Intelligent liquid chromatograph fault monitoring system

Abstract

The invention discloses an intelligent liquid chromatograph fault monitoring system, and relates to the technical field of liquid chromatograph fault monitoring, and the system comprises a data collection module which is used for collecting parameter data of a liquid chromatograph in real time through a sensor; the operation verification module is used for establishing an operation-equipment association model by collecting historical operation data and marking corresponding fault or bad experiment results to obtain an operation mode influence coefficient; the grading fault monitoring model building module takes the operation mode influence coefficient and the real-time data of the sensor as model input, builds a grading fault monitoring model and outputs fault reminding; and the fault early warning module is used for generating a real-time fault diagnosis report according to the monitoring result and marking the fault type. According to the method, the problems that the traditional liquid chromatograph fault detection and monitoring depend on experience judgment and are easily interfered by human factors, misoperation is caused by equipment faults by mistake, or potential faults are ignored are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of liquid chromatograph fault monitoring, and specifically to an intelligent liquid chromatograph fault monitoring system. Background Art

[0002] A liquid chromatograph involves multiple subsystems, such as sample injection, pump control system, temperature control module, detector, and data processing module, etc. The coupling effect between each module makes the equipment state extremely vulnerable to factors such as environmental temperature, unstable flow rate, and improper operation, resulting in a decline in detection accuracy and frequent failures;

[0003] Currently, the fault monitoring of liquid chromatographs mostly relies on the experience judgment of operators and regular maintenance; traditional methods mainly focus on the static threshold monitoring of a single parameter. There are multiple sensors and components inside the liquid chromatograph, and the single-threshold method is difficult to fully capture the complex dynamic correlations between various parts of the equipment and give early warnings in the initial stage of faults. Once a equipment fault occurs, it may cause deviations in analysis data and even affect the experimental results; moreover, the operator's reliance on experience judgment is easily interfered by environmental noise and human factors. It is difficult to solve the problem that the misjudgment of faults caused by operation errors during the training of the fault prediction model will affect the training input data of the model, resulting in errors, and it is difficult to achieve real-time and accurate fault warnings, affecting the accuracy of the model output; due to the fact that the small anomalies that occur during the continuous operation of the equipment are often difficult to detect in time, delaying the repair of faults will not only affect the reliability of the data, but may also cause instrument downtime and reduced production efficiency, and further lead to misjudgments.

[0004] Therefore, the present invention provides an intelligent liquid chromatograph fault monitoring system. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent liquid chromatograph fault monitoring system to solve the existing problems mentioned in the above background art.

[0006] To achieve the above purpose, the present invention provides the following technical solution: An intelligent liquid chromatograph fault monitoring system, comprising:

[0007] A data acquisition module, which is used to collect parameter data of the liquid chromatograph in real time through sensors;

[0008] An operation verification module, which is used to collect historical operation data, mark the corresponding faults or poor experimental results, establish an operation-equipment association model, and obtain the operation mode influence coefficient;

[0009] A hierarchical fault monitoring model establishment module, which takes the operation mode influence coefficient and the real-time data of the sensor as model inputs, establishes a hierarchical fault monitoring model, and outputs a fault reminder;

[0010] The fault warning module generates a real-time fault diagnosis report based on the monitoring results, indicating the fault type.

[0011] A further improvement of the present invention lies in that the operation verification module includes collecting the historical sample injection volume and injection speed of the operator as operation characteristics, collecting the data of the equipment fault type corresponding to the operator's operation when the equipment fails, marking the historical operation data and the equipment fault data type to obtain a fault-operation label data set; recording the time when the operation occurs, and collecting the corresponding equipment status, combining the operation characteristics and the equipment status to form a status-operation characteristic, which is used as the training data of the operation-equipment association model. Based on the neural network model, taking the current operation behavior as the input, calculating the SHAP of the current operation behavior, and outputting the influence gain of the current operation behavior on the occurrence probability of each fault, which is recorded as the influence coefficient of the operation mode of this fault type.

