A plant survey becomes useful for model fitting when each measurement can be assigned to a stream, a time window, a physical phase and an uncertainty. A slurry density reading, a sampled solids mass fraction and a conveyor dry solids flow are different observations even when all three describe one stream. Their relationship follows from the stream basis. Reconciliation uses such relationships to estimate a consistent bank of stream states while retaining a record of what was actually measured.
Design the survey around the question
Begin at a boundary with a decision attached to it. A cyclone survey meant to distinguish water recovery and corrected cut needs samples or calibrated measurements of the feed, overflow and underflow. Product PSDs alone leave the product mass split uncertain. A plant water survey needs the fresh, recycle, entrained and discarded water paths. Flow to an internal pond is an internal transfer when the pond is included within the boundary. A measured net change in the pond volume belongs in the accumulation term.
Start and stop sampling windows on a common clock. A ten-minute feed composite paired with a ten-minute concentrate composite may represent different feed parcels if the process residence time is 25 minutes. Either work over a window in which throughput and inventories are stable or align signals with the residence-time structure and include accumulation. Material in transit and material held in sumps account for legitimate differences between inflow and outflow over short windows.
For each observation, record sampling interval, location, equipment operating state, measurement basis, units, uncertainty estimate and an identifier for any derived value. Retain full precision until reporting. If wet throughput is 200 t/h and solids are 60% by mass, the derived phases are 120 t/h dry solids and 80 t/h water. Both estimates share the wet-throughput measurement and the solids fraction. Treating the phase estimates as independent inflates the information content.
Write balances with a measurable basis
For a stream containing dry solids flow S, water flow W, component dry masses Mⱼ and retained size masses mᵢ, the identities are
where water and solid specific volumes are assumed additive and Q is volumetric flow. For independent class-by-component data, a size-by-component matrix Mᵢⱼ provides both marginal totals. A survey with only class masses and bulk grades does not identify every element of that matrix; any completion of the matrix requires additional assumptions or observations. Distinguish a modeled distribution of liberation from a measured assay in each size class.
A balance matrix A describes the independent steady-state constraints. For one splitting node with an inlet F and outlets U and O, choose A = [1, −1, −1]. Replicate balance rows for conservative components and phases. At circuits with reactions or intentional reagent addition, include source terms where their physical basis warrants them. For dry solids and water in a grinding–classification circuit, size reduction transfers material between size intervals while total dry solids and elemental masses are conserved through an inert mill.
Size assays specified as cumulative passing require care. If four cumulative passing values are used to form three adjacent differences and a composite mass, their errors are correlated. For example, t₁ = P₁ − P₂ and t₂ = P₂ − P₃ share P₂ with opposite signs. With independent original determinations, Cov(t₁,t₂) = −Var(P₂). When the starting cumulative data already share numerator samples or denominators, use their full covariance. A normalized PSD with k bins has k − 1 independent fractions. Fit a consistent subset or work with the appropriate rank. A nil reading has a defined detection rule and a different meaning from an unsampled port.
Weighted least squares and the balance residual
Let y be the measurement vector and x the unknown consistent vector. Suppose y = x + e, with unbiased measurement error and positive definite covariance Σ. Set Ax = b, where b represents known sources or accumulation. The weighted linear problem minimizes (x − y)ᵀΣ⁻¹(x − y) / 2 subject to the balances. With Lagrange multiplier λ the stationarity condition is
Substituting x̂ = y − ΣAᵀλ into the constraints gives
Implement the inverse notation as a linear solve. If balance rows are linearly dependent, retain a basis of independent constraints or use rank-aware factorization. The projection form follows from conservation and the measurement covariance; Crowe, García Campos and Hrymak applied it to steady-state process flow measurements and unmeasured streams. Crowe, García Campos and Hrymak, 1983
An individual correction is influenced by measurement uncertainty and topology. A less precise flow can carry a larger correction in physical units. A graph with no redundant observation around a particular branch may allow many balances to close without testing its values. A missing quantity is observable if all admissible data-conforming states give it the same value. Algebraically, with measurement matrix H, an individual coordinate is observable when its direction is orthogonal to every vector in the joint nullspace of A and H.
