoring method for the production process of high-purity lithium salts. The method bypasses the high-speed data bus of the distributed control system (DCS) by deploying a computing node with a multi-dimensional feature time encoder and a deep attention regression network to obtain a thermodynamic and hydrodynamic state matrix covering the entire residence cycle of the material through high-frequency sampling. Based on the state matrix, the computing node performs nonlinear mapping operations to generate a real-time continuous concentration prediction vector for a preset quality marker, thereby realizing real-time feedforward intervention.
[0023] Figure 1 This is a flowchart of an online monitoring method for a high-purity lithium salt production process according to an embodiment of the present invention. This method can be applied to electronic devices, such as industrial-grade rack servers, embedded industrial control computers, or edge computing gateways integrating tensor processing units (TPUs). Figure 1 As shown, the process includes the following steps: S100 collects multiple process variables in real time within a preset time window to form a time-series feature matrix. These process variables include at least thermodynamic and hydrodynamic parameters associated with the high-purity lithium salt production process.
[0024] The data acquisition module is equipped with a high-speed network interface card. This module responds to timer interrupts and communicates with the programmable logic controller (PLC) via an industrial Ethernet protocol (such as OPC-UA or Profinet). Its sampling frequency is 1Hz, acquiring readings from the sensor array on the process pipeline. Multiple process variables are used as high-dimensional vector space coordinates of the system state, and thermodynamic parameters are mapped to include the inlet temperatures of the heat transfer medium in the multi-stage reactor jacket. Temperature of the slurry inside the reactor The floating-point arrays of temperature control instrument feedback values for each temperature zone of the drying kiln and the refrigerant outlet temperature of the vacuum condenser are used; the fluid dynamic parameters are mapped to include the instantaneous flow rate of the gaseous reactant mass flow meter. Torque feedback value of axial flow agitator in reactor Centrifuge separation speed And a scalar sequence of absolute pressure distribution in the system's piping network.
[0025] Preset time window The length must be greater than the maximum residence time (MRT) of the material in the entire process flow; this value is hard-coded in the system; a preset time window is set based on the 450-minute maximum residence time parameter calibrated by impulse response experiments. Set the corresponding number of seconds for 480 minutes in the storage register.
[0026] The data acquisition module maintains a circular buffer in memory. When the data acquisition module receives a new sampling vector, the memory controller moves the address pointer to overwrite the oldest data and writes the new data to the current address. As the buffer is fully filled, the memory data is serialized to construct a temporal feature matrix. ; where, sequence length Based on a resampling strategy set to 480 (corresponding to a 1-minute resampling step size); feature dimension Set to 128, representing the total number of independent process variables; correspondingly, matrix elements... Stored in single-precision floating-point format, used to represent backtracking to the specified time scale. minutes ago, number The physical quantity measured by sensor number 1.
[0027] For startup scenarios lacking initial data labels, the system executes the following control branches: S110: For newly built production lines, acquire process data and corresponding offline test data during the trial production phase to train the initial surrogate model, thereby generating the initial baseline model.
[0028] In response to the production line initialization command, during the trial production phase of the first preset cycle, the system retrieves offline moisture titration results and their corresponding historical process time-series data from the Laboratory Information Management System (LIMS) via the application programming interface. Subsequently, the control processor performs singular value decomposition (SVD) on the high-dimensional collinear independent and dependent variable matrices using the Partial Least Squares (PLS) algorithm engine, projecting them into an orthogonal latent variable space containing three principal components. The principal axes are then extracted by maximizing the feature covariance. Based on this projection matrix, the control processor serializes and stores the weight coefficients of the linear regression equation in read-only memory, instantiating it as the initial baseline model. .
[0029] S120: When the accumulated production data reaches the preset data volume threshold, use the accumulated production data to train a new deep time series prediction model.
