An industrial wastewater treatment control decision system based on quantum sensing data analysis

Abstract

The present application relates to the field of wastewater treatment, and discloses an industrial wastewater treatment control decision system based on quantum sensing data analysis, comprising: a quantum sensing monitoring module is deployed in multiple points and integrated with temperature compensation, and real-time collection of multiple parameters such as heavy metal ion concentration and COD; a data processing module is composed of an on-site edge computing unit and a cloud processing unit, the edge computing unit completes data preprocessing, the cloud processing unit combines an LSTM neural network water quality prediction model and a random forest fault diagnosis model, dynamically generates an LSTM memory decay factor Q and a random forest split depth coefficient G through wavelet packet decomposition and Mahalanobis distance calculation, realizes adaptive adjustment of model parameters, and simultaneously establishes a causal relationship between water quality exceeding the standard and equipment failure through cross-validation; an operation and maintenance decision module generates early warning information containing fault information and treatment suggestions. The present application realizes real-time and accurate monitoring of industrial wastewater treatment, water quality prediction, and improves the efficiency and quality of wastewater treatment.

Description

Technical Field

[0001] This invention relates to the field of wastewater treatment, and more specifically to an industrial wastewater treatment control and decision-making system based on quantum sensing data analysis. Background Technology

[0002] Industrial wastewater has a complex composition and fluctuates greatly in pollutant concentration. Its efficient treatment is a key link in environmental protection and sustainable industrial development, which places high demands on real-time and accurate monitoring of water quality parameters, timely diagnosis of equipment operation status, and intelligent control of the treatment process.

[0003] Existing industrial wastewater treatment control systems have many technical shortcomings: Firstly, traditional sensing and monitoring technologies have low accuracy, weak anti-interference capabilities, and lack temperature compensation mechanisms. The collected parameters such as heavy metal ion concentration, COD, and pressure have large errors. At the same time, most sensors are deployed in a single point and are scattered, making it impossible to achieve synchronous monitoring of multiple nodes throughout the entire process. Secondly, data processing mostly adopts a centralized cloud processing mode, with no preprocessing unit set up on site. The direct transmission of raw data leads to high cloud processing pressure, high data transmission latency, and poor real-time performance. Third, both the water quality prediction and equipment fault diagnosis models are designed with fixed parameters, which cannot dynamically adjust the model structure according to the degree of deviation from the on-site working conditions. When the working conditions are abnormal or in a transitional state, the prediction and diagnosis accuracy drops significantly.

[0004] The above problems make it difficult for existing systems to adapt to the complex dynamic conditions of industrial wastewater treatment, thus hindering the development of the wastewater treatment process. Summary of the Invention

[0005] The purpose of this invention is to provide an industrial wastewater treatment control and decision-making system based on quantum sensing data analysis, which solves at least one of the above-mentioned technical problems.

[0006] The objective of this invention can be achieved through the following technical solutions: An industrial wastewater treatment control and decision-making system based on quantum sensing data analysis includes: The quantum sensing monitoring module is installed at the inlet, reaction tank, sedimentation tank, outlet, dosing pump, aeration fan and filter press of the wastewater treatment site to collect heavy metal ion concentration, COD, BOD, turbidity, pressure and flow parameters in real time. The data processing module includes: The edge computing unit, set up on-site, is connected to the quantum sensing monitoring module to filter, reduce noise, and convert the format of the collected data before transmitting it to the cloud. The cloud-based processing unit includes an LSTM neural network water quality prediction model and a random forest fault diagnosis model, which are used to analyze and process the received data. The operation and maintenance decision module, located in the cloud and connected to the processing unit, is used to generate early warning information based on the analysis results of the processing unit.

[0007] As a further technical solution, the LSTM neural network water quality prediction model is used to predict the trend of water quality parameter changes; the random forest fault diagnosis model is used to identify equipment fault types, the fault types including at least pump leakage and fan abnormal noise.

[0008] As a further technical solution, the working process of the processing unit is as follows; Obtain the raw time-series data sequence collected by each sensor in the quantum sensing monitoring module within N consecutive sampling periods; The original time-series data sequence is decomposed into wavelet packet coefficients for multiple frequency bands. Calculate the energy value of the wavelet packet coefficients for each frequency band and construct an energy feature vector; Calculate the Mahalanobis distance D between the energy eigenvector and the preset standard operating condition energy eigenvector; The Mahalanobis distance D is input into the first mapping function and the second mapping function respectively to generate the LSTM memory decay factor Q and the random forest split depth coefficient G; The LSTM memory decay factor Q is applied to the forget gate of the LSTM neural network water quality prediction model, so that the LSTM model can dynamically adjust the retention time of historical information according to the degree of deviation from the current operating conditions. The random forest split depth coefficient G is applied to the decision tree growth process of the random forest fault diagnosis model, so that the random forest model dynamically adjusts the maximum depth of the decision tree according to the deviation of the current working condition. The LSTM neural network water quality prediction model, adjusted by the LSTM memory decay factor Q, analyzes real-time monitoring data, predicts the future trend of water quality parameter changes, and generates water quality prediction results. The random forest fault diagnosis model, adjusted by the random forest split depth coefficient G, uses the water quality prediction results as one of the input features and combines real-time monitoring data to diagnose equipment faults. The output of the LSTM neural network water quality prediction model is cross-validated with the output of the random forest fault diagnosis model to generate the final diagnosis result.