[0012] A further improvement of the present invention lies in that the parameter data of the liquid chromatograph includes establishing a fault parameter data set, standardizing and weighted calculating the changes in the equipment status corresponding to each fault type to obtain a fault result influence factor, marking the fault result influence factor, and sending it together with the corresponding fault type into the fault parameter data set.

[0013] A further improvement of the present invention lies in that the hierarchical fault monitoring model establishment module includes a fast threshold screening unit, a time series pattern analysis unit, and a deep fusion diagnosis unit;

[0014] The fast threshold screening unit is used to automatically adjust the static threshold according to the historical equipment data, output the abnormal marks in the fault parameter data set, and output the preliminary confidence level in combination with the influence coefficient of the operation mode of this fault type; when the preliminary confidence level is greater than 0.8, directly alarm, and when the preliminary confidence level ∈ [0.4, 0.8] or the fault combination has low fluctuations, trigger the time series pattern analysis unit;

[0015] The time series pattern analysis unit is used to set a lightweight LSTM predictor, extract the abnormal marks in the fast threshold screening unit, take the equipment status data as the input, predict the change trend of the equipment data within the next 5s, output the potential fault type and distribution probability, and when the maximum value of the distribution probability is greater than or equal to 75%, output the corresponding fault type and alarm, and when the maximum value of the distribution probability is less than 75%, trigger the deep fusion diagnosis unit;

[0016] The deep fusion diagnosis unit is used to inherit the data of the fast threshold screening unit and the time series pattern analysis unit, introduce a deep learning model, and output the accurate fault type.

[0017] A further improvement of the present invention lies in that the fast threshold screening unit is equipped with a confidence calculation strategy. The confidence calculation strategy calculates the mean value μFa and the standard deviation σFa of the corresponding fault result impact factors of any fault type in the fault parameter dataset with a time interval of t for the historical N times, and obtains the dynamic threshold TFa of any fault type as TFa = μFa + k·σFa, where k represents the confidence interval coefficient. When the result impact factor RFa corresponding to any current fault type is greater than the current dynamic threshold corresponding to this fault type, it is determined as an abnormal mark, and this fault type is output, and the preliminary confidence is calculated. It represents the influence coefficient of the operation mode of this fault type after normalization.

[0018] A further improvement of the present invention lies in that the fast threshold screening unit is also equipped with a low-fluctuation data judgment strategy. The low-fluctuation data judgment strategy obtains the low-fluctuation judgment threshold TFa + Tnn by setting the fluctuation threshold Tnn. When there are more than or equal to 3 groups of data in the abnormal marks that are less than their corresponding low-fluctuation judgment thresholds, it is judged as a fault combined with low fluctuation, and the timing pattern analysis unit is triggered.

[0019] A further improvement of the present invention lies in that the timing pattern analysis unit includes extracting the device state x t from the abnormal marks from the fast threshold screening unit as the main input, and using the operation influence coefficient as an additional feature to construct a combined input. At this time represents the operation influence coefficients of all fault types at time t. The LSTM model automatically learns the influence of the operation influence coefficients on the change of the device state during the time step calculation. During the model operation, the current error is added to the input of the current time step, and a raw score output layer is added after the hidden state h T of the last layer to output the raw scores z i = W z ·h T + b z where W z represents the weight matrix, represents the current predicted value, and b z represents the bias term. Subsequently, the raw scores are converted into a probability distribution to obtain the probabilities of each fault type.

[0020]

[0021] A further improvement of the present invention lies in that the deep fusion diagnosis unit is equipped with a dynamic fault prediction model. The dynamic fault prediction model is based on a convolutional neural network and extracts the input data from the timing pattern analysis unit as the input, including an input layer, a processing layer, an operation influence prediction layer, an influence threshold check layer, a dynamic regulation layer, and an output layer.

