Research method explained

How fuzzy logic ranked industrial pollution

A complete walkthrough of Srinivas and Singh's Ganges River basin assessment, from field evidence to final ranking.

0.066

Final score for CETP Unnao at Industrial site-2. Lower score means greater pollution priority.

Study at a glance

The model combines monitored evidence with judgments from three institutions. It assesses impact; it does not forecast future water quality.

33industrial sources

10decision criteria

3expert decision makers

109water and effluent samples

Entire research pipeline

Each block produces the input required by the next block.

Evidence

Collect effluent and groundwater observations in pre-monsoon and post-monsoon periods.

Structure

Define 33 alternatives, ten impact criteria, and one ranking objective.

Judgment

Three experts rate criteria and industries with five linguistic levels.

AHP

Convert ratings into reciprocal pairwise comparison matrices.

Fuzzify

Combine expert comparisons into interval-valued triangular fuzzy numbers.

Weight

Use fuzzy geometric means and normalization to calculate relative weights.

Rank

Defuzzify, multiply local weights by criterion weights, and order industries.

Map

Display critical pollutants in ArcGIS and propose management actions.

Evidence and study design

UPPCB monitoring data guided expert ratings and later supported qualitative comparison with model rankings.

Industrial site-1 and Banthar
Industries, CETP, and drains

Industrial site-2
Industries, CETP, and drains

Magarwara-I
Industrial sources and drains

Magarwara-II
Industrial sources

Sampling

54 wastewater or effluent samples and 55 groundwater samples. Deep and shallow aquifers. Pre-monsoon and post-monsoon observations from 2012.

Measured pollutants

pH, BOD, COD, TSS, TDS, chromium, fluoride, lead, nickel, cadmium, arsenic, iron, copper, and zinc.

Ten criteria

C10 receives greatest final weight. Ratings for C1, C3, and C4 require reverse interpretation because high values mean worse pollution.

C1Effluent discharge quantity0.139

C2ETP or CETP efficiency0.072

C3Critical parameters, pre-monsoon0.148

C4Critical parameters, post-monsoon0.096

C5Effluent or waste type0.096

C6River impact, pre-monsoon0.139

C7River impact, post-monsoon0.030

C8Groundwater impact, pre-monsoon0.096

C9Groundwater impact, post-monsoon0.029

C10Public-health impact0.235

Fuzzy method, step by step

Open each step for inputs, operation, and output.

01Collect expert ratingsQuestionnaire

Three experts, five levels

Experts rate criterion importance and every industry under every criterion. Linguistic answers become crisp values.

1Very low

3Low

5Medium

7High

9Very high

02Build pairwise AHP matricesRelative preference

Compare every pair

Equal alternatives receive 1. One to four linguistic-level gaps receive 3, 5, 7, or 9. Worse direction receives the reciprocal.

equal = 1 one level = 3 two levels = 5 reverse = 1/3 or 1/5

03Check consistencyIR below 0.1

Reject inconsistent judgments

They calculate matrix inconsistency from the largest eigenvalue. Experts should revise ratings when the ratio reaches 0.1.

II = (lambda_max - n) / (n - 1) IR = II / RII accept when IR < 0.1

04Combine experts into one fuzzy numberIVTFN

Preserve disagreement

Minimum and maximum retain the expert range. Geometric mean gives the center. Two midpoints form inner bounds.

q = geometric_mean(Z1, Z2, Z3) p' = min(Z) p = (min(Z) + q) / 2 r = (q + max(Z)) / 2 r' = max(Z) Z_fuzzy = [(p', p); q; (r, r')]

05Build group fuzzy comparison matrixReciprocal matrix

One matrix represents all experts

Every pair contains five fuzzy components. The opposite comparison uses reciprocal values in reverse order.

if Zij = [(p',p); q; (r,r')] Zji = [(1/r',1/r); 1/q; (1/p,1/p')]

06Calculate fuzzy relative weightsGeometric mean

Aggregate each matrix row

Multiply every fuzzy comparison in one row, take the nth root, then normalize against the sum of all rows.