[0030] The offline training server polls the database to obtain matching records of high-frequency sequences and labels. This server has distributed database mount points. The system-level scheduler wakes up the computing kernels in the graphics processing unit (GPU) cluster in response to the record count counter overflowing a preset data volume threshold (1000 valid batches). The training process of the deep temporal prediction model uses a set of key hyperparameters to ensure convergence efficiency and generalization performance. Specifically, the training batch size is set to 64, meaning that each gradient update is calculated based on 64 temporal samples. The initial learning rate uses the Adam optimizer and is set to 0.001, with an exponential decay strategy. Every 20 training epochs, the learning rate is multiplied by a decay factor of 0.8. To prevent overfitting, all fully connected layers of the model are configured with L2 regularization terms, with regularization coefficients... It is set to 0.01, and a random deactivation mechanism with a ratio of 0.2 is applied during training.
[0031] The training server then slices the historical tensor dataset to form micro-batch tensor streams, loads them into the memory bus, and performs backpropagation iterative derivative calculations within the internal computation graph using a stochastic gradient descent optimizer and the root mean square error loss function. After gradient updates set to 200 epochs, an early stopping mechanism is triggered, at which point the server treats the computation graph weights in memory as containing the new deep time series prediction model. The checkpoint status is obtained and exported as a binary weight block file.
[0032] S130: Run the new deep time series prediction model and the initial baseline model for the first time within a preset transition period.
[0033] To avoid sudden output changes affecting the control system during model switching, the inference engine runs parallel computations of the old and new models in an isolated environment. The weight file is loaded into the sandbox memory; in response to the input stream of the real-time temporal feature matrix, the TPU performs matrix multiplication-addition forward inference to output the prediction vector. The output stream is transmitted to a background log device for physically isolated read and write operations; in parallel, the initial baseline model... Continuously map output in its own thread The dual-track state of this sandbox is controlled by a transition period timer. Transition period timer Set the countdown to 168 hours.
[0034] S140: During the first run, a time-based smooth weighted fusion mechanism is used to calculate the weighted sum of the output of the new deep time series prediction model and the output of the initial baseline model, and this weighted sum is used as the final prediction value.
[0035] During the transition period timer activation interrupt, the arithmetic logic center (ALU) of the output fusion unit obtains the elapsed time scalar from the system clock. And execute a division instruction to calculate the strictly monotonically increasing mixing coefficients: Then calculate the prediction vector in the core loading floating-point format. and The data is transferred to a temporary register, where scalar multiplication and addition vector operations are performed via a floating-point pipeline. This generates a smooth fused prediction vector. And directly redirect it to the main control output register.
[0036] S150: Execution context information embedding and feature matrix fusion.
[0037] The network coprocessor initiates a call to the manufacturing execution system via the HTTP protocol and receives the returned JSON data. The coprocessor then parses the data and extracts production status tags (such as the production grade string of the currently fed material). Subsequently, the tensor preprocessing unit responds to the string comparator hitting a specific context tag (e.g., the vocabulary corresponding to the complete set of 10 grades). The "brand name C" in the text is used to convert it into a one-dimensional format. sparse binary one-hot column vectors .
[0038] The embedding engine then extracts this one-hot column vector. And call the dense embedding weight matrix maintained in the on-chip cache. And the basic linear algebra subroutine library performs dense matrix multiplication: This projects the original sparse representation into a one-dimensional representation. Dense continuous feature column vectors The embedded engine is coupled to the main system bus. Then, the vector diffusion module moves along a preset time step reference dimension (i.e., ), for dense feature vectors Send a tensor tiling instruction to allocate a build dimension of in the memory pool. Extended context matrix Finally, the data alignment unit performs tensor concatenation instructions along the feature channel indices on the sensor feature stream and the context matrix: In order to make the original Feature matrix structured dimension expansion to A novel time-series feature matrix Load it into the L3 cache for inference graph traversal.
[0039] S200: Input the time series feature matrix into a pre-trained deep time series prediction model to obtain the predicted value of at least one target product quality indicator corresponding to the production process.
[0040] The high-frequency edge server loads the static computation graph of the Time Fusion Transformer (TFT) in the storage subsystem into local high-speed memory based on the incoming interrupt signal, and triggers the serialized data flow graph execution engine to process the timing feature matrix pushed into memory. The tensor forward propagation operation sequence is executed sequentially, wherein the high-frequency edge server is equipped with a neural network computing unit; specifically, the process includes the following sub-steps: S210: Perform network computation by selecting variables.