[0022] The input layer includes extracting the operation mode influence coefficients corresponding to all fault types and real-time device status data;

[0023] The processing layer processes the input data for each data channel using a fully connected layer and then fuses and splices the input data into a vector X through a splicing layer de = Concatenate(vs', vo'), where Concatenate() represents the splicing function, vs' represents the device status vector, and vo' represents the operation mode influence coefficient vector;

[0024] The operation influence prediction layer downsamples the convolutional features using convolutional and pooling layers and predicts the fault influence coefficient FIC and the operation influence coefficient OIC through a fully connected layer;

[0025] And the influence threshold check layer checks whether the predicted operation mode influence coefficient is greater than the set maximum operation influence value, and judges whether to truncate or send it to the dynamic regulation layer;

[0026] The dynamic regulation layer uses the operator behavior vector vo' as a compensation factor, dynamically adjusts the processing path of the model, and then outputs it by the output layer, and outputs the fault type probability in the same way as the timing pattern analysis unit.

[0027] A further improvement of the present invention lies in that the influence threshold check layer includes setting the maximum operation influence value O max , if the predicted operation mode influence coefficient is greater than the set maximum operation influence value, then perform truncation processing on OIC adj = min(OIC, O max ), if the predicted operation mode influence coefficient is less than or equal to the set maximum operation influence value, then input it to the dynamic regulation layer.

[0028] A further improvement of the present invention lies in that the dynamic regulation layer adopts a gating mechanism combined with the compensation factor vo', calculates the dynamic routing weight G = sig(vo'), and the dynamically regulated feature representation is F route = G Θ FIC + (1 - G) Θ FIC', where Θ represents element-wise multiplication, and FIC' represents the output without considering operator behavior.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] 1. First, through the operation - device association model established by the operation verification module, the training data of the multi - level hierarchical fault monitoring model is adjusted by the operation mode influence coefficient, and the confidence level is regulated in real - time according to the operation behavior. Subsequently, the data error generated by the operation mode influence coefficient is used as training data, enabling the model to "know" the potential interference of current improper operations on the results when predicting future states. In the deep fusion diagnosis unit, the operator behavior vector is used as a compensation factor to dynamically adjust the processing path of the model, adjust the processing strategy, reduce the false alarm risk caused by operations, effectively weaken the interference of human operations on sensor data, reduce the false alarm risk caused by operation errors, and improve the accuracy of fault diagnosis.

[0031] 2. Through the multi - level hierarchical fault monitoring model, the real - time monitoring of the device state is achieved, and the static threshold and dynamic threshold can be automatically adjusted to ensure the accurate identification of potential faults in the shortest time, save computing resources, and send out alarms in a timely manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a framework diagram of an intelligent liquid chromatography instrument fault monitoring system of the present invention;

[0033] Figure 2 It is a flow chart of a hierarchical fault monitoring model of an intelligent liquid chromatography instrument fault monitoring system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] The technical solution of the present invention will be described in detail below with reference to the drawings and specific embodiments. It should be understood that the specific features in the embodiments of the present invention and the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. Without conflict, the technical features in the embodiments of the present invention and the embodiments can be combined with each other.

[0035] The term "and / or" merely describes the associated relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " generally represents an "or" relationship between the associated objects before and after.

[0036] Embodiment 1

[0037] Figure 1 A framework diagram of an intelligent liquid chromatography instrument fault monitoring system disclosed in this embodiment is shown, including:

[0038] A data acquisition module, which is used to collect the parameter data of the liquid chromatography instrument in real time through sensors; since each fault type of the liquid chromatography instrument is not determined by a single sensor parameter, the parameter data of the liquid chromatography instrument includes establishing a fault parameter data set, standardizing and weighted calculating the changes in the device state corresponding to each fault type to obtain a fault result influence factor, marking the fault result influence factor, and sending it together with the corresponding fault type into the fault parameter data set; that is, which parameters can affect this fault type.