Xi = (Zi1 x Zi2 x ... x Zin)^(1/n) Wi = Xi x (X1 + X2 + ... + Xn)^(-1)

07DefuzzifyOne crisp weight

Collapse five values into one

The central component receives double weight. Result becomes one local performance value.

W* = (w1' + w1 + 2w2 + w3 + w3') / 6

08Calculate final score and rankingWeighted sum

Combine criteria

Each industry local weight is multiplied by corresponding criterion importance. The ten products are added.

Final(i) = sum[ CriterionWeight(j) x LocalWeight(i,j) ] lower score = more critical source higher score = less critical source

Worked example

S1 compared with S2 under C3 at Industrial site-2.

Expert judgments

E1: 1/3
E2: 1/3
E3: 1

q = (1/3 x 1/3 x 1)^(1/3) q = 0.481 fuzzy comparison: [(0.333, 0.407); 0.481; (0.740, 1.000)]

After row weighting

S1 fuzzy weight: [(0.031, 0.037); 0.044; (0.057, 0.070)] crisp weight: (0.031 + 0.037 + 2x0.044 + 0.057 + 0.070) / 6

S1 under C3 = 0.047

What drove the ranking

Public health received greatest expert weight. Post-monsoon river and groundwater criteria received least.

C100.235

C30.148

C10.139

C60.139

C40.096

C50.096

C80.096

C20.072

C70.030

C90.029

Reported area results

Each area is normalized and ranked separately. Low score means high intervention priority.

Industrial area	Most critical	Score	Least critical	Score
Site-1 and Banthar	S4 in Table 10	0.046	S6 and S8	0.233
Industrial site-2	S1, CETP Unnao	0.066	S8, Omega International	0.258
Magarwara-I	S1, City Jail Drain	0.073	S4, Sadaf Enterprises	0.267
Magarwara-II	S1, Global Rimjhim	0.087	S4, ACI Oils	0.193

GIS came after fuzzy ranking

ArcGIS mapped BOD, COD, pH, TSS, fluoride, and chromium. Spatial maps supported interpretation but did not enter fuzzy equations.

monitoring data -> expert ratings expert ratings -> fuzzy ranking pollutant values -> GIS maps

Complete ML replacement

Replace expert estimates of environmental impact with predictions learned from repeated measurements. Keep legal limits and public-health priorities as explicit policy rules.

The current data are not enough

The paper reports 33 sources and 109 samples, but not a reusable source-by-date dataset with paired outcomes. We can choose candidate algorithms now. We can choose the final model only after collecting and auditing the data.

fuzzy model: expert opinion -> rank ML replacement: measurements -> predicted impact -> rank

What replaces each fuzzy component

Paper component	What it did	ML replacement
Expert linguistic ratings	Estimated severity and importance	Measured features and measured outcomes
Pairwise AHP matrix	Converted opinions into relative preferences	Training examples linking conditions to outcomes
Interval fuzzy number	Preserved disagreement among three experts	Prediction intervals from bootstrap or quantile models
Defuzzified local weight	Gave each source one score under each criterion	Predicted load, environmental change, or violation probability
Weighted final score	Combined ten criteria into a ranking	Transparent risk rule applied to model outputs
GIS maps	Displayed measured spatial patterns	Map predictions and uncertainty at monitored locations

The targets the model must learn

Train one model per pollutant and outcome first. This is easier to validate and explain than one large model that hides different pollutant behaviour.

outlet_load = outlet_concentration x discharge_flow river_change = downstream_concentration - upstream_concentration groundwater_change = near_well_concentration - control_well_concentration exceedance_ratio = observed_concentration / permitted_limit violation = 1 when observed_concentration > permitted_limit

Final intervention score

The model predicts physical outcomes. A separate documented rule combines exceedance, pollutant toxicity, predicted river change, predicted groundwater change, exposed population, and prediction uncertainty.