[0041] The first cluster of nodes in the directed acyclic graph of the TFT tensor is defined as the variable selection network module; at this point, for the input feature tensor in the time slice... Feature vector slices at the location The module passes its reference address to the gated residual network core, which then initiates the multi-branch execution engine: Main circuit pair Perform parallel affine transformations; The auxiliary loop calculates the gated linear mask function by executing the activation function. Mapping information filtering vector, the range of which is ,in The instruction for element-wise multiplication of tensors.
[0042] Subsequently, the floating-point adder core extracts the filtering result and compares it with the input vector temporarily stored in the register. Element-wise summation is performed to close the residual feedback loop, and layer normalization instructions are used to smooth the characteristic variance output. .
[0043] The main controller targets the column dimension of the feature matrix. Each independent sensing process variable is used to iterate through the gated residual network nodes using a loop unrolling technique. The output array is then pushed into a Softmax activation kernel, and an exponentialization and summation algorithm is executed to output a dynamic feature selection weight scalar set. The sum of this set is always equal to 1. Subsequently, a hardware-level floating-point multiplier synchronously extracts the feature embeddings and the weight scalar, and performs the multiplication operation. This process removes dimensions with weights below a specific threshold, generating a time-step cleaning sequence with a high signal-to-noise ratio.
[0044] S220: Time-dimensional encoding of the multi-head self-attention mechanism.
[0045] The denoised sequence feature blocks are fed into a multi-head self-attention topology layer. The Matrix Arithmetic Logic Unit (ALU) loads three independent weight tensors, with length... Sequences perform linear kernel mapping calculations, resulting in query tensors. Key tensors Sum tensor Three different memory page views in video memory. Control flow is based on the set number of attention heads (e.g., ...). The kernel function is then distributed to four independent stream processors. The stream processors execute the scaled dot product kernel function in parallel, based on memory-mapped addresses. This computational core will generate The fully connected time thermal similarity matrix, through (constant The division operation performs numerical truncation to prevent gradient vanishing, and is ultimately multiplied by the merged value tensor. After the kernel computation threads converge, feature reconstruction is performed through a fully connected neural network layer, generating a predicted value vector at the computation output of the output layer nodes. The predicted value vector consists of the target product quality index values, and its memory view is a two-element array in a fixed single-precision floating-point format, such as containing the numerical scalar 12.45 (moisture indicator in ppm) and the numerical scalar 0.32 (iron indicator in ppm).
[0046] S300: Compares the predicted value with the preset quality specifications. When the predicted value meets the warning conditions, it generates and outputs an alarm message.
[0047] The alarm generation module adopts a microservice architecture. This module listens to the message pipeline of the S200 inference process, decodes the predicted value in the data packet, and sends the predicted value to the decision engine. The read-only storage medium area contains key-value pair structured data blocks, which internally store a static set of quality specification thresholds. Specifically, it includes the following sub-steps: S310: Execute the numerical boundary determination protocol.
[0048] The digital comparator extracts a specific channel scalar (such as moisture content) from the predicted value vector. This value is compared to a preset threshold, including: Product control specifications upper limit ; Control chart calculation triggers early warning response boundary .
[0049] It should be noted that the early warning response boundary The updates are controlled by a background statistical algorithm coroutine, which is based on SQL aggregate functions and extracts sample clusters from high-quality batch test tables from the most recent 90 days, calculating their unbiased sample mean. Standard deviation of the sample .
[0050] The coroutine calls a statistics function library to calculate the Z-value constant based on a confidence level of 0.95. The instruction set sequence is then used to obtain the warning boundary through SIMD operations. The system bus then writes the result 15.5 into memory as the warning trigger level.
[0051] S320: Generate a status response.
[0052] Field-programmable gate arrays (FPGAs) transmit execution action packets via the carry flag of a comparator using multiplexed pin responses.
[0053] If the predicted value is determined to be ≤15.5, the HTTP POST method is triggered to refresh the normal operating condition data field to the monitoring terminal. At this time, in response to the detection signal that the predicted value is in the range of (15.5, 20.0], the logic module encodes and generates a message stream with an orange warning level constant, and pushes it to the field monitoring screen via UDP protocol, so that it displays a bright warning flashing signal to call the operation terminal to intervene. If the predicted value is greater than 20.0, in response to the corresponding high-level overflow interrupt, the hardware layer directly sets the multi-channel digital output (DO) relay coil channel connected to the programmable logic controller, drives the hard-wired control loop to close, generates a millisecond-level valve shut-off action to block the output of unqualified fluid, and simultaneously sends a high-power alarm siren pulse signal.