[0039] An operation verification module, which is used to collect historical operation data and mark the corresponding faults or bad experimental results;

[0040] Since the operation mode will affect the health state of the liquid chromatography instrument, but this influence is not absolute, and the magnitude of the error caused by operation errors will not play a key role, but it will affect the system's judgment of faults. To make up for this influence, an operation-device association model is established to obtain an operation mode influence coefficient; it includes collecting the historical sample injection volume and injection speed of the operator as operation characteristics, and collecting the device fault type data corresponding to each operation. The fault types include pump seal leakage, proportional valve tremor, abnormal friction of the plunger rod, sensor drift, liquid leakage, temperature abnormality, detector signal abnormality, automatic sampler position abnormality, etc. Mark the historical operation data and the device fault data types to obtain a fault-operation label data set; record the time when the operation occurs, and collect the corresponding device state. The device state includes pressure, temperature, flow rate, position and signal. Combine the operation characteristics and the device state to form a state-operation characteristic, which is used as the training data of the operation-device association model. Based on the neural network model, use the current operation behavior as the input, calculate the SHAP of the current operation behavior, and output the influence gain of the current operation behavior on the occurrence probability of each fault, which is recorded as the operation mode influence coefficient of this fault type;

[0041] By collecting the historical sample injection volume and injection speed operation characteristics of the operator and marking them with the corresponding device fault types, a fault-operation label data set is constructed, realizing a quantitative analysis of the risk contribution of operation behaviors.

[0042] A hierarchical fault monitoring model establishment module, which takes the operation mode influence coefficient and the real-time data of the sensor as the model input, establishes a hierarchical fault monitoring model, and outputs a fault reminder;

[0043] A fault warning module, which generates a real-time fault diagnosis report according to the monitoring results and indicates the fault type.

[0044] Embodiment 2

[0045] Figure 2The figure shows the flow chart of the hierarchical fault monitoring model of an intelligent liquid chromatography fault monitoring system according to the present invention. Based on the inventive concept of Embodiment 1, the present invention provides a hierarchical fault monitoring model, including:

[0046] A fast threshold screening unit, a timing pattern analysis unit, and a deep fusion diagnosis unit;

[0047] Although the error caused by a single operation error may be small, its cumulative effect will interfere with fault diagnosis. Therefore, after introducing the operation influence coefficient into the fault prediction model, the system can compensate for these interferences during the prediction process, thereby avoiding misjudgment caused by operation deviation; the specific implementation methods include:

[0048] The fast threshold screening unit is used to automatically adjust the static threshold according to historical device data, output the abnormal marks in the fault parameter dataset, and output the preliminary confidence level in combination with the operation mode influence coefficient of this fault type;

[0049] The fast threshold screening unit is equipped with a confidence calculation strategy. The confidence calculation strategy calculates the mean μFa and standard deviation σFa of the corresponding fault result influence factors of any fault type in the fault parameter dataset with a historical N - time interval of t, and obtains the dynamic threshold TFa of any fault type = μFa + k·σFa, where k represents the confidence interval coefficient. When the current result influence factor RFa of any fault type is greater than the current dynamic threshold corresponding to this fault type, it is determined as an abnormal mark, and this fault type is output, and the preliminary confidence level is calculated α~ represents the normalized operation mode influence coefficient of this fault type. When the preliminary confidence level is greater than 0.8, an alarm is directly issued;

[0050] The fast threshold screening unit is also equipped with a low - fluctuation data judgment strategy. The low - fluctuation data judgment strategy obtains the low - fluctuation judgment threshold TFa + Tnn by setting the fluctuation threshold Tnn. When there are 3 or more groups of data in the abnormal marks that are less than their corresponding low - fluctuation judgment thresholds, it is judged as a fault combined with low - fluctuation, and the timing pattern analysis unit is triggered;

[0051] When the preliminary confidence level ∈[0.4, 0.8] or there is a fault combined with low - fluctuation, the timing pattern analysis unit is triggered;

[0052] By combining sensor data with the operation mode, the confidence level can be adjusted in real - time according to the operation behavior. When is large (the operation behavior has a large influence): the confidence level will decrease, avoiding misjudging the abnormality caused by operation errors as a device fault; when When it is small (the operation impact is weak): The confidence level is mainly determined by the device sensor data, that is, the impact of sensor anomalies on the confidence level is greater, reducing the risk of false alarms; achieving a more accurate calculation of the fault confidence level, reducing false alarms, and improving the diagnostic accuracy.

