Anomaly Detection Method for Sludge Bulking in Urban Sewage Treatment Process Integrating Multi-Step Prediction

Abstract

The present invention discloses a method for detecting sludge bulking anomalies in the urban sewage treatment process by integrating multi-step prediction, belonging to the technical field of sewage treatment, and realizing the real-time detection of sludge bulking anomalies under intermittent measurement; this anomaly detection method first completes the reconstruction of sludge bulking samples through a multi-step prediction method based on an adaptive fuzzy neural network to obtain sludge bulking samples with a shorter sampling interval; then designs a sludge bulking anomaly detection strategy based on trend analysis to complete the extraction of sludge bulking trend features, constructs an anomaly evaluation strategy based on the l2 norm of trend features, and realizes the detection of sludge bulking anomalies; finally solves the problem that it is difficult to detect anomalies in the urban sewage treatment process in real time and accurately, and can lay a foundation for the safe and stable operation of the urban sewage treatment process.

Description

Technical Field

[0001] The present invention belongs to the technical field of sewage treatment, and particularly relates to a method for detecting sludge bulking anomalies in the urban sewage treatment process by integrating multi-step prediction. Background Art

[0002] In the urban sewage treatment process, the activated sludge method is mainly used to promote the adsorption, decomposition, and oxidation of organic pollutants by microorganisms to achieve sewage purification. For the urban sewage treatment process using the activated sludge method, normal sludge-water separation is the key to ensuring its safe and stable operation. However, sludge bulking is a frequent abnormal phenomenon in the urban sewage treatment process. The problem of sludge bulking, that is, the phenomenon of loose structure and difficult sedimentation separation, persists, which will not only lead to a decline in sludge sedimentation performance and deterioration of the effluent quality, but may even lead to the failure of the urban sewage treatment process in severe cases. In addition, sludge bulking is characterized by a high incidence rate, a wide occurrence range, and a long recovery time. Implementing an effective method for detecting sludge bulking anomalies is the key to improving the sewage treatment effect and ensuring the effective operation of urban sewage treatment plants, and has good environmental and social benefits. Therefore, the research results of the present invention have broad application prospects.

[0003] Sludge bulking is one of the inevitable and intractable problems in the activated sludge method. Its inducing factors are numerous, including water quality conditions, environmental factors, operating conditions, etc.; the operating mechanism of sludge bulking is complex and is affected by factors such as microorganism concentration, substrate concentration, and microorganism concentration, making it difficult to establish an accurate mechanism model. At the same time, the severity of sludge bulking is mainly evaluated by the sludge volume index, and the standard method for detecting the sludge volume index is laboratory testing. This method has a long measurement cycle, and its measurement frequency is much lower than that of relevant process variables. Therefore, how to establish a multi-step prediction model for sludge bulking based on process variable samples with a high sampling frequency and achieve multi-step prediction of sludge bulking is of great significance for real-time detection of sludge bulking anomalies. In addition, urban sewage treatment is a typical non-stationary process, and the characteristics of sludge bulking will change with the occurrence of operating conditions, anomalies, etc. How to accurately obtain the trend characteristics of sludge bulking and judge the occurrence of anomalies in real time is an urgent problem to be solved to ensure sewage treatment performance. Therefore, it is necessary to establish an effective multi-step prediction method for sludge bulking to increase the number of samples of sludge bulking evaluation indicators for easy real-time detection of anomalies; design a sludge bulking anomaly detection strategy based on trend analysis to obtain the trend characteristics of sludge bulking and achieve accurate detection of sludge bulking to ensure the effective operation of urban sewage treatment plants. Summary of the Invention

[0004] To solve the above problems, the present invention proposes a method for detecting sludge bulking anomalies in the urban sewage treatment process that integrates multi-step prediction. A sludge bulking model that integrates multi-step direct prediction and multi-step recursive prediction is constructed based on an adaptive fuzzy neural network to achieve multi-step prediction of sludge bulking, facilitating the acquisition of sludge bulking samples with shorter sampling intervals. An anomaly detection method based on trend features is designed to achieve real-time and accurate judgment of sludge bulking anomalies, solve the problem of anomaly detection in the urban sewage treatment process under intermittent measurement, improve sewage treatment efficiency, and ensure the safe and stable operation of urban sewage treatment plants.

[0005] The technical solution of the present invention is as follows:

[0006] A method for detecting sludge bulking anomalies in the urban sewage treatment process that integrates multi-step prediction, which performs real-time anomaly detection of sludge bulking by constructing a multi-step prediction model for sludge bulking, specifically including the following steps:

[0007] Step 1, construct a multi-step prediction model for sludge bulking;

[0008] Step 2, design an SVI anomaly detection strategy based on trend features;

[0009] Step 3, collect process variable data and sludge bulking data generated during the urban sewage treatment process in real time, design a sludge bulking evaluation index based on the constructed multi-step prediction model for sludge bulking, and then determine whether sludge bulking occurs at the current moment.