[0054] S400: Online monitoring steps for model health.
[0055] The statistics daemon, deployed on the management node of the high-performance computing cluster, is configured with an event timer that periodically breaks out of the suspended state and performs calculations to evaluate the fitting and approximation error of the current neural network topology to the new process data distribution.
[0056] S410: Calculate the multidimensional prediction residual vector.
[0057] The system listens to the TCP data stream of the laboratory information management system to receive actual test values. (Including truth elements) Then, the corresponding predicted values are extracted from the time series database. (Right now The arithmetic processor loads these two register families in parallel and executes vector subtraction instructions. The calculated multidimensional prediction residual vector, which encompasses the discrete error results of each physical channel, is then used to calculate the result. Push it into the pending queue.
[0058] S420: Nonparametric dimensionality reduction of residual vectors.
[0059] Analyze the persistent inverse covariance static tensor in microservice calls The analysis microservice is equipped with a linear algebra acceleration package; wherein, the covariance matrix and its inverse matrix Instead of being a static constant, it is updated periodically using a rolling time window strategy. Specifically, the system is configured with a background batch processing task that automatically extracts all valid prediction residual vector samples from the past 30 days every 24 hours, recalculates the sample covariance matrix, and performs matrix inversion. To ensure the numerical stability of the inversion process, the system uses an inversion method with Tikhonov regularization, i.e., calculating... Where I is the identity matrix and the regularization parameter is... For a positive number close to zero (for example, This avoids the singular matrix problem caused by sample collinearity. The updated inverse covariance matrix. The atomic replacement of the old matrix that resides in memory ensures the dynamic adaptability of the Mahalanobis distance calculation benchmark.
[0060] Then, through the General Matrix Multiplication (GEMM) core program interface, the controller sequentially triggers a series of command set operations for vector transpose, matrix multiplication, and subsequent multiplication: Due to the physical cancellation mechanism of matrix algebra, this computational pipeline completely folds multidimensional real vector inputs containing diverse physical units into a single dimensionless quadratic pure numerical output. The coprocessor then invokes the fast square root hardware acceleration instruction to calculate and output the Mahalanobis distance (i.e., the first distance) as a single scalar: This makes the error system isomorphic across dimensions.
[0061] S430: Perform smoothing control statistic generation.
[0062] The controller imports this single scalar stream and performs exponentially weighted moving average state estimation via a first-order infinite impulse response (IIR) low-pass filter model. The controller is equipped with an accumulator storage module. Subsequently, the internal arithmetic kernel iterates according to a preset discrete difference equation. The controller then extracts a smoothing factor, which is a fixed configuration within the register and characterizes the response damping property. (e.g., 0.2) and the temporary state of the previous operation cycle. The system executes a series of multiplication and register addition instructions to calculate the current periodic smoothing control statistics after the filter converges. .
[0063] S440: Determine the mismatch condition and trigger the model update.
[0064] Low-level monitoring trigger loop extraction The value and the mismatch control limit fixed in the register area A constant (e.g., quantized as a scalar 4.5) performs a hard comparison operation; wherein the mismatch control limit... It is determined by statistical principles. Specifically, for a p-dimensional multivariate normal distribution, the square of its Mahalanobis distance is... It follows a chi-square distribution with p degrees of freedom Therefore, control limits Set to correspond to the preset significance level (For example The upper quantile of the chi-square distribution of ). For example, for a two-dimensional residual vector of p=2 (two quality indicators, moisture and iron ions), in At the given confidence level, the chi-square distribution table or numerical calculation library can be consulted to obtain the result. Therefore, mismatch control limits Set as .
[0065] when Value exceeds When the system receives a signal greater than the interrupt signal generated by the breakdown, it sets the model mismatch flag. The flipped state of this flag directly serves as the activation source, sending a RESTful task trigger instruction with a unique model identifier (UUID) to the backend container orchestration engine (such as the Kubernetes API). After that, the external engine takes over the resources and performs incremental data training.