[0053] The timing pattern analysis unit is used to set up a lightweight LSTM predictor, extract the anomaly markers in the fast threshold screening unit, take the device status data as input, predict the change trend of the device data within the next 5s, and output the potential fault types and distribution probabilities, including extracting the device status x from the anomaly markers of the fast threshold screening unit t as the main input, and taking the operation impact coefficient as an additional feature to construct a combined input At this time represents the operation impact coefficient of all fault types at time t. The LSTM model automatically learns the impact of the operation impact coefficient on the change of the device status during the time step calculation. During the model operation, the current error is calculated added to the input of the current time step, and an original score output layer is added after the hidden state h of the last layer to output the original scores z of each fault type T z = W i =W z ·h T +b z W z represents the weight matrix, represents the current predicted value, and b z represents the bias term. Subsequently, the original scores are converted into a probability distribution to obtain the probabilities of each fault type

[0054]

[0055] Combining the operation impact coefficient with the device status data enables the model to "know" the potential interference of current improper operations on the result when predicting the future state. During the training process, the model will automatically learn how to adjust the impact on the state prediction if the operation impact coefficient is high (such as improper operation), so as to make corresponding adjustments to the prediction output.

[0056] When the maximum value of the distribution probability is greater than or equal to 75%, the corresponding fault type is output and an alarm is issued. When the maximum value of the distribution probability is less than 75%, it means that all fault possibilities are not high enough and more detailed analysis is needed to confirm the fault type. Therefore, the deep fusion diagnosis unit is triggered;

[0057] The deep fusion diagnosis unit is used to inherit the data of the fast threshold screening unit and the time series pattern analysis unit, introduce a deep learning model, and carry a dynamic fault prediction model. The dynamic fault prediction model is based on a convolutional neural network, and extracts the input data in the time series pattern analysis unit as input, including an input layer, a processing layer, an operation impact prediction layer, an impact threshold check layer, a dynamic regulation layer, and an output layer;

[0058] The input layer includes extracting the operation mode impact coefficients corresponding to all fault types and the real-time device status data;

[0059] The processing layer uses a fully connected layer to process the input data for each data channel, and then fuses and splices the input data into a vector X through a splicing layer de = Concatenate(vs', vo'), where Concatenate() represents the splicing function, vs' represents the device status vector, and vo' represents the operation mode impact coefficient vector;

[0060] The operation impact prediction layer downsamples the convolutional features by using a convolutional layer and a pooling layer, and predicts the fault impact coefficient FIC and the operation impact coefficient OIC through a fully connected layer;

[0061] And the impact threshold check layer checks whether the predicted operation mode impact coefficient is greater than the set maximum operation impact value, and judges whether to truncate or send it to the dynamic regulation layer; the impact threshold check layer includes setting the maximum operation impact value O max , if the predicted operation mode impact coefficient is greater than the set maximum operation impact value, then perform truncation processing OIC adj = min(OIC, O max ), to prevent excessive outliers from interfering with subsequent judgments. If the predicted operation mode impact coefficient is less than or equal to the set maximum operation impact value, it is input to the dynamic regulation layer.

[0062] The dynamic regulation layer uses the operator behavior vector vo' as a compensation factor, dynamically adjusts the processing path of the model, and then outputs through the output layer, and outputs the fault type probability in the same way as the time series pattern analysis unit. The dynamic regulation layer adopts a gating mechanism combined with the compensation factor vo', calculates the dynamic routing weight G = sig(vo'), then the dynamically regulated feature representation is F route = G Θ FIC + (1 - G) Θ FIC', where Θ represents element-wise multiplication, and FIC’ represents the output without considering the operation behavior.