[0010] Further, the specific process of Step 1 is as follows:

[0011] Step 1.1, analyze the operating characteristics of the urban sewage treatment process, obtain the sludge volume index SVI, an index that can evaluate the severity of sludge bulking, and related process variables; the related process variables include the influent flow rate Q in , dissolved oxygen concentration S O , hydraulic retention time SRT, sludge return ratio SRR, sludge loading rate F / M, temperature T;

[0012] Among them, the SVI value is obtained through laboratory tests, and its measurement period is 2 hours; the Q in value is measured by an influent flow meter; the S O concentration value is obtained by an on-line dissolved oxygen analyzer; the SRT value is obtained by the ratio of the volume of the biochemical reaction tank to Q in ; the SRR value is obtained by the ratio of the mixed liquor sludge concentration to the difference between the return sludge concentration and the mixed liquor sludge concentration, and the sludge concentration is obtained by a sludge concentration detector; the F / M value is obtained by Q inObtained by the ratio of the product of the chemical oxygen demand COD to the product of the aeration tank volume and the sludge concentration, the COD value is obtained by a chemical oxygen demand detector; the T value is obtained by a temperature sensor, and the sampling period of all process variables is 40 minutes;

[0013] Step 1.2: Establish a one-step recursive prediction model for sludge bulking based on a fuzzy neural network. The model inputs are Q in , S O , SRT, SRR, F / M, and T, and the model output is SVI, specifically expressed as:

[0014]

[0015] where represents the one-step recursive prediction value of SVI at time k, and f r (·) represents the non-linear relationship between SVI and the inputs during the recursive prediction process, represents the one-step recursive prediction value of SVI at time (k - 1), which is used as the input of the recursive prediction model at time k, and x k represents the relevant process variables, and x k = [Q in , S O , SRT, SRR, F / M, T], ω d,j,k represents the output weight of the jth hidden layer of the recursive prediction model at time k, J represents the number of hidden layers, and υ j,k represents the output of the jth hidden layer of the recursive prediction model at time k, and υ j,k is specifically expressed as:

[0016]

[0017] where l represents the lth relevant process variable, l = 1, …, 6; x l,k represents the lth relevant process variable at time k; c l,j,k represents the center of the jth hidden layer corresponding to the lth relevant process variable at time k; ω r,j,k represents the recursive weight of the jth hidden layer at time k; σ l,j,k represents the width of the jth hidden layer corresponding to the lth relevant process variable at time k;

[0018] Step 1.3: Establish a one-step direct prediction model for sludge bulking based on a fuzzy neural network, specifically expressed as:

[0019]

[0020] where represents the one-step direct prediction value of SVI at time k; f d(·) represents the non - linear relationship between the SVI value and the input in a one - step direct prediction process; w j,k represents the output weight of the j - th hidden layer of the direct prediction model at time k; v j,k represents the output of the j - th hidden layer of the direct prediction model at time k, v j,k Specifically, it is expressed as:

[0021]

[0022] Among them, represents the center of the j - th hidden layer corresponding to the l - th relevant process variable of the direct prediction model at time k; represents the width of the j - th hidden layer corresponding to the l - th relevant process variable of the direct prediction model at time k;

[0023] Step 1.4. Design a multi - step recursive prediction strategy based on the one - step recursive prediction model, specifically expressed as:

[0024]

[0025] Among them, represents the recursive prediction value of SVI at time (k + i), i represents the prediction step, i = 2,…,h, and h represents the maximum value of the prediction step; represents the recursive prediction value of SVI at time (k + i - 1); y k represents the actual value of SVI at time k; x k+i represents the relevant process variable at time (k + i); Θ k+i represents the parameter to be optimized at time (k + i) in the multi - step recursive prediction strategy, Θ k+i =[ω d,k+i ,ω r,k+i ,c k+i ,σ k+i , ω d,k+i represents the output weight vector of the recursive prediction model at time (k + i), ω d,k+i =[ω d,1,k+i ,…,ω d,J,k+i , ω r,k+i represents the recursive weight vector of the hidden layer at time (k + i), ω r,k+i =[ω r,1,k+i ,…,ω r,J,k+i , c k+i represents the center vector of the hidden layer in the recursive prediction model at time (k + i), c k+i =[c 1,k+i ,…,c J,k+i , c J,k+i represents the center vector of the J - th hidden layer, c J,k+i =[c 1,J,k+i ,…,c6,J,k+i , σ k+i represents the width vector of the hidden layer in the recursive prediction model at the (k + i) - th moment, σ k+i = [σ 1,k+i , …, σ J,k+i , σ J,k+i represents the width vector of the J - th hidden layer at the (k + i) - th moment, σ J,k+i = [σ 1,J,k+i , …, σ 6,J,k+i ; Considering that the sampling period of the actual SVI samples is higher than that of the relevant process variables, Θ k+h is optimized and adjusted only when the sampling period of the SVI samples is reached, which is specifically expressed as:

[0026] Θ k+h = Θ k +(Ψ k +λ k I) -1 ×Ω k (6)

[0027] where, Θ k+h represents the parameter to be optimized at the (k + h) - th moment of the multi - step recursive prediction strategy, Θ k represents the optimized parameter at the k - th moment of the multi - step recursive prediction strategy, Ψ k represents the quasi - Hessian matrix at the k - th moment of the multi - step recursive prediction strategy, λ k represents the learning rate at the k - th moment, I represents the identity matrix, and Ω k represents the gradient vector at the k - th moment of the multi - step recursive prediction strategy;