Worked example: one measured split
Take three simultaneous dry-mass measurements in t/h. Assume independent standard deviations and negligible accumulation.
| Stream | Measured y | Standard deviation | Reconciled x̂ | Correction x̂ − y |
|---|---|---|---|---|
| Feed F | 100.000 | 2.000 | 98.667 | −1.333 |
| Underflow U | 64.000 | 1.000 | 64.333 | +0.333 |
| Overflow O | 34.000 | 1.000 | 34.333 | +0.333 |
The imbalance is 100 − 64 − 34 = 2 t/h. Its variance is 4 + 1 + 1 = 6 (t/h)². Thus λ = 2/6 = 1/3 and corrections are −(4, −1, −1)/3. After adjustment, 98.6667 − 64.3333 − 34.3333 = 0 to machine precision. The one-degree-of-freedom statistic is 2²/6 = 0.667. If errors are Gaussian with the stated variance and the assumed balance is correct, this is below the 95% χ² threshold of 3.841. The measurement difference is compatible with the assumed random errors. The conclusion depends on both the time window and the uncertainty estimates.
After reconciliation the adjusted quantities are correlated. The covariance is
The variance of U becomes 5/6 (t/h)² and Cov(U,O) = −1/6 (t/h)². If only U and O were measured, the estimate F̂ = U + O would close exactly, but the balance would supply an inferred feed rather than a redundant test of instrument consistency.
Now suppose the ore contains a valuable phase. With U assay 5.0% and O assay 1.0% by dry mass, the inferred feed assay is (64.333 × 0.05 + 34.333 × 0.01)/98.667 ≈ 3.608%. If a feed assay is also independently observed, reconcile the component mass balance using measurement uncertainties. Reserve the original 5.0% and 1.0% assay results as observations even if estimated stream assays move slightly. For a mass product y₁y₂, linearization at the measurements gives derivatives y₂ and y₁ and carries uncertainty from both inputs into the derived mass.
Handling discrepancies and difficult samples
A high global balance statistic signals possible bias, a missed stream, inventory movement or variance assumptions that are too narrow. Inspect residuals and raw records before tuning a process model. A representative survey report keeps the original data, estimated values, corrections, the covariance basis, excluded observations and reasons, and the results of redundant balance checks. Assign an explicit variable to an unmeasured evaporative or pond loss when the boundary requires it. A closure imposed by defining a missing flow is algebraic closure, with uncertainty set by its measured dependencies.
For nonlinear observations, such as slurry density measured separately from mass flows and assays, use measured basis equations y ≈ h(x). Gauss–Newton or sequential-constrained methods use a local Jacobian to update the estimates while respecting equality and inequality constraints. Test the final original equations. Bounds can ensure nonnegative flows and physically admissible grades. Robust losses may limit the influence of outliers, but their use and effect on estimated uncertainty must be stated. For a censored assay, a defensible treatment may retain an interval or a detection-limit likelihood rather than substituting a fabricated zero.
Exercises and answers
- A measured split uses identical standard deviations. Its feed is 86 t/h, and its two outlets are 48 t/h and 35 t/h. Reconcile it. The imbalance is 3 t/h, so the equal-weights correction is −1, +1, +1 t/h. The consistent flows are 85, 49 and 36 t/h.
- A three-bin PSD is entered as three adjusted retained fractions summing to one. How many independent coordinates exist? Two. The third is the complement of the other two, so the covariance of all three normalized fractions has a null direction.
- One survey shows inflow 180 t/h and outflow 168 t/h while a surge bin gains 6 t in 30 minutes. What residual belongs in the dynamic balance? The inventory increase is 12 t/h. Inflow minus outflow minus accumulation is 0 t/h.