[0063] If G is relatively large, it indicates that the operation behavior intervention is obvious, and the model automatically reduces the false alarm risk caused thereby. In the dynamic regulation layer, the operator behavior vector is used as a compensation factor to dynamically adjust the processing path of the model. When a significant impact of the operation behavior on the device state is detected, the model can adjust the processing strategy to reduce the false alarm risk caused by the operation. This dynamic regulation mechanism effectively distinguishes the short-term fluctuations caused by human operations from the true fault signals of the device, reduces the false alarm rate, and improves the accuracy of fault detection.

[0064] The setting of the threshold and weight can be based on the default settings of the present invention or can be set by the operator himself.

[0065] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0066] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the processes and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0067] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured product including an instruction device, and the instruction device implements the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0068] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide means for implementing the specified functions in Figure 1steps of one or more processes and / or boxes Figure 1 steps of functions specified in one or more boxes.

[0069] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative rather than restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit of the present invention and the scope protected by the claims. All of these are within the protection scope of the present invention.

Claims

1. An intelligent liquid chromatograph fault monitoring system, characterized in that: include: A data acquisition module, used for collecting parameter data of the liquid chromatograph in real time through sensors; The operation verification module is used to collect historical operation data and mark the corresponding faults or bad experimental results, establish an operation-equipment association model, and obtain the operation mode influence coefficient; The hierarchical fault monitoring model building module takes the operation mode influence coefficient and the real-time data of the sensor as the model input, builds the hierarchical fault monitoring model, and outputs fault reminders; The fault warning module generates a real-time fault diagnosis report based on the monitoring results, indicating the fault type.

2. The intelligent liquid chromatograph fault monitoring system according to claim 1, characterized in that: The operation verification module includes collecting the operator's historical sample injection volume and injection speed as operation characteristics, and collecting equipment failure type data corresponding to the operator's operation when the equipment fails, marking the historical operation data and the equipment failure data type to obtain a failure-operation label data set; recording the time when the operation occurred, and collecting the corresponding equipment status, combining the operation characteristics with the equipment status to form a state-operation characteristic as the operation-equipment association model training data, based on the neural network model, taking the current operation behavior as input, calculating the SHAP of the current operation behavior, and outputting the impact gain of the current operation behavior on the probability of each failure, which is recorded as the impact coefficient of the operation mode of the failure type.

3. The intelligent liquid chromatograph fault monitoring system according to claim 1, characterized in that: The parameter data of the liquid chromatograph includes establishing a fault parameter data set, standardizing and weighting the changes in the equipment status corresponding to each fault type, obtaining the fault result influencing factor, marking the fault result influencing factor, and sending it to the fault parameter data set together with the corresponding fault type.

4. The intelligent liquid chromatograph fault monitoring system according to claim 1, characterized in that: The hierarchical fault monitoring model building module includes a fast threshold screening unit, a timing pattern analysis unit and a deep fusion diagnosis unit; The fast threshold screening unit is used to automatically adjust the static threshold according to the historical equipment data, output the abnormal mark in the fault parameter data set, and output the preliminary confidence in combination with the influence coefficient of the fault type operation mode; when the preliminary confidence is greater than 0.8, an alarm is directly issued, and when the preliminary confidence ∈ [0.4, 0.8] or the fault is combined with low fluctuation, the timing pattern analysis unit is triggered; The time series pattern analysis unit is used to set a lightweight LSTM predictor, extract abnormal marks in the fast threshold screening unit, take the equipment status data as input, predict the equipment data change trend within the next 5 seconds, output the potential fault type and distribution probability, and when the maximum value of the distribution probability is greater than or equal to 75%, output the corresponding fault type and alarm, and when the maximum value of the distribution probability is less than 75%, trigger the deep fusion diagnosis unit; The deep fusion diagnosis unit is used to inherit the data of the fast threshold screening unit and the timing pattern analysis unit, introduce a deep learning model, and output an accurate fault type.