[0009] As a further technical solution, the first mapping function and the second mapping function are obtained as follows: Collect quantum sensing monitoring data under historical operating conditions and calculate the corresponding historical Mahalanobis distance sequence; Based on historical fault records, the historical Mahalanobis distance sequence is divided into steady-state interval, transitional interval, and abnormal interval. Using the historical Mahalanobis distance as input and the preset target attenuation factor and target splitting depth coefficient as output, the first mapping function and the second mapping function are respectively fitted with parameters. The first mapping function is a monotonically decreasing Sigmoid function Q = 1 / (1 + e^(-1 / (1+e ... R*(D-D0) ); The second mapping function is a monotonically increasing exponential function G = 1 - e^(-t / t). -TD .

[0010] As a further technical solution, the specific process by which the LSTM memory decay factor Q acts on the forget gate of the LSTM neural network water quality prediction model is as follows: Each LSTM unit of the LSTM neural network water quality prediction model contains a forget gate; The output ft is determined by the current input xt, the hidden state h(t-1) from the previous time step, and the memory decay factor Q. The calculation formula is as follows: ft=Y*(Q⊙(Wf·[h(t-1),xt]+bf)); Where Y is the Sigmoid activation function, ⊙ represents element-wise multiplication, and Wf and bf are the weight matrix and bias term of the forget gate, respectively.

[0011] As a further technical solution, the random forest split depth coefficient G acts on the decision tree growth process of the random forest fault diagnosis model as follows: When constructing each decision tree in the random forest, the maximum allowed depth dmax of the decision tree is dynamically adjusted according to the split depth coefficient G: dmax = dc × (1 + G); Where dc is the preset base depth, G ranges from [0, Δmax], and Δmax is the maximum depth increment.

[0012] As a further technical solution, cross-validation is performed to generate the final diagnostic results. The specific process is as follows: Extract the predicted water quality parameters within a preset time window from the output of the LSTM neural network water quality prediction model, identify predicted exceedance events that exceed preset thresholds, and record the occurrence time and degree of each predicted exceedance event. Identify the equipment failure events at the current moment from the output of the random forest failure diagnosis model, and obtain the failure type, confidence level and preset failure impact time window for each equipment failure event; Align and match the occurrence time of each predicted out-of-range event with the time window of the failure impact of each equipment failure event on the time axis. If the occurrence time of any predicted excess event falls within the fault impact time window of one of the equipment failure events, it is determined that there is a causal relationship between the two, a comprehensive diagnostic result containing the correlation between the predicted excess event and the equipment failure event is generated, and a warning message based on the causal relationship determination is output. If the predicted exceedance event cannot be matched with the equipment failure event, then based on the degree of exceedance of the predicted exceedance event and the confidence level of the equipment failure event, incremental training of the LSTM neural network water quality prediction model or the random forest fault diagnosis model is triggered to update the model parameters.

[0013] As a further technical solution, the process of correcting the LSTM memory decay factor Q and the random forest split depth coefficient G is as follows: Obtain the first prediction result output by the LSTM neural network water quality prediction model after adjusting the LSTM memory decay factor Q, and the first diagnosis result output by the random forest fault diagnosis model after adjusting the random forest split depth coefficient G. Calculate the correlation coefficient between the first prediction result and the first diagnosis result; When the correlation coefficient is greater than the first preset value, Q is decreased by the first step length and G is decreased by the second step length; wherein the first step length is greater than the second step length, and the difference between the first step length and the second step length increases as the difference between the correlation coefficient and the first preset value increases; When the correlation coefficient is less than the second preset value, Q is increased by the third step size and G is increased by the fourth step size; wherein the third step size is less than the fourth step size, and the difference between the third step size and the fourth step size increases as the difference between the correlation coefficient and the second preset value increases; The length of the first step is greater than the length of the third step, and the length of the fourth step is greater than the length of the second step; The adjusted Q and G are then reapplied to the forget gate of the LSTM neural network water quality prediction model and the decision tree growth process of the random forest fault diagnosis model, respectively.

[0014] The beneficial effects of this invention are: (1) The quantum sensing monitoring module integrates multiple types of quantum sensing units and integrates a temperature compensation module, which improves the accuracy and stability of parameter acquisition. It is deployed at key nodes in the entire wastewater treatment process to achieve real-time monitoring in multiple dimensions and all scenarios. The on-site edge computing unit preprocesses the collected data, reduces the data processing pressure in the cloud, reduces data transmission delay, ensures the real-time performance of data processing, and provides a reliable data foundation for subsequent model analysis and decision-making. (2) Energy feature vectors are constructed by wavelet packet decomposition and Mahalanobis distance is calculated to quantify the deviation of on-site working conditions. Based on the Q and G parameters generated by this distance, the retention time of historical information of LSTM model and the maximum depth of decision tree of random forest model can be adjusted respectively to realize the adaptive optimization of model according to different working conditions. Water quality prediction results are used as input features for fault diagnosis. Cross-validation of the results of the two models is completed by time axis alignment to establish the causal relationship between water quality exceeding the standard and equipment failure, so that the early warning information is traceable. The model supports incremental training and the Q and G parameters can be dynamically corrected to realize iterative optimization of the model and improve the accuracy of water quality prediction and equipment fault diagnosis. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a system structure block diagram of the present invention. Detailed Implementation

[0017] The technical solutions of 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.