[0028] Step 1.5. Design a multi - step direct prediction strategy based on the one - step direct prediction model, which is specifically expressed as:

[0029]

[0030] where, represents the SVI direct prediction value at the (k + i) - th moment; Υ k+i represents the parameter to be optimized at the (k + i) - th moment of the multi - step direct prediction strategy, w k+i represents the output weight vector of the direct prediction model at the (k + i) - th moment, w k+i = [w 1,k+i , …, w J,k+i , represents the center vector of the hidden layer in the direct prediction model at the (k + i) - th moment, represents the center vector of the J - th hidden layer, Denote the width vector of the hidden layer in the direct prediction model at time (k+i). Denote the width vector of the J-th hidden layer at time (k+i). Considering that the actual SVI sample sampling period is higher than the relevant process variables, Υ is optimized and adjusted only when the SVI sample sampling period is reached, specifically expressed as: k+h For optimization and adjustment, specifically expressed as:

[0031] Υ k+h = Υ k +(Γ k + λ k I) -1 × Ξ k (8)

[0032] Where, Υ k+h Denote the parameter to be optimized at time (k+h) of the multi-step direct prediction strategy, Υ k Denote the optimized parameter at time k of the multi-step direct prediction strategy, Γ k Denote the quasi-Hessian matrix of the multi-step direct prediction strategy at time k, Ξ k Denote the gradient vector of the multi-step direct prediction strategy at time k;

[0033] Step 1.6. Design a fusion strategy for multi-step recursive prediction and multi-step direct prediction, specifically expressed as:

[0034]

[0035] Where, Denote the SVI predicted value at time (k+h), θ k+h Denote the weight parameter, Denote the SVI predicted value of multi-step recursive prediction at time (k+h), Denote the SVI predicted value of multi-step direct prediction at time (k+h); θ k+h The value of is defined as:

[0036]

[0037] Where, e r,k+h Denote the multi-step recursive prediction error, y k+h Denote the actual value of SVI at time (k+h); e d,k+h Denote the multi-step direct prediction error,

[0038] Step 1.7. According to the SVI prediction result, the SVI sample is reconstructed as:

[0039]

[0040] Among them, represents the reconstructed SVI sample, m represents the number of actually obtained SVI samples, and n represents the number of reconstructed SVI samples. represents the SVI predicted value at time (k + h - 1), y k+h represents the true SVI value at time (k + h), represents the SVI sample at time (k + 1) after reconstruction.

[0041] Furthermore, the specific process of step 2 is as follows:

[0042] Step 2.1: Obtain the trend characteristics of the reconstructed SVI;

[0043] Step 2.1.1: Rewrite the reconstructed SVI sample in the form of a sliding window, specifically expressed as:

[0044]

[0045] Among them, ζ represents the length of the sliding window sample; for the k-th sliding window sample It is expressed as:

[0046]

[0047] Among them, p s represents the s-th sliding window sample and the change trend between the (s + 1)-th sliding window sample ε k represents the residual at time k;

[0048] Step 2.1.2: Design an optimization algorithm based on the adaptive alternating direction method of multipliers to obtain the trend characteristics of SVI. The optimization objective is described as:

[0049]

[0050] Among them, P = [p1, p2,..., p ζ-1 , P represents the trend characteristics, 1 represents the all-ones vector, and η represents the regularization parameter.

[0051]

[0052] According to the adaptive alternating direction method of multipliers, we have:

[0053]

[0054] Among them, P k+1 represents the trend characteristics at time (k + 1), ρ represents the penalty parameter, Q k represents the change in trend characteristics at time k, and R k represents the residual vector at time k, S = (2UUT +ρVV T ) -1 ; Q k+1 represents the change amount of the trend feature at the (k + 1) - th moment, represents the soft - measurement operation operator, and are respectively two custom - defined trend feature vectors at the (k + 1) - th moment, R k+1 represents the residual vector at the (k + 1) - th moment;

[0055] Step 2.2: Design a sludge bulking anomaly detection strategy based on the l2 - norm of the trend feature, specifically expressed as:

[0056]

[0057] where, represents the sludge bulking evaluation index, and e a represents the unit vector of the a - th trend feature. The threshold of the evaluation index is designed as:

[0058]

[0059] where, represents the threshold, represents the sludge bulking evaluation index at the n - ζ+1 - th moment; The anomaly evaluation logic is:

[0060]

[0061] When the evaluation index is greater than the threshold, sludge bulking occurs; otherwise, there is no sludge bulking phenomenon.