This policy layer replaces the paper's weighted sum. It must be approved by domain and regulatory experts; ML cannot learn what society should value from chemistry alone.

Models to compare

Model	Role	Use when	Main caution
Elastic Net	Required baseline	Small data; approximately linear effects	Misses complex interactions
XGBoost	Primary candidate	Structured nonlinear data with enough repeated records	Can overfit small site counts
Random Forest	Tree comparison	Nonlinear effects and robust tabular baseline	Poor extrapolation outside training range
SVR	Small-data comparison	Clean, scaled data with moderate feature count	Sensitive to scaling and tuning

Do not choose the winner now. Start with Elastic Net. Compare XGBoost, Random Forest, and SVR on identical held-out sites and future dates. Select the simplest model with stable error and useful uncertainty.

Do not use deep learning with this expected sample size. Do not use reinforcement learning: no repeated control action, reward, or safe simulator exists in the present study.

How to build and test it

Define outcomes

Choose the pollutant, location, prediction horizon, and physical target before sampling.

Sample repeatedly

Measure sources and paired environmental locations on matching dates across seasons.

Audit data

Standardize units, retain detection limits, flag missingness, and prevent impossible values.

Split correctly

Hold out entire sources and future dates. Never split repeated rows randomly.

Train baseline

Fit Elastic Net first, then compare three nonlinear candidates using the same folds.

Measure error

Use MAE and RMSE for quantities; recall, precision, and calibration for violations.

Quantify uncertainty

Report prediction intervals. A rank without uncertainty is incomplete.

Rank and monitor

Apply the policy rule, map risk, then retrain only when new validated data improve it.

Data the paper used but did not publish

The paper prints examples and final weights, not the complete source data. These missing tables are required to reproduce the fuzzy calculation exactly.

Required table	What the paper prints	What is missing
Sampling-station register	Table 1 covers site-1, Banthar, and site-2	Magarwara-I and Magarwara-II station details
Laboratory measurements	Table 2 shows only CETP Unnao inlet and outlet	Complete 54 effluent and 55 groundwater sample records
Expert source ratings	Table 6 shows nine site-2 sources for C3 only	Ratings for every source under every criterion
Pairwise matrices	Table 7 shows site-2 under C3 only	Remaining criteria, sites, and criterion-importance matrices
Detailed fuzzy calculations	Tables 8 and 9 show site-2 under C3 only	Fuzzy matrices and weights for every remaining criterion and site

AComplete sampling-station registerPartly published

What is absent

The authors explicitly omit station information for Magarwara-I and Magarwara-II. We need the identity and location of every source, inlet, outlet, drain, river station, and groundwater well.

source_id industrial_area industry_or_drain_name sampling_station_id station_type inlet_or_outlet latitude, longitude receiving_drain_or_river

BComplete laboratory measurement tableHighest-value missing table

All 109 samples

The study used 54 effluent samples and 55 groundwater samples in pre-monsoon and post-monsoon 2012. Only one CETP example is printed. We need every row, including nondetects and missing results.

sample_id, source_id, station_id sampling_date, season, sample_type inlet_or_outlet, discharge_flow pH, TDS, TSS, BOD, COD Cu, Ni, Pb, Cd, Zn, Cr, As, F, Fe unit, detection_limit, QA_flag

CComplete expert-rating tableAbout 963 ratings absent

Original questionnaire answers

A complete table requires 33 sources x 10 criteria x 3 experts, or 990 source-level ratings. Table 6 prints only 27 ratings for nine site-2 sources under C3, leaving about 963 unprinted.

expert_id expert_role industrial_area source_id criterion_id linguistic_rating numeric_rating questionnaire_date

DPairwise-comparison matricesCan be regenerated

Derived, not essential to request

The complete AHP matrices and inconsistency ratios are absent. If we receive the original expert ratings and the exact comparison rule, we can calculate these again.

expert_id industrial_area criterion_id source_i, source_j pairwise_value matrix_IR

EFuzzy matrices and intermediate weightsCan be regenerated

Derived, not raw data

The full interval-valued matrices, row geometric means, normalized fuzzy weights, and defuzzified weights are not printed. They can be reproduced after obtaining the expert ratings and resolving the paper's pairwise-rule ambiguity.

p_prime, p, q, r, r_prime row_geometric_mean normalized_fuzzy_weight defuzzified_local_weight final_weight

Minimum request to reproduce the paper: the complete station register, all 109 laboratory records, all 990 expert source ratings, and the exact rule or software export used to convert rating differences into pairwise values.