[0018] This embodiment uses a comprehensive industrial wastewater treatment station in a chemical industrial park as an application scenario. This treatment station mainly treats comprehensive industrial wastewater containing heavy metals and high COD discharged by chemical enterprises in the park. The wastewater treatment process includes influent regulation, pH adjustment, Fenton reaction, sedimentation and filtration, and effluent treatment to meet standards. It is equipped with core equipment such as dosing pumps, aeration fans, and plate and frame filter presses. The industrial wastewater treatment control and decision-making system based on quantum sensing data analysis of this invention is applied to this treatment station to realize real-time monitoring, water quality prediction, fault diagnosis, and intelligent operation and maintenance decision-making of the entire wastewater treatment process. The following is a detailed implementation method.

[0019] Please see Figure 1 As shown, this invention is an industrial wastewater treatment control and decision-making system based on quantum sensing data analysis, comprising: The quantum sensing monitoring module consists of a quantum magnetic sensing unit, a quantum optical sensing unit, and a quantum pressure / flow sensing unit. All sensing units integrate a platinum resistance temperature compensation module and are deployed in a distributed, multi-point manner at key nodes throughout the wastewater treatment plant process: all types of sensing units are deployed at the inlet, pH adjustment tank, Fenton reaction tank, sedimentation tank, and outlet, while quantum pressure / flow sensing units and vibration sensing accessories are specifically deployed at the equipment ends of the dosing pump, aeration blower, and plate and frame filter press.

[0020] Among them, the quantum magnetic sensing unit is used to collect heavy metal ion concentration parameters in real time, the quantum optical sensing unit is used to collect COD, BOD and turbidity parameters in real time, the quantum pressure / flow sensing unit is used to collect water pressure, water flow rate and equipment operating pressure and flow parameters in real time at each node; the temperature compensation module collects the ambient temperature around the sensing unit in real time and performs real-time compensation and correction on the collected data of each sensing unit according to the temperature change.

[0021] The temperature compensation module can reduce the interference of ambient temperature fluctuations on sensor acquisition and reduce the parameter acquisition error caused by temperature in traditional sensing technology, thereby improving the accuracy and stability of water quality and equipment parameter acquisition. The distributed multi-point deployment method enables synchronous monitoring of key nodes and core equipment in the entire wastewater treatment process, reducing the one-sidedness of traditional single-point scattered monitoring and providing effective raw time-series data for subsequent data processing and model analysis.

[0022] The data processing module includes: The edge computing unit, set up on-site, is connected to the quantum sensing monitoring module to filter, reduce noise, and convert the format of the collected data before transmitting it to the cloud. The edge computing unit adopts an industrial-grade edge computing gateway, which is wired to each sensing unit of the quantum sensing monitoring module through an industrial bus. It performs three-level preprocessing operations on the collected raw time-series data: first, the median filtering algorithm is used to remove impulse interference in the raw data; then, the wavelet denoising algorithm is used to eliminate high-frequency noise; finally, the denoised non-standardized data is uniformly converted into JSON standardized data format. After the preprocessing is completed, the data is transmitted to the cloud processing unit with low latency through a 5G industrial wireless gateway.

[0023] The preprocessing operations of the on-site edge computing unit can remove redundant information and noise interference from the original data, reduce the amount of data transmitted to the cloud, and reduce the overall data processing load of the cloud. The 5G industrial wireless transmission method avoids the latency problem of traditional wired transmission, ensures the real-time transmission of data, and avoids the lag in model analysis and decision-making due to data delay, laying the foundation for real-time analysis of the cloud processing unit.

[0024] The cloud-based processing unit includes an LSTM neural network water quality prediction model and a random forest fault diagnosis model, used to analyze and process the received data. The LSTM neural network water quality prediction model is used to predict the trend of water quality parameter changes. The random forest fault diagnosis model is used to identify equipment fault types, which include at least pump leakage and fan abnormal noise.

[0025] The working process of the processing unit is as follows; Obtain the raw time-series data sequence collected by each sensor in the quantum sensing monitoring module within N consecutive sampling periods; The original time-series data sequence is decomposed using wavelet packet decomposition to obtain wavelet packet coefficients for multiple frequency bands; for example, if the decomposition layer is 3, wavelet packet coefficients for 8 frequency bands are obtained. Calculate the energy value of the wavelet packet coefficients in each frequency band and construct an 8-dimensional energy feature vector; Calculate the Mahalanobis distance D between the 8-dimensional energy eigenvector and the preset standard operating condition energy eigenvector; The Mahalanobis distance D is input into the first mapping function and the second mapping function respectively to generate the LSTM memory decay factor Q and the random forest split depth coefficient G; The LSTM memory decay factor Q is applied to the forget gate of the LSTM neural network water quality prediction model, so that the LSTM model can dynamically adjust the retention time of historical information according to the degree of deviation from the current operating conditions. The random forest split depth coefficient G is applied to the decision tree growth process of the random forest fault diagnosis model, so that the random forest model dynamically adjusts the maximum depth of the decision tree according to the deviation of the current working condition. The LSTM neural network water quality prediction model, adjusted by the LSTM memory decay factor Q, analyzes real-time monitoring data, predicts the future trend of water quality parameter changes, and generates water quality prediction results. The random forest fault diagnosis model, adjusted by the random forest split depth coefficient G, uses the water quality prediction results as one of the input features and combines real-time monitoring data to diagnose equipment faults. The output of the LSTM neural network water quality prediction model is cross-validated with the output of the random forest fault diagnosis model to generate the final diagnosis result.