[0062] The beneficial technical effects brought by the present invention:

[0063] 1. According to the process variable samples with high sampling frequency and the sludge bulking samples with low sampling frequency, the present invention establishes a multi - step prediction model that combines multi - step direct prediction and multi - step recursive prediction to realize the multi - step prediction of sludge bulking, and further obtains sludge bulking samples with shorter sampling intervals; uses the trend feature extraction method based on the alternating direction multiplier method to obtain the trend change of sludge bulking, and designs an anomaly evaluation index based on the trend feature to realize the real - time and accurate detection of sludge bulking;

[0064] 2. The present invention adopts an abnormal detection method for urban sewage treatment process that integrates multi-step prediction to achieve real-time and accurate judgment of sludge bulking. This abnormal detection method reconstructs the sludge bulking samples under intermittent measurement through the multi-step prediction method, facilitating the real-time detection of the occurrence of sludge bulking abnormalities. At the same time, an abnormal detection method based on trend features is designed, and an evaluation index based on trend features is established to reduce the problem of high false alarm rate of sludge bulking abnormalities caused by non-stationary features, etc., realizing the effective detection of sludge bulking and ensuring the stable operation of the urban sewage treatment process. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is the flowchart of the method for detecting sludge bulking abnormalities in the urban sewage treatment process that integrates multi-step prediction of the present invention;

[0066] Figure 2 is the test effect diagram of multi-step prediction of sludge bulking under normal samples in the experiment of the present invention;

[0067] Figure 3 is the test error diagram of multi-step prediction of sludge bulking under normal samples in the experiment of the present invention;

[0068] Figure 4 is the test effect diagram of multi-step prediction of sludge bulking under abnormal samples in the experiment of the present invention;

[0069] Figure 5 is the test error diagram of multi-step prediction of sludge bulking under abnormal samples in the experiment of the present invention;

[0070] Figure 6 is the test effect diagram of detecting sludge bulking abnormalities in the experiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0071] The present invention will be further described in detail below in conjunction with the drawings and the specific embodiments:

[0072] The present invention proposes an abnormal detection method for sludge bulking in the urban sewage treatment process that integrates multi-step prediction. This method constructs a sludge bulking model that integrates multi-step direct prediction and multi-step recursive prediction to achieve multi-step prediction of sludge bulking, facilitating the acquisition of sludge bulking samples with shorter sampling intervals; designs an abnormal detection method based on trend features to achieve real-time and accurate judgment of sludge bulking abnormalities; solves the problem of abnormal detection in the urban sewage treatment process under intermittent measurement, improves the sewage treatment efficiency, and ensures the safe and stable operation of urban sewage treatment plants.

[0073] As Figure 1 shown, an abnormal detection method for sludge bulking in the urban sewage treatment process that integrates multi-step prediction performs real-time abnormal detection of sludge bulking by constructing a multi-step prediction model for sludge bulking, and specifically includes the following steps:

[0074] Step 1: Construct a multi-step prediction model for sludge bulking; the specific process is as follows:

[0075] Step 1.1: Analyze the operating characteristics of the urban sewage treatment process, obtain the sludge volume index SVI, which can evaluate the severity of sludge bulking, and related process variables; the related process variables include the influent flow rate Q in , dissolved oxygen concentration S O , hydraulic retention time SRT, sludge return ratio SRR, sludge loading rate F / M, temperature T;

[0076] Among them, the SVI value is mainly obtained through laboratory tests, and its measurement period is 2 hours; Q in value is mainly measured by an influent flow meter; S O concentration value is mainly obtained by an on-line dissolved oxygen analyzer; SRT value is mainly obtained by the ratio of the volume of the biochemical reaction tank to Q in ; SRR value is obtained by the ratio of the mixed liquor sludge concentration to the difference between the return sludge concentration and the mixed liquor sludge concentration, and the sludge concentration can be obtained by a sludge concentration detector; F / M value is mainly obtained by the ratio of the product of Q in and chemical oxygen demand COD to the product of the volume of the aeration tank and the sludge concentration, and the COD value can be obtained by a chemical oxygen demand detector; T value is mainly obtained by a temperature sensor, and the sampling period of all process variables is 40 minutes;

[0077] Step 1.2: Establish a one-step recursive prediction model for sludge bulking based on a fuzzy neural network. The model inputs are Q in , S O , SRT, SRR, F / M and T, and the model output is SVI, which can be specifically expressed as:

[0078]

[0079] Among them, represents the one-step recursive prediction value of SVI at time k, f r (·) represents the non-linear relationship between SVI and the input during the recursive prediction process, represents the one-step recursive prediction value of SVI at time (k - 1), which is used as the input of the recursive prediction model at time k, x k represents the related process variables, x k = [Q in , S O , SRT, SRR, F / M, T], ω d,j,k represents the output weight of the j-th hidden layer of the recursive prediction model at time k, J represents the number of hidden layers, υ j,k represents the output of the j-th hidden layer of the recursive prediction model at time k, υ j,k can be specifically expressed as:

[0080]

[0081] Among them, \(l\) represents the \(l\)-th relevant process variable, where \(l = 1,\cdots,6\); \(x\) l,k represents the \(l\)-th relevant process variable at time \(k\); \(c\) l,j,k represents the center of the \(j\)-th hidden layer corresponding to the \(l\)-th relevant process variable at time \(k\), and the value range of \(c\) l,j,k is \([-1,1]\); \(\omega\) r,j,k represents the recurrent weight of the \(j\)-th hidden layer at time \(k\), and the value range of \(\omega\) r,j,k is \((-1,1)\); \(\sigma\) l,j,k represents the width of the \(j\)-th hidden layer corresponding to the \(l\)-th relevant process variable at time \(k\), and the value range of \(\sigma\) l,j,k is \((0,2)\);