[0066] S500: System reset and robust recovery under extreme transient conditions.
[0067] The main control circuit is equipped with an extreme transient protection architecture, which handles sensor data interruption and recovery caused by external power loss and prevents signal overload.
[0068] Specifically, the microcontroller polls the data packet markers; if the proportion of NaN exceeds 30% within 5 minutes, the system circuit breaker is triggered. At this time, the interrupt controller blocks the S200 thread and sends a "network blind spot circuit breaker" alarm through the gateway.
[0069] If the packet loss rate is lower than the threshold, the circuit breaker will not be triggered. At this time, the system will transfer the data stream to the auxiliary coprocessor to perform Lagrange interpolation calculation. The coprocessor executes a loop search algorithm to address the five nearest data page storage nodes with valid numerical stamps, solves the second-order discrete fitting function matrix, and after the operation, overwrites the obtained algebraic result in place to the unused memory location to bridge the feature dimension gap.
[0070] Furthermore, to prevent abnormal data from affecting monitoring during the initial warm start, a clamping layer is set between S420 and S430, which limits the maximum value of the residual vector. Subsequently, the microcontroller output pin is forced to output the minimum scalar value between the two variables: .
[0071] This operator forcibly intercepts any divergent scalar signal greater than 6.0 at the memory read / write level (e.g., an abnormal peak value surging to 45.0 due to physical environment oscillations), saturating and truncating it. After assembly-level interception, the values substituted into the IIR filter execution cycle are forcibly constrained, ensuring the weighted operations performed at the operator nodes (such as...) The output result will never impact the mismatch control limit of 4.5. At the pure computing power control logic level, the path of the sudden impact pulse to the trigger retirement interlock control terminal is cut off, thus maintaining the long lifespan of the effective computing weight.
[0072] For example, the effect of this mechanism can be demonstrated by performing a micro-simulation: when the system encounters extreme residuals during power outage recovery, the original residuals are calculated. The data stream enters the nonlinear clamping layer. After calculation and operator determination, 45.0 is found to be greater than the clamping threshold of 6.0. Therefore, the values of subsequent sequences are forcibly truncated. Subsequently, in the calculation of the exponentially weighted moving average, it is assumed that the historical state quantity at the previous time step is... Smoothing factor The currently suppressed smoothing statistic is then updated as follows: Since 2.16 is still safely within the mismatch control limit. (For example, 3.03) The clamping mechanism successfully absorbed the transient ultra-large residual pulse impact caused by this known cause.
[0073] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.
[0074] This embodiment also provides an online monitoring system for the production process of high-purity lithium salts. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.
[0075] Figure 2 This is a structural block diagram of an online monitoring system for high-purity lithium salt production processes according to an embodiment of the present invention, as shown below. Figure 2 As shown, it is configured as an industrial-grade edge computing node device deployed within the factory information control network.
[0076] The core printed circuit board of the system is coupled with multiple high-bandwidth communication buses. Specifically, the data acquisition module 210 physically corresponds to a microprocessor cluster configured with a dedicated network interface controller (NIC). This controller receives multiplexed electrical signals from the underlying sensors through the PCI-Express (PCIe) bus and allocates consecutive address pages in the system dynamic random access memory (DRAM) to construct a circular buffer data structure in order to complete the assembly of the timing feature matrix.
[0077] The quality prediction module 220 physically includes at least one embedded tensor processing unit (TPU) or general-purpose graphics processing unit (GPU) that reads the feature matrix tensor from DRAM with zero copy via a direct memory access (DMA) channel across the northbridge chip. The arithmetic logic array inside the processing unit is configured to execute matrix multiplication and activation function instruction cycles of each layer of the TFT model in parallel, and finally writes the prediction result vector in floating-point format back to the register file.
[0078] The alarm generation module 230 physically involves a field-programmable gate array (FPGA) or application-specific integrated circuit (ASIC) connected to the system bus. The chip internally programs a parallel comparator digital circuit to perform a real-time XOR comparison between the predicted value in the register file and the threshold in the on-chip static random access memory (SRAM). In response to an over-limit signal, it directly pulls down its peripheral general-purpose input / output (GPIO) pin, issuing a hard command to the external solid-state relay to cut off the circuit.