[0026] The cloud processing unit first acquires the raw time-series data sequences collected by each sensor in the quantum sensing monitoring module within 60 consecutive sampling periods (N=60). It then performs a 3-level wavelet packet decomposition on the raw time-series data sequences to obtain wavelet packet coefficients for 8 frequency bands. The energy values ​​of the wavelet packet coefficients for each of the 8 frequency bands are calculated, and an 8-dimensional energy feature vector is constructed based on these energy values. Simultaneously, it retrieves the standard operating condition data from the stable operation phase of the wastewater treatment plant, performs the same wavelet packet decomposition and energy calculation, and obtains the 8-dimensional energy feature vector for the standard operating condition. The Mahalanobis distance D between the real-time constructed energy feature vector and the standard operating condition energy feature vector is calculated to quantify the deviation between the on-site operating conditions and the standard operating conditions.

[0027] The standard operating condition energy feature vector is obtained by calculating the arithmetic mean of the energy values ​​of each frequency band after decomposing the historical monitoring data of the treatment station for more than 30 consecutive days with no equipment failure records and stable effluent water quality. The linear correlation between the feature dimensions is eliminated in the Mahalanobis distance calculation process, and only the actual deviation features of the operating condition are reflected.

[0028] Wavelet packet decomposition can effectively extract the frequency domain features of time series monitoring data, realizing feature dimensionality reduction and effective extraction of the original time series data; the Mahalanobis distance calculation method eliminates the interference between various feature dimensions, and can accurately quantify the deviation between the actual working conditions on site and the standard working conditions, providing a precise and unique quantitative basis for the subsequent adaptive adjustment of parameters of LSTM and random forest models.

[0029] The operation and maintenance decision module, located in the cloud and connected to the processing unit, is used to generate early warning information including fault location, fault type, and treatment suggestions based on the analysis results of the processing unit, and to calculate the optimal dosage of chemicals and equipment operating parameters based on water quality prediction results. The execution module includes a PLC and a dosing valve, a variable frequency pump, and an aeration valve connected to the PLC. The PLC is connected to the cloud and is used to receive the optimal dosage of chemicals and equipment operating parameters and control the actuators to operate, while feeding back the execution status to the cloud.

[0030] After receiving the comprehensive diagnostic results from the cloud processing unit, the operation and maintenance decision module, in conjunction with the water quality prediction results, calculates the optimal dosage of reagents and equipment operating parameters for wastewater treatment using a process simulation algorithm. In this embodiment, based on the fault diagnosis results of abnormal noise from the aeration blower and the prediction results of moderate COD exceeding the standard at the outlet in the next 8 hours, the process simulation algorithm determines that the dosage of Fenton reagent will be appropriately increased, and the aeration rate of the standby aeration blower will be adjusted to 80% of its rated aeration rate.

[0031] The process simulation algorithm used in this system is an existing technology in the field of industrial wastewater treatment. It relies on water quality prediction and equipment fault diagnosis results, and combines material balance, energy balance and wastewater treatment reaction kinetic model to simulate the operation of each process link in order to solve the optimal reagent dosage and equipment operating parameters.

[0032] The operation and maintenance decision module transmits the optimal parameters to the on-site PLC via cloud communication. As the core of the execution module, the PLC controls the opening of the dosing valve to adjust the dosage of the reagent and controls the frequency of the variable frequency pump to adjust the aeration rate of the standby blower, thus completing the intelligent control of the wastewater treatment process. At the same time, the PLC feeds back execution information such as the opening of the dosing valve, the aeration rate of the blower, and the operating status of the equipment to the cloud processing unit and the operation and maintenance decision module in real time, forming a closed-loop control of monitoring, analysis, decision-making, execution, and feedback.

[0033] The operation and maintenance decision module generates optimal process parameters based on the comprehensive diagnostic results of causal relationships, realizing intelligent and precise control of the wastewater treatment process and avoiding the subjectivity and lag of traditional manual control. The rapid response and real-time status feedback of the execution module construct a closed-loop control system for wastewater treatment, improve the automation and intelligence level of industrial wastewater treatment, and effectively ensure the efficiency of wastewater treatment and the quality of effluent.

[0034] The first and second mapping functions are obtained as follows: Collect quantum sensing monitoring data under historical operating conditions and calculate the corresponding historical Mahalanobis distance sequence; Based on historical fault records, the historical Mahalanobis distance sequence is divided into steady-state interval, transitional interval, and abnormal interval. Using the historical Mahalanobis distance as input and the preset target attenuation factor and target splitting depth coefficient as output, the first mapping function and the second mapping function are respectively fitted with parameters. The first mapping function is a monotonically decreasing Sigmoid function Q = 1 / (1 + e^(-1 / (1+e ... R*(D-D0) The parameters R and D0 are obtained by training a loss function that minimizes Q to be close to 1 in the steady state interval and Q to be close to 0 in the abnormal state interval. The second mapping function is a monotonically increasing exponential function G = 1 - e^(-t / t). -TD The parameter T is obtained by training the model to minimize the difference between the diagnostic accuracy of the model and the diagnostic accuracy of the benchmark model in the transition state interval and the abnormal state interval.