5. The intelligent liquid chromatograph fault monitoring system according to claim 4, characterized in that: The fast threshold screening unit is equipped with a confidence calculation strategy. The confidence calculation strategy calculates the mean μFa and standard deviation σFa of the corresponding fault result influencing factors of any fault type in the fault parameter data set with a historical time interval of N times t, and obtains the dynamic threshold TFa=μFa+k·σFa of any fault type, where k represents the confidence interval coefficient. When the result influencing factor RFa corresponding to any current fault type is greater than the dynamic threshold corresponding to the current fault type, it is judged as an abnormal mark, and the fault type is output, and the preliminary confidence is calculated. Represents the normalized impact coefficient of the operation mode of this fault type.

6. The intelligent liquid chromatograph fault monitoring system according to claim 5, characterized in that: The rapid threshold screening unit is also equipped with a low fluctuation data judgment strategy. The low fluctuation data judgment strategy obtains the low fluctuation judgment threshold TFa+Tnn by setting the fluctuation threshold Tnn. When there are greater than or equal to 3 groups of data in the abnormal mark that are less than the corresponding low fluctuation judgment threshold, it is judged as a fault combined with low fluctuation, and the timing pattern analysis unit is triggered.

7. The intelligent liquid chromatograph fault monitoring system according to claim 4, characterized in that: The timing pattern analysis unit includes extracting the device status x from the abnormal mark of the fast threshold screening unit t As the main input, and the operation impact coefficient as an additional feature, construct a joint input at this time Represents the operational impact coefficient of all fault types at time t. The LSTM model automatically learns the impact of the operational impact coefficient on the change of the device state during the time step calculation. During the model operation, the current error is calculated. Added to the input of the current time step and added to the hidden state h of the last layer T Then add a raw score output layer to output the raw score z of each fault type i =W z ·h T +b z , W z represents the weight matrix, represents the current predicted value, b z represents the bias term, and then the original score is converted into a probability distribution to obtain the probability of each fault type 8. The intelligent liquid chromatograph fault monitoring system according to claim 4, characterized in that: The deep fusion diagnosis unit is equipped with a dynamic fault prediction model, which is based on a convolutional neural network and extracts input data from the timing pattern analysis unit as input, including an input layer, a processing layer, an operation impact prediction layer, an impact threshold check layer, a dynamic control layer, and an output layer; The input layer includes extracting the operation mode influence coefficients corresponding to all fault types and the real-time data of the equipment status; The processing layer uses a fully connected layer to process the input data for each data channel and then fuses and splices the input data into a vector X through a splicing layer. de =Concatenate(vs',vo'), Concatenate() represents the concatenation function, vs' represents the device state vector, and vo' represents the operation mode influence coefficient vector; The operation impact prediction layer downsamples the convolution features by using the convolution layer and the pooling layer, and predicts the fault impact coefficient FIC and the operation impact coefficient OIC through the fully connected layer; The impact threshold checking layer checks whether the predicted operation mode impact coefficient is greater than the set operation impact maximum value, and determines whether to truncate or send to the dynamic control layer; The dynamic control layer uses the operator behavior vector vo' as a compensation factor, dynamically adjusts the processing path of the model, and then uses the output layer as the output, outputting the fault type probability in the same way as the timing pattern analysis unit.

9. The intelligent liquid chromatograph fault monitoring system according to claim 8, characterized in that: The impact threshold check layer includes setting the maximum value of the operation impact. max If the predicted operation mode impact coefficient is greater than the set maximum operation impact, truncation processing OIC is performed adj =min(OIC,O max ), if the predicted operation mode influence coefficient is less than or equal to the set operation influence maximum value, it is input to the dynamic control layer.

10. The intelligent liquid chromatograph fault monitoring system according to claim 9, characterized in that: The dynamic control layer uses a gating mechanism combined with a compensation factor vo' to calculate the dynamic routing weight G = sig (vo'), and the characteristic after dynamic control is expressed as F route =GΘFIC+(1-G)ΘFIC', where Θ represents element-by-element multiplication and FIC' represents the output without considering the operation behavior.

Images