[0082] Step 1.3. Establish a one-step direct prediction model for sludge bulking based on a fuzzy neural network, which can be specifically expressed as:

[0083]

[0084] Among them, represents the one-step direct prediction value of SVI at time \(k\); \(f\) d (·) represents the non-linear relationship between the SVI value and the input during the one-step direct prediction process; \(w\) j,k represents the output weight of the \(j\)-th hidden layer of the direct prediction model at time \(k\), and the value range of \(w\) j,k is \((-3,3)\); \(v\) j,k represents the output of the \(j\)-th hidden layer of the direct prediction model at time \(k\), and \(v\) j,k can be specifically expressed as:

[0085]

[0086] Among them, represents the center of the \(j\)-th hidden layer corresponding to the \(l\)-th relevant process variable of the direct prediction model at time \(k\), and the value range of each element in it is \([-1,1]\); represents the width of the \(j\)-th hidden layer corresponding to the \(l\)-th relevant process variable of the direct prediction model at time \(k\), and the value range of each element in it is \((0,2)\);

[0087] Step 1.4. Design a multi-step recursive prediction strategy based on the one-step recursive prediction model, which can be specifically expressed as:

[0088]

[0089] Among them, Denotes the recursive prediction value of SVI at time (k + i), where i represents the prediction step size, i = 2, …, h, and h represents the maximum prediction step size, h = 3; Denotes the recursive prediction value of SVI at time (k + i - 1); y k Denotes the actual value of SVI at time k; x k+i Denotes the relevant process variable at time (k + i); Θ k+i Denotes the parameter to be optimized at time (k + i) for the multi-step recursive prediction strategy, Θ k+i =[ω d,k+i , ω r,k+i , c k+i , σ k+i , where ω d,k+i Denotes the output weight vector of the recursive prediction model at time (k + i), ω d,k+i =[ω d,1,k+i , …, ω d,J,k+i , where ω r,k+i Denotes the recursive weight vector of the hidden layer at time (k + i), ω r,k+i =[ω r,1,k+i , …, ω r,J,k+i , c k+i Denotes the center vector of the hidden layer in the recursive prediction model at time (k + i), c k+i =[c 1,k+i , …, c J,k+i , where c J,k+i Denotes the center vector of the J-th hidden layer, c J,k+i =[c 1,J,k+i , …, c 6,J,k+i , σ k+i Denotes the width vector of the hidden layer in the recursive prediction model at time (k + i), σ k+i =[σ 1,k+i , …, σ J,k+i , where σ J,k+i Denotes the width vector of the J-th hidden layer at time (k + i), σ J,k+i =[σ 1,J,k+i , …, σ 6,J,k+i ; Considering that the sampling period of the actual SVI samples is higher than that of the relevant process variables, Θ k+h is optimized and adjusted only when the sampling period of the SVI samples is reached, which can be specifically expressed as:

[0090] Θ k+h = Θ k +(Ψ k + λ k I) -1 × Ω k (6)

[0091] where, Θ k+hDenote the parameter to be optimized at time (k+h) for the multi-step recursive prediction strategy, Θ k Denote the parameter optimized at time k for the multi-step recursive prediction strategy, Ψ k Denote the quasi-Hessian matrix at time k for the multi-step recursive prediction strategy, λ k Denote the learning rate at time k, I denote the identity matrix, Ω k Denote the gradient vector at time k for the multi-step recursive prediction strategy;

[0092] Step 1.5, Design a multi-step direct prediction strategy based on the one-step direct prediction model, which can be specifically expressed as:

[0093]

[0094] Among them, Denote the SVI direct prediction value at time (k+i); Υ k+i Denote the parameter to be optimized at time (k+i) for the multi-step direct prediction strategy, w k+i Denote the output weight vector of the direct prediction model at time (k+i), w k+i =[w 1,k+i ,…,w J,k+i , Denote the center vector of the hidden layer in the direct prediction model at time (k+i), Denote the center vector of the Jth hidden layer, Denote the width vector of the hidden layer in the direct prediction model at time (k+i), Denote the width vector of the Jth hidden layer at time (k+i), Considering that the actual SVI sample sampling period is higher than the relevant process variables, only when the SVI sample sampling period is reached, Υ k+h is optimized and adjusted, which can be specifically expressed as:

[0095] Υ k+h =Υ k +(Γ k +λ k I) -1 ×Ξ k (8)

[0096] Among them, Γ k Denote the quasi-Hessian matrix of the multi-step direct prediction strategy at time k, Ξ k Denote the gradient vector of the multi-step direct prediction strategy at time k;

[0097] Step 1.6, Design a fusion strategy for multi-step recursive prediction and multi-step direct prediction, which can be specifically expressed as:

[0098]