[0035] Example: One year of historical quantum sensing monitoring data from the wastewater treatment station in the chemical industrial park was collected, and the historical Mahalanobis distance sequence was calculated. Based on the historical equipment failure records during this period, such as pump leakage and abnormal fan noise, the historical Mahalanobis distance sequence is divided into three intervals: D ≤ 0.8 is the steady-state interval, 0.8 < D ≤ 1.5 is the transitional interval, and D > 1.5 is the abnormal interval. Using the historical Mahalanobis distance as input and preset target attenuation factor and target splitting depth coefficient as output, the steady-state interval Q approaches 1, the transitional interval 0.3 < Q < 1, and the abnormal interval Q approaches 0; the steady-state interval G approaches 0, the transitional interval 0 < G < 0.6, and the abnormal interval G approaches 0.6. Gradient descent is used to fit the parameters of the first and second mapping functions, respectively, resulting in the first mapping function Q = 1 / (1 + e^(-1 / (1+e ... R*(D-D0) The parameters R=2.5, D0=0.8, and the second mapping function G=1-e -TD The parameter T = 1.2.

[0036] Substituting the calculated real-time Mahalanobis distance D into the two fitted mapping functions, we can generate the LSTM memory decay factor Q and the random forest split depth coefficient G, which are adapted to the current working conditions.

[0037] Based on the actual historical monitoring data and fault records on site, the mapping function parameters are fitted so that the mapping function can accurately match the operating characteristics and equipment operation rules of the wastewater treatment plant. The generated Q and G parameters can truly and accurately reflect the deviation of the current operating conditions, avoiding the mismatch between general model parameters and actual on-site operating conditions, and laying a precise parameter foundation for the adaptive adjustment of the two subsequent models.

[0038] The specific process by which the LSTM memory decay factor Q acts on the forget gate of the LSTM neural network water quality prediction model is as follows: Each LSTM unit of the LSTM neural network water quality prediction model contains a forget gate; The output ft is determined by the current input xt, the hidden state h(t-1) from the previous time step, and the memory decay factor Q. The calculation formula is as follows: ft=Y*(Q⊙(Wf·[h(t-1),xt]+bf)); Where Y is the Sigmoid activation function, ⊙ represents element-wise multiplication, and Wf and bf are the weight matrix and bias term of the forget gate, respectively. When Q approaches 1, the forget gate retains a complete memory of historical information. When Q approaches 0, the forget gate is closed, historical information is reset, and the model only relies on the current input for prediction.

[0039] In this embodiment, the LSTM neural network water quality prediction model is constructed as a network structure with three hidden layers, an input layer, and an output layer. Each hidden layer has 64 neurons. The input layer consists of an 8-dimensional energy feature vector and six types of real-time monitored water quality / equipment parameters. The output layer contains the predicted values ​​of six core water quality parameters, such as COD, heavy metal ion concentration, and BOD, for the next 24 hours.

[0040] Each LSTM unit of the LSTM model contains a forget gate. The weight matrix Wf of the forget gate is set to 128×128, and the bias term bf is set to 128×1. The output ft of the forget gate is determined by the current input xt, the hidden state h(t-1) of the previous time step, and the memory decay factor Q. The calculation formula is ft=Sigmoid*(Q⊙(Wf·[h(t-1),xt]+bf)).

[0041] When the real-time calculated Mahalanobis distance D=1.6 (abnormal interval), substituting it into the mapping function yields Q=0.05. At this point, the output of the forget gate is controlled by the element-wise multiplication of Q and approaches 0. The forget gate is close to closed, and the model discards historical information, relying only on the current real-time monitoring data for water quality prediction. When the real-time calculated Mahalanobis distance D=0.5 (steady-state interval), substituting it into the mapping function yields Q=0.98. The forget gate is close to fully open, and the model retains complete historical monitoring information, combining historical and real-time data for water quality prediction.

[0042] By dynamically controlling the forget gate of the LSTM model using Q, the model can flexibly adjust the retention time of historical information according to the degree of deviation from the operating conditions. Under steady-state conditions, historical information is fully utilized to improve the accuracy and stability of water quality prediction. Under abnormal operating conditions, invalid historical information is discarded to avoid its interference with the prediction results, thus greatly improving the predictive adaptability and accuracy of the LSTM model under complex dynamic operating conditions.

[0043] The random forest splitting depth coefficient G acts on the decision tree growth process of the random forest fault diagnosis model as follows: When constructing each decision tree in the random forest, the maximum allowable depth dmax of the decision tree is dynamically adjusted according to the split depth coefficient G: dmax=dc×(1+G); where dc is the preset base depth, the value range of G is [0,Δmax], and Δmax is the maximum depth increment. When G=0, the decision tree stops growing at the basic depth; when G>0, the decision tree is allowed to continue splitting until it reaches the dynamically adjusted maximum depth. During the splitting process, the minimum number of samples for each node decreases dynamically with the increase of depth to enhance the fitting ability of complex nonlinear patterns under abnormal operating conditions.