[0079] It should be noted that the above modules can be implemented by software or hardware. For the latter, they can be implemented in the following ways, but are not limited to: all the above modules are located in the same processor; or, the above modules are located in different processors in any combination.
[0080] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to perform the steps in any of the above method embodiments when executed.
[0081] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.
[0082] Embodiments of the present invention also provide an electronic device including a memory and a processor, the memory storing a computer program and the processor being configured to run the computer program to perform the steps in any of the above method embodiments.
[0083] In one exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.
[0084] Through the above description of the embodiments, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.
[0085] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another apparatus, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0086] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units; that is, it can be located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0087] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0088] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0089] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. An online monitoring system for the production process of high-purity lithium salts, characterized in that, include: The data acquisition module is used to collect process variables in real time and construct a time-series feature matrix based on the process variables; The quality prediction module is used to input the time series feature matrix into a pre-trained deep time series prediction model to obtain the predicted value of the target product quality index. The alarm generation module is used to compare the predicted value with the preset quality specifications, and generate and output alarm information when the predicted value meets the warning conditions.
2. The system according to claim 1, characterized in that, The variables acquired during the real-time acquisition process include: A health assessment is performed on the real-time data of the process variables to generate health labels; When the health status label is determined to be abnormal, a sensor maintenance alarm is triggered.
3. The system according to claim 2, characterized in that, The health assessment of the real-time data of the process variables includes: A flatness check is performed on the real-time data, wherein when the change in the value of the real-time data within a preset check time is less than a preset flatness threshold, the health label is determined to be abnormal. And / or, A change rate check is performed on the real-time data, wherein when the instantaneous change rate of the real-time data is greater than a preset change rate threshold, the health label is determined to be abnormal.
4. The system according to claim 1, characterized in that, After generating and outputting the alarm information, the following steps are also performed: Obtain the actual test values of the target product quality indicators; Based on the predicted value and the actual test value, calculate the multidimensional prediction residual vector; The multidimensional prediction residual vector is normalized based on the inverse of the historical residual covariance matrix. An exponentially weighted moving average is calculated on the first distance sequence within a preset time period to obtain a smoothing control statistic. The first distance is obtained through a normalization process. When the smoothing control statistic is greater than the preset mismatch control limit, the model mismatch condition is determined to be met, and a model update instruction is triggered.
5. A method for online monitoring of high-purity lithium salt production process, characterized in that, include: Real-time acquisition of process variables; Construct a time-series feature matrix based on the process variables; The time-series feature matrix is input into a pre-trained deep time-series prediction model to obtain the predicted values of the target product quality indicators. The predicted value is compared with the preset quality specification, and when the predicted value meets the warning conditions, an alarm message is generated and output.
6. The method according to claim 5, characterized in that, The variables acquired during the real-time acquisition process include: A health assessment is performed on the real-time data of the process variables to generate health labels; When the health status label is determined to be abnormal, a sensor maintenance alarm is triggered.
7. The method according to claim 6, characterized in that, The health assessment of the real-time data of the process variables includes: A flatness check is performed on the real-time data, wherein when the change in the value of the real-time data within a preset check time is less than a preset flatness threshold, the health label is determined to be abnormal. And / or, A change rate check is performed on the real-time data, wherein when the instantaneous change rate of the real-time data is greater than a preset change rate threshold, the health label is determined to be abnormal.
8. The method according to claim 5, characterized in that, After generating and outputting the alarm information, the method further includes: Obtain the actual test values of the target product quality indicators; Based on the predicted value and the actual test value, calculate the multidimensional prediction residual vector; The multidimensional prediction residual vector is normalized based on the inverse of the historical residual covariance matrix. An exponentially weighted moving average is calculated on the first distance sequence within a preset time period to obtain a smoothing control statistic. The first distance is obtained through a normalization process. When the smoothing control statistic is greater than the preset mismatch control limit, the model mismatch condition is determined to be met, and a model update instruction is triggered.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program is configured to perform the method described in any one of claims 5 to 8 when executed.
10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method as described in any one of claims 5 to 8.