[0099] Among them, represents the predicted value of SVI at time (k + h), and θ k+h represents the weight parameter, represents the predicted value of SVI for multi-step recursive prediction at time (k + h), represents the predicted value of SVI for multi-step direct prediction at time (k + h); θ k+h is defined as:

[0100]

[0101] Among them, e r,k+h represents the multi-step recursive prediction error, y k+h represents the actual value of SVI at time (k + h); e d,k+h represents the multi-step direct prediction error,

[0102] Step 1.7. According to the SVI prediction result, the SVI sample can be reconstructed as:

[0103]

[0104] Among them, represents the reconstructed SVI sample, m represents the number of actually obtained SVI samples, and n represents the number of reconstructed SVI samples, represents the predicted value of SVI at time (k + h - 1), and y k+h represents the true value of SVI at time (k + h), represents the (k + 1)-th sample in the reconstructed SVI sample;

[0105] Step 2. Design an SVI anomaly detection strategy based on trend features.

[0106] Step 2.1. Obtain the trend features of the reconstructed SVI;

[0107] Step 2.1.1. Rewrite the reconstructed SVI sample in the form of a sliding window, which can be specifically expressed as:

[0108]

[0109] Among them, ζ represents the length of the sliding window sample, and ζ = 5; for the k-th sliding window sample it can be expressed as:

[0110]

[0111] Among them, p sDenote the s-th sliding window sample and the (s + 1)-th sliding window sample The change trend between them, ε k denotes the residual at time k;

[0112] Step 2.1.2: Design an optimization algorithm based on the adaptive alternating direction method of multipliers to obtain the trend characteristics of SVI. The optimization objective can be described as:

[0113]

[0114] where, P = [p1, p2, …, p ζ-1 , P represents the trend characteristics, 1 represents the all-ones vector, η represents the regularization parameter, η = 0.005,

[0115]

[0116] According to the adaptive alternating direction method of multipliers, we have:

[0117]

[0118] where, P k+1 represents the trend characteristics at time (k + 1), ρ represents the penalty parameter, Q k represents the change in trend characteristics at time k, R k represents the residual vector at time k, S = (2UU T + ρVV T ) -1 ; Q k+1 represents the change in trend characteristics at time (k + 1), represents the soft sensing operation operator, and are respectively two custom trend feature vectors at time (k + 1), R k+1 represents the residual vector at time (k + 1);

[0119] Step 2.2: Design a sludge bulking anomaly detection strategy based on the l2 norm of trend characteristics, which can be specifically expressed as:

[0120]

[0121] where, represents the sludge bulking evaluation index, e a represents the unit vector of the a-th trend characteristic, and the threshold of the evaluation index is designed as:

[0122]

[0123] where, represents a threshold value, represents the sludge bulking evaluation index at the (n - ζ + 1)th moment; the abnormal evaluation logic is as follows:

[0124]

[0125] When the evaluation index is greater than the threshold value, sludge bulking has occurred; otherwise, no sludge bulking phenomenon occurs.

[0126] Step 3: Real-time collect the process variable data and sludge bulking data generated during the urban sewage treatment process, obtain the sludge bulking evaluation index based on the constructed multi-step prediction model of sludge bulking, and then judge whether sludge bulking occurs at the current moment.

[0127] To prove the feasibility and superiority of the present invention, 400 groups of normal process variable samples, 133 groups of normal SVI samples, 475 groups of abnormal process variable samples, and 158 groups of abnormal SVI samples from an actual urban sewage treatment plant were collected. During the multi-step prediction process, 275 groups of normal process variable samples were used for training, 125 groups of normal process variable samples were used for testing, 330 groups of abnormal samples were used for training, and 145 groups of abnormal samples were used for testing. The sludge bulking abnormal detection effect is as Figures 2-6 shown.

[0128] Figure 2 shows the three-step prediction effect of sludge bulking under normal samples. X-axis: number of samples, unit is piece; Y-axis: sludge volume index value, unit is mL / g. The solid line is the reconstructed sludge volume index value, and the dashed line is the sludge volume index value after three-step prediction. Figure 3 shows the three-step prediction error effect diagram of the sludge volume index under normal samples. X-axis: number of samples, unit is piece; Y-axis: sludge volume index prediction error, unit is mL / g. From Figure 2 and Figure 3 it can be seen that the proposed multi-step fusion prediction method can accurately predict the sludge volume index under normal samples, and the prediction error can be maintained between (-3, 2).

[0129] Figure 4 shows the three-step prediction effect of sludge bulking under abnormal samples. X-axis: number of samples, unit is piece; Y-axis: sludge volume index value, unit is mL / g. The solid line is the reconstructed sludge volume index value, and the dashed line is the sludge volume index value after three-step prediction; Figure 5 shows the three-step prediction error effect diagram of the sludge volume index under abnormal samples. X-axis: number of samples, unit is piece; Y-axis: sludge volume index prediction error, unit is mL / g. From Figure 4 and Figure 5It can be seen that the proposed multi-step fusion prediction method can achieve accurate prediction of the sludge volume index under abnormal samples.