[0044] In this embodiment, the random forest fault diagnosis model is constructed by integrating 50 decision trees. The preset base depth of the decision trees is dc=8, the maximum depth increment is Δmax=0.6, and the value range of G is [0,0.6]. When constructing each decision tree, the maximum allowable depth of the decision tree is dynamically adjusted according to the split depth coefficient G using the formula dmax=dc×(1+G). At the same time, the minimum number of samples for each node of the decision tree dynamically decreases as the depth increases.

[0045] When the real-time calculated Mahalanobis distance D=1.6 (out-of-state interval), substituting it into the mapping function yields G=0.58. After calculation, dmax=8×(1+0.58)=12.64. After rounding, the maximum allowable depth of the decision tree is 12, and the minimum number of samples for each node gradually decreases from the basic 5 to 2. When the real-time calculated Mahalanobis distance D=0.5 (steady-state interval), substituting it into the mapping function yields G=0.02. After calculation, dmax=8×1.02=8.16. After rounding, the decision tree grows to the basic depth of 8, and the minimum number of samples for each node remains unchanged at 5.

[0046] By dynamically adjusting the maximum depth of the decision tree in the random forest model using G, the model can adapt the growth scale of the decision tree according to the degree of deviation from the operating conditions. Under abnormal operating conditions, the depth of the decision tree is increased and the minimum number of samples per node is reduced, which enhances the model's ability to fit complex nonlinear fault modes under abnormal operating conditions and improves the accuracy of equipment fault diagnosis. Under steady-state operating conditions, the tree grows at the basic depth, which effectively avoids the problem of model overfitting and ensures the efficiency of fault diagnosis.

[0047] The LSTM neural network water quality prediction model, dynamically adjusted by the LSTM memory decay factor Q, performs feature extraction and time series analysis on the preprocessed real-time monitoring data to predict the changing trends of water quality parameters at key nodes within the next 24 hours and generate water quality prediction results. The random forest fault diagnosis model, dynamically adjusted by the random forest split depth coefficient G, uses the above water quality prediction results as one of the core input features. It combines real-time monitored equipment operating parameters and water quality parameters to perform multi-feature fusion analysis, completes fault diagnosis of core equipment such as dosing pumps and aeration fans, identifies fault types, and calculates fault confidence.

[0048] By using water quality prediction results as input features for fault diagnosis, equipment fault diagnosis not only relies on real-time monitoring data but also incorporates future trends in water quality parameters. This makes the fault diagnosis results more consistent with the actual operating conditions of wastewater treatment, breaking through the limitations of traditional fault diagnosis that relies solely on real-time data and improving the foresight and comprehensiveness of fault diagnosis.

[0049] Cross-validation is performed to generate the final diagnostic results. The specific process is as follows: Extract the predicted water quality parameters within a preset time window from the output of the LSTM neural network water quality prediction model, identify predicted exceedance events that exceed preset thresholds, and record the occurrence time and degree of each predicted exceedance event. Identify the equipment failure events at the current moment from the output of the random forest failure diagnosis model, and obtain the failure type, confidence level and preset failure impact time window for each equipment failure event; Align and match the occurrence time of each predicted out-of-range event with the time window of the failure impact of each equipment failure event on the time axis. If the occurrence time of any predicted excess event falls within the fault impact time window of one of the equipment failure events, it is determined that there is a causal relationship between the two, a comprehensive diagnostic result containing the correlation between the predicted excess event and the equipment failure event is generated, and a warning message based on the causal relationship determination is output. If the predicted exceedance event cannot be matched with the equipment failure event, then based on the degree of exceedance of the predicted exceedance event and the confidence level of the equipment failure event, incremental training of the LSTM neural network water quality prediction model or the random forest fault diagnosis model is triggered to update the model parameters.

[0050] For example, the LSTM model predicts that the COD concentration at the effluent outlet will exceed the emission standard (moderate exceedance) in the next 8 hours. The random forest model diagnoses that the current aeration blower has an abnormal noise fault (confidence level of 95%). The preset impact time window of this fault is 4-12 hours after the fault occurs. By aligning and matching the predicted occurrence time of the exceedance event (in the next 8 hours) with the impact time window of the blower abnormal noise fault, it is found that the predicted exceedance event falls within the fault impact time window, and it is determined that there is a causal relationship between the two.

[0051] The cloud-based processing unit generates a comprehensive diagnostic result based on this causal relationship and transmits the result to the operation and maintenance decision module. The operation and maintenance decision module generates early warning information based on this. The early warning information includes the location of the fault, namely the aeration blower at the outlet; the type of fault, namely abnormal noise from the blower; the impact of the fault, which will cause the COD at the outlet to exceed the standard moderately in the next 8 hours; and the handling suggestion, namely, to immediately shut down the blower for maintenance and temporarily increase the aeration volume of the standby blower.

[0052] By using cross-validation aligned to the timeline, a causal relationship between water quality exceeding standards and equipment failure is established, making early warning information clearly traceable. This provides maintenance personnel with targeted troubleshooting suggestions, avoiding the blind troubleshooting of traditional maintenance and improving the efficiency and accuracy of maintenance decisions.