For ML: do not use expert ratings, pairwise matrices, fuzzy weights, or Table 10 final scores as training labels. Use repeated measured environmental outcomes from the next section.

Exact data needed

Priority runs from low to highest. One observation must join source conditions and environmental outcomes from the same sampling event.

Priority 1
Low

Administrative context

Stable source ID, industry type, products, applicable legal limits, inspection history, complaints, and enforcement history.

Priority 2
Low-medium

Spatial context

Outlet coordinates, river reach, flow direction, distances, upstream and downstream stations, nearby and control wells, settlements, drinking-water intakes, drainage, and aquifer information.

Priority 3
Medium

Time, weather, and river conditions

Date and time, season, recent rainfall, river level, river discharge, and water temperature where relevant.

Priority 4
Medium-high

Treatment and production

ETP or CETP type, operating status, design capacity, actual inflow, inlet and outlet flow, downtime, maintenance, production rate, chemical use, and process settings.

Priority 5
High

Repeated effluent measurements

At each outlet: discharge flow, pH, BOD, COD, TSS, TDS, chromium, fluoride, lead, nickel, cadmium, arsenic, iron, copper, zinc, and industry-specific toxic pollutants.

For every result: unit, laboratory method, detection limit, quantification limit, replicate or quality-control flag, and missing-value reason.

Priority 6
Very high

Paired measured outcomes

On matching dates, measure the same pollutant panel upstream and downstream, in a nearby well and a comparable control well. These paired differences are the labels used to train the environmental-impact models.

Priority 7
Highest

Repeated observations over time

Repeat the complete sampling event at fixed sources and stations, monthly if practical, across all seasons and preferably multiple years. Keep source and station IDs unchanged.

Without repeated paired measurements, there is no defensible supervised model for environmental impact.

Minimum machine-learning row

one row = one source x one sampling event x one pollutant IDs: source_id, outlet_id, station_ids, sample_id time: date_time, season source: industry_type, production_rate, treatment_status hydrology: rainfall, river_flow, discharge_flow chemistry: inlet, outlet, upstream, downstream, near_well, control_well quality: unit, method, detection_limit, QA_flag, missing_reason target: load, river_change, groundwater_change, or violation

Collection rule: concentration without discharge flow cannot estimate load. Downstream without upstream cannot isolate river change. A nearby well without a control well cannot estimate local groundwater change. A single visit cannot teach seasonal or future behaviour.

What to treat carefully

These issues matter if the method is reproduced or replaced with machine learning.

Missing raw judgments

Only one complete C3 example is shown. Full expert ratings for all 33 sources and ten criteria are absent.

Pairwise formula mismatch

Equation 2 states a ratio. Table 7 behaves like linguistic-level distance on Saaty's scale.

Weights do not sum to one

Published criterion weights sum to 1.08. Final score remains relative, not a probability.

Ranking contradiction

Site-1 prose names S3 as worst. Table 10 gives S4 a lower score.

Validation is not independent

Monitoring observations informed experts, then similar observations were used to support rankings.

No prediction

Model aggregates expert preferences. It does not learn temporal patterns or forecast concentrations.

Exact role of fuzzy logic

Fuzzy logic preserves disagreement between three experts while AHP calculates relative importance. It replaces a simple average with an interval around the group judgment. Defuzzification then collapses that interval into one ranking weight.

It does not model laboratory uncertainty, pollutant transport, future water quality, or cause and effect.