[0130] Figure 6 The effect diagram of abnormal detection of the sludge volume index is shown. X-axis: number of samples, unit is piece; Y-axis: value of abnormal detection evaluation index, unit is mL / g. The solid line is the evaluation value of abnormal detection of the sludge volume index, and the dotted line is the evaluation control limit of the sludge volume index. From Figure 6 It can be seen that the proposed abnormal detection method based on trend features can effectively detect sludge bulking.

[0131] Certainly, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for detecting abnormal sludge bulking in the urban sewage treatment process by integrating multi-step prediction, characterized in that, Real-time anomaly detection of sludge bulking is carried out by constructing a multi-step prediction model for sludge bulking, which specifically includes the following steps: Step 1: Construct a multi-step prediction model for sludge bulking; Step 2: Design an SVI anomaly detection strategy based on trend features; Step 3: Real-time collect the process variable data and sludge bulking data generated during the urban sewage treatment process, design a sludge bulking evaluation index based on the constructed multi-step prediction model for sludge bulking, and judge whether sludge bulking occurs at the current moment; The specific process of Step 1 is as follows: Step 1.1: Analyze the operation characteristics of the urban sewage treatment process, obtain the sludge volume index SVI, an index for evaluating the severity of sludge bulking, and related process variables; The relevant process variables include the inlet flow rate Q in , dissolved oxygen concentration S O , hydraulic retention time SRT, sludge return ratio SRR, sludge loading rate F / M, temperature T; Among them, the SVI value is obtained through laboratory tests, and its measurement period is 2 hours; the Q in value is measured by an influent flowmeter; the S O concentration value is obtained by an on-line dissolved oxygen analyzer; the SRT value is obtained by the ratio of the volume of the biochemical reaction tank to Q in ; the SRR value is obtained by the ratio of the mixed liquor sludge concentration to the difference between the return sludge concentration and the mixed liquor sludge concentration, and the sludge concentration is obtained by a sludge concentration detector; the F / M value is obtained by the ratio of the product of Q in and the chemical oxygen demand COD to the product of the volume of the aeration tank and the sludge concentration, and the COD value is obtained by a chemical oxygen demand detector; the T value is obtained by a temperature sensor, and the sampling period of all process variables is 40 minutes; Step 1.2: Establish a one-step recursive prediction model for sludge bulking based on a fuzzy neural network. The model inputs are Q in , S O , SRT, SRR, F / M, and T, and the model output is SVI, which is specifically expressed as: Among them, represents the one-step recursive prediction value of SVI at time k, and f r (·) represents the non-linear relationship between SVI and the input during the recursive prediction process. represents the one-step recursive prediction value of SVI at time (k - 1), which is used as the input of the recursive prediction model at time k. x k represents the relevant process variable, x k = [Q in , S O , SRT, SRR, F / M, T], ω d,j,k represents the output weight of the j-th hidden layer of the recursive prediction model at time k. J represents the number of hidden layers. υ j,k represents the output of the j-th hidden layer of the recursive prediction model at time k. υ j,k Specifically, it is expressed as: where, l represents the l-th relevant process variable, l = 1, …, 6; x l,k represents the l-th relevant process variable at time k; c l,j,k represents the center of the j-th hidden layer corresponding to the l-th relevant process variable at time k; ω r,j,k represents the recurrent weight of the j-th hidden layer at time k; σ l,j,k represents the width of the j-th hidden layer corresponding to the l-th relevant process variable at time k; Step 1.3: Establish a one-step direct prediction model for sludge bulking based on a fuzzy neural network, which is specifically expressed as: Among them, represents the one-step direct prediction value of SVI at time k; f d (·) represents the non-linear relationship between the SVI value and the input during the one-step direct prediction process; w j,k represents the output weight of the j-th hidden layer of the direct prediction model at time k; v j,k represents the output of the j-th hidden layer of the direct prediction model at time k, v j,k Specifically expressed as: Among them, represents the center of the j-th hidden layer corresponding to the l-th relevant process variable at the k-th moment of the direct prediction model; represents the width of the j-th hidden layer corresponding to the l-th relevant process variable at the k-th moment of the direct prediction model; Step 1.4: Design a multi-step recursive prediction strategy based on the one-step recursive prediction model, which is specifically expressed as: Among them, represents the recursive prediction value of SVI at time (k + i), i represents the prediction step size, i = 2, …, h, and h represents the maximum value of the prediction step size; represents the recursive prediction value of SVI at time (k + i - 1); y k represents the actual value of SVI at time k; x k+i represents the relevant process variable at time (k + i); Θ k+i represents the parameter to be optimized at time (k + i) for the multi-step recursive prediction strategy, Θ k+i = [ω d,k+i , ω r,k+i , c k+i , σ k+i , ω d,k+i represents the output weight vector of the recursive prediction model at time (k + i), ω d,k+i = [ω d,1,k+i , …, ω d,J,k+i , ω r,k+i represents the recursive weight vector of the hidden layer at time (k + i), ω r,k+i = [ω r,1,k+i , …, ω r,J,k+i , c k+i represents the center vector of the hidden layer in the recursive prediction model at time (k + i), c k+i = [c 1,k+i , …, c J,k+i , c J,k+i represents the center vector of the J-th hidden layer, c J,k+i = [c 1,J,k+i , …, c 6,J,k+i , σ k+i represents the width vector of the hidden layer in the recursive prediction model at time (k + i), σ k+i = [σ 1,k+i , …, σ J,k+i , σ J,k+i represents the width vector of the J-th hidden layer at time (k + i), σ J,k+i = [σ 1,J,k+i , …, σ 6,J,k+i ; Considering that the sampling period of the actual SVI samples is higher than that of the relevant process variables, Θ k+h is optimized and adjusted only when the sampling period of the SVI samples is reached, which is specifically expressed as: Θ k+h = Θ k +(Ψ k + λ k I) -1 × Ω k (6) Among them, Θ k+h represents the parameter to be optimized at the (k + h) - th moment of the multi - step recursive prediction strategy, Θ k represents the parameter optimized at the k - th moment of the multi - step recursive prediction strategy, Ψ k represents the quasi - Hessian matrix at the k - th moment of the multi - step recursive prediction strategy, λ k represents the learning rate at the k - th moment, I represents the identity matrix, Ω k represents the gradient vector at the k - th moment of the multi - step recursive prediction strategy; Step 1.5: Design a multi-step direct prediction strategy based on the one-step direct prediction model, which is specifically expressed as: Among them, represents the direct prediction value of SVI at the (k + i)th moment; γ k+i represents the parameter to be optimized at the (k + i)th moment of the multi-step direct prediction strategy, w k+i represents the output weight vector of the direct prediction model at the (k + i)th moment, w k+i = [w 1,k+i , …, w J,k+i , represents the center vector of the hidden layer in the direct prediction model at the (k + i)th moment, represents the center vector of the Jth hidden layer, represents the width vector of the hidden layer in the direct prediction model at the (k + i)th moment, represents the width vector of the Jth hidden layer at the (k + i)th moment, Considering that the actual SVI sample sampling period is higher than the relevant process variables, only when the SVI sample sampling period is reached, Υ k+h is optimized and adjusted, specifically expressed as: Υ k+h = Υ k +(Γ k + λ k I) -1 × Ξ k (8) Among them, Υ k+h represents the parameter to be optimized at the (k + h) - moment of the multi - step direct prediction strategy, Υ k represents the parameter optimized at the k - moment of the multi - step direct prediction strategy, Γ k represents the quasi - Hessian matrix of the multi - step direct prediction strategy at the k - moment, Ξ k represents the gradient vector of the multi - step direct prediction strategy at the k - moment; Step 1.6: Design a fusion strategy for multi-step recursive prediction and multi-step direct prediction, which is specifically expressed as: Among them, represents the predicted value of SVI at time (k + h), and θ k+h represents the weight parameter, represents the predicted value of SVI for multi-step recursive prediction at time (k + h), represents the predicted value of SVI for multi-step direct prediction at time (k + h); the value of θ k+h is defined as: Among them, e r,k+h represents the multi-step recursive prediction error, y k+h represents the actual value of SVI at time (k + h); e d,k+h represents the multi-step direct prediction error, Step 1.7: According to the SVI prediction result, the SVI sample is reconstructed as: Among them, represents the reconstructed SVI sample, m represents the number of actually obtained SVI samples, and n represents the number of reconstructed SVI samples. represents the SVI predicted value at time (k + h - 1), y k+h represents the true SVI value at time (k + h), represents the SVI sample at time (k + 1) after reconstruction; The specific process of Step 2 is as follows: Step 2.1: Obtain the trend features of the reconstructed SVI; Step 2.1.1: Rewrite the reconstructed SVI sample into a sliding window form, which is specifically expressed as: where ζ represents the length of the sliding window sample; for the k-th sliding window sample it is expressed as: Among them, p s represents the change trend between the s-th sliding window sample and the (s + 1)-th sliding window sample , and ε k represents the residual at time k; Step 2.1.2: Design an optimization algorithm based on the adaptive alternating direction multiplier method to obtain the trend features of SVI, and the optimization objective is described as: Among them, P = [p1, p2, …, p ζ-1 , P represents the trend feature, 1 represents the all-1 vector, and η represents the regularization parameter. According to the adaptive alternating direction multiplier method: Among them, P k+1 represents the trend feature at the (k + 1)th moment, ρ represents the penalty parameter, Q k represents the change in the trend feature at the kth moment, R k represents the residual vector at the kth moment, S = (2UU T + ρVV T ) -1 ; Q k+1 represents the change in the trend feature at the (k + 1)th moment, represents the soft-sensing operation operator, and are respectively two custom trend feature vectors at the (k + 1)th moment, R k+1 represents the residual vector at the (k + 1)th moment; Step 2.2: Design a sludge bulking anomaly detection strategy based on trend features which is specifically expressed as: Among them, represents the sludge bulking evaluation index, e a represents the unit vector of the a-th trend feature, and the threshold of the evaluation index is designed as: Among them, represents the threshold value, represents the sludge bulking evaluation index at the (n - ζ + 1)-th moment; the abnormal evaluation logic is as follows: When the evaluation index is greater than the threshold value, sludge bulking occurs; otherwise, no sludge bulking phenomenon occurs.

Images