[0053] If the predicted exceedance events in the water quality prediction results and the equipment failure events in the fault diagnosis results cannot be aligned on the time axis (e.g., the LSTM model predicts severe COD exceedance but the random forest model fails to diagnose any equipment failure, or diagnoses equipment failure but fails to predict water quality exceedance), then incremental training of the corresponding model is triggered based on the degree of exceedance of the predicted exceedance event and the confidence level of the equipment failure event. If it is a water quality prediction deviation, the latest real-time monitoring data and actual water quality detection data are added to the training set of the LSTM model, incremental training is performed, and the model parameters are updated. If it is a fault diagnosis deviation, the latest equipment operation data and fault investigation data are added to the training set of the random forest model, incremental training is performed, and the model parameters are updated. This incremental training mechanism enables dynamic iterative optimization of the model, allowing it to continuously adapt to the dynamic changes in wastewater treatment conditions. This overcomes the limitation of traditional fixed-parameter models that cannot be updated, ensuring the model maintains high prediction and diagnostic accuracy throughout long-term operation.

[0054] The process of correcting the LSTM memory decay factor Q and the random forest split depth coefficient G is as follows: Obtain the first prediction result output by the LSTM neural network water quality prediction model after adjusting the LSTM memory decay factor Q, and the first diagnosis result output by the random forest fault diagnosis model after adjusting the random forest split depth coefficient G. Calculate the correlation coefficient between the first prediction result and the first diagnosis result; When the correlation coefficient is greater than the first preset value, Q is decreased by the first step length and G is decreased by the second step length; wherein the first step length is greater than the second step length, and the difference between the first step length and the second step length increases as the difference between the correlation coefficient and the first preset value increases; When the correlation coefficient is less than the second preset value, Q is increased by the third step size and G is increased by the fourth step size; wherein the third step size is less than the fourth step size, and the difference between the third step size and the fourth step size increases as the difference between the correlation coefficient and the second preset value increases; The length of the first step is greater than the length of the third step, and the length of the fourth step is greater than the length of the second step; The adjusted Q and G are then reapplied to the forget gate of the LSTM neural network water quality prediction model and the decision tree growth process of the random forest fault diagnosis model, respectively.

[0055] Example: The first preset value is 0.8, the second preset value is 0.3, the initial value of the first step length is 0.1, the initial value of the second step length is 0.05, the initial value of the third step length is 0.05, and the initial value of the fourth step length is 0.1. The first step length is always greater than the third step length, and the fourth step length is always greater than the second step length.

[0056] The cloud processing unit first obtains the first prediction result output by the Q-adjusted LSTM model and the first diagnostic result output by the G-adjusted random forest model, and calculates the correlation coefficient between the two: The LSTM water quality prediction sequence and the random forest fault confidence sequence, which are strictly aligned on the time axis (normalized to 0-1 first when the numerical ranges differ greatly), are used to calculate the coefficient representing the degree of linear correlation between the two using the Pearson method. This coefficient is the correlation coefficient required for the system.

[0057] When the correlation coefficient is 0.85, which is greater than the first preset value of 0.8, and the difference between the correlation coefficient and the first preset value is 0.05, the first step size is adjusted to 0.12 and the second step size is adjusted to 0.06. Q and G are reduced respectively according to the adjusted step sizes, and the step size difference increases as the difference increases. When the correlation coefficient is 0.25, which is less than the second preset value of 0.3, and the difference between the correlation coefficient and the second preset value is 0.05, the third step size is adjusted to 0.04 and the fourth step size is adjusted to 0.11. Q and G are increased according to the adjusted step sizes, and the step size difference increases as the difference increases.

[0058] The corrected Q and G are then applied again to the forget gate of the LSTM model and the decision tree growth process of the random forest model to complete the secondary optimization of the model parameters.

[0059] Based on the dynamic correction of Q and G of the correlation coefficients of the output results of the two models, the output results of the LSTM model and the random forest model are kept reasonably correlated, avoiding the distortion of the overall analysis results caused by the parameter bias of a single model, realizing the synergistic optimization of the two models, and further improving the accuracy and stability of the model combination analysis.

[0060] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. An industrial wastewater treatment control and decision-making system based on quantum sensing data analysis, characterized in that, include: The quantum sensing monitoring module is used to collect heavy metal ion concentration, COD, BOD, turbidity, pressure and flow parameters in real time. The data processing module includes: The edge computing unit, set up on-site, is connected to the quantum sensing monitoring module to filter, reduce noise, and convert the format of the collected data before transmitting it to the cloud. The cloud-based processing unit includes an LSTM neural network water quality prediction model and a random forest fault diagnosis model, which are used to analyze and process the received data. The operation and maintenance decision module, located in the cloud and connected to the processing unit, is used to generate early warning information based on the analysis results of the processing unit.

2. The industrial wastewater treatment control and decision-making system based on quantum sensing data analysis according to claim 1, characterized in that, The LSTM neural network water quality prediction model is used to predict the trend of water quality parameter changes; the random forest fault diagnosis model is used to identify equipment fault types, which include at least pump leakage and fan noise.

3. The industrial wastewater treatment control and decision-making system based on quantum sensing data analysis according to claim 2, characterized in that, The working process of the processing unit is as follows; Obtain the raw time-series data sequence collected by each sensor in the quantum sensing monitoring module within N consecutive sampling periods; The original time-series data sequence is decomposed into wavelet packet coefficients for multiple frequency bands. Calculate the energy value of the wavelet packet coefficients for each frequency band and construct an energy feature vector; Calculate the Mahalanobis distance D between the energy eigenvector and the preset standard operating condition energy eigenvector; The Mahalanobis distance D is input into the first mapping function and the second mapping function respectively to generate the LSTM memory decay factor Q and the random forest split depth coefficient G; The LSTM memory decay factor Q is applied to the forget gate of the LSTM neural network water quality prediction model, so that the LSTM model can dynamically adjust the retention time of historical information according to the degree of deviation from the current operating conditions. The random forest split depth coefficient G is applied to the decision tree growth process of the random forest fault diagnosis model, so that the random forest model dynamically adjusts the maximum depth of the decision tree according to the degree of deviation of the current working condition. The LSTM neural network water quality prediction model, adjusted by the LSTM memory decay factor Q, analyzes real-time monitoring data, predicts the future trend of water quality parameter changes, and generates water quality prediction results. The random forest fault diagnosis model, adjusted by the random forest split depth coefficient G, uses the water quality prediction results as one of the input features and combines real-time monitoring data to diagnose equipment faults. The output of the LSTM neural network water quality prediction model is cross-validated with the output of the random forest fault diagnosis model to generate the final diagnosis result.

4. The industrial wastewater treatment control and decision-making system based on quantum sensing data analysis according to claim 3, characterized in that, The first and second mapping functions are obtained as follows: Collect quantum sensing monitoring data under historical operating conditions and calculate the corresponding historical Mahalanobis distance sequence; Based on historical fault records, the historical Mahalanobis distance sequence is divided into steady-state interval, transitional interval, and abnormal interval. Using the historical Mahalanobis distance as input and the preset target attenuation factor and target splitting depth coefficient as output, the first mapping function and the second mapping function are respectively fitted with parameters. The first mapping function is a monotonically decreasing Sigmoid function Q = 1 / (1 + e^(-1 / (1+e ... R*(D-D0) ); The second mapping function is a monotonically increasing exponential function G = 1 - e^(-t / t). -TD .

5. The industrial wastewater treatment control and decision-making system based on quantum sensing data analysis according to claim 3, characterized in that, The specific process by which the LSTM memory decay factor Q acts on the forget gate of the LSTM neural network water quality prediction model is as follows: Each LSTM unit of the LSTM neural network water quality prediction model contains a forget gate; The output ft is determined by the current input xt, the hidden state h(t-1) from the previous time step, and the memory decay factor Q. The calculation formula is as follows: ft=Y*(Q⊙(Wf·[h(t-1),xt]+bf)); Where Y is the Sigmoid activation function, ⊙ represents element-wise multiplication, and Wf and bf are the weight matrix and bias term of the forget gate, respectively.

6. The industrial wastewater treatment control and decision-making system based on quantum sensing data analysis according to claim 3, characterized in that, The random forest splitting depth coefficient G acts on the decision tree growth process of the random forest fault diagnosis model as follows: When constructing each decision tree in the random forest, the maximum allowed depth dmax of the decision tree is dynamically adjusted according to the split depth coefficient G: dmax = dc × (1 + G); Where dc is the preset base depth, G ranges from [0, Δmax], and Δmax is the maximum depth increment.

7. The industrial wastewater treatment control and decision-making system based on quantum sensing data analysis according to claim 3, characterized in that, Cross-validation is performed to generate the final diagnostic results. The specific process is as follows: Extract the predicted water quality parameters within a preset time window from the output of the LSTM neural network water quality prediction model, identify predicted exceedance events that exceed preset thresholds, and record the occurrence time and degree of each predicted exceedance event. Identify the equipment failure events at the current moment from the output of the random forest failure diagnosis model, and obtain the failure type, confidence level and preset failure impact time window for each equipment failure event; Align and match the occurrence time of each predicted out-of-range event with the time window of the failure impact of each equipment failure event on the time axis. If the occurrence time of any predicted excess event falls within the fault impact time window of one of the equipment failure events, it is determined that there is a causal relationship between the two, a comprehensive diagnostic result containing the correlation between the predicted excess event and the equipment failure event is generated, and a warning message based on the causal relationship determination is output. If the predicted exceedance event cannot be matched with the equipment failure event, then based on the degree of exceedance of the predicted exceedance event and the confidence level of the equipment failure event, incremental training of the LSTM neural network water quality prediction model or the random forest fault diagnosis model is triggered to update the model parameters.

8. The industrial wastewater treatment control and decision-making system based on quantum sensing data analysis according to claim 7, characterized in that, The process of correcting the LSTM memory decay factor Q and the random forest split depth coefficient G is as follows: Obtain the first prediction result output by the LSTM neural network water quality prediction model after adjusting the LSTM memory decay factor Q, and the first diagnosis result output by the random forest fault diagnosis model after adjusting the random forest split depth coefficient G. Calculate the correlation coefficient between the first prediction result and the first diagnosis result; When the correlation coefficient is greater than the first preset value, Q is decreased by the first step and G is decreased by the second step. The first step length is greater than the second step length, and the difference between the first step length and the second step length increases as the difference between the correlation coefficient and the first preset value increases; When the correlation coefficient is less than the second preset value, Q is increased by the third step and G is increased by the fourth step. The third step length is less than the fourth step length, and the difference between the third step length and the fourth step length increases as the difference between the correlation coefficient and the second preset value increases. The length of the first step is greater than the length of the third step, and the length of the fourth step is greater than the length of the second step; The adjusted Q and G are then reapplied to the forget gate of the LSTM neural network water quality prediction model and the decision tree growth process of the random forest fault diagnosis model, respectively